Welcome back. Over the next few lessons and the wider course, we'll be covering storage a lot, and the exam expects you to know the appropriate type of storage to pick for a given situation. So before we move on to the AWS-specific storage lessons, I wanted to quickly do a refresher. So let's get started.

Let's start by covering some key storage terms. First is direct attached or local attached storage. This is storage, so physical disks, which are connected directly to a device, so a laptop or a server. In the context of EC2, this storage is directly connected to the EC2 hosts and it's called the instance store. Directly attached storage is generally super fast because it's directly attached to the hardware, but it suffers from a number of problems. If the disk fails, the storage can be lost. If the hardware fails, the storage can be lost. If an EC2 instance moves between hosts, the storage can be lost.

The alternative is network attached storage, which is where volumes are created and attached to a device over the network. In on-premises environments, this uses protocols such as iSCSI or Fiber Channel. In AWS, it uses a product called Elastic Blockstore known as EBS. Network storage is generally highly resilient and is separate from the instance hardware, so the storage can survive issues which impact the EC2 host.

The next term is ephemeral storage and this is just temporary storage, storage which doesn't exist long-term, storage that you can't rely on to be persistent. And persistent storage is the next point, storage which exists as its own thing. It lives on past the lifetime of the device that it's attached to, in this case, EC2 instances. So an example of ephemeral storage, so temporary storage, is the instance store, so the physical storage that's attached to an EC2 host. This is ephemeral storage. You can't rely on it, it's not persistent. An example of persistent storage in AWS is the network attached storage delivered by EBS.

Remember that, it's important for the exam. You will get questions testing your knowledge of which types of storage are ephemeral and persistent. Okay, next I want to quickly step through the three main categories of storage available within AWS. The category of storage defines how the storage is presented either to you or to a server and also what it can be used for.

Now the first type is block storage. With block storage, you create a volume, for example, inside EBS and the red object on the right is a volume of block storage and a volume of block storage has a number of addressable blocks, the cubes with the hash symbol. It could be a small number of blocks or a huge number, that depends on the size of the volume, but there's no structure beyond that. Block storage is just a collection of addressable blocks presented either logically as a volume or as a blank physical hard drive.

Generally when you present a unit of block storage to a server, so a physical disk or a volume, on top of this, the operating system creates a file system. So it takes the raw block storage, it creates a file system on top of this, for example, NTFS or EXT3 or many other different types of file systems and then it mounts that, either as a C drive in Windows operating systems or the root volume in Linux.

Now block storage comes in the form of spinning hard disks or SSDs, so physical media that's block storage or delivered as a logical volume, which is itself backed by different types of physical storage, so hard disks or SSDs. In the physical world, network attached storage systems or storage area network systems provide block storage over the network and a simple hard disk in a server is an example of physical block storage. The key thing is that block storage has no inbuilt structure, it's just a collection of uniquely addressable blocks. It's up to the operating system to create a file system and then to mount that file system and that can be used by the operating system.

So with block storage in AWS, you can mount a block storage volume, so you can mount an EBS volume and you can also boot off an EBS volume. So most EC2 instances use an EBS volume as their boot volume and that's what stores the operating system, and that's what's used to boot the instance and start up that operating system.

Now next up, we've got file storage and file storage in the on-premises world is provided by a file server. It's provided as a ready-made file system with a structure that's already there. So you can take a file system, you can browse to it, you can create folders and you can store files on there. You access the files by knowing the folder structure, so traversing that structure, locating the file and requesting that file.

You cannot boot from file storage because the operating system doesn't have low-level access to the storage. Instead of accessing tiny blocks and being able to create your own file system as the OS wants to, with file storage, you're given access to a file system normally over the network by another product. So file storage in some cases can be mounted, but it cannot be used for booting. So inside AWS, there are a number of file storage or file system-style products. And in a lot of cases, these can be mounted into the file system of an operating system, but they can't be used to boot.

Now lastly, we have object storage and this is a very abstract system where you just store objects. There is no structure, it's just a flat collection of objects. And an object can be anything, it can have attached metadata, but to retrieve an object, you generally provide a key and in return for providing the key and requesting to get that object, you're provided with that object's value, which is the data back in return.

And objects can be anything, there can be binary data, they can be images, they can be movies, they can be cat pictures, like the one in the middle here that we've got of whiskers. If they can be any data really that's stored inside an object. The key thing about object storage though is it is just flat storage. It's flat, it doesn't have a structure. You just have a container. In AWS's case, it's S3 and inside that S3 bucket, you have objects. But the benefits of object storage is that it's super scalable. It can be accessed by thousands or millions of people simultaneously, but it's generally not mountable inside a file system and it's definitely not bootable.

So that's really important, you understand the differences between these three main types of storage. So generally in the on-premises world and in AWS, if you want to utilize storage to boot from, it will be block storage. If you want to utilize high performance storage inside an operating system, it will also be block storage. If you want to share a file system across multiple different servers or clients or have them accessed by different services, that can often be file storage. If you want large access to read and write object data at scale. So if you're making a web scale application, you're storing the biggest collection of cat pictures in the world, that is ideal for object storage because it is almost infinitely scalable.

Now let's talk about storage performance. There are three terms which you'll see when anyone's referring to storage performance. There's the IO or block size, the input output operations per second, pronounced IOPS, and then the throughput. So the amount of data that can be transferred in a given second, generally expressed in megabytes per second.

Now these things cannot exist in isolation. You can think of IOPS as the speed at which the engine of a race car runs at, the revolutions per second. You can think of the IO or block size as the size of the wheels of the race car. And then you can think of the throughput as the end speed of the race car. So the engine of a race car spins at a certain revolutions, whether you've got some transmission that affect that slightly, but that transmission, that power is delivered to the wheels and based on their size, that causes you to go at a certain speed.

In theory in isolation, if you increase the size of the wheels or increase the revolutions of the engine, you would go faster. For storage and the analogy I just provided, they're all related to each other. The possible throughput a storage system can achieve is the IO or the block size multiplied by the IOPS.

As we talk about these three performance aspects, keep in mind that a physical storage device, a hard disk or an SSD, isn't the only thing involved in that chain of storage. When you're reading or writing data, it starts with the application, then the operating system, then the storage subsystem, then the transport mechanism to get the data to the disk, the network or the local storage bus, such as SATA, and then the storage interface on the drive, the drive itself and the technology that the drive uses. There are all components of that chain. Any point in that chain can be a limiting factor and it's the lowest common denominator of that entire chain that controls the final performance.

Now IO or block size is the size of the blocks of data that you're writing to disk. It's expressed in kilobytes or megabytes and it can range from pretty small sizes to pretty large sizes. An application can choose to write or read data of any size and it will either take the block size as a minimum or that data can be split up over multiple blocks as it's written to disk. If your storage block size is 16 kilobytes and you write 64 kilobytes of data, it will use four blocks.

Now IOPS measures the number of IO operations the storage system can support in a second. So how many reads or writes that a disk or a storage system can accommodate in a second? Using the car analogy, it's the revolutions per second that the engine can generate given its default wheel size. Now certain media types are better at delivering high IOPS versus other media types and certain media types are better at delivering high throughput versus other media types. If you use network storage versus local storage, the network can also impact how many IOPS can be delivered. Higher latency between a device that uses network storage and the storage itself can massively impact how many operations you can do in a given second.

Now throughput is the rate of data a storage system can store on a particular piece of storage, either a physical disk or a volume. Generally this is expressed in megabytes per second and it's related to the IO block size and the IOPS but it could have a limit of its own. If you have a storage system which can store data using 16 kilobyte block sizes and if it can deliver 100 IOPS at that block size, then it can deliver a throughput of 1.6 megabytes per second. If your application only stores data in four kilobyte chunks and the 100 IOPS is a maximum, then that means you can only achieve 400 kilobytes a second of throughput.

Achieving the maximum throughput relies on you using the right block size for that storage vendor and then maximizing the number of IOPS that you pump into that storage system. So all of these things are related. If you want to maximize your throughput, you need to use the right block size and then maximize the IOPS. And if either of these three are limited, it can impact the other two. With the example on screen, if you were to change the 16 kilobyte block size to one meg, it might seem logical that you can now achieve 100 megabytes per second. So one megabyte times 100 IOPS in a second, 100 megabytes a second, but that's not always how it works. A system might have a throughput cap, for example, or as you increase the block size, the IOPS that you can achieve might decrease.

As we talk about the different AWS types of storage, you'll become much more familiar with all of these different values and how they relate to each other. So you'll start to understand the maximum IOPS and the maximum throughput levels that different types of storage in AWS can deliver. And you might face exam questions where you need to answer what type of storage you will pick for a given level of performance demands. So it's really important as we go through the next few lessons that you pay attention to these key levels that I'll highlight.

It might be, for example, that a certain type of storage can only achieve 1000 IOPS or 64000 IOPS. Or it might be that certain types of storage cap at certain levels of throughput. And you need to know those values for the exam so that you can know when to use a certain type of storage.

Now, this is a lot of theory and I'm talking in the abstract and I'm mindful that I don't want to make this boring and it probably won't sink in and you won't start to understand it until we focus on some AWS specifics. So I am going to end this lesson here. I wanted to give you the foundational understanding, but over the next few lessons, you'll start to be exposed to the different types of storage available in AWS and you will start to paint a picture of when to pick particular types of storage versus others.

So with that being said, that's everything I wanted to cover. I know this has been abstract, but it will be useful if you do the rest of the lessons in this section. I promise you this is going to be really valuable for the exam. So thanks for watching. Go ahead and complete the video. When you're ready, you can join me in the next.

Welcome back—this is part two of this lesson, and we're going to continue immediately from the end of part one, so let's get started.

Now, this is an overview of all of the different categories of instances, and then for each category, the most popular or current generation types that are available; I created this with the hope that it will help you retain this information.

This is the type of thing that I would generally print out or keep an electronic copy of and refer to constantly as we go through the course—by doing so, whenever we talk about particular size and type and generation of instance, if you refer to the details in the notes column, you'll be able to start making a mental association between the type and then what additional features you get.

So, for example, if we look at the general purpose category, we've got three main entries in that category: we've got the A1 and M6G types, and these are a specific type of instance that are based on ARM processors—so the A1 uses the AWS-designed Graviton ARM processor, and the M6G uses the generation 2, so Graviton 2 ARM-based processor.

And using ARM-based processors, as long as you've got operating systems and applications that can run under the architecture, they can be very efficient—so you can use smaller instances with lower cost and achieve really great levels of performance.

The T3 and T3A instance types are burstable instances, so the assumption with those types of instances is that your normal CPU load will be fairly low, and you have an allocation of burst credits that allows you to burst up to higher levels occasionally but then return to that normally low CPU level.

So this type of instance—T3 and T3A—are really good for machines which have low normal loads with occasional bursts, and they're a lot cheaper than the other types of general purpose instances.

Then we've got M5, M5A, and M5N—so M5 is your starting point, M5A uses the AMD architecture whereas normal M5s just use Intel, and these are your steady-state general instances.

So if you don't have a burst requirement and you're running a certain type of application server which requires consistent steady-state CPU, then you might use the M5 type—maybe a heavily used Exchange email server that runs normally at 60% CPU utilization might be a good candidate for M5.

But if you've got a domain controller or an email relay server that normally runs maybe at 2%, 3% with occasional bursts up to 20%, 30%, or 40%, then you might want to run a T-type instance.

Now, not to go through all of these in detail, we've got the compute optimized category with the C5 and C5N, and they go for media encoding, scientific modeling, gaming servers, general machine learning.

For memory optimized, we start off with R5 and R5A; if you want to use really large in-memory applications, you've got the X1 and the X1E; if you want the highest memory of all A-to-the-U instances, you've got the high memory series; and you've got the Z1D, which comes with large memory and NVMe storage.

Then, Accelerated Computing—these are the ones that come with these additional capabilities, so the P3 type and G4 type come with different types of GPUs: the P type is great for parallel processing and machine learning, while the G type is kind of okay for machine learning and much better for graphics-intensive requirements.

You've got the F1 type, which comes with field programmable gate arrays, which is great for genomics, financial analysis, and big data—anything where you want to program the hardware to do specific tasks.

You've got the Inf1 type, which is relatively new, custom-designed for machine learning—so recommendation forecasting, analysis, voice conversation, anything machine learning-related, look at using that type.

And then, storage-optimized instances—these come with high-speed local storage, and depending on the type you pick, you can get high throughput or maximum I/O or somewhere in between.

So, keep this somewhere safe, print it out, keep it electronically, and as we go through the course and use the different types of instances, refer to this and start making the mental association between what a category is, what instance types are in that category, and then what benefits they provide.

Now again, don't worry about memorizing all of this for the exam—you don't need it—I'll draw out anything specific that you need as we go through the course, but just try to get a feel for which letters are in which categories.

If that's the minimum that you can do—if I can give you a letter like the T type, or the C type, or the R type—and you can try and understand the mental association with which category that goes into, that will be a great step.

And there are ways we can do this—we can make these associations—so C stands for compute, R stands for RAM (which is a way for describing memory), we've got I which stands for I/O, D which stands for dense storage, G which stands for GPU, P which stands for parallel processing; there's lots of different mind tricks and mental associations that we can do, and as we go through the course, I'll try and help you with that.

But as a minimum, either print this out or store it somewhere safe and refer to it as we go through the course.

The key thing to understand though is how picking an instance type is specific to a particular type of computing scenario—so if you've got an application that requires maximum CPU, look at compute optimized; if you need memory, look at memory optimized; if you've got a specific type of acceleration, look at accelerated computing; start off in the general purpose instance types and then go out from there as you've got a particular requirement to.

Now before we finish up, I did want to demonstrate two really useful sites that I refer to constantly—I'll include links to both of these in the lesson text.

The first one is the Amazon documentation site for Amazon EC2 instance types—this gives you a follow-up view of all the different categories of EC2 instances.

You can look in a category, a particular family and generation of instance—so T3—and then in there you can see the use cases that this is suited to, any particular features, and then a list of each instance size and exactly what allocation of resources that you get and then any particular notes that you need to be aware of.

So this is definitely something you should refer to constantly, especially if you're selecting instances to use for production usage.

This other website is something similar—it’s EC2instances.info—and it provides a really great sortable list which can be filtered and adjusted with different attributes and columns, which give you an overview of exactly what each instance provides.

So you can either search for a particular type of instance—maybe a T3—and then see all the different sizes and capabilities of T3; as well as that, you can see the different costings for those instance types—so Linux on-demand, Linux reserved, Windows on-demand, Windows reserved—and we’ll talk about what this reserved column is later in the course.

You can also click on columns and show different data for these different instance types, so if I scroll down, you can see which offer EBS optimization, you can see which operating systems these different instances are compatible with, and you've got a lot of options to manipulate this data.

I find this to be one of the most useful third-party sites—I always refer back to this when I’m doing any consultancy—so this is a really great site.

And again, it will go into the lesson text, so definitely as you’re going through the course, experiment and have a play around with this data, and just start to get familiar with the different capabilities of the different types of EC2 instances.

With that being said, that’s everything I wanted to cover in this lesson—you’ve done really well, and there’s been a lot of theory, but it will come in handy in the exam and real-world usage.

So go ahead, complete this video, and when you’re ready, you can join me in the next.

Welcome back. In this lesson, now that we've covered virtualization at a high level, I want to focus on the architecture of the EC2 product in more detail. EC2 is one of the services you'll use most often in AWS since one which features on a lot of exam questions, so let's get started.

First thing, let's cover some key, high level architectural points about EC2. EC2 instances are virtual machines, so this means an operating system plus an allocation of resources such as virtual CPU, memory, potential some local storage, maybe some network storage, and access to other hardware such as networking and graphics processing units. EC2 instances run on EC2 hosts, and these are physical servers hardware which AWS manages. These hosts are either shared hosts or dedicated hosts.

Shared hosts are hosts which are shared across different AWS customers, so you don't get any ownership of the hardware and you pay for the individual instances based on how long you run them for and what resources they have allocated. It's important to understand, though, that every customer when using shared hosts are isolated from each other, so there's no visibility of it being shared, there's no interaction between different customers, even if you're using the same shared host, and shared hosts are the default.

With dedicated hosts, you're paying for the entire host, not the instances which run on it. It's yours, it's dedicated to your account, and you don't have to share it with any other customers. So if you pay for a dedicated host, you pay for that entire host, you don't pay for any instances running on it, and you don't share it with other AWS customers.

EC2 is an availability zone resilient service. The reason for this is that hosts themselves run inside a single availability zone, so if that availability zone fails, the hosts inside that availability zone could fail, and any instances running on any hosts that fail will themselves fail. So as a solutions architect, you have to assume if an AZ fails, then at least some and probably all of the instances that are running inside that availability zone will also fail or be heavily impacted.

Now let's look at how this looks visually. So this is a simplification of the US East One region, I've only got two AZs represented, AZA and AZB, and in AZA, I've represented that I've got two subnet, subnet A and subnet B. Now inside each of these availability zones is an EC2 host. Now these EC2 hosts, they run within a single AZ, I'm going to keep repeating that because it's critical for the exam and you're thinking about EC2 in the exam.

Keep thinking about it being an AZ resilient service, if you see EC2 mentioned in an exam, see if you can locate the availability zone details because that might factor into the correct answer. Now EC2 hosts have some local hardware, logically CPU and memory, which you should be aware of, but also they have some local storage called the instance store. The instance store is temporary, if an instance is running on a particular host, depending on the type of the instance, it might be able to utilize this instance store, but if the instance moves off this host to another one, then that storage is lost.

And they also have two types of networking, storage networking and data networking. When instances are provisioned into a specific subnet within a VPC, what's actually happening is that a primary elastic network interface is provisioned in a subnet, which maps to the physical hardware on the EC2 host. Remember, subnets are also in one specific availability zone. Instances can have multiple network interfaces, even in different subnets, as long as they're in the same availability zone. Everything about EC2 is focused around this architecture, the fact that it runs in one specific availability zone.

Now EC2 can make use of remote storage so an EC2 host can connect to the elastic block store, which is known as EBS. The elastic block store service also runs inside a specific availability zone, so the service running inside availability zone A is different than the one running inside availability zone B, and you can't access them cross zone. EBS lets you allocate volumes and volumes of portions of persistent storage, and these can be allocated to instances in the same availability zone, so again, it's another area where the availability zone matters.

What I'm trying to do by keeping repeating availability zone over and over again is to paint a picture of a service which is very reliant on the availability zone that it's running in. The host is in an availability zone, the network is per availability zone, the persistent storage is per availability zone, even availability zone in AWS experiences major issues, it impacts all of those things.

Now an instance runs on a specific host, and if you restart the instance, it will stay on a host. Instances stay on a host until one of two things happen: firstly, the host fails or is taken down for maintenance for some reason by AWS; or secondly, if an instance is stopped and then started, and that's different than just restarting, so I'm focusing on an instance being stopped and then being started, so not just a restart. If either of those things happen, then an instance will be relocated to another host, but that host will also be in the same availability zone.

Instances cannot natively move between availability zones. Everything about them, their hardware, networking and storage is locked inside one specific availability zone. Now there are ways you can do a migration, but it essentially means taking a copy of an instance and creating a brand new one in a different availability zone, and I'll be covering that later in this section where I talk about snapshots and AMIs.

What you can never do is connect network interfaces or EBS storage located in one availability zone to an EC2 instance located in another. EC2 and EBS are both availability zone services, they're isolated, you cannot cross AZs with instances or with EBS volumes. Now instances running on an EC2 host share the resources of that host. And instances of different sizes can share a host, but generally instances of the same type and generation will occupy the same host.

And I'll be talking in much more detail about instance types and sizes and generations in a lesson that's coming up very soon. But when you think about an EC2 host, think that it's from a certain year and includes a certain class of processor and a certain type of memory and a certain type and configuration of storage. And instances are also created with different generations, different versions that you apply specific types of CPU memory and storage, so it's logical that if you provision two different types of instances, they may well end up on two different types of hosts.

So a host generally has lots of different instances from different customers of the same type, but different sizes. So before we finish up this lesson, I want to answer a question. That question is what's EC2 good for? So what types of situations might you use EC2 for? And this is equally valuable when you're evaluating a technical architecture while you're answering questions in the exam.

So first, EC2 is great when you've got a traditional OS and application compute need, so if you've got an application that requires to be running on a certain operating system at a certain runtime with certain configuration, maybe your internal technical staff are used to that configuration, or maybe your vendor has a certain set of support requirements, EC2 is a perfect use case for this type of scenario.

And it's also great for any long running compute needs. There are lots of other services inside AWS that provide compute services, but many of these have got runtime limits, so you can't leave these things running consistently for one year or two years. With EC2, it's designed for persistent, long running compute requirements. So if you have an application that runs constantly 24/7, 365, and needs to be running on a normal operating system, Linux or Windows, then EC2 is the default and obvious choice for this.

If you have any applications, which is server style applications, so traditional applications they expect to be running in an operating system, waiting for incoming connections, then again, EC2 is a perfect service for this. And it's perfect for any applications or services that need burst requirements or steady state requirements. There are different types of EC2 instances, which are suitable for low levels of normal loads with occasional bursts, as well as steady state load.

So again, if your application needs an operating system, and it's not bursty needs or consistent steady state load, then EC2 should be the first thing that you review. EC2 is also great for monolithic application stack, so if your monolithic application requires certain components, a stack, maybe a database, maybe some middleware, maybe other runtime based components, and especially if it needs to be running on a traditional operating system, EC2 should be the first thing that you look at.

And EC2 is also ideally suited for migrating application workloads, so application workloads, which expect a traditional virtual machine or server style environment, or if you're performing disaster recovery. So if you have existing traditional systems which run on virtual servers, and you want to provision a disaster recovery environment, then EC2 is perfect for that.

In general, EC2 tends to be the default compute service within AWS. There are lots of niche requirements that you might have, and if you do have those, there are other compute services such as the elastic container service or Lambda. But generally, if you've got traditional style workloads, or you're looking for something that's consistent, or if it requires an operating system, or if it's monolithic, or if you migrated into AWS, then EC2 is a great default first option.

Now in this section of the course, I'm covering the basic architectural components of EC2, so I'm gonna be introducing the basics and let you get some exposure to it, and I'm gonna be teaching you all the things that you'll need for the exam.

Welcome back and in this first lesson of the EC2 section of the course, I want to cover the basics of virtualization as briefly as possible. EC2 provides virtualization as a service. It's an infrastructure as a service or I/O product. To understand all the value it provides and why some of the features work the way that they do, understanding the fundamentals of virtualization is essential. So that's what this lesson aims to do.

Now, I want to be super clear about one thing. This is an introduction level lesson. There's a lot more to virtualization than I can talk about in this brief lesson. This lesson is just enough to get you started, but I will include a lot of links in the lesson description if you want to learn more. So let's get started.

We do have a fair amount of theory to get through, but I promise when it comes to understanding how EC2 actually works, this lesson will be really beneficial. Virtualization is the process of running more than one operating system on a piece of physical hardware, a server. Before virtualization, the architecture looked something like this. A server had a collection of physical resources, so CPU and memory, network cards and maybe other logical devices such as storage. And on top of this runs a special piece of software known as an operating system.

That operating system runs with a special level of access to the hardware. It runs in privilege mode, or more specifically, a small part of the operating system runs in privilege mode, known as the kernel. The kernel is the only part of the operating system, the only piece of software on the server that's able to directly interact with the hardware. Some of the operating system doesn't need this privilege level of access, but some of it does. Now, the operating system can allow other software to run such as applications, but these run in user mode or unprivileged mode. They cannot directly interact with the hardware, they have to go through the operating system.

So if Bob or Julie are attempting to do something with an application, which needs to use the system hardware, that application needs to go through the operating system. It needs to make a system call. If anything but the operating system attempts to make a privileged call, so tries to interact with the hardware directly, the system will detect it and cause a system-wide error, generally crashing the whole system or at minimum the application. This is how it works without virtualization.

Virtualization is how this is changed into this. A single piece of hardware running multiple operating systems. Each operating system is separate, each runs its own applications. But there's a problem, CPU at least at this point in time, could only have one thing running as privileged. A privileged process member has direct access to the hardware. And all of these operating systems, if they're running in their unmodified state, they expect to be running on their own in a privileged state. They contain privileged instructions. And so trying to run three or four or more different operating systems in this way will cause system crashes.

Virtualization was created as a solution to this problem, allowing multiple different privileged applications to run on the same hardware. But initially, virtualization was really inefficient, because the hardware wasn't aware of it. Virtualization had to be done in software, and it was done in one of two ways. The first type was known as emulated virtualization or software virtualization. With this method, a host operating system still ran on the hardware and included additional capability known as a hypervisor. The software ran in privileged mode, and so it had full access to the hardware on the host server.

Now, around the multiple other operating systems, which we'll now refer to as guest operating systems, were wrapped a container of sorts called a virtual machine. Each virtual machine was an unmodified operating system, such as Windows or Linux, with a virtual allocation of resources such as CPU, memory and local disk space. Virtual machines also had devices mapped into them, such as network cards, graphics cards and other local devices such as storage. The guest operating systems believed these to be real. They had drivers installed, just like physical devices, but they weren't real hardware. They were all emulated, fake information provided by the hypervisor to make the guest operating systems believe that they were real.

The crucial thing to understand about emulator virtualization is that the guest operating systems still believe that they were running on real hardware, and so they still attempt to make privileged calls. They tried to take control of the CPU, they tried to directly read and write to what they think of as their memory and their disk, which are actually not real, they're just areas of physical memory and disk that have been allocated to them by the hypervisor. Without special arrangements, the system would at best crash, and at worst, all of the guests would be overriding each other's memory and disk areas.

So the hypervisor, it performs a process known as binary translation. Any privileged operations which the guests attempt to make, they're intercepted and translated on the fly in software by the hypervisor. Now, the binary translation in software is the key part of this. It means that the guest operating systems need no modification, but it's really, really slow. It can actually halve the speed of the guest operating systems or even worse. Emulated virtualization was a cool set of features for its time, but it never achieved widespread adoption for demanding workloads because of this performance penalty.

But there was another way that virtualization was initially handled, and this is called para-virtualization. With para-virtualization, the guest operating systems are still running in the same virtual machine containers with virtual resources allocated to them, but instead of the slow binary translation which is done by the hypervisor, another approach is used. Para-virtualization only works on a small subset of operating systems, operating systems which can be modified. Because with para-virtualization, there are areas of the guest operating systems which attempt to make privileged calls, and these are modified. They're modified to make them user calls, but instead of directly calling on the hardware, they're calls to the hypervisor called hypercalls.

So areas of the operating systems which would traditionally make privileged calls directly to the hardware, they're actually modified. So the source code of the operating system is modified to call the hypervisor rather than the hardware. So the operating systems now need to be modified specifically for the particular hypervisor that's in use. It's no longer just generic virtualization, the operating systems are modified for the particular vendor performing this para-virtualization. By modifying the operating system this way, and using para-virtual drivers in the operating system for network cards and storage, it means that the operating system became almost virtualization aware, and this massively improved performance. But it was still a set of software processors designed to trick the operating system and/or the hardware into believing that nothing had changed.

The major improvement in virtualization came when the physical hardware started to become virtualization aware. This allows for hardware virtualization, also known as hardware assisted virtualization. With hardware assisted virtualization, hardware itself has become virtualization aware. The CPU contains specific instructions and capabilities so that the hypervisor can directly control and configure this support, so the CPU itself is aware that it's performing virtualization. Essentially, the CPU knows that virtualization exists.

What this means is that when guest operating systems attempt to run any privileged instructions, they're trapped by the CPU, which knows to expect them from these guest operating systems, so the system as a whole doesn't halt. But these instructions can't be executed as is because the guest operating system still thinks that it's running directly on the hardware, and so they're redirected to the hypervisor by the hardware. The hypervisor handles how these are executed. And this means very little performance degradation over running the operating system directly on the hardware.

The problem, though, is while this method does help a lot, what actually matters about a virtual machine tends to be the input/output operation, so network transfer and disk I/O. The virtual machines, they have what they think is physical hardware, for example, a network card. But these cards are just logical devices using a driver, which actually connect back to a single physical piece of hardware which sits in the host. The hardware, everything is running on.

Unless you have a physical network card per virtual machine, there's always going to be some level of software getting in the way, and when you're performing highly transactional activities such as network I/O or disk I/O, this really impacts performance, and it consumes a lot of CPU cycles on the host.

The final iteration that I want to talk about is where the hardware devices themselves become virtualization aware, such as network cards. This process is called S-R-I-O-V, single root I/O virtualization. Now, I could talk about this process for hours about exactly what it does and how it works, because it's a very complex and feature-rich set of standards. But at a very high level, it allows a network card or any other add-on card to present itself, not just one single card, but almost a several mini-cards.

Because this is supported in hardware, these are fully unique cards, as far as the hardware is concerned, and these are directly presented to the guest operating system as real cards dedicated for its use. And this means no translation has to happen by the hypervisor. The guest operating system can directly use its card whenever it wants. Now, the physical card which supports S-R-I-O-V, it handles this process end-to-end. It makes sure that when the guest operating system is used, there are logical mini-network cards that they have physical access to the physical network connection when required.

In EC2, this feature is called enhanced networking, and it means that the network performance is massively improved. It means faster speeds. It means lower latency. And more importantly, it means consistent lower latency, even at high loads. It means less CPU usage for the host CPU, even when all of the guest operating systems are consuming high amounts of consistent I/O.

Many of the features that you'll see EC2 using are actually based on AWS implementing some of the more advanced virtualization techniques that have been developed across the industry. AWS do have their own hypervisor stack now called Nitro, and I'll be talking about that in much more detail in an upcoming lesson, because that's what enables a lot of the higher-end EC2 features.

But that's all the theory I wanted to cover. I just wanted to introduce virtualization at a high level and get you to the point where you understand what S-R-I-O-V is, because S-R-I-O-V is used for enhanced networking right now, but it's also a feature that can be used outside of just network cards. It can help hardware manufacturers design cards, which, whilst they're a physical single card, can be split up into logical cards that can be presented to guest operating systems. It essentially makes any hardware virtualization aware, and any of the advanced EC2 features that you'll come across within this course will be taking advantage of S-R-I-O-V.

At this point, though, we've completed all of the theory I wanted to cover, so go ahead, complete the slicing when you're ready. You can join me in the next.

Author response:

The following is the authors’ response to the original reviews

eLife Assessment

Examination of (a)periodic brain activity has gained particular interest in the last few years in the neuroscience fields relating to cognition, disorders, and brain states. Using large EEG/MEG datasets from younger and older adults, the current study provides compelling evidence that age-related differences in aperiodic EEG/MEG signals can be driven by cardiac rather than brain activity. Their findings have important implications for all future research that aims to assess aperiodic neural activity, suggesting control for the influence of cardiac signals is essential.

We want to thank the editors for their assessment of our work and highlighting its importance for the understanding of aperiodic neural activity. Additionally, we want to thank the three present and four former reviewers (at a different journal) whose comments and ideas were critical in shaping this manuscript to its current form. We hope that this paper opens up many more questions that will guide us - as a field - to an improved understanding of how “cortical” and “cardiac” changes in aperiodic activity are linked and want to invite readers to engage with our work through eLife’s comment function.

Public Reviews:

Reviewer #1 (Public review):

Summary:

The present study addresses whether physiological signals influence aperiodic brain activity with a focus on age-related changes. The authors report age effects on aperiodic cardiac activity derived from ECG in low and high-frequency ranges in roughly 2300 participants from four different sites. Slopes of the ECGs were associated with common heart variability measures, which, according to the authors, shows that ECG, even at higher frequencies, conveys meaningful information. Using temporal response functions on concurrent ECG and M/EEG time series, the authors demonstrate that cardiac activity is instantaneously reflected in neural recordings, even after applying ICA analysis to remove cardiac activity. This was more strongly the case for EEG than MEG data. Finally, spectral parameterization was done in large-scale resting-state MEG and ECG data in individuals between 18 and 88 years, and age effects were tested. A steepening of spectral slopes with age was observed particularly for ECG and, to a lesser extent, in cleaned MEG data in most frequency ranges and sensors investigated. The authors conclude that commonly observed age effects on neural aperiodic activity can mainly be explained by cardiac activity.

Strengths:

Compared to previous investigations, the authors demonstrate the effects of aging on the spectral slope in the currently largest MEG dataset with equal age distribution available. Their efforts of replicating observed effects in another large MEG dataset and considering potential confounding by ocular activity, head movements, or preprocessing methods are commendable and valuable to the community. This study also employs a wide range of fitting ranges and two commonly used algorithms for spectral parameterization of neural and cardiac activity, hence providing a comprehensive overview of the impact of methodological choices. Based on their findings, the authors give recommendations for the separation of physiological and neural sources of aperiodic activity.

Weaknesses:

While the aim of the study is well-motivated and analyses rigorously conducted, the overall structure of the manuscript, as it stands now, is partially misleading. Some of the described results are not well-embedded and lack discussion.

We want to thank the reviewer for their comments focussed on improving the overall structure of the manuscript. We agree with their suggestions that some results could be more clearly contextualized and restructured the manuscript accordingly.

Reviewer #2 (Public review):

I previously reviewed this important and timely manuscript at a previous journal where, after two rounds of review, I recommended publication. Because eLife practices an open reviewing format, I will recapitulate some of my previous comments here, for the scientific record.

In that previous review, I revealed my identity to help reassure the authors that I was doing my best to remain unbiased because I work in this area and some of the authors' results directly impact my prior research. I was genuinely excited to see the earlier preprint version of this paper when it first appeared. I get a lot of joy out of trying to - collectively, as a field - really understand the nature of our data, and I continue to commend the authors here for pushing at the sources of aperiodic activity!

In their manuscript, Schmidt and colleagues provide a very compelling, convincing, thorough, and measured set of analyses. Previously I recommended that the push even further, and they added the current Figure 5 analysis of event-related changes in the ECG during working memory. In my opinion this result practically warrants a separate paper its own!

The literature analysis is very clever, and expanded upon from any other prior version I've seen.

In my previous review, the broadest, most high-level comment I wanted to make was that authors are correct. We (in my lab) have tried to be measured in our approach to talking about aperiodic analyses - including adopting measuring ECG when possible now - because there are so many sources of aperiodic activity: neural, ECG, respiration, skin conductance, muscle activity, electrode impedances, room noise, electronics noise, etc. The authors discuss this all very clearly, and I commend them on that. We, as a field, should move more toward a model where we can account for all of those sources of noise together. (This was less of an action item, and more of an inclusion of a comment for the record.)

I also very much appreciate the authors' excellent commentary regarding the physiological effects that pharmacological challenges such as propofol and ketamine also have on non-neural (autonomic) functions such as ECG. Previously I also asked them to discuss the possibility that, while their manuscript focuses on aperiodic activity, it is possible that the wealth of literature regarding age-related changes in "oscillatory" activity might be driven partly by age-related changes in neural (or non-neural, ECG-related) changes in aperiodic activity. They have included a nice discussion on this, and I'm excited about the possibilities for cognitive neuroscience as we move more in this direction.

Finally, I previously asked for recommendations on how to proceed. The authors convinced me that we should care about how the ECG might impact our field potential measures, but how do I, as a relative novice, proceed. They now include three strong recommendations at the end of their manuscript that I find to be very helpful.

As was obvious from previous review, I consider this to be an important and impactful cautionary report, that is incredibly well supported by multiple thorough analyses. The authors have done an excellent job responding to all my previous comments and concerns and, in my estimation, those of the previous reviewers as well.

We want to thank the reviewer for agreeing to review our manuscript again and for recapitulating on their previous comments and the progress the manuscript has made over the course of the last ~2 years. The reviewer's comments have been essential in shaping the manuscript into its current form. Their feedback has made the review process truly feel like a collaborative effort, focused on strengthening the manuscript and refining its conclusions and resulting recommendations.

Reviewer #3 (Public review):

Summary:

Schmidt et al., aimed to provide an extremely comprehensive demonstration of the influence cardiac electromagnetic fields have on the relationship between age and the aperiodic slope measured from electroencephalographic (EEG) and magnetoencephalographic (MEG) data.

Strengths:

Schmidt et al., used a multiverse approach to show that the cardiac influence on this relationship is considerable, by testing a wide range of different analysis parameters (including extensive testing of different frequency ranges assessed to determine the aperiodic fit), algorithms (including different artifact reduction approaches and different aperiodic fitting algorithms), and multiple large datasets to provide conclusions that are robust to the vast majority of potential experimental variations.

The study showed that across these different analytical variations, the cardiac contribution to aperiodic activity measured using EEG and MEG is considerable, and likely influences the relationship between aperiodic activity and age to a greater extent than the influence of neural activity.

Their findings have significant implications for all future research that aims to assess aperiodic neural activity, suggesting control for the influence of cardiac fields is essential.

We want to thank the reviewer for their thorough engagement with our work and the resultant substantive amount of great ideas both mentioned in the section of Weaknesses and Authors Recommendations below. Their suggestions have sparked many ideas in us on how to move forward in better separating peripheral- from neuro-physiological signals that are likely to greatly influence our future attempts to better extract both cardiac and muscle activity from M/EEG recordings. So we want to thank them for their input, time and effort!

Weaknesses:

Figure 4I: The regressions explained here seem to contain a very large number of potential predictors. Based on the way it is currently written, I'm assuming it includes all sensors for both the ECG component and ECG rejected conditions?

I'm not sure about the logic of taking a complete signal, decomposing it with ICA to separate out the ECG and non-ECG signals, then including these latent contributions to the full signal back into the same regression model. It seems that there could be some circularity or redundancy in doing so. Can the authors provide a justification for why this is a valid approach?

After observing significant effects both in the MEG_{ECG component} and MEG_{ECG rejected} conditions in similar frequency bands we wanted to understand whether or not these age-related changes are statistically independent. To test this we added both variables as predictors in a regression model (thereby accounting for the influence of the other in relation to age). The regression models we performed were therefore actually not very complex. They were built using only two predictors, namely the data (in a specific frequency range) averaged over channels on which we noticed significant effects in the ECG rejected and ECG components data respectively (Wilkinson notation: age ~ 1 + ECG rejected + ECG components). This was also described in the results section stating that: “To see if MEG_{ECG rejected} and MEG_{ECG component} explain unique variance in aging at frequency ranges where we noticed shared effects, we averaged the spectral slope across significant channels and calculated a multiple regression model with MEG_{ECG component} and MEG_{ECG rejected} as predictors for age (to statistically control for the effect of MEG_{ECG component}s and MEG_{ECG rejected} on age). This analysis was performed to understand whether the observed shared age-related effects (MEG_{ECG rejected} and MEG_{ECG component}) are in(dependent).”  

We hope this explanation solves the previous misunderstanding.

I'm not sure whether there is good evidence or rationale to support the statement in the discussion that the presence of the ECG signal in reference electrodes makes it more difficult to isolate independent ECG components. The ICA algorithm will still function to detect common voltage shifts from the ECG as statistically independent from other voltage shifts, even if they're spread across all electrodes due to the referencing montage. I would suggest there are other reasons why the ICA might lead to imperfect separation of the ECG component (assumption of the same number of source components as sensors, non-Gaussian assumption, assumption of independence of source activities).

The inclusion of only 32 channels in the EEG data might also have reduced the performance of ICA, increasing the chances of imperfect component separation and the mixing of cardiac artifacts into the neural components, whereas the higher number of sensors in the MEG data would enable better component separation. This could explain the difference between EEG and MEG in the ability to clean the ECG artifact (and perhaps higher-density EEG recordings would not show the same issue).

The reviewer is making a good argument suggesting that our initial assumption that the presence of cardiac activity on the reference electrode influences the performance of the ICA may be wrong. After rereading and rethinking upon the matter we think that the reviewer is correct and that their assumptions for why the ECG signal was not so easily separable from our EEG recordings are more plausible and better grounded in the literature than our initial suggestion. We therefore now highlight their view as a main reason for why the ECG rejection was more challenging in EEG data. However, we also note that understanding the exact reason probably ends up being an empirical question that demands further research stating that:

“Difficulties in removing ECG related components from EEG signals via ICA might be attributable to various reasons such as the number of available sensors or assumptions related to the non-gaussianity of the underlying sources. Further understanding of this matter is highly important given that ICA is the most widely used procedure to separate neural from peripheral physiological sources. ”

In addition to the inability to effectively clean the ECG artifact from EEG data, ICA and other component subtraction methods have also all been shown to distort neural activity in periods that aren't affected by the artifact due to the ubiquitous issue of imperfect component separation (https://doi.org/10.1101/2024.06.06.597688). As such, component subtraction-based (as well as regression-based) removal of the cardiac artifact might also distort the neural contributions to the aperiodic signal, so even methods to adequately address the cardiac artifact might not solve the problem explained in the study. This poses an additional potential confound to the "M/EEG without ECG" conditions.

The reviewer is correct in stating that, if an “artifactual” signal is not always present but appears and disappears (like e.g. eye-blinks) neural activity may be distorted in periods where the “artifactual” signal is absent. However, while this plausibly presents a problem for ocular activity, there is no obvious reason to believe that this applies to cardiac activity. While the ECG signal is non-stationary in nature, it is remarkably more stable than eye-movements in the healthy populations we analyzed (especially at rest). Therefore, the presence of the cardiac “artifact” was consistently present across the entirety of the MEG recordings we visually inspected.

Literature Analysis, Page 23: was there a method applied to address studies that report reducing artifacts in general, but are not specific to a single type of artifact? For example, there are automated methods for cleaning EEG data that use ICLabel (a machine learning algorithm) to delete "artifact" components. Within these studies, the cardiac artifact will not be mentioned specifically, but is included under "artifacts".

The literature analysis was largely performed automatically and solely focussed on ECG related activity as described in the methods section under Literature Analysis, if no ECG related terms were used in the context of artifact rejection a study was flagged as not having removed cardiac activity. This could have been indeed better highlighted by us and we apologize for the oversight on our behalf. We now additionally link to these details stating that:

“However, an analysis of openly accessible M/EEG articles (N_Articles=279; see Methods - Literature Analysis for further details) that investigate aperiodic activity revealed that only 17.1% of EEG studies explicitly mention that cardiac activity was removed and only 16.5% measure ECG (45.9% of MEG studies removed cardiac activity and 31.1% of MEG studies mention that ECG was measured; see Figure 1EF).”

The reviewer makes a fair point that there is some uncertainty here and our results probably present a lower bound of ECG handling in M/EEG research as, when I manually rechecked the studies that were not initially flagged in studies it was often solely mentioned that “artifacts” were rejected. However, this information seemed too ambiguous to assume that cardiac activity was in fact accounted for. However, again this could have been mentioned more clearly in writing and we apologize for this oversight. Now this is included as part of the methods section Literature Analysis stating that:

“All valid word contexts were then manually inspected by scanning the respective word context to ensure that the removal of “artifacts” was related specifically to cardiac and not e.g. ocular activity or the rejection of artifacts in general (without specifying which “artifactual” source was rejected in which case the manuscript was marked as invalid). This means that the results of our literature analysis likely present a lower bound for the rejection of cardiac activity in the M/EEG literature investigating aperiodic activity.”

Statistical inferences, page 23: as far as I can tell, no methods to control for multiple comparisons were implemented. Many of the statistical comparisons were not independent (or even overlapped with similar analyses in the full analysis space to a large extent), so I wouldn't expect strong multiple comparison controls. But addressing this point to some extent would be useful (or clarifying how it has already been addressed if I've missed something).

In the present study we tried to minimize the risk of type 1 errors by several means, such as A) weakly informative priors, B) robust regression models and C) by specifying a region of practical equivalence (ROPE, see Methods Statistical Inference for further Information) to define meaningful effects.

Weakly informative priors can lower the risk of type 1 errors arising from multiple testing by shrinking parameter estimates towards zero (see e.g. Lemoine, 2019). Robust regression models use a Student T distribution to describe the distribution of the data. This distribution features heavier tails, meaning it allocates more probability to extreme values, which in turn minimizes the influence of outliers. The ROPE criterion ensures that only effects exceeding a negligible size are considered meaningful, representing a strict and conservative approach to interpreting our findings (see Kruschke 2018, Cohen, 1988).

Furthermore, and more generally we do not selectively report “significant” effects in the situations in which multiple analyses were conducted on the same family of data (e.g. Figure 2 & 4). Instead we provide joint inference across several plausible analysis options (akin to a specification curve analysis, Simonsohn, Simmons & Nelson 2020) to provide other researchers with an overview of how different analysis choices impact the association between cardiac and neural aperiodic activity.

Lemoine, N. P. (2019). Moving beyond noninformative priors: why and how to choose weakly informative priors in Bayesian analyses. Oikos, 128(7), 912-928.

Simonsohn, U., Simmons, J. P., & Nelson, L. D. (2020). Specification curve analysis. Nature Human Behaviour, 4(11), 1208-1214.

Methods:

Applying ICA components from 1Hz high pass filtered data back to the 0.1Hz filtered data leads to worse artifact cleaning performance, as the contribution of the artifact in the 0.1Hz to 1Hz frequency band is not addressed (see Bailey, N. W., Hill, A. T., Biabani, M., Murphy, O. W., Rogasch, N. C., McQueen, B., ... & Fitzgerald, P. B. (2023). RELAX part 2: A fully automated EEG data cleaning algorithm that is applicable to Event-Related-Potentials. Clinical Neurophysiology, result reported in the supplementary materials). This might explain some of the lower frequency slope results (which include a lower frequency limit <1Hz) in the EEG data - the EEG cleaning method is just not addressing the cardiac artifact in that frequency range (although it certainly wouldn't explain all of the results).

We want to thank the reviewer for suggesting this interesting paper, showing that lower high-pass filters may be preferable to the more commonly used >1Hz high-pass filters for detection of ICA components that largely contain peripheral physiological activity. However, the results presented by Bailey et al. contradict the more commonly reported findings by other researchers that >1Hz high-pass filter is actually preferable (e.g. Winkler et al. 2015; Dimingen, 2020 or Klug & Gramann, 2021) and recommendations in widely used packages for M/EEG analysis (e.g. https://mne.tools/1.8/generated/mne.preprocessing.ICA.html). Yet, the fact that there seems to be a discrepancy suggests that further research is needed to better understand which type of high-pass filtering is preferable in which situation. Furthermore, it is notable that all the findings for high-pass filtering in ICA component detection and removal that we are aware of relate to ocular activity. Given that ocular and cardiac activity have very different temporal and spectral patterns it is probably worth further investigating whether the classic 1Hz high-pass filter is really also the best option for the detection and removal of cardiac activity. However, in our opinion this requires a dedicated investigation on its own..

We therefore highlight this now in our manuscript stating that:

“Additionally, it is worth noting that the effectiveness of an ICA crucially depends on the quality of the extracted components(63,64) and even widely suggested settings e.g. high-pass filtering at 1Hz before fitting an ICA may not be universally applicable (see supplementary material of (64)).

Winkler, S. Debener, K. -R. Müller and M. Tangermann, "On the influence of high-pass filtering on ICA-based artifact reduction in EEG-ERP," 2015 37th Annual International Conference of the IEEE Engineering in Medicine and Biology Society (EMBC), Milan, Italy, 2015, pp. 4101-4105, doi: 10.1109/EMBC.2015.7319296.

Dimigen, O. (2020). Optimizing the ICA-based removal of ocular EEG artifacts from free viewing experiments. NeuroImage, 207, 116117.

Klug, M., & Gramann, K. (2021). Identifying key factors for improving ICA‐based decomposition of EEG data in mobile and stationary experiments. European Journal of Neuroscience, 54(12), 8406-8420.

It looks like no methods were implemented to address muscle artifacts. These can affect the slope of EEG activity at higher frequencies. Perhaps the Riemannian Potato addressed these artifacts, but I suspect it wouldn't eliminate all muscle activity. As such, I would be concerned that remaining muscle artifacts affected some of the results, particularly those that included high frequency ranges in the aperiodic estimate. Perhaps if muscle activity were left in the EEG data, it could have disrupted the ability to detect a relationship between age and 1/f slope in a way that didn't disrupt the same relationship in the cardiac data (although I suspect it wouldn't reverse the overall conclusions given the number of converging results including in lower frequency bands). Is there a quick validity analysis the authors can implement to confirm muscle artifacts haven't negatively affected their results?

I note that an analysis of head movement in the MEG is provided on page 32, but it would be more robust to show that removing ICA components reflecting muscle doesn't change the results. The results/conclusions of the following study might be useful for objectively detecting probable muscle artifact components: Fitzgibbon, S. P., DeLosAngeles, D., Lewis, T. W., Powers, D. M. W., Grummett, T. S., Whitham, E. M., ... & Pope, K. J. (2016). Automatic determination of EMG-contaminated components and validation of independent component analysis using EEG during pharmacologic paralysis. Clinical neurophysiology, 127(3), 1781-1793.

We thank the reviewer for their suggestion. Muscle activity can indeed be a potential concern, for the estimation of the spectral slope. This is precisely why we used head movements (as also noted by the reviewer) as a proxy for muscle activity. We also agree with the reviewer that this is not a perfect estimate. Additionally, also the riemannian potato would probably only capture epochs that contain transient, but not persistent patterns of muscle activity.

The paper recommended by the reviewer contains a clever approach of using the steepness of the spectral slope (or lack thereof) as an indicator whether or not an independent component (IC) is driven by muscle activity. In order to determine an optimal threshold Fitzgibbon et al. compared paralyzed to temporarily non paralyzed subjects. They determined an expected “EMG-free” threshold for their spectral slope on paralyzed subjects and used this as a benchmark to detect IC’s that were contaminated by muscle activity in non paralyzed subjects.

This is a great idea, but unfortunately would go way beyond what we are able to sensibly estimate with our data for the following reasons. The authors estimated their optimal threshold on paralyzed subjects for EEG data and show that this is a feasible threshold to be applied across different recordings. So for EEG data it might be feasible, at least as a first shot, to use their threshold on our data. However, we are measuring MEG and as alluded to in our discussion section under “Differences in aperiodic activity between magnetic and electric field recordings” the spectral slope differs greatly between MEG and EEG recordings for non-trivial reasons. Furthermore, the spectral slope even seems to also differ across different MEG devices. We noticed this when we initially tried to pool the data recorded in Salzburg with the Cambridge dataset. This means we would need to do a complete validation of this procedure for the MEG data recorded in Cambridge and in Salzburg, which is not feasible considering that we A) don’t have direct access to one of the recording sites and B) would even if we had access face substantial hurdles to get ethical approval for the experiment performed by Fitzgibbon et al..

However, we think the approach brought forward by Fitzgibbon and colleagues is a clever way to remove muscle activity from EEG recordings, whenever EMG was not directly recorded. We therefore suggested in the Discussion section that ideally also EMG should be recorded stating that:

“It is worth noting that, apart from cardiac activity, muscle activity can also be captured in (non-)invasive recordings and may drastically influence measures of the spectral slope(72). To ensure that persistent muscle activity does not bias our results we used changes in head movement velocity as a control analysis (see Supplementary Figure S9). However, it should be noted that this is only a proxy for the presence of persistent muscle activity. Ideally, studies investigating aperiodic activity should also be complemented by measurements of EMG. Whenever such measurements are not available creative approaches that use the steepness of the spectral slope (or the lack thereof) as an indicator to detect whether or not e.g. an independent component is driven by muscle activity are promising(72,73). However, these approaches may require further validation to determine how well myographic aperiodic thresholds are transferable across the wide variety of different M/EEG devices.”

Recommendations for the authors:

Reviewer #1 (Recommendations for the authors):

(1) As outlined above, I recommend rephrasing the last section of the introduction to briefly summarize/introduce all main analysis steps undertaken in the study and why these were done (for example, it is only mentioned that the Cam-CAN dataset was used to study the impact of cardiac on MEG activity although the author used a variety of different datasets). Similarly, I am missing an overview of all main findings in the context of the study goals in the discussion. I believe clarifying the structure of the paper would not only provide a red thread to the reader but also highlight the efforts/strength of the study as described above.

This is a good call! As suggested by the reviewer we now try to give a clearer overview of what was investigated why. We do that both at the end of the introduction stating that: “Using the publicly available Cam-CAN dataset(28,29), we find that the aperiodic signal measured using M/EEG originates from multiple physiological sources. In particular, significant portions of age-related changes in aperiodic activity –normally attributed to neural processes– can be better explained by cardiac activity. This observation holds across a wide range of processing options and control analyses (see Supplementary S1), and was replicable on a separate MEG dataset. However, the extent to which cardiac activity accounts for age-related changes in aperiodic activity varies with the investigated frequency range and recording site. Importantly, in some frequency ranges and sensor locations, age-related changes in neural aperiodic activity still prevail. But does the influence of cardiac activity on the aperiodic spectrum extend beyond age? In a preliminary analysis, we demonstrate that working memory load modulates the aperiodic spectrum of “pure” ECG recordings. The direction of this working memory effect mirrors previous findings on EEG data(5) suggesting that the impact of cardiac activity goes well beyond aging. In sum, our results highlight the complexity of aperiodic activity while cautioning against interpreting it as solely “neural“ without considering physiological influences.”

and at the beginning of the discussion section:

“Difficulties in removing ECG related components from EEG signals via ICA might be attributable to various reasons such as the number of available sensors or assumptions related to the non-gaussianity of the underlying sources. Further understanding of this matter is highly important given that ICA is the most widely used procedure to separate neural from peripheral physiological sources (see Figure 1EF). Additionally, it is worth noting that the effectiveness of an ICA crucially depends on the quality of the extracted components(63,64) and even widely suggested settings e.g. high-pass filtering at 1Hz before fitting an ICA may not be universally applicable (see supplementary material of (64)). “

(2) I found it interesting that the spectral slopes of ECG activity at higher frequency ranges (> 10 Hz) seem mostly related to HRV measures such as fractal and time domain indices and less so with frequency-domain indices. Do the authors have an explanation for why this is the case? Also, the analysis of the HRV measures and their association with aperiodic ECG activity is not explained in any of the method sections.

We apologize for the oversight in not mentioning the HRV analysis in more detail in our methods section. We added a subsection to the Methods section entitled ECG Processing - Heart rate variability analysis to further describe the HRV analyses.

“ECG Processing - Heart rate variability analysis

Heart rate variability (HRV) was computed using the NeuroKit2 toolbox, a high level tool for the analysis of physiological signals. First, the raw electrocardiogram (ECG) data were preprocessed, by highpass filtering the signal at 0.5Hz using an infinite impulse response (IIR) butterworth filter(order=5) and by smoothing the signal with a moving average kernel with the width of one period of 50Hz to remove the powerline noise (default settings of neurokit.ecg.ecg_clean). Afterwards, QRS complexes were detected based on the steepness of the absolute gradient of the ECG signal. Subsequently, R-Peaks were detected as local maxima in the QRS complexes (default settings of neurokit.ecg.ecg_peaks; see (98) for a validation of the algorithm). From the cleaned R-R intervals, 90 HRV indices were derived, encompassing time-domain, frequency-domain, and non-linear measures. Time-domain indices included standard metrics such as the mean and standard deviation of the normalized R-R intervals , the root mean square of successive differences, and other statistical descriptors of interbeat interval variability. Frequency-domain analyses were performed using power spectral density estimation, yielding for instance low frequency (0.04-0.15Hz) and high frequency (0.15-0.4Hz) power components. Additionally, non-linear dynamics were characterized through measures such as sample entropy, detrended fluctuation analysis and various Poincaré plot descriptors. All these measures were then related to the slopes of the low frequency (0.25 – 20 Hz) and high frequency (10 – 145 Hz) aperiodic spectrum of the raw ECG.”

With regards to association of the ECG’s spectral slopes at high frequencies and frequency domain indices of heart rate variability. Common frequency domain indices of heart rate variability fall in the range of 0.01-.4Hz. Which probably explains why we didn’t notice any association at higher frequency ranges (>10Hz).

This is also stated in the related part of the results section:

“In the higher frequency ranges (10 - 145 Hz) spectral slopes were most consistently related to fractal and time domain indices of heart rate variability, but not so much to frequency-domain indices assessing spectral power in frequency ranges < 0.4 Hz.”

(3) Related to the previous point - what is being reflected in the ECG at higher frequency ranges, with regard to biological mechanisms? Results are being mentioned, but not further discussed. However, this point seems crucial because the age effects across the four datasets differ between low and high-frequency slope limits (Figure 2C).

This is a great question that definitely also requires further attention and investigation in general (see also Tereshchenko & Josephson, 2015). We investigated the change of the slope across frequency ranges that are typically captured in common ECG setups for adults (0.05 - 150Hz, Tereshchenko & Josephson, 2015; Kusayama, Wong, Liu et al. 2020). While most of the physiological significant spectral information of an ECG recording rests between 1-50Hz (Clifford & Azuaje, 2006), meaningful information can be extracted at much higher frequencies. For instance, ventricular late potentials have a broader frequency band (~40-250Hz) that falls straight in our spectral analysis window. However, that’s not all, as further meaningful information can be extracted at even higher frequencies (>100Hz). Yet, the exact physiological mechanisms underlying so-called high-frequency QRS remain unclear (HF-QRS; see Tereshchenko & Josephson, 2015; Qiu et al. 2024 for a review discussing possible mechanisms). Yet, at the same time the HF-QRS seems to be highly informative for the early detection of myocardial ischemia and other cardiac abnormalities that may not yet be evident in the standard frequency range (Schlegel et al. 2004; Qiu et al. 2024). All optimism aside, it is also worth noting that ECG recordings at higher frequencies can capture skeletal muscle activity with an overlapping frequency range up to 400Hz (Kusayama, Wong, Liu et al. 2020). We highlight all of this now when introducing this analysis in the results sections as outstanding research question stating that:

“However, substantially less is known about aperiodic activity above 0.4Hz in the ECG. Yet, common ECG setups for adults capture activity at a broad bandwidth of 0.05 - 150Hz(33,34).

Importantly, a lot of the physiological meaningful spectral information rests between 1-50Hz(35), similarly to M/EEG recordings. Furthermore, meaningful information can be extracted at much higher frequencies. For instance, ventricular late potentials have a broader frequency band (~40-250Hz(35)). However, that’s not all, as further meaningful information can be extracted at even higher frequencies (>100Hz). For instance, the so-called high-frequency QRS seems to be highly informative for the early detection of myocardial ischemia and other cardiac abnormalities that may not yet be evident in the standard frequency range(36,37). Yet, the exact physiological mechanisms underlying the high-frequency QRS remain unclear (see (37) for a review discussing possible mechanisms). ”

Tereshchenko, L. G., & Josephson, M. E. (2015). Frequency content and characteristics of ventricular conduction. Journal of electrocardiology, 48(6), 933-937.

Kusayama, T., Wong, J., Liu, X. et al. Simultaneous noninvasive recording of electrocardiogram and skin sympathetic nerve activity (neuECG). Nat Protoc 15, 1853–1877 (2020). https://doi.org/10.1038/s41596-020-0316-6

Clifford, G. D., & Azuaje, F. (2006). Advanced methods and tools for ECG data analysis (Vol. 10). P. McSharry (Ed.). Boston: Artech house.

Qiu, S., Liu, T., Zhan, Z., Li, X., Liu, X., Xin, X., ... & Xiu, J. (2024). Revisiting the diagnostic and prognostic significance of high-frequency QRS analysis in cardiovascular diseases: a comprehensive review. Postgraduate Medical Journal, qgae064.

Schlegel, T. T., Kulecz, W. B., DePalma, J. L., Feiveson, A. H., Wilson, J. S., Rahman, M. A., & Bungo, M. W. (2004, March). Real-time 12-lead high-frequency QRS electrocardiography for enhanced detection of myocardial ischemia and coronary artery disease. In Mayo Clinic Proceedings (Vol. 79, No. 3, pp. 339-350). Elsevier.

(4) Page 10: At first glance, it is not quite clear what is meant by "processing option" in the text. Please clarify.

Thank you for catching this! Upon re-reading this is indeed a bit oblivious. We now swapped “processing options” with “slope fits” to make it clearer that we are talking about the percentage of effects based on the different slope fits.

(5) The authors mention previous findings on age effects on neural 1/f activity (References Nr 5,8,27,39) that seem contrary to their own findings such as e.g., the mostly steepening of the slopes with age. Also, the authors discuss thoroughly why spectral slopes derived from MEG signals may differ from EEG signals. I encourage the authors to have a closer look at these studies and elaborate a bit more on why these studies differ in their conclusions on the age effects. For example, Tröndle et al. (2022, Ref. 39) investigated neural activity in children and young adults, hence, focused on brain maturation, whereas the CamCAN set only considers the adult lifespan. In a similar vein, others report age effects on 1/f activity in much smaller samples as reported here (e.g., Voytek et al., 2015).

I believe taking these points into account by briefly discussing them, would strengthen the authors' claims and provide a more fine-grained perspective on aging effects on 1/f.

The reviewer is making a very important point. As age-related differences in (neuro-)physiological activity are not necessarily strictly comparable and entirely linear across different age-cohorts (e.g. age-related changes in alpha center frequency). We therefore, added the suggested discussion points to the discussion section.

“Differences in electric and magnetic field recordings aside, aperiodic activity may not change strictly linearly as we are ageing and studies looking at younger age groups (e.g. <22; (44) may capture different aspects of aging (e.g. brain maturation), than those looking at older subjects (>18 years; our sample). A recent report even shows some first evidence of an interesting putatively non-linear relationship with age in the sensorimotor cortex for resting recordings(59)”

(6) The analysis of the working memory paradigm as described in the outlook-section of the discussion comes as a bit of a surprise as it has not been introduced before. If the authors want to convey with this study that, in general, aperiodic neural activity could be influenced by aperiodic cardiac activity, I recommend introducing this analysis and the results earlier in the manuscript than only in the discussion to strengthen their message.

The reviewer is correct. This analysis really comes a bit out of the blue. However, this was also exactly the intention for placing this analysis in the discussion. As the reviewer correctly noted, the aim was to suggest “that, in general, aperiodic neural activity could be influenced by aperiodic cardiac activity”. We placed this outlook directly after the discussion of “(neuro-)physiological origins of aperiodic activity”, where we highlight the potential challenges of interpreting drug induced changes to M/EEG recordings. So the aim was to get the reader to think about whether age is the only feature affected by cardiac activity and then directly present some evidence that this might go beyond age.

However, we have been rethinking this approach based on the reviewers comments and moved that paragraph to the end of the results section accordingly and introduce it already at the end of the introduction stating that:

“But does the influence of cardiac activity on the aperiodic spectrum extend beyond age? In a preliminary analysis, we demonstrate that working memory load modulates the aperiodic spectrum of “pure” ECG recordings. The direction of this working memory effect mirrors previous findings on EEG data(5) suggesting that the impact of cardiac activity goes well beyond aging.”

(7) The font in Figure 2 is a bit hard to read (especially in D). I recommend increasing the font sizes where necessary for better readability.

We agree with the Reviewer and increased the font sizes accordingly.

(8) Text in the discussion: Figure 3B on page 10 => shouldn't it be Figure 4?

Thank you for catching this oversight. We have now corrected this mistake.

(9) In the third section on page 10, the Figure labels seem to be confused. For example, Figure 4 E is supposed to show "steepening effects", which should be Figure 4B I believe.

Please check the figure labels in this section to avoid confusion.

Thank you for catching this oversight. We have now corrected this mistake.

(10) Figure Legend 4 I), please check the figure labels in the text

Thank you for catching this oversight. We have now corrected this mistake.

Reviewer #3 (Recommendations for the authors):

I have a number of suggestions for improving the manuscript, which I have divided by section in the following:

ABSTRACT:

I would suggest re-writing the first sentences to make it easier to read for non-expert readers: "The power of electrophysiologically measured cortical activity decays with an approximately 1/fX function. The slope of this decay (i.e. the spectral exponent, X) is modulated..."

Thank you for the suggestion. We adjusted the sentence as suggested to make it easier for less technical readers to understand that “X” refers to the exponent.

Including the age range that was studied in the abstract could be informative.

Done as suggested.

As an optional recommendation, I think it would increase the impact of the article if the authors note in the abstract that the current most commonly applied cardiac artifact reduction approaches don't resolve the issue for EEG data, likely due to an imperfect ability to separate the cardiac artifact from the neural activity with independent component analysis. This would highlight to the reader that they can't just expect to address these concerns by cleaning their data with typical cleaning methods.

I think it would also be useful to convey in the abstract just how comprehensive the included analyses were (in terms of artifact reduction methods tested, different aperiodic algorithms and frequency ranges, and both MEG and EEG). Doing so would let the reader know just how robust the conclusions are likely to be.

This is a brilliant idea! As suggested we added a sentence highlighting that simply performing an ICA may not be sufficient to separate cardiac contributions to M/EEG recordings and refer to the comprehensiveness of the performed analyses.

INTRODUCTION:

I would suggest re-writing the following sentence for readability: "In the past, aperiodic neural activity, other than periodic neural activity (local peaks that rise above the "power-law" distribution), was often treated as noise and simply removed from the signal"

To something like: "In the past, aperiodic neural activity was often treated as noise and simply removed from the signal e.g. via pre-whitening, so that analyses could focus on periodic neural activity (local peaks that rise above the "power-law" distribution, which are typically thought to reflect neural oscillations).

We are happy to follow that suggestion.

Page 3: please provide the number of articles that were included in the examination of the percentage that remove cardiac activity, and note whether the included articles could be considered a comprehensive or nearly comprehensive list, or just a representative sample.

We stated the exact number of articles in the methods section under Literature Analysis. However, we added it to the Introduction on page 3 as suggested by the reviewer. The selection of articles was done automatically, dependent on a list of pre-specified terms and exclusively focussed on articles that had terms related to aperiodic activity in their title (see Literature Analysis). Therefore, I would personally be hesitant in calling it a comprehensive or nearly comprehensive list of the general M/EEG literature as the analysis of aperiodic activity is still relatively niche compared to the more commonly investigated evoked potentials or oscillations. I think whether or not a reader perceives our analysis as comprehensive should be up to them to decide and does not reflect something I want to impose on them. This is exacerbated by the fact that the analysis of neural aperiodic activity has rapidly gained traction over the last years (see Figure 1D orange) and the literature analysis was performed almost 2 years ago and therefore, in my eyes, only represents a glimpse in the rapidly evolving field related to the analysis of aperiodic activity.

Figure 1E-F: It's not completely clear that the "Cleaning Methods" part of the figure indicates just methods to clean the cardiac artifact (rather than any artifact). It also seems that ~40% of EEG studies do not apply any cleaning methods even from within the studies that do clean the cardiac artifact (if I've read the details correctly). This seems unlikely. Perhaps there should be a bar for "other methods", or "unspecified"? Having said that, I'm quite familiar with the EEG artifact reduction literature, and I would be very surprised if ~40% of studies cleaned the cardiac artifact using a different method to the methods listed in the bar graph, so I'm wondering if I've misunderstood the figure, or whether the data capture is incomplete / inaccurate (even though the conclusion that ICA is the most common method is almost certainly accurate).

The cleaning is indeed only focussed on cardiac activity specifically. This was however also mentioned in the caption of Figure 1: “We were further interested in determining which artifact rejection approaches were most commonly used to remove cardiac activity, such as independent component analysis (ICA(22)), singular value decomposition (SVD(23)), signal space separation (SSS(24)), signal space projections (SSP(25)) and denoising source separation (DSS(26)).” and in the methods section under Literature Analysis. However, we adjusted figure 1EF to make it more obvious that the described cleaning methods were only related to the ECG. Aside from using blind source separation techniques such as ICA a good amount of studies mentioned that they cleaned their data based on visual inspection (which was not further considered). Furthermore, it has to be noted that only studies were marked as having separated cardiac from neural activity, when this was mentioned explicitly.

RESULTS:

Page 6: I would delete the "from a neurophysiological perspective" clause, which makes the sentence more difficult to read and isn't so accurate (frequencies 13-25Hz would probably more commonly be considered mid-range rather than low or high). Additionally, both frequency ranges include 15Hz, but the next sentence states that the ranges were selected to avoid the knee at 15Hz, which seems to be a contradiction. Could the authors explain in more detail how the split addresses the 15Hz knee?

We removed the “from a neurophysiological perspective” clause as suggested. With regards to the “knee” at ~15Hz I would like to defer the reviewer to Supplementary Figure S1. The Knee Frequency varies substantially across subjects so splitting the data at only 1 exact Frequency did not seem appropriate. Additionally, we found only spurious significant age-related variations in Knee Frequency (i.e. only one out of the 4 datasets; not shown).

Furthermore, we wanted to better connect our findings to our MEG results in Figure 4 and also give the readers a holistic overview of how different frequency ranges in the aperiodic ECG would be affected by age. So to fulfill all of these objectives we decided to fit slopes with respective upper/lower bounds around a range of 5Hz above and below the average 15Hz Knee Frequency across datasets.

The later parts of this same paragraph refer to a vast amount of different frequency ranges, but only the "low" and "high" frequency ranges were previously mentioned. Perhaps the explanation could be expanded to note that multiple lower and upper bounds were tested within each of these low and high frequency windows?

This is a good catch we adjusted the sentence as suggested. We now write: “.. slopes were fitted individually to each subject's power spectrum in several lower (0.25 – 20 Hz) and higher (10-145 Hz) frequency ranges.”

The following two sentences seem to contradict each other: "Overall, spectral slopes in lower frequency ranges were more consistently related to heart rate variability indices(> 39.4% percent of all investigated indices)" and: "In the lower frequency range (0.25 - 20Hz), spectral slopes were consistently related to most measures of heart rate variability; i.e. significant effects were detected in all 4 datasets (see Figure 2D)." (39.4% is not "most").

The reviewer is correct in stating that 39.4% is not most. However, the 39.4% is the lowest bound and only refers to 1 dataset. In the other 3 datasets the percentage of effects was above 64% which can be categorized as “most” i.e. above 50%. We agree that this was a bit ambiguous in the sentence so we added the other percentages as well as a reference to Figure 2D to make this point clearer.

Figure 2D: it isn't clear what the percentages in the semi-circles reflect, nor why some semi-circles are more full circles while others are only quarter circles.

The percentages in the semi-circles reflect the amount of effects (marked in red) and null effects (marked in green) per dataset, when viewed as average across the different measures of HRV. Sometimes less effects were found for some frequency ranges resulting in quarters instead of semi circles.

Page 8: I think the authors could make it more clear that one of the conditions they were testing was the ECG component of the EEG data (extracted by ICA then projected back into the scalp space for the temporal response function analysis).

As suggested by the reviewer we adjusted our wording and replaced the arguably a bit ambiguous “... projected back separately” with “... projected back into the sensor space”. We thank the reviewer for this recommendation, as it does indeed make it easier to understand the procedure.

“After pre-processing (see Methods) the data was split in three conditions using an ICA(22). Independent components that were correlated (at r > 0.4; see Methods: MEG/EEG Processing - pre-processing) with the ECG electrode were either not removed from the data (Figure 3ABCD - blue), removed from the data (Figure 2ABCD - orange) or projected back into the sensor space (Figure 3ABCD - green).”

Figure 4A: standardized beta coefficients for the relationship between age and spectral slope could be noted to provide improved clarity (if I'm correct in assuming that is what they reflect).

This was indeed shown in Figure 4A and noted in the color bar as “average beta (standardized)”. We do not specifically highlight this in the text, because the exact coefficients would depend on both on the analyzed frequency range and the selected electrodes.

Figure 4I: The regressions explained at this point seems to contain a very large number of potential predictors, as I'm assuming it includes all sensors for both the ECG component and ECG rejected conditions? (if that is not the case, it could be explained in greater detail). I'm also not sure about the logic of taking a complete signal, decomposing it with ICA to separate out the ECG and non-ECG signals, then including them back into the same regression model. It seems that there could be some circularity or redundancy in doing so. However, I'm not confident that this is an issue, so would appreciate the authors explaining why it this is a valid approach (if that is the case).

After observing significant effects both in the MEG_{ECG component} and MEG_{ECG rejected} conditions in similar frequency bands we wanted to understand whether or not these age-related changes are statistically independent. To test this we added both variables as predictors in a regression model (thereby accounting for the influence of the other in relation to age). The regression models we performed were therefore actually not very complex. They were built using only two predictors, namely the data (in a specific frequency range) averaged over channels on which we noticed significant effects in the ECG rejected and ECG components data respectively (Wilkinson notation: age ~ 1 + ECG rejected + ECG components). This was also described in the results section stating that: “To see if MEG_{ECG rejected} and MEG_{ECG component} explain unique variance in aging at frequency ranges where we noticed shared effects, we averaged the spectral slope across significant channels and calculated a multiple regression model with MEG_{ECG component} and MEG_{ECG rejected} as predictors for age (to statistically control for the effect of MEG_{ECG component}s and MEG_{ECG rejected} on age). This analysis was performed to understand whether the observed shared age-related effects (MEG_{ECG rejected} and MEG_{ECG component}) are in(dependent).”  

We hope this explanation solves the previous misunderstanding.

The explanation of results for relationships between spectral slopes and aging reported in Figure 4 refers to clusters of effects, but the statistical inference methods section doesn't explain how these clusters were determined.

The wording of “cluster” was used to describe a “category” of effects e.g. null effects. We changed the wording from “cluster” to “category” to make this clearer stating now that: “This analysis, which is depicted in Figure 4, shows that over a broad amount of individual fitting ranges and sensors, aging resulted in a steepening of spectral slopes across conditions (see Figure 4E) with “steepening effects” observed in 25% of the processing options in MEG_{ECG not rejected} , 0.5% in MEG_{ECG rejected}, and 60% for MEG_{ECG components}. The second largest category of effects were “null effects” in 13% of the options for MEG_{ECG not rejected} , 30% in MEG_{ECG rejected}, and 7% for MEG_{ECG components}. ”

Page 12: can the authors clarify whether these age related steepenings of the spectral slope in the MEG are when the data include the ECG contribution, or when the data exclude the ECG? (clarifying this seems critical to the message the authors are presenting).

We apologize for not making this clearer. We now write: “This analysis also indicates that a vast majority of observed effects irrespective of condition (ECG components, ECG not rejected, ECG rejected) show a steepening of the spectral slope with age across sensors and frequency ranges.”

Page 13: I think it would be useful to describe how much variance was explained by the MEG-ECG rejected vs MEG-ECG component conditions for a range of these analyses, so the reader also has an understanding of how much aperiodic neural activity might be influenced by age (vs if the effects are really driven mostly by changes in the ECG).

With regards to the explained variance I think that the very important question of how strong age influences changes in aperiodic activity is a topic better suited for a meta analysis. As the effect sizes seems to vary largely depending on the sample e.g. for EEG in the literature results were reported at r=-0.08 (Cesnaite et al. 2023), r=-0.26 (Cellier et al. 2021), r=-0.24/r=-0.28/r=-0.35 (Hill et al. 2022) and r=0.5/r=0.7 (Voytek et al. 2015). I would defer the reader/reviewer to the standardized beta coefficients as a measure of effect size in the current study that is depicted in Figure 4A.

Cellier, D., Riddle, J., Petersen, I., & Hwang, K. (2021). The development of theta and alpha neural oscillations from ages 3 to 24 years. Developmental cognitive neuroscience, 50, 100969.

Cesnaite, E., Steinfath, P., Idaji, M. J., Stephani, T., Kumral, D., Haufe, S., ... & Nikulin, V. V. (2023). Alterations in rhythmic and non‐rhythmic resting‐state EEG activity and their link to cognition in older age. NeuroImage, 268, 119810.

Hill, A. T., Clark, G. M., Bigelow, F. J., Lum, J. A., & Enticott, P. G. (2022). Periodic and aperiodic neural activity displays age-dependent changes across early-to-middle childhood. Developmental Cognitive Neuroscience, 54, 101076.

Voytek, B., Kramer, M. A., Case, J., Lepage, K. Q., Tempesta, Z. R., Knight, R. T., & Gazzaley, A. (2015). Age-related changes in 1/f neural electrophysiological noise. Journal of Neuroscience, 35(38), 13257-13265.

Also, if there are specific M/EEG sensors where the 1/f activity does relate strongly to age, it would be worth noting these, so future research could explore those sensors in more detail.

I think it is difficult to make a clear claim about this for MEG data, as the exact location or type of the sensor may differ across manufacturers. Such a statement could be easier made for source projected data or in case EEG electrodes were available, where the location would be normed eg. according to the 10-20 system.

DISCUSSION:

Page 15: Please change the wording of the following sentence, as the way it is currently worded seems to suggest that the authors of the current manuscript have demonstrated this point (which I think is not the case): "The authors demonstrate that EEG typically integrates activity over larger volumes than MEG, resulting in differently shaped spectra across both recording methods."

Apologies for the oversight! The reviewer is correct we in fact did not show this, but the authors of the cited manuscript. We correct the sentence as suggested stating now that:

“Bénar et al. demonstrate that EEG typically integrates activity over larger volumes than MEG, resulting in differently shaped spectra across both recording methods.”

Page 16: The authors mention the results can be sensitive to the application of SSS to clean the MEG data, but not ICA. I think it would be sensitive to the application of either SSS or ICA?

This is correct and actually also supported by Figure S7, as differences in ICA thresholds affect also the detection of age-related effects. We therefore adjusted the related sentences stating now that:

“ In case of the MEG signal this may include the application of Signal-Space-Separation algorithms (SSS(24,55)), different thresholds for ICA component detection (see Figure S7), high and low pass filtering, choices during spectral density estimation (window length/type etc.), different parametrization algorithms (e.g. IRASA vs FOOOF) and selection of frequency ranges for the aperiodic slope estimation.”

It would be worth clarifying that the linked mastoid re-reference alone has been proposed to cancel out the ECG signal, rather than that a linked-mastoid re-reference improves the performance of the ICA separation (which could be inferred by the explanation as it's currently written).

This is correct and we adjusted the sentence accordingly! Stating now that:

“ Previous work(12,56) has shown that a linked mastoid reference alone was particularly effective in reducing the impact of ECG related activity on aperiodic activity measured using EEG. “

The issue of the number of EEG channels could probably just be noted as a potential limitation, as could the issue of neural activity being mixed into the ECG component (although this does pose a potential confound to the M/EEG without ECG condition, I suspect it wouldn't be critical).

This is indeed a very fair point as a higher amount of electrodes would probably make it easier to better isolate ECG components in the EEG, which may be the reason why the separation did not work so well in our case. However, this is ultimately an empirical question so we highlighted it in the discussion section stating that: “Difficulties in removing ECG related components from EEG signals via ICA might be attributable to various reasons such as the number of available sensors or assumptions related to the non-gaussianity of the underlying sources. Further understanding of this matter is highly important given that ICA is the most widely used procedure to separate neural from peripheral physiological sources. ”

OUTLOOK:

Page 19: Although there has been a recent trend to control for 1/f activity when examining oscillatory power, recent research suggests that this should only be implemented in specific circumstances, otherwise the correction causes more of a confound than the issue does. It might be worth considering this point with regards to the final recommendation in the Outlook section: Brake, N., Duc, F., Rokos, A., Arseneau, F., Shahiri, S., Khadra, A., & Plourde, G. (2024). A neurophysiological basis for aperiodic EEG and the background spectral trend. Nature Communications, 15(1), 1514.

We want to thank the reviewer for recommending this very interesting paper! The authors of said paper present compelling evidence showing that, while peak detection above an aperiodic trend using methods like FOOOF or IRASA is a prerequisite to determine the presence of oscillatory activity, it’s not necessarily straightforward to determine which detrending approach should be applied to determine the actual power of an oscillation. Furthermore, the authors suggest that wrongfully detrending may cause larger errors than not detrending at all. We therefore added a sentence stating that: “However, whether or not periodic activity (after detection) should be detrended using approaches like FOOOF or IRASA still remains disputed, as incorrectly detrending the data may cause larger errors than not detrending at all(75).”

RECOMMENDATIONS:

Page 20: "measure and account for" seems like it's missing a word, can this be re-written so the meaning is more clear?

Done as suggested. The sentence now states: “To better disentangle physiological and neural sources of aperiodic activity, we propose the following steps to (1) measure and (2) account for physiological influences.”

I would re-phrase "doing an ICA" to "reducing cardiac artifacts using ICA" (this wording could be changed in other places also).

I do not like to describe cardiac or ocular activity as artifactual per se. This is also why I used hyphens whenever I mention the word “artifact” in association with the ECG or EOG. However, I do understand that the wording of “doing an ICA” is a bit sloppy. We therefore reworded it accordingly throughout the manuscript to e.g. “separating cardiac from neural sources using an ICA” and “separating physiological from neural sources using an ICA”.

I would additionally note that even if components are identified as unambiguously cardiac, it is still likely that neural activity is mixed in, and so either subtracting or leaving the component will both be an issue (https://doi.org/10.1101/2024.06.06.597688). As such, even perfect identification of whether components are cardiac or not would still mean the issue remains (and this issue is also consistent across a considerable range of component based methods). Furthermore, current methods including wavelet transforms on the ICA component still do not provide good separation of the artifact and neural activity.

This is definitely a fair point and we also highlight this in our recommendations under 3 stating that:

“However, separating physiological from neural sources using an ICA is no guarantee that peripheral physiological activity is fully removed from the cortical signal. Even more sophisticated ICA based methods that e.g. apply wavelet transforms on the ICA components may still not provide a good separation of peripheral physiological and neural activity76,77. This turns the process of deciding whether or not an ICA component is e.g. either reflective of cardiac or neural activity into a challenging problem. For instance, when we only extract cardiac components using relatively high detection thresholds (e.g. r > 0.8), we might end up misclassifying residual cardiac activity as neural. In turn, we can’t always be sure that using lower thresholds won’t result in misinterpreting parts of the neural effects as cardiac. Both ways of analyzing the data can potentially result in misconceptions.”

Castellanos, N. P., & Makarov, V. A. (2006). Recovering EEG brain signals: Artifact suppression with wavelet enhanced independent component analysis. Journal of neuroscience methods, 158(2), 300-312.

Bailey, N. W., Hill, A. T., Godfrey, K., Perera, M. P. N., Rogasch, N. C., Fitzgibbon, B. M., & Fitzgerald, P. B. (2024). EEG is better when cleaning effectively targets artifacts. bioRxiv, 2024-06.

METHODS:

Pre-processing, page 24: I assume the symmetric setting of fastica was used (rather than the deflation setting), but this should be specified.

Indeed the reviewer is correct, we used the standard setting of fastICA implemented in MNE python, which is calling the FastICA implementation in sklearn that is per default using the “parallel” or symmetric algorithm to compute an ICA. We added this information to the text accordingly, stating that:

“For extracting physiological “artifacts” from the data, 50 independent components were calculated using the fastica algorithm(22) (implemented in MNE-Python version 1.2; with the parallel/symmetric setting; note: 50 components were selected for MEG for computational reasons for the analysis of EEG data no threshold was applied).”

Temporal response functions, page 26: can the authors please clarify whether the TRF is computed against the ECG signal for each electrode or sensory independently, or if all electrodes/sensors are included in the analysis concurrently? I'm assuming it was computed for each electrode and sensory separately, since the TRF was computed in both the forward and backwards direction (perhaps the meaning of forwards and backwards could be explained in more detail also - i.e. using the ECG to predict the EEG signal, or using the EEG signal to predict the ECG signal?).

A TRF can also be conceptualized as a multiple regression model over time lags. This means that we used all channels to compute the forward and backward models. In the case of the forward model we predicted the signal of the M/EEG channels in a multivariate regression model using the ECG electrode as predictor. In case of the backward model we predicted the ECG electrode based on the signal of all M/EEG channels. The forward model was used to depict the time window at which the ECG signal was encoded in the M/EEG recording, which appears at 0 time lags indicating volume conduction. The backward model was used to see how much information of the ECG was decodable by taking the information of all channels.

We tried to further clarify this approach in the methods section stating that:

“We calculated the same model in the forward direction (encoding model; i.e. predicting M/EEG data in a multivariate model from the ECG signal) and backward direction (decoding model; i.e. predicting the ECG signal using all M/EEG channels as predictors).”

Page 27: the ECG data was fit using a knee, but it seems the EEG and MEG data was not.

Does this different pose any potential confound to the conclusions drawn? (having said this, Figure S4 suggests perhaps a knee was tested in the M/EEG data, which should perhaps be explained in the text also).

This was indeed tested in a previous review round to ensure that our results are not dependent on the presence/absence of a knee in the data. We therefore added figure S4, but forgot to actually add a description in the text. We are sorry for this oversight and added a paragraph to S1 accordingly:

“Using FOOOF(5), we also investigated the impact of different slope fitting options (fixed vs. knee model fits) on the aperiodic age relationship (see Supplementary Figure S4). The results that we obtained from these analyses using FOOOF offer converging evidence with our main analysis using IRASA.”

Page 32: my understanding of the result reported here is that cleaning with ICA provided better sensitivity to the effects of age on 1/f activity than cleaning with SSS. Is this accurate? I think this could also be reported in the main manuscript, as it will be useful to researchers considering how to clean their M/EEG data prior to analyzing 1/f activity.

The reviewer is correct in stating that we overall detected slightly more “significant” effects, when not additionally cleaning the data using SSS. However, I am a bit wary of recommending omitting the use of SSS maxfilter solely based on this information. It can very well be that the higher quantity of effects (when not employing SSS maxfilter) stems from other physiological sources (e.g. muscle activity) that are correlated with age and removed when applying SSS maxfiltering. I think that just conditioning the decision of whether or not maxfilter is applied based on the amount or size of effects may not be the best idea. Instead I think that the applicability of maxfilter for research questions related to aperiodic activity should be the topic of additional methodological research. We therefore now write in Text S1:

“Considering that we detected less and weaker aperiodic effects when using SSS maxfilter is it advisable to omit maxfilter, when analyzing aperiodic signals? We don’t think that we can make such a judgment based on our current results. This is because it's unclear whether or not the reduction of effects stems from an additional removal of peripheral information (e.g. muscle activity; that may be correlated with aging) or is induced by the SSS maxfiltering procedure itself. As the use of maxfilter in detecting changes of aperiodic activity was not subject of analysis that we are aware of, we suggest that this should be the topic of additional methodological research.”

Page 39, Figure S6 and Figure S8: Perhaps the caption could also briefly explain the difference between maxfilter set to false vs true? I might have missed it, but I didn't gain an understanding of what varying maxfilter would mean.

Figure S6 shows the effect of ageing on the spectral slope averaged across all channels. The maxfilter set to false in AB) means that no maxfiltering using SSS was performed vs. in CD) where the data was additionally processed using the SSS maxfilter algorithm. We now describe this more clearly by writing in the caption:

“Supplementary Figure S6: Age-related changes in aperiodic brain activity are most prominent on explained by cardiac components irrespective of maxfiltering the data using signal space separation (SSS) or not AC) Age was used to predict the spectral slope (fitted at 0.1-145Hz) averaged across sensors at rest in three different conditions (ECG components not rejected [blue], ECG components rejected [orange], ECG components only [green].”

Welcome back to this lesson where I want to talk briefly about VPC Flow Logs, which are a useful networking feature of AWS VPCs, providing details of traffic flow within the private network. The most important thing to know about VPC Flow Logs is that they only capture packet metadata; they don't capture packet contents. If you need to capture the contents of packets, then you need a packet sniffer, something which you might install on an EC2 instance. So just to be really clear on this point, VPC Flow Logs only capture metadata, which means things like the source IP, the destination IP, the source and destination ports, packet size, and so on — anything which conceptually you could observe from outside, anything to do with the flow of data through the VPC.

Now Flow Logs work by attaching virtual monitors within a VPC and these can be applied at three different levels. We can apply them at the VPC level, which monitors every network interface in every subnet within that VPC; at the subnet level, which monitors every interface within that specific subnet; and directly to interfaces, where they only monitor that one specific network interface.

Now Flow Logs aren't real time — there's a delay between traffic entering or leaving monitored interfaces and showing up within VPC Flow Logs. This often comes up as an exam question, so this is something that you need to be aware of: you can't rely on Flow Logs to provide real-time telemetry on network packet flow, as there's a delay between that traffic flow occurring and that data showing up within the Flow Logs product.

Now Flow Logs can be configured to go to multiple destinations — currently this is S3 and CloudWatch Logs. It's a preference thing, and each of these comes with their own trade-offs. If you use S3, you're able to access the log files directly and can integrate that with either a third-party monitoring solution or something that you design yourself. If you use CloudWatch Logs, then obviously you can integrate that with other products, stream that data into different locations, and access it either programmatically or using the CloudWatch Logs console. So that's important — that distinction you need to understand for the exam.

You can also use Athena if you want to query Flow Logs stored in S3 using a SQL-like querying method. This is important if you have an existing data team and a more formal, rigorous review process of your Flow Logs. You can use Athena to query those logs in S3 and only pay for the amount of data read. Athena, remember, is an ad hoc querying engine which uses a schema-on-read architecture, so you're only billed for the data as it's read through the product and the data that's stored on S3 — that's critical to understand.

Now visually, this is how the Flow Logs product is architected. We start with a VPC with two subnets — a public one on the right in green and a private one on the left in blue. This architecture is running the Categorum application and this specific implementation has an application server in the public subnet, which is accessed by our user Bob. The application uses a database within the private subnet, which has a primary instance as well as a replicated standby instance.

Flow Logs can be captured, as I just mentioned, at a few different points — at the VPC level, at the subnet level, and directly on specific elastic network interfaces — and it's important to understand that Flow Logs capture from that point downwards. So any Flow Logs enabled at the VPC level will capture traffic metadata from every network interface in every subnet in that VPC; anything enabled at the subnet level is going to capture metadata for any network interfaces in that specific subnet, and so on.

Flow Logs can be configured to capture metadata on only accepted connections, only on rejected connections, or on all connections. Visually, this is an example of a Flow Log configuration at the network interface level — it captures metadata from the single elastic network interface of the application instance within the public subnet. If we created something at the subnet level, for example the private subnet, then metadata from both of the database instances is captured as part of that configuration. Anything captured can be sent to a destination, and the current options are S3 and CloudWatch Logs.

Now I'm going to be discussing this in detail in a moment, but the Flow Logs product captures what are known as Flow Log Records, and architecturally these look something like this. I'm going to be covering this next in detail — I'm going to step through all of the different fields just to give you a level of familiarity before you get the experience practically in a demo lesson. A VPC Flow Log is a collection of rows and each row has the following fields. All of the fields are important in different situations, but I've highlighted the ones that I find are used most often — source and destination IP address, source and destination port, the protocol, and the action.

Consider this example: Bob is running a ping against an application instance inside AWS. Bob sends a ping packet to the instance and it responds — this is a common way to confirm connectivity and to assess the latency, so this is a good indication of the performance between two different internet-connected services. The Flow Log for this particular interaction might look something like this — I've highlighted Bob's IP address in pink and the server's private IP address in blue. This shows outward traffic from Bob to the EC2 instance — remember the order: source and destination, and that’s for both the IP addresses and the port numbers. Normally you would have a source and destination port number directly after that, but this is ping, so ICMP, which doesn't use ports, so that’s empty.

The one highlighted in pink is the protocol number — ICMP is 1, TCP is 6, and UDP is 17. Now you don't really need to know this in detail for the exam, but it definitely will help you if you use VPC Flow Logs day to day, and it might feature as a point of elimination in an exam question, so do your best to remember the number for ICMP, TCP, and UDP.

The second to last item indicates if the traffic was accepted or rejected — this indicates if it was blocked or not by a security group or a network access control list. If it's a security group, then generally only one line will show in the Flow Logs — remember security groups are stateful, so if the request is allowed, then the response is automatically allowed in return. What you might see is something like this, where you have one Flow Log record which accepts traffic and then another which rejects the response to that conversation.

If you have an EC2 instance inside a subnet where the instance has a security group allowing pings from an external IP address, then the response will be automatically allowed. But if you have a network ACL on that instance's subnet which allows the ping inbound but doesn't allow it outbound, then it can cause a second line — a reject. It's important that you look out for both of these types of things in the exam, so if you see an accept and then a reject, and these look to be for the same flow of traffic, then you're going to be able to tell that both a security group and a network ACL are used and they're potentially restricting the flow of traffic between the source and the destination.

Flow Logs show the results of traffic flows as they're evaluated — security groups are stateful and so they only evaluate the conversation itself, which includes the request and the response, while network ACLs are stateless and consider traffic flows as two separate parts, request and response, both of which are evaluated separately, so you might see two log entries within VPC Flow Logs.

Now one thing before I finish up with this lesson: VPC Flow Logs don't log all types of traffic — there are some things which are excluded. This includes things such as the metadata service (so any accesses to the metadata service running inside the EC2 instance), time server requests, any DHCP requests which are running inside the VPC, and any communications with the Amazon Windows license server — obviously this applies only for Windows EC2 instances — so you need to be aware that certain types of traffic are not actually recorded using Flow Logs.

Now we are going to have some demos elsewhere in the course where you are going to get some practical experience of working with Flow Logs, but this is all of the theory which I wanted to introduce within this lesson. At this point go ahead and complete this video, and when you're ready, I'll look forward to you joining me in the next.

Generate one English language annotation for a provided text passage, using a simple annotation mode.

On a Friday afternoon, the school board and the teachers’ union reached an agreement, and students were told to return to school on Monday. [Resolution of conflict: End of strike/dispute] Two weeks of uncertainty, duality, and opposition had come to an end. [Duration of conflict] Some positions and relations had become hardened in the meantime—between the school board and the schools, the teachers and families, and the community among one another. [Consequences of conflict: Strained relationships] Remarks and political theater about who “won” and “lost,” who was “right” and “wrong,” were already circulating in the news, on social media, and among neighbors. [Post-conflict rhetoric and division]

Binaries again. [Observation: Reinforcement of opposing viewpoints]

But also, some unstuckness. [Observation: Potential for positive change]

A few days before the agreement was announced, the parent-teacher organizations of two schools extended an open invitation to community members to be in relation together in a public space. Approximately fifty families attended. [Positive community initiative: Dialogue and collaboration] This “invitation for dialogue” between families, parents who were teachers, parents who were school board members, and city councilors—some of whom were also parents in the school system—came to matter as it created the conditions for unpredictable newness and difference to emerge in a way that finally began to usher in hope. [Impact of initiative: Building bridges and fostering hope]

A children’s book can also be that invitational vibrant matter that brings people together and creates the conditions to be outside of some of the expired, stuck stories. [Metaphor: Children's books as agents of positive change] Children’s books can indeed be mirrors, windows, and sliding glass doors for readers, and they are also powerful agents in themselves, affecting space, a moment in time, and a community in visceral ways to become something different and new. [Concluding statement: The transformative power of children's literature]

Welcome back and this lesson will be one of a number of lessons in this section of the course where I'll be covering identity federation within AWS. Now identity federation is the process of using an identity from another identity provider to access AWS resources.

In AWS though, this can't be direct. Only AWS credentials can be used to access AWS resources. And so some form of exchange is required. And that's what I'll be covering in this lesson. So let's jump in and get started.

Now before we talk about the architecture of SAML 2.0 Identity Federation, let's talk about SAML itself. So SAML stands for Security Assertion Markup Language. And SAML 2.0 is version two of this standard.

Now it's an open standard which is used by many identity providers such as Microsoft with their Active Directory Federation Services known as ADFS, but many other on-premise identity providers utilize the SAML standard. And SAML 2.0 based federation allows you to indirectly use on-premises identities to access the AWS console and AWS command line interface.

Now I want to stress the important point here and that's the word indirectly. You can't access AWS resources using anything but AWS credentials. And so the process of federation within AWS involves exchanging or swapping these external identities for valid AWS credentials.

For the exam it's important that you both know how the architecture works as well as when you should select it versus the other identity federation options available within AWS.

Now SAML 2.0 based identity federation is used when you currently use an enterprise-based identity provider which is also SAML 2.0 compatible. Both of these need to be true. You wouldn't use it with a Google identity provider and you wouldn't use it with one which isn't SAML 2.0 compatible. So focus on those selection criteria for the exam.

Secondly, if you have an existing identity management team and you want to maintain that function allowing them to manage access to AWS as well, then SAML 2.0 based federation is ideal. Or if you're looking to maintain a single source of identity truth within your business and/or you have more than 5,000 users, then you should also look at using SAML 2.0 based federation.

Now if you read a question in the exam which mentions Google, Facebook, Twitter, Web or anything which suggests that SAML 2.0 is not supported, then this is not the right type of identity federation to use. So keep that in mind when you're reviewing exam questions. If it mentions any of those terms, you should probably assume that SAML 2.0 identity federation is not the right thing to select.

Now federation within AWS uses IAM roles and temporary credentials and in the case of SAML 2.0 based identity federation, these temporary credentials generally have up to a 12-hour validity. So keep that in mind for the exam.

Now at this point I want to step through the architectural flow of exactly how SAML 2.0 based identity federation works, both when you're accessing using the API or CLI as well as the console UI. So let's do that next and we'll start with API or CLI access.

So we start with an on-premises environment on the left using a SAML 2.0 compatible identity provider and identity store. And on the right is an AWS environment configured to allow SAML 2.0 based identity federation.

What this actually means from an infrastructure perspective is that an identity provider exists on the left side and a SAML identity provider has been created within IAM on the right and a two-way trust has been established between the two. So the instance of IAM on the right has been configured to trust the identity provider on-premises within the environment on the left.

Then we have an application, in this case the enterprise version of the Categorum application. And this is something that's developed internally within the Animals for Life organization and it initiates this process by communicating with the identity provider to request access.

Once this process is initiated, the identity provider accesses the identity store and authenticates the request, pulling a list of roles which the identity used by the Categorum application has access to. Inside the identity provider you've mapped identities onto one or more roles and one identity might be able to use multiple different roles. And so the Categorum application will have a selection from which to pick.

Now once this authentication process completes and the Categorum application has selected a role, what it gets back is what's known as a SAML assertion. And think of this as a token proving that the identity has authenticated. This SAML assertion is trusted by AWS.

Now the key concept to realize so far is that the application has initiated this process by communicating with the identity provider. This is an application-initiated process.

The next step is that the application communicates with AWS, specifically STS using the STS Assume Role with SAML operation. And it passes in the SAML assertion with that request. Now STS accepts the call and allows the role assumption. And this generates temporary AWS credentials which are returned to the application.

And this is another critical stage because what's happened here is the SAML assertion has essentially been exchanged for valid temporary AWS credentials. And remember only AWS credentials can be used to access AWS resources. And so the Categorum application can now use these temporary credentials to interact with AWS services such as DynamoDB.

So at a high level the process does require some upfront configuration. There needs to be a bidirectional trust created between IAM and the identity provider that's being used for this architecture. And once this trust has been established, AWS will respect these SAML assertions and use them as proof of authentication. And allow whatever identity is performing this process to gain access to temporary credentials which can be used to interact with AWS.

Now this process occurs behind the scenes. This is the architecture that's used if you're developing these applications in a bespoke way inside your business. So if you're a developer and you're looking to utilize identity federation to access AWS, this is the type of architecture that you'll use.

Now you can also use SAML 2.0 based identity federation to grant access to the AWS console for internal enterprise users. And on the whole it uses a very similar architectural flow. So let's have a look at that next.

When we're using SAML based identity federation to provide console access, we still have the on-premises environment on the left and AWS on the right. We still have the same identity provider, for example ADFS, but this time it's a user who wants to access the AWS console rather than an application interacting with AWS products and services.

There still needs to be a trust configured, this time between the identity provider and an SSO endpoint, also known as the AWS SAML endpoint. And this is configured within IAM inside the AWS account.

Now to begin this process, our user Bob browsers to our identity provider portal. And this might look something like this URL. So this URL is for the Animals for Life ADFS server. Bob browsers to this URL and he sees a portal which he needs to interact with.

Now when Bob loads this portal, before he can interact with it in any way, behind the scenes, the identity provider authenticates the request. Now this might mean explicitly logging in to the identity provider portal, or it might use the fact that you're already logged in to an active directory domain on your local laptop or workstation and use this authentication instead of asking you to log in again.

But in either case, you'll be presented with a list of roles that you can use based on your identity. So there might be an admin role and normal role or even an auditing role and all of these provide different permissions to AWS.

So once you've been authenticated or once you've selected a role, the identity provider portal returns a SAML assertion and instructions to point at a SAML endpoint that operates inside AWS.

Now once the client receives this information, it sends the SAML assertion to the SAML endpoint that you've configured in advance within AWS. And this assertion proves that you've authenticated as your identity and it provides details on the access rights that you should receive.

Now on your behalf in the background, the IAM role that you selected in the identity provider portal is now assumed using STS and the endpoint receives temporary security credentials on your behalf.

Then at this point it creates a sign-in URL for the AWS console which includes those credentials and it delivers those back to the client which the client then uses to access the AWS console UI.

So at a high level this is a fairly similar process. You're being authenticated by an identity provider that exists on premises. You're getting a SAML assertion. This is delivered to AWS. It's exchanged for temporary security credentials which are valid to interact with AWS resources.

The only difference in this case is that the SAML endpoint is constructing a URL which can be used to access the AWS console UI which includes those credentials. And then once that URL has been created it's passed back to your client and your client uses it to access the AWS management console. And this all happens behind the scenes without you having any visibility of it.

The key concept to understand and the really important thing about AWS Identity Federation is that you cannot use external identities directly to access AWS resources. They have to be exchanged for valid AWS credentials. And this is how that process happens when you're using a SAML 2.0 compatible identity provider.

Now that's all of the architectural theory that I wanted to cover in this lesson. In other lessons in this section of the course you're going to get some additional exposure to other types of identity federation within AWS. But SAML 2.0 based identity federation is the one that tends to be used in larger enterprises, especially those with a Windows based identity provider.

So that's everything I wanted to cover. So go ahead complete this video and when you're ready I'll look forward to you joining me in the next.

use for top display as well

Author response:

The following is the authors’ response to the original reviews

Public Reviews:

Reviewer #1 (Public review):

Summary:

This manuscript by Kaya et al. studies the effect of food consumption on hippocampal sharp wave ripples (SWRs) in mice. The authors use multiple foods and forms of food delivery to show that the frequency and power of SWRs increase following food intake, and that this effect depends on the caloric content of food. The authors also studied the effects of the administration of various food-intake-related hormones on SWRs during sleep, demonstrating that ghrelin negatively affects SWR rate and power, but not GLP1, insulin, or leptin. Finally, the authors use fiber photometry to show that GABAergic neurons in the lateral hypothalamus, increase activity during a SWR event.

Strengths:

The experiments in this study seem to be well performed, and the data are well presented, visually. The data support the main conclusions of the manuscript that food intake enhances hippocampal SWRs. Taken together, this study is likely to be impactful to the study of the impact of feeding on sleep behavior, as well as the phenomena of hippocampal SWRs in metabolism.

Weaknesses:

Details of experiments are missing in the text and figure legends. Additionally, the writing of the manuscript could be improved.

We thank the reviewer for their favorable assessment of the work and its potential impact. We have added all requested details in the text and figure legends and revised the wording of the manuscript to improve its clarity.

Reviewer #2 (Public review):

Summary:

Kaya et al uncover an intriguing relationship between hippocampal sharp wave-ripple production and peripheral hormone exposure, food intake, and lateral hypothalamic function. These findings significantly expand our understanding of hippocampal function beyond mnemonic processes and point a direction for promising future research.

Strengths:

Some of the relationships observed in this paper are highly significant. In particular, the inverse relationship between GLP1/Leptin and Insulin/Ghrelin are particularly compelling as this aligns well with opposing hormone functions on satiety.

Weaknesses:

I would be curious if there were any measurable behavioral differences that occur with different hormone manipulations.

We thank the reviewer for their favorable assessment of the work and its contribution to our understanding of non-mnemonic hippocampal function. Whether there are behavioral differences that occur following administration of the different hormones is a great question, yet unfortunately our study design did not include fine behavioral monitoring to the degree that would allow answering it. While some previous studies have partially addressed the behavioral consequences of the delivery of these hormones (and we reference these studies in our Discussion), how these changes may interact with the hippocampal and hypothalamic effects we observe is a very interesting next step.

Reviewer #3 (Public review):

Summary:

The manuscript by Kaya et al. explores the effects of feeding on sharp wave-ripples (SWRs) in the hippocampus, which could reveal a better understanding of how metabolism is regulated by neural processes. Expanding on prior work that showed that SWRs trigger a decrease in peripheral glucose levels, the authors further tested the relationship between SWRs and meal consumption by recording LFPs from the dorsal CA1 region of the hippocampus before and after meal consumption. They found an increase in SWR magnitude during sleep after food intake, in both food restricted and ad libitum fed conditions. Using fiber photometry to detect GABAergic neuron activity in the lateral hypothalamus, they found increased activity locked to the onset of SWRs. They conclude that the animal's satiety state modulates the amplitude and rate of SWRs, and that SWRs modulate downstream circuits involved in regulating feeding. These experiments provide an important step forward in understanding how metabolism is regulated in the brain. However, currently, the paper lacks sufficient analyses to control for factors related to sleep quality and duration; adding these analyses would further support the claim that food intake itself, as opposed to sleep quality, is primarily responsible for changes in SWR activity. Adding this, along with some minor clarifications and edits, would lead to a compelling case for SWRs being modulated by a satiety state. The study will likely be of great interest in the field of learning and memory while carrying broader implications for understanding brain-body physiology.

Strengths:

The paper makes an innovative foray into the emerging field of brain-body research, asking how sharp wave-ripples are affected by metabolism and hunger. The authors use a variety of advanced techniques including LFP recordings and fiber photometry to answer this question. Additionally, they perform comprehensive and logical follow-up experiments to the initial food-restricted paradigm to account for deeper sleep following meal times and the difference between consumption of calories versus the experience of eating. These experiments lay the groundwork for future studies in this field, as the authors pose several follow-up questions regarding the role of metabolic hormones and downstream brain regions.

We thank the reviewer for their appreciation and constructive review of the work.

Weaknesses:

Major comments:

(1) The authors conclude that food intake regulates SWR power during sleep beyond the effect of food intake on sleep quality. Specifically, they made an attempt to control for the confounding effect of delta power on SWRs through a mediation analysis. However, a similar analysis is not presented for SWR rate. Moreover, this does not seem to be a sufficient control. One alternative way to address this confound would be to subsample the sleep data from the ad lib and food restricted conditions (or high calorie and low calorie, etc), to match the delta power in each condition. When periods of similar mean delta power (i.e. similar sleep quality) are matched between datasets, the authors can then determine if a significant effect on SWR amplitude and rate remains in the subsampled data.

This is an important point that we believe we addressed in a few complementary ways. First, the mediation analysis we implemented measures the magnitude and significance of the contribution of food on SWR power after accounting for the effects of delta power, showing a highly significant food-SWR contribution. While the objective of subsampling is similar, mediation is a more statistically robust approach as it models the relationship between food, SWR power, and delta power in a way that explicitly accounts for the interdependence of these variables. Further, subsampling introduces the risk of losing statistical power by reducing the sample size, due to exclusion of data that might contain relevant and valuable information. Mediation analysis, on the other hand, uses the full dataset and retains statistical power while modeling the relationships between variables more holistically. However, as we were not satisfied with a purely analytical approach to test this issue, we carried out a new set of experiments in ad-libitum fed mice, where there is no concern of food restriction impairing sleep quality in the presleep session. In these conditions food amount also significantly correlated with, and showed significant mediation of, the SWR power change. Finally, we acknowledge and discuss this point in the Discussion, highlighting that given the known relationship between cortical delta and SWRs, it is challenging to fully disentangle these signals. 

(2) Relatedly, are the animals spending the same amount of time sleeping in the ad lib vs. food restricted conditions? The amount of time spent sleeping could affect the probability of entering certain stages of sleep and thus affect SWR properties. A recent paper (Giri et al., Nature, 2024) demonstrated that sleep deprivation can alter the magnitude and frequency of SWRs. Could the authors quantify sleep quantity and control for the amount of time spent sleeping by subsampling the data, similar to the suggestion above?

Following the reviewer’s comment, we have quantified and compared the amount of time spent in NREM sleep in the Pre and Post session pairs in which the animals were food restricted, with 0-1.5 g of chow given between the sleep sessions. We found that there was no significant difference in the amount of time spent in NREM sleep in the Pre and Post sessions. We have added this result to the Results section of the manuscript and as a new Supplementary Fig. 1. 

Additionally, we have added details to the Methods section that were missing in the original submission that are relevant to this point. Specifically, within the sleep sessions, the ongoing sleep states were scored using the AccuSleep toolbox (https://github.com/zekebarger/AccuSleep) using the EEG and EMG signals. NREM periods were detected based on high EEG delta power and low EMG power, REM periods were detected based on high EEG theta power and low EMG power, and Wake periods were detected based on high EMG power. Importantly, only NREM periods were included for subsequent SWR detection, quantification and analyses (in particular, reported SWR rates reflect the number of SWRs per second of NREM sleep). 

(3) Plot 5I only reports significance but does not clearly show the underlying quantification of LH GABAergic activity. Upon reading the methods for how this analysis was conducted, it would be informative to see a plot of the pre-SWR and post-SWR integral values used for the paired t-test whose p-values are currently shown. For example, these values could be displayed as individual points overlaid on a pair of boxand-whisker plots of the pre- and post-distribution within the session (perhaps for one example session per mouse with the p-value reported, to supplement a plot of the distribution of p-values across sessions and mice). If these data are non-normal, the authors should also use a non-parametric statistical test.

We have generated the summary plots the reviewer requested and have now included them in Supplementary Fig. 2. 

Minor comments:

(4) A brief explanation (perhaps in the discussion) of what each change in SWR property (magnitude, rate, duration) could indicate in the context of the hypothesis may be helpful in bridging the fields of metabolism and memory. For example, by describing the hypothesized mechanistic consequence of each change, could the authors speculate on why ripple rate may not increase in all the instances where ripple power increases after feeding? Why do the authors speculate that ripple duration does not increase, given that prior work (Fernandez-Ruiz et al. 2019) has shown that prolonged ripples support enhanced memory?

This is an interesting point and we have added a section to the Discussion to discuss it (pg. 17, last paragraph)

(5) The authors suggest that "SWRs could modulate peripheral metabolism" as a future implication of their work. However, the lack of clear effects from GLP-1, leptin and insulin complicates this interpretation. It might be informative for readers if the authors expanded their discussion of what specific role they speculate that SWRs could play in regulating metabolism, given these negative results.

We have added a section to the Discussion proposing potential reasons for this point (pg. 16, last paragraph)

Recommendations for the authors:  

Reviewer #1 (Recommendations for the authors):

Major Comments:

(1) The experiments involve very precise windows of time for sleeping and eating that seem impossible to control. For example, the authors state that for the experiments in Figure 1, there was a 2-h sleep period, followed by a 1-h feeding period, followed by another 2-h sleep period. Without sleep deprivation procedures or other environmental manipulations, how can these periods be so well-defined? Even during the inactive period, mice typically don't sleep for 2-h bouts at once, and the addition of food would not likely lead to an exact 1-h period of wakefulness in the middle. The validity of these experimental times would be more believable if the authors provided much more data on these sessions. For example, the authors could provide a table or visual display of data for the actual timing of the pre-sleep, eating, and post-sleep phases with exact time measurements and/or visual display of sleep versus wakefulness.

This is an important point, which we were not clear enough about in the original submission. While the durations of the Pre-sleep, Wake and Post-sleep sessions were indeed 2 h, 1 h and 2 h respectively, the animals did not actually sleep during the entirety of the sleep sessions. Importantly, we performed sleep state scoring on all sessions, and only analyzed identified NREM sleep for all SWR analyses. Following the reviewer’s comment (and that of Reviewer 1), we have quantified and compared the amount of time spent in NREM sleep in the Pre and Post session pairs in which the animals were food restricted and 0-1.5 g of chow were given between the sleep sessions. We found that there was no significant difference in the amount of time spent in NREM sleep in the Pre and Post sessions. We have added this result to the Results section of the manuscript and as a new Supplementary Fig. 1. 

Additionally, we have added details to the Methods section that were missing in the original submission that are relevant to this point. Specifically, within the sleep sessions, the ongoing sleep states were scored using the AccuSleep toolbox (https://github.com/zekebarger/AccuSleep) using the EEG and EMG signals. NREM periods were detected based on high EEG delta power and low EMG power, REM periods were detected based on high EEG theta power and low EMG power, and Wake periods were detected based on high EMG power. Importantly, only NREM periods were included for subsequent SWR detection, quantification and analyses (in particular, reported SWR rates reflect the number of SWRs per second of NREM sleep). 

(2) I may have missed this (although I tried searching in the text and figure legend), but the authors did not state the difference between green versus red bar colors in Figure 1 C-E. For Figures 1 F-J, do the individual dots represent both the test (fed) animals and control animals, or just the test animals?

We thank the reviewer for the opportunity to clarify these points. Red bars in Fig. 1C-E represent the SWR changes observed following delivery of equal or more than 0.5 g of chow, while the green bars represent the changes observed following delivery of less than 0.5 g. Fig. 1F-J includes both the experimental and control animals- the control animals appearing as having received 0 food amount. This information has now been added to the figure legend.

(3) For the jello experiments in Figure 3, was there only 1 trial per animal? Previous studies show that animals learn the caloric value of jello after subsequent trials, so whether or not multiple trials took place in each animal is important for interpretation of the results.

In Figure 3, the datapoints within each panel represent different animals and this information has now been added to the figure legend. Nevertheless, the animals were previously habituated to all foods, including regular jello, sugar-free jello and chocolate. While we consider it unlikely that this prior experience was sufficient to underlie the differential effects on SWRs, we cannot fully rule out the possibility that it provided some ability to predict the caloric value and consequences of the different foods. We have added details to the acknowledgement of this point in the Discussion (pg. 17, second paragraph).

(4) The experiments in Figure 5 are informative but don't relate to the experiments in the rest of the study. It is difficult to interpret their meaning given that these experiments take place over seconds while the other experiments take place over hours. Some attempt should be made to bridge these experiments over the timescales relevant for the behaviors studied in Figures 1-4.

We have now further acknowledged and discussed the point that our investigation is limited to the timescale of seconds around SWRs, and thus identified a potential communication channel, but whether and how this communication changes across hours following feeding remains for future studies (pg. 18, second paragraph).

(5) Figure 5B should depict the x-axis in seconds, not an arbitrary set of times from a recording.

We have replaced these with a time scale bar.

Minor Comments:

(6) The writing of the manuscript can be improved in many places:

Sometimes the writing could be more precise. For example, the Abstract states: "hippocampal sharp wave ripples (SWRs)... have been shown to influence peripheral glucose metabolism." Could this be written in a more informative way, rather than just staying "has been shown to influence?" A few more words would provide a lot more information. Similarly, at the end of the Introduction: "we set out to test the hypothesis that SWRs are modulated following meal times as part of the systems-level response to changing metabolic needs." This is not a strong hypothesis... could it be written to boldly state how the SWRs will be modulated (increase or decrease) and provide more assertive information?

The writing can be grandiose at times. Phrases such as "life is a continuous journey" or "the hypothalamus is a master regulator of homeostasis" are a bit sophomoric and too colloquial.

Finally, a representative recording should be referred to as just that-a "representative recording," as opposed to a "snippet," which is also colloquial. This word is used in the figure legends to Figures 1 and 5, and misspelled as "sinpper" in Figure 1

We have reworded all these sentences and phrases to make them clearer, more concrete and more formal.

(7) The methods state that the study used both male and female mice. Were they used in equal numbers across experiments?

Only one female was used in the final dataset, and we have corrected the wording accordingly.

Reviewer #2 (Recommendations for the authors):

Great paper!

Thanks!

Reviewer #3 (Recommendations for the authors):

Below are some minor requests for clarification, including in figures:

(1) Fig. 5H y-axis should say "normalized dF/F."

Done

(2) Fig. 1B is missing a y-axis label. It may be clearer to display separate y-axis scale bars for each component (SWR envelope, ripple-filtered amplitude, etc).

Done

(3) Please include labels for brain areas and methodological components in Fig. 5A.

Done

(4) Should Fig. 5B have the same y-axis or scale bars as 1B?

We have edited the figure labels and legends to be visually similar

(5) In Fig. 5J, is the y-axis a count of sessions?

Yes, we have added that to the y-axis label

(6) Could the authors please clarify whether the sugar-free jello was sweetened with an artificial sweetener? If so, this is a robust control for the rewarding nature of the two jellos, so a quick clarification would highlight this strength of the experiment.

We thank the reviewer for this great point. Indeed, the sugar free jello contained artificial sweeteners (Aspartame and Acesulfame Potassium). We have added this information to the Results and Methods.

(7) It appears in Fig. 5 that there may be a reliable dip in activity **at** the time of SWR onset, followed by the increase afterward, as shown in the example FP trace and the individual ripple-triggered traces. Is this indeed the case, and does this dip fall significantly below baseline? This characterization would be interesting, but I acknowledge is not necessarily crucial to the study to include.

This would indeed be an interesting finding, but upon examination and statistical testing, we found that this is not the case. We believe this may appear as such due to the normalization of the traces.

(8) The authors mention a reduction in ripple rate following insulin under food restriction as the only significant effect for insulin, GLP-1, and leptin, yet there was also a significant increase (at p<0.05) in ripple duration for GLP-1 in the ab lib condition. Is this not considered noteworthy?

This is a fair point and we have reworded the description of this result to simply state that there were no robust, consistent, dose-dependent effects of GLP-1, leptin and insulin on SWR attributes.

Voici un sommaire de la vidéo avec des indications temporelles basées sur le déroulement du contenu :

• Introduction (Début de la vidéo) : L'introduction est faite par Elena, la fondatrice de Toadhouse Games. Elle explique que ce tutoriel est conçu pour les débutants qui n'ont aucune connaissance en codage et que les premières vidéos seront gratuites sur YouTube. Elle présente Ren'Py comme un moteur de roman visuel utilisé par des milliers de créateurs.

• Qu'est-ce que Ren'Py ? (Environ 0:00 - 1:00) : Ren'Py est un moteur pour créer des romans visuels et de la fiction interactive. Bien qu'il fonctionne avec du code Python, il n'est pas nécessaire de savoir coder pour l'utiliser. Le logiciel fournit tout ce dont vous avez besoin, y compris des éditeurs de texte.

• Téléchargement de Ren'Py (Environ 1:00 - 2:00) : Il faut se rendre sur ri.org et cliquer sur le bouton de téléchargement. Différentes versions sont disponibles pour Windows, Mac, Linux, Android et iOS. Une fois le fichier téléchargé, il faut l'exécuter et extraire les fichiers dans le dossier de votre choix.

• Ouverture et présentation du lanceur Ren'Py (Environ 2:00 - 4:00) : Dans le dossier extrait, double-cliquez sur l'application Ren'Py (l'icône avec un anime) pour ouvrir le lanceur. Le lanceur affiche les projets ouverts (tutoriel et question par défaut) et les fichiers associés à chaque projet. Sur la droite, l'option "script" permet d'accéder aux fichiers de code, qui peuvent être édités dans un éditeur de texte comme Atom. Ren'Py peut télécharger et installer Atom pour vous.

• Exploration des fichiers du projet (Environ 4:00 - 5:00) : Le dossier "game" contient tous les fichiers du jeu (audio, musique, images, etc.). Un raccourci vers le dossier "images" est également disponible. Le fichier "script" contient le code du jeu, y compris les dialogues, les transitions, la musique et les scènes. Les options et les écrans (screens) permettent de personnaliser l'apparence du jeu.

• Construction et distribution du jeu (Environ 5:00 - 5:30) : L'option "build distributions" permet de créer une version jouable de votre jeu pour la partager avec d'autres sur différentes plateformes comme PC, Linux, Mac, itch.io ou Steam.

• Exercice pratique avec le projet "The Question" (Environ 5:30 - 8:00) : Il est recommandé de sélectionner le projet "the question" et de lancer le projet pour jouer au jeu. Ensuite, ouvrez le script du projet "the question". L'exercice consiste à jouer au jeu tout en regardant le code correspondant dans l'éditeur de texte. Cela permet de comprendre comment le code contrôle le déroulement du jeu (musique, scènes, dialogues, choix). Il est possible de faire de petites modifications dans le script et de recharger le jeu pour voir les changements.

• Présentation de Scrivener (Environ 8:00 - 9:00) : Scrivener est un logiciel optionnel qui peut être utilisé pour écrire le dialogue et organiser le contenu de votre roman visuel. Un modèle Ren'Py pour Scrivener créé par Toadhouse Games est disponible. Scrivener propose des conseils d'écriture de base et des modèles pour les profils de personnages et le code Ren'Py.

• Conclusion (Environ 9:00 - Fin de la vidéo) : Elena encourage les spectateurs à commencer à expérimenter avec Ren'Py en modifiant le projet "the question". Des tutoriels plus avancés sur les "flags" et les choix seront proposés ultérieurement. Des ressources d'aide sont disponibles sur Twitter, par e-mail (teamtoadhouse@gmail.com), sur les subreddits et les forums Ren'Py, ainsi que sur le Discord de Toadhouse Games.

for a country that is always talking about freedom and equality, this is super ironic. black people were belittled and hated on so much.

safari parks

Reviewer #2 (Public review):

Summary:

The current dataset utilized a 2x2 factorial shuttle-escape task in combination with extracellular single-unit recording in the anterior cingulate cortex (ACC) of mice to determine ACC action coding. The contributions of neocortical signaling to action-outcome learning as assessed by behavioral tasks outside of the prototypical reward versus non-reward or punished vs non-punished is an important and relevant research topic, given that ACC plays a clear role in several human neurological and psychiatric conditions. The authors present useful findings regarding the role of ACC in action monitoring and learning. The core methods themselves - electrophysiology and behavior - are adequate; however, the analyses are incomplete since ruling out alternative explanations for neural activity, such as movement itself, requires substantial control analyses, and details on statistical methods are not clear.

Strengths:

(1) The factorial design nicely controls for sensory coding and value coding, since the same stimulus can signal different actions and values.

(2) The figures are mostly well-presented, labeled, and easy to read.

(3) Additional analyses, such as the 2.5/7.5s windows and place-field analysis, are nice to see and indicate that the authors were careful in their neural analyses.

(4) The n-trial + 1 analysis where ACC activity was higher on trials that preceded correct responses is a nice addition, since it shows that ACC activity predicts future behavior, well before it happens.

(5) The authors identified ACC neurons that fire to shuttle crossings in one direction or to crossings in both directions. This is very clear in the spike rasters and population-scaled color images. While other factors such as place fields, sensory input, and their integration can account for this activity, the authors discuss this and provide additional supplemental analyses.

Weaknesses:

(1) The behavioral data could use slightly more characterization, such as separating stay versus shuttle trials.

(2) Some of the neural analyses could use the necessary and sufficient comparisons to strengthen the authors' claims.

(3) Many of the neural analyses seem to utilize long time windows, not leveraging the very real strength of recording spike times. Specifics on the exact neural activity binning/averaging, tests, classifier validation, and methods for quantification are difficult to find.

(4) The neural analyses seem to suggest that ACC neurons encode one variable or the other, but are there any that multiplex? Given the overwhelming evidence of multiplexing in the ACC a bit more discussion of its presence or absence is warranted.

I know that many people prefer different versions of software and various browsers. But is there any real significant difference between these pieces of software/browsers or are they all relatively the same? What make iOS so much different than what android uses? I often hear that Google Chrome is the best search engine to use. But why is this the case and are there other browsers that are just as good?

network operating system (NOS) provides basic functionalities like remote login and file transfer, allowing users to access resources on other machines within a network. It forms the foundation of many systems, enabling communication and resource sharing over a network. The system often requires users to manually interact with remote machines, using commands such as SSH for remote login or FTP for file transfers. While these systems offer basic functionality, they may lack the transparency and integration provided by distributed operating systems. Network operating systems, such as UNIX and Windows Server, often require explicit user commands to interact with remote systems, making them more complex for users. As technology advances, the lines between network and distributed systems blur, with modern network operating systems incorporating some features of distributed systems, such as cloud storage integration or remote access through APIs, which streamlines user interactions and enhances accessibility across distributed resources.

VMware Workstation is a widely used Type 2 hypervisor that enables multiple virtual machines to run on a host operating system like Linux or Windows. It abstracts hardware resources, allowing guest operating systems to function independently with dedicated virtual components such as CPU, memory, and disk. This virtualization enhances resource utilization, system efficiency, and disaster recovery. A key feature is the ability to copy guest operating system files, enabling easy backups and migrations. The architecture consists of a virtualization layer between the host OS and hardware, ensuring seamless integration of multiple environments while maintaining isolation and performance across virtual machines.

Type 2 hypervisors run as applications on a host operating system, making them less efficient but more accessible. They allow users to run different operating systems without modifying the host OS, making them useful for testing and educational purposes. However, they have limited hardware access and higher overhead due to the need to run within a general-purpose OS. Performance is lower compared to Type 0 and Type 1 hypervisors. The section highlights how Type 2 hypervisors are convenient for individual users but are rarely used in large-scale virtualization environments.

Type 1 hypervisors, also known as "bare-metal" hypervisors, function as specialized operating systems that manage guest VMs directly on hardware. They run in kernel mode, handling CPU scheduling, memory management, and I/O operations while offering APIs for external tools. They improve data-center efficiency by consolidating workloads, enabling live migration, and simplifying backup and replication. However, their complexity and cost, especially for commercial options like VMware ESX, pose challenges. Some general-purpose operating systems, such as Red Hat Enterprise Linux, incorporate Type 1 hypervisors, treating VMs like processes but providing special execution handling.

Virtual machines first became commercially available with IBM VM/370 in 1972, enabling mainframe partitioning into multiple VMs. A major challenge was disk allocation, which IBM resolved using minidisks, allowing virtual machines to have dedicated yet flexible storage. Initially, virtualization was exclusive to IBM, but broader adoption emerged when Intel 80x86 CPUs advanced in the 1990s. Technologies like Xen and VMware enabled virtualization on commodity hardware, making it mainstream. Today, projects like VirtualBox support multiple architectures and operating systems. The core principles—fidelity, performance, and safety—continue guiding virtualization developments, reinforcing its role in modern computing.

There are multiple types of VMMs, categorized based on how they integrate with hardware and software:

Type 0 Hypervisors: Implemented in firmware, commonly used in enterprise servers (e.g., IBM LPARs, Oracle LDOMs). Type 1 Hypervisors: Installed directly on hardware, functioning as an OS substitute (e.g., VMware ESX, Citrix XenServer, Joyent SmartOS). Type 2 Hypervisors: Run as applications on an existing OS (e.g., VMware Workstation, Parallels, VirtualBox). Paravirtualization: Guest OS is modified for optimized interaction with the hypervisor. Programming-Environment Virtualization: Uses a custom virtual system instead of real hardware (e.g., Java Virtual Machine (JVM), Microsoft .NET). Emulators: Allow applications for one architecture to run on a different CPU type. Application Containment: Not true virtualization but isolates applications for security and management (e.g., Solaris Zones, BSD Jails).

To improve file system performance, storage devices use on-board caches to store tracks or blocks, reducing delays. The operating system also caches blocks in main memory for faster access. Some systems use a buffer cache for frequently used blocks, while others use a page cache that stores file data as virtual memory pages, making access more efficient. Systems like Solaris, Linux, and Windows use unified virtual memory to cache both process pages and file data.

Some operating systems, like UNIX, treat directories as special types of files, while others, like Windows, handle files and directories separately with different system calls. Regardless of the approach, the logical file system maps directory operations to storage blocks, which are then managed by the basic file system and I/O control system.

Performance optimization involves caching strategies, memory management, and efficient write operations. Modern storage controllers use on-board caches, while operating systems employ buffer caches or page caches to reduce I/O latency. Unified caches prevent redundant data storage, minimizing memory waste and CPU overhead. Solaris and Linux leverage unified virtual memory for efficiency. Least Recently Used (LRU) algorithms manage cache replacement, but their implementation varies. Asynchronous writes improve speed by allowing immediate process continuation, whereas synchronous writes ensure data integrity. Balancing caching, write strategies, and metadata updates is crucial for maximizing storage performance while maintaining reliability.

File types define their structure, which some operating systems enforce for proper execution. While supporting multiple file structures adds flexibility, it can make the system complex, so many OSs, like UNIX, treat files as simple byte sequences, leaving interpretation to applications.

The example of file locking in Java demonstrates how developers can manage concurrent file access. Java’s FileChannel class allows acquiring locks on specific portions of a file, supporting both shared and exclusive locks. This is particularly useful in applications that involve collaborative editing or log file management. However, improper use of file locks can lead to deadlocks, where two processes indefinitely wait for each other to release a lock. This example reinforces the importance of synchronization in file handling. Additionally, understanding how different operating systems implement file locking (such as Windows using mandatory locks and UNIX using advisory locks) can help developers write cross-platform applications that handle concurrency efficiently.

File structures help the operating system and applications organize and manage data effectively. Some files, like source and object files, follow a defined structure expected by the programs that use them. Operating systems often impose structures on certain files, such as executable files, to facilitate their loading and execution. However, supporting multiple file structures increases system complexity and can restrict flexibility. Some systems, like UNIX, adopt a minimal approach, treating all files as byte sequences without enforcing structure. This approach enhances flexibility but places the burden on applications to interpret data. A rigid file structure system can create issues when new types of files, such as encrypted files, do not conform to predefined formats, requiring workarounds or alternative storage methods.

Unlike Unix, Windows manages ACLs through a graphical user interface, making it more user-friendly. The section highlights an example from Windows 7 (NTFS file system) where a user named "guest" is explicitly denied access to a file. Windows permissions allow detailed configurations, such as granting or restricting access based on individual users or groups. This flexibility is useful in corporate environments where different departments require different levels of access. However, managing ACLs in Windows can be complex due to overlapping permissions and precedence rules. The section also discusses how operating systems prioritize permissions when conflicts arise, with Solaris and other Unix-like systems granting precedence to ACLs over standard permissions. This approach ensures that more specific rules override general access settings.

Shared Memory in the Windows API Windows provides built-in support for memory-mapped files to enable shared memory between processes. The process begins with the creation of a file mapping using CreateFile() and CreateFileMapping(), followed by mapping a view of the file into a process’s address space with MapViewOfFile(). A second process can then access the same memory-mapped file, allowing seamless data sharing. This mechanism is particularly useful for producer-consumer scenarios, where one process writes to shared memory while another reads from it. The example code in this section demonstrates how a producer writes a message to a shared-memory object, which a consumer then reads. This implementation avoids the overhead of traditional IPC methods like pipes or message queues, leveraging efficient memory management instead.

Just to be clear, I would change it to the below (or something similar) just in case people think they only need to do it on the control machine.

During installation, it is also recommended to add the Nuke Stage Listener shortcut to your Windows startup app on all machines used for Nuke Stage.

We only ship on Windows.

Odd examples.

Reviewer #1 (Public review):

Summary:

The authors of this study set out to find RNA binding proteins in the CNS in cell-type specific sequencing data and discover that the cardiomyopathy-associated protein RBM20 is selectively expressed in olfactory bulb glutamatergic neurons and PV+ GABAergic neurons. They make an HA-tagged RBM20 allele to perform CLIP-seq to identify RBM20 binding sites and find direct targets of RBM20 in olfactory bulb glutmatergic neurons. In these neurons, RBM20 binds intronic regions. RBM20 has previously been implicated in splicing, but when they selectively knockout RBM20 in glutamatergic neurons they do not see changes in splicing, but they do see changes in RNA abundance, especially of long genes with many introns, which are enriched for synapse-associated functions. These data show that RBM20 has important functions in gene regulation in neurons, which was previously unknown, and they suggest it acts through a mechanism distinct from what has been studied before in cardiomyocytes.

Strengths:

The study finds expression of the cardiomyopathy-associated RNA binding protein RBM20 in specific neurons in the brain, opening new windows into its potential functions there.

The study uses CLIP-seq to identify RBM20 binding RNAs in olfactory bulb neurons.

Conditional knockout of RBM20 in glutamatergic or PV neurons allows the authors to detect mRNA expression that is regulated by RBM20.

The data include substantial controls and quality control information to support the rigor of the findings.

Weaknesses:

The authors do not fully identify the mechanism by which RBM20 acts to regulate RNA expression in neurons, though they do provide data suggesting that neuronal RBM20 does not regulate alternate splicing in neurons, which is an interesting contrast to its proposed mechanism of function in cardiomyocytes. Discovery of the RNA regulatory functions of RBM20 in neurons is left as a question for future studies.

The study does not identify functional consequences of the RNA changes in the conditional knockout cells, so this is also a question for the future.

Reviewer #2 (Public review):

van Vliet and colleagues present results of a study correlating internal states of a convolutional neural network trained on visual word stimuli with evoked MEG potentials during reading.

In this study, a standard deep learning image recognition model (VGG-11) trained on a large natural image set (ImageNet) that begins illiterate but is then further trained on visual word stimuli, is used on a set of predefined stimulus images to extract strings of characters from "noisy" words, pseudowords and real words. This methodology is used in hopes of creating a model which learns to apply the same nonlinear transforms that could be happening in different regions of the brain - which would be validated by studying the correlations between the weights of this model and neural responses. Specifically, the aim is that the model learns some vector embedding space, as quantified by the spread of activations across a layer's weights (L2 Norm prior to ReLu Activation Function), for the different kinds of stimuli, that creates a parameterized decision boundary that is similar to amplitude changes at different times for a MEG signal. More importantly, the way that the stimuli are ordered or ranked in that space should be separable to the degree we see separation in neural activity. This study does show that the layer weights corresponding to five different broad classes of stimuli do statistically correlate with three specific components in the ERP. However, I believe there are fundamental theoretical issues that limit the implications of the results of this study.

As has been shown over many decades, there are many potential computational algorithms, with varied model architectures, that can perform the task of text recognition from an image. However, there is no evidence presented here that this particular algorithm has comparable performance to human behavior (i.e. similar accuracy with a comparable pattern of mistakes). This is a fundamental prerequisite before attempting to meaningfully correlate these layer activations to human neural activations. Therefore, it is unlikely that correlating these derived layer weights to neural activity provides meaningful novel insights into neural computation beyond what is seen using traditional experimental methods.

One example of a substantial discrepancy between this model and neural activations is that, while incorporating frequency weighting into the training data is shown to slightly increase neural correlation with the model, Figure 7 shows that no layer of the model appears directly sensitive to word frequency. This is in stark contrast to the strong neural sensitivity to word frequency seen in EEG (e.g. Dambacher et al 2006 Brain Research), fMRI (e.g. Kronbichler et al 2004 NeuroImage), MEG (e.g. Huizeling et al 2021 Neurobio. Lang.), and intracranial (e.g. Woolnough et al 2022 J. Neurosci.) recordings. Figure 7 also demonstrates that late stages of the model show a strong negative correlation with font size, whereas later stages of neural visual word processing are typically insensitive to differences in visual features, instead showing sensitivity to lexical factors.

Another example of the mismatch between this model and visual cortex is the lack of feedback connections in the model. Within visual cortex there are extensive feedback connections, with later processing stages providing recursive feedback to earlier stages. This is especially evident in reading, where feedback from lexical level processes feeds back to letter level processes (e.g. Heilbron et al 2020 Nature Comms.). This feedback is especially relevant for reading of words in noisy conditions, as tested in the current manuscript, as lexical knowledge enhances letter representation in visual cortex (the word superiority effect). This results in neural activity in multiple cortical areas varying over time, changing selectivity within a region at different measured time points (e.g. Woolnough et al 2021 Nature Human Behav.), which in the current study is simplified down to three discrete time windows, each attributed to different spatial locations.

The presented model needs substantial further development to be able to replicate, both behaviorally and neurally, many of the well-characterized phenomena seen in human behavior and neural recordings that are fundamental hallmarks of human visual word processing. Until that point it is unclear what novel contributions can be gleaned from correlating low dimensional model weights from these computational models with human neural data.

The revised version of this manuscript has not addressed these concerns.

Author response:

The following is the authors’ response to the original reviews.

We thank the reviewers for their efforts. They have pointed out several shortcomings and made very helpful suggestions. Based on their feedback, we have substantially revised the manuscript and feel the paper has been much improved because of it.

Notable changes are:

(1) As our model does not contain feed-back connections, the focus of the study is now more clearly communicated to be on feed-forward processes only, with appropriate justifications for this choice added to the Introduction and Discussion sections. Accordingly, the title has been changed to include the term “feed-forward”.

(2) The old Figure 5 has been removed in favor of reporting correlation scores to the right of the response profiles in other figures.

(3) We now discuss changes to the network architecture (new Figure 5) and fine-tuning of the hyperparameters (new Figure 6) in the main text instead of only the Supplementary Information.

(4) The discussion on qualitative versus quantitative analysis has been extended and given its own subsection entitled “On the importance of experimental contrasts and qualitative analysis of the model”.

Below, we address each point that the reviewers brought up in detail and outline what improvements we have made in the revision to address them.

Reviewer #1 (Public Review):

Summary:

This study trained a CNN for visual word classification and supported a model that can explain key functional effects of the evoked MEG response during visual word recognition, providing an explicit computational account from detection and segmentation of letter shapes to final word-form identification.

Strengths:

This paper not only bridges an important gap in modeling visual word recognition, by establishing a direct link between computational processes and key findings in experimental neuroimaging studies, but also provides some conditions to enhance biological realism.

Weaknesses:

The interpretation of CNN results, especially the number of layers in the final model and its relationship with the processing of visual words in the human brain, needs to be further strengthened.

We have experimented with the number of layers and the number of units in each layer. In the previous version of the manuscript, these results could be found in the supplementary information. For the revised version, we have brought some of these results into the main text and discuss them more thoroughly.

We have added a figure (Figure 5 in the revised manuscript) showing the impact of the number of convolution and fully-connected layers on the response profiles of the layers, as well as the correlation with the three MEG components.

We discuss the figure in the Results section as follows:

“Various variations in model architecture and training procedure were evaluated. We found that the number of layers had a large impact on the response patterns produced by the model (Figure 5). The original VGG-11 architecture defines 5 convolution layers and 3 fully connected layers (including the output layer). Removing a convolution layer (Figure 5, top row), or removing one of the fully connected layers (Figure 5, second row), resulted in a model that did exhibit an enlarged response to noisy stimuli in the early layers that mimics the Type-I response. However, such models failed to show a sufficiently diminished response to noisy stimuli in the later layers, hence failing to produce responses that mimic the Type-II or N400m, a failure which also showed as low correlation scores.

Adding an additional convolution layer (Figure 5, third row) resulted in a model where none of the layer response profiles mimics that of the Type-II response. The Type-II response is characterized by a reduced response to both noise and symbols, but an equally large response to consonant strings, real and pseudo words. However, in the model with an additional convolution layer, the consonant strings evoked a reduced response already in the first fully connected layer, which is a feature of the N400m rather than the Type-II. These kind of subtleties in the response pattern, which are important for the qualitative analysis, generally did not show quantitatively in the correlation scores, as the fully connected layers in this model correlate as well with the Type-II response as models that did show a response pattern that mimics the Type-II.

Adding an additional fully connected layer (Figure 5, fourth row) resulted in a model with similar response profiles and correlation with the MEG components as the original VGG-11 architecture (Figure 5, bottom row) The N400m-like response profile is now observed in the third fully connected layer rather than the output layer. However, the decrease in response to consonant strings versus real and pseudo words, which is typical of the N400m, is less distinct than in the original VGG-11 architecture.”

And in the Discussion section:

“In the model, convolution units are followed by pooling units, which serve the purpose of stratifying the response across changes in position, size and rotation within the receptive field of the pooling unit. Hence, the effect of small differences in letter shape, such as the usage of different fonts, was only present in the early convolution layers, in line with findings in the EEG literature (Chauncey et al., 2008; Grainger & Holcomb, 2009; Hauk & Pulvermüller, 2004). However, the ability of pooling units to stratify such differences depends on the size of their receptive field, which is determined by the number of convolution-and-pooling layers. As a consequence, the response profiles of the subsequent fully connected layers was also very sensitive to the number of convolution-and-pooling layers. The optimal number of such layers is likely dependent on the input size and pooling strategy. Given the VGG-11 design of doubling the receptive field after each layer, combined with an input size of 225×225 pixels, the optimal number of convolution-andpooling layers for our model was five, or the model would struggle to produce response profiles mimicking those of the Type-II component in the subsequent fully connected layers (Figure 5).”

Reviewer #1 (Recommendations For The Authors):

(1) The similarity between CNNs and human MEG responses, including type-I (100ms), type-II (150ms), and N400 (400ms) components, looks like separately, lacking the sequential properties among these three components. Is the recurrent neural network (RNN), which can be trained to process and convert a sequential data input into a specific sequential data output, a better choice?

When modeling sequential effects, meaning that the processing of the current word is influenced by the word that came before it, such as priming and top-down modulations, we agree that such a model would indeed require recurrency in its architecture. However, we feel that the focus of modeling efforts in reading has been overwhelmingly on the N400 and such priming effects, usually skipping over the pixel-to-letter process. So, for this paper, we were keen on exploring more basic effects such as noise and symbols versus letters on the type-I and type-II responses. And for these effects, a feed-forward model turns out to be sufficient, so we can keep the focus of this particular paper on bottom-up processes during single word reading, on which there is already a lot to say.

To clarify our focus on feed-forward process, we have modified the title of the paper to be:

“Convolutional networks can model the functional modulation of the MEG responses associated with feed-forward processes during visual word recognition” furthermore, we have revised the Introduction to highlight this choice, noting:

“Another limitation is that these models have primarily focused on feed-back lexicosemantic effects while oversimplifying the initial feed-forward processing of the visual input.

[…]

For this study, we chose to focus on modeling the early feed-forward processing occurring during visual word recognition, as the experimental setup in Vartiainen et al. (2011) was designed to demonstrate.

[…]

By doing so, we restrict ourselves to an investigation of how well the three evoked components can be explained by a feed-forward CNN in an experimental setting designed to demonstrate feed-forward effects. As such, the goal is not to present a complete model of all aspects of reading, which should include feed-back effects, but rather to demonstrate the effectiveness of using a model that has a realistic form of input when the aim is to align the model with the evoked responses observed during visual word recognition.”

And in the Discussion section:

“In this paper we have restricted our simulations to feed-forward processes. Now, the way is open to incorporate convolution-and-pooling principles in models of reading that simulate feed-back processes as well, which should allow the model to capture more nuance in the Type-II and N400m components, as well as extend the simulation to encompass a realistic semantic representation.”

(2) There is no clear relationship between the layers that signal needs to traverse in the model and the relative duration of the three components in the brain.

While some models offer a tentative mapping between layers and locations in the brain, none of the models we are aware of actually simulate time accurately and our model is no exception.

While we provide some evidence that the three MEG components are best modeled with different types of layers, and the type-I becomes somewhere before type-II and N400m is last in our model, the lack of timing information is a weakness of our model we have not been able to address. In our previous version, this already was the main topic of our “Limitations of the model” section, but since this weakness was pointed out by all reviewers, we have decided to widen our discussion of it:

“One important limitation of the current model is the lack of an explicit mapping from the units inside its layers to specific locations in the brain at specific times. The temporal ordering of the components is simulated correctly, with the response profile matching that of the type-I occurring the layers before those matching the type-II, followed by the N400m. Furthermore, every component is best modeled by a different type of layer, with the type-I best described by convolution-and-pooling, the type-II by fully-connected linear layers and the N400m by a one-hot encoded layer. However, there is no clear relationship between the number of layers the signal needs to traverse in the model to the processing time in the brain. Even if one considers that the operations performed by the initial two convolution layers happen in the retina rather than the brain, the signal needs to propagate through three more convolution layers to reach the point where it matches the type-II component at 140-200 ms, but only through one more additional layer to reach the point where it starts to match the N400m component at 300-500 ms. Still, cutting down on the number of times convolution is performed in the model seems to make it unable to achieve the desired suppression of noise (Figure 5). It also raises the question what the brain is doing during the time between the type-II and N400m component that seems to take so long. It is possible that the timings of the MEG components are not indicative solely of when the feed-forward signal first reaches a certain location, but are rather dictated by the resolution of feed-forward and feedback signals (Nour Eddine et al., 2024).”

See also our response to the next comment of the Reviewer, in which we dive more into the effect of the number of layers, which could be seen as a manipulation of time.

(3) I am impressed by the CNN that authors modified to match the human brain pattern for the visual word recognition process, by the increase and decrease of the number of layers. The result of this part was a little different from the author’s expectation; however, the author didn’t explain or address this issue.

We are glad to hear that the reviewer found these results interesting. Accordingly, we now discuss these results more thoroughly in the main text.

We have moved the figure from the supplementary information to the main text (Figure 5 in the revised manuscript). And describe the results in the Results section:

“Various variations in model architecture and training procedure were evaluated. We found that the number of layers had a large impact on the response patterns produced by the model (Figure 5). The original VGG-11 architecture defines 5 convolution layers and 3 fully connected layers (including the output layer). Removing a convolution layer (Figure 5, top row), or removing one of the fully connected layers (Figure 5, second row), resulted in a model that did exhibit an enlarged response to noisy stimuli in the early layers that mimics the Type-I response. However, such models failed to show a sufficiently diminished response to noisy stimuli in the later layers, hence failing to produce responses that mimic the Type-II or N400m, a failure which also showed as low correlation scores.

Adding an additional convolution layer (Figure 5, third row) resulted in a model where none of the layer response profiles mimics that of the Type-II response. The Type-II response is characterized by a reduced response to both noise and symbols, but an equally large response to consonant strings, real and pseudo words. However, in the model with an additional convolution layer, the consonant strings evoked a reduced response already in the first fully connected layer, which is a feature of the N400m rather than the Type-II. These kind of subtleties in the response pattern, which are important for the qualitative analysis, generally did not show quantitatively in the correlation scores, as the fully connected layers in this model correlate as well with the Type-II response as models that did show a response pattern that mimics the Type-II.

Adding an additional fully connected layer (Figure 5, fourth row) resulted in a model with similar response profiles and correlation with the MEG components as the original VGG-11 architecture (Figure 5, bottom row) The N400m-like response profile is now observed in the third fully connected layer rather than the output layer. However, the decrease in response to consonant strings versus real and pseudo words, which is typical of the N400m, is less distinct than in the original VGG-11 architecture.”

We also incorporated these results in the Discussion:

“However, the ability of pooling units to stratify such differences depends on the size of their receptive field, which is determined by the number of convolution-andpooling layers. This might also explain why, in later layers, we observed a decreased response to stimuli where text was rendered with a font size exceeding the receptive field of the pooling units (Figure 8). Hence, the response profiles of the subsequent fully connected layers was very sensitive to the number of convolution-and-pooling layers. This number is probably dependent on the input size and pooling strategy. Given the VGG11 design of doubling the receptive field after each layer, combined with an input size of 225x225 pixels, the optimal number of convolution-and-pooling layers for our model was five, or the model would struggle to produce response profiles mimicking those of the type-II component in the subsequent fully connected layers (Figure 5).

[…]

A minimum of two fully connected layers was needed to achieve this in our case, and adding more fully connected layers would make them behave more like the component (Figure 5).”

(4) Can the author explain why the number of layers in the final model is optimal by benchmarking the brain hierarchy?

We have incorporated the figure describing the correlation between each model and the MEG components (previously Figure 5) with the figures describing the response profiles (Figures 4 and 5 in the revised manuscript and Supplementary Figures 2-6). This way, we (and the reader) can now benchmark every model qualitatively and quantitatively.

As we stated in our response to the previous comment, we have added a more thorough discussion on the number of layers, which includes the justification for our choice for the final model. The benchmark we used was primarily whether the model shows the same response patterns as the Type I, Type II and N400 responses, which disqualifies all models with fewer than 5 convolution and 3 fully connected layers. Models with more layers also show the proper response patterns, however we see that there is actually very little difference in the correlation scores between different models. Hence, our justification for sticking with the original VGG11 architecture is that it produces the qualitative best response profiles, while having roughly the same (decently high) correlation with the MEG components. Furthermore, by sticking to the standard architecture, we make it slightly easier to replicate our results as one can use readily available pre-trained ImageNet weights.

As well as always discussing the correlation scores in tandem with the qualitative analysis, we have added the following statement to the Results:

“Based on our qualitative and quantitative analysis, the model variant that performed best overall was the model that had the original VGG11 architecture and was preinitialized from earlier training on ImageNet, as depicted in the bottom rows of Figure 4 and Figure 5.”

Reviewer #2 (Public Review):

As has been shown over many decades, many potential computational algorithms, with varied model architectures, can perform the task of text recognition from an image. However, there is no evidence presented here that this particular algorithm has comparable performance to human behavior (i.e. similar accuracy with a comparable pattern of mistakes). This is a fundamental prerequisite before attempting to meaningfully correlate these layer activations to human neural activations. Therefore, it is unlikely that correlating these derived layer weights to neural activity provides meaningful novel insights into neural computation beyond what is seen using traditional experimental methods.

We very much agree with the reviewer that a qualitative analysis of whether the model can explain experimental effects needs to happen before a quantitative analysis, such as evaluating model-brain correlation scores. In fact, this is one of the intended key points we wished to make.

As we discuss at length in the Introduction, “traditional” models of reading (those that do not rely on deep learning) are not able to recognize a word regardless of exact letter shape, size, and (up to a point) rotation. In this study, our focus is on these low-level visual tasks rather than high-level tasks concerning semantics. As the Reviewer correctly states, there are many potential computational algorithms able to perform these visual task at a human level and so we need to evaluate the model not only on its ability to mimic human accuracy but also on generating a comparable pattern of mistakes. In our case, we need a pattern of behavior that is indicative of the visual processes at the beginning of the reading pipeline. Hence, rather than relying on behavioral responses that are produced at the very end, we chose the evaluate the model based on three MEG components that provide “snapshots” of the reading process at various stages. These components are known to manifest a distinct pattern of “behavior” in the way they respond to different experimental conditions (Figure 2), akin to what to Reviewer refers to as a “pattern of mistakes”. The model was first evaluated on its ability to replicate the behavior of the MEG components in a qualitative manner (Figure 4). Only then do we move on to a quantitative correlation analysis. In this manner, we feel we are in agreement with the approach advocated by the Reviewer.

In the Introduction, we now clarify:

“Another limitation is that these models have primarily focused on feed-back lexicosemantic effects while oversimplifying the initial feed-forward processing of the visual input.

[…]

We sought to construct a model that is able to recognize words regardless of length, size, typeface and rotation, as well as humans can, so essentially perfectly, whilst producing activity that mimics the type-I, type-II, and N400m components which serve as snapshots of this process unfolding in the brain.

[…]

These variations were first evaluated on their ability to replicate the experimental effects in that study, namely that the type-I response is larger for noise embedded words than all other stimuli, the type-II response is larger for all letter strings than symbols, and that the N400m is larger for real and pseudowords than consonant strings. Once a variation was found that could reproduce these effects satisfactorily, it was further evaluated based on the correlation between the amount of activation of the units in the model and MEG response amplitude.”

To make this prerequisite more clear, we have removed what was previously Figure 5, which showed the correlation between the various models the MEG components out of the context of their response patterns. Instead, these correlation values are now always presented next to the response patterns (Figures 4 and 5, and Supplementary Figures 2-6 in the revised manuscript). This invites the reader to always consider these metrics in relation to one another.

One example of a substantial discrepancy between this model and neural activations is that, while incorporating frequency weighting into the training data is shown to slightly increase neural correlation with the model, Figure 7 shows that no layer of the model appears directly sensitive to word frequency. This is in stark contrast to the strong neural sensitivity to word frequency seen in EEG (e.g. Dambacher et al 2006 Brain Research), fMRI (e.g. Kronbichler et al 2004 NeuroImage), MEG (e.g. Huizeling et al 2021 Neurobio. Lang.), and intracranial (e.g. Woolnough et al 2022 J. Neurosci.) recordings. Figure 7 also demonstrates that the late stages of the model show a strong negative correlation with font size, whereas later stages of neural visual word processing are typically insensitive to differences in visual features, instead showing sensitivity to lexical factors.

We are glad the reviewer brought up the topic of frequency balancing, as it is a good example of the importance of the qualitative analysis. Frequency balancing during training only had a moderate impact on correlation scores and from that point of view does not seem impactful. However, when we look at the qualitative evaluation, we see that with a large vocabulary, a model without frequency balancing fails to properly distinguish between consonant strings and (pseudo)words (Figure 4, 5th row). Hence, from the point of view of being able to reproduce experimental effects, frequency balancing had a large impact. We now discuss this more explicitly in the revised Discussion section:

“Overall, we found that a qualitative evaluation of the response profiles was more helpful than correlation scores. Often, a deficit in the response profile of a layer that would cause a decrease in correlation on one condition would be masked by an increased correlation in another condition. A notable example is the necessity for frequency-balancing the training data when building models with a vocabulary of 10 000. Going by correlation score alone, there does not seem to be much difference between the model trained with and without frequency balancing (Figure 4A, fifth row versus bottom row). However, without frequency balancing, we found that the model did not show a response profile where consonant strings were distinguished from words and pseudowords (Figure 4A, fifth row), which is an important behavioral trait that sets the N400m component apart from the Type-II component (Figure 2D). This underlines the importance of the qualitative evaluation in this study, which was only possible because of a straightforward link between the activity simulated within a model to measurements obtained from the brain, combined with the presence of clear experimental conditions.”

It is true that the model, even with frequency balancing, only captures letter- and bigramfrequency effects and not the word-frequency effects that we know the N400m is sensitive to. Since our model is restricted to feed-forward processes, this finding adds to the evidence that frequency-modulated effects are driven by feed-back effects as modeled by Nour Eddine et al. (2024, doi:10.1016/j.cognition.2024.105755). See also our response to the next comment by the Reviewer where we discuss feed-back connections. We have added the following to the section about model limitations in the revised Discussion:

“The fact that the model failed to simulate the effects of word-frequency on the N400m (Figure 8), even after frequency-balancing of the training data, is additional evidence that this effect may be driven by feed-back activity, as for example modeled by Nour Eddine et al. (2024).”

Like the Reviewer, we initially thought that later stages of neural visual word processing would be insensitive to differences in font size. When diving into the literature to find support for this claim, we found only a few works directly studying the effect of font size on evoked responses, but, surprisingly, what we did find seemed to align with our model. We have added the following to our revised Discussion:

“The fully connected linear layers in the model show a negative correlation with font size. While the N400 has been shown to be unaffected by font size during repetition priming (Chauncey et al., 2008), it has been shown that in the absence of priming, larger font sizes decrease the evoked activity in the 300–500 ms window (Bayer et al., 2012; Schindler et al., 2018). Those studies refer to the activity within this time window, which seems to encompass the N400, as early posterior negativity (EPN). What possibly happens in the model is that an increase in font size causes an initial stronger activation in the first layers, due to more convolution units receiving input. This leads to a better signal-to-noise ratio (SNR) later on, as the noise added to the activation of the units remains constant whilst the amplitude of the input signal increases. A better SNR translates ultimately in less co-activation of units corresponding to orthographic neighbours in the final layers, hence to a decrease in overall layer activity.”

Another example of the mismatch between this model and the visual cortex is the lack of feedback connections in the model. Within the visual cortex, there are extensive feedback connections, with later processing stages providing recursive feedback to earlier stages. This is especially evident in reading, where feedback from lexical-level processes feeds back to letter-level processes (e.g. Heilbron et al 2020 Nature Comms.). This feedback is especially relevant for the reading of words in noisy conditions, as tested in the current manuscript, as lexical knowledge enhances letter representation in the visual cortex (the word superiority effect). This results in neural activity in multiple cortical areas varying over time, changing selectivity within a region at different measured time points (e.g. Woolnough et al 2021 Nature Human Behav.), which in the current study is simplified down to three discrete time windows, each attributed to different spatial locations.

We agree with the Reviewer that a full model of reading in the brain must include feed-back connections and share their sentiment that these feed-back processes play an important role and are a fascinating topic to study. The intent for the model presented in our study is very much to be a stepping stone towards extending the capabilities of models that do include such connections.

However, there is a problem of scale that cannot be ignored.

Current models of reading that do include feedback connections fall into the category we refer to in the paper as “traditional models” and all only a few layers deep and operate on very simplified inputs, such as pre-defined line segments, a few pixels, or even a list of prerecognized letters. The Heilbron et al. 2020 study that the Reviewer refers to is a good example of such a model. (This excellent and relevant work was somehow overlooked in our literature discussion in the Introduction. We thank the Reviewer for pointing it out to us.) Models incorporating realistic feed-back activity need these simplifications, because they have a tendency to no longer converge when there are too many layers and units. However, in order for models of reading to be able to simulate cognitive behavior such as resolving variations in font size or typeface, or distinguish text from non-text, they need to operate on something close to the pixel-level data, which means they need many layers and units.

Hence, as a stepping stone, it is reasonable to evaluate a model that has the necessary scale, but lacks the feed-back connections that would be problematic at this scale, to see what it can and cannot do in terms of explaining experimental effects in neuroimaging studies. This was the intended scope of our study. For the revision, we have attempted to make this more clear.

We have changed the title to be:

“Convolutional networks can model the functional modulation of the MEG responses associated with feed-forward processes during visual word recognition” and added the following to the Introduction:

“The simulated environments in these models are extremely simplified, partly due to computational limitations and partly due to the complex interaction of feed-forward and feed-back connectivity that causes problems with convergence when the model grows too large. Consequently, these models have primarily focused on feed-back lexico-semantic effects while oversimplifying the initial feed-forward processing of the visual input. 

[…]

This rather high level of visual representation sidesteps having to deal with issues such as visual noise, letters with different scales, rotations and fonts, segmentation of the individual letters, and so on. More importantly, it makes it impossible to create the visual noise and symbol string conditions used in the MEG study to modulate the type-I and type-II components. In order to model the process of visual word recognition to the extent where one may reproduce neuroimaging studies such as Vartiainen et al. (2011), we need to start with a model of vision that is able to directly operate on the pixels of a stimulus. We sought to construct a model that is able to recognize words regardless of length, size, typeface and rotation with very high accuracy, whilst producing activity that mimics the type-I, type-II, and N400m components which serve as snapshots of this process unfolding in the brain. For this model, we chose to focus on the early feed-forward processing occurring during visual word recognition, as the experimental setup in the MEG study was designed to demonstrate, rather than feed-back effects

[…]

By doing so, we restrict ourselves to an investigation of how well the three evoked components can be explained by a feed-forward CNN in an experimental setting designed to demonstrate feed-forward effects. > As such, the goal is not to present a complete model of all aspects of reading, which should include feed-back effects, but rather to demonstrate the effectiveness of using a model that has a realistic form of input when the aim is to align the model with the evoked responses observed during visual word recognition.”

And we have added the following to the Discussion section:

“In this paper we have restricted our simulations to feed-forward processes. Now, the way is open to incorporate convolution-and-pooling principles in models of reading that simulate feed-back processes as well, which should allow the model to capture more nuance in the Type-II and N400m components, as well as extend the simulation to encompass a realistic semantic representation. A promising way forward may be to use a network architecture like CORNet (Kubilius et al., 2019), that performs convolution multiple times in a recurrent fashion, yet simultaneously propagates activity forward after each pass. The introduction of recursion into the model will furthermore align it better with traditional-style models, since it can cause a model to exhibit attractor behavior (McLeod et al., 2000), which will be especially important when extending the model into the semantic domain.

Furthermore, convolution-and-pooling has recently been explored in the domain of predictive coding models (Ororbia & Mali, 2023), a type of model that seems particularly well suited to model feed-back processes during reading (Gagl et al., 2020; Heilbron et al., 2020; Nour Eddine et al., 2024).”

We also would like to point out to the Reviewer that we did in fact perform a correlation between the model and the MNE-dSPM source estimate of all cortical locations and timepoints (Figure 7B). Such a brain-wide correlation map confirms that the three dipole groups are excellent summaries of when and where interesting effects occur within this dataset.

The presented model needs substantial further development to be able to replicate, both behaviorally and neurally, many of the well-characterized phenomena seen in human behavior and neural recordings that are fundamental hallmarks of human visual word processing. Until that point, it is unclear what novel contributions can be gleaned from correlating low-dimensional model weights from these computational models with human neural data.

We hope that our revisions have clarified the goals and scope of this study. The CNN model we present in this study is a small but, we feel, essential piece in a bigger effort to employ deep learning techniques to further enhance already existing models of reading. In our revision, we have extended our discussion where to go from here and outline our vision on how these techniques could help us better model the phenomena the reviewer speaks of. We agree with the reviewer that there is a long way to go, and we are excited to be a part of it.

In addition to the changes described above, we now end the Discussion section as follows: 

“Despite its limitations, our model is an important milestone for computational models of reading that leverages deep learning techniques to encompass the entire computational process starting from raw pixels values to representations of wordforms in the mental lexicon. The overall goal is to work towards models that can reproduce the dynamics observed in brain activity observed during the large number of neuroimaging experiments performed with human volunteers that have been performed over the last few decades. To achieve this, models need to be able to operate on more realistic inputs than a collection of predefined lines or letter banks (for example: Coltheart et al., 2001; Heilbron et al., 2020; Laszlo & Armstrong, 2014; McClelland & Rumelhart, 1981; Nour Eddine et al., 2024). We have shown that even without feed-back connections, a CNN can simulate the behavior of three important MEG evoked components across a range of experimental conditions, but only if unit activations are noisy and the frequency of occurrence of words in the training dataset mimics their frequency of use in actual language.”

Reviewer #3 (Public Review):

The paper is rather qualitative in nature. In particular, the authors show that some resemblance exists between the behavior of some layers and some parts of the brain, but it is hard to quantitively understand how strong the resemblances are in each layer, and the exact impact of experimental settings such as the frequency balancing (which seems to only have a very moderate effect according to Figure 5).

The large focus on a qualitative evaluation of the model is intentional. The ability of the model to reproduce experimental effects (Figure 4) is a pre-requisite for any subsequent quantitative metrics (such as correlation) to be valid. The introduction of frequency balancing is a good example of this. As the reviewer points out, frequency balancing during training has only a moderate impact on correlation scores and from that point of view does not seem impactful. However, when we look at the qualitative evaluation, we see that with a large vocabulary, a model without frequency balancing fails to properly distinguish between consonant strings and (pseudo)words (Figure 4, 5th row). Hence, from the point of view of being able to reproduce experimental effects, frequency balancing has a large impact.

That said, the reviewer is right to highlight the value of quantitative analysis. An important limitation of the “traditional” models of reading that do not employ deep learning is that they operate in unrealistically simplified environments (e.g. input as predefined line segments, words of a fixed length), which makes a quantitative comparison with brain data problematic. The main benefit that deep learning brings may very well be the increase in scale that makes more direct comparisons with brain data possible. In our revision we attempt to capitalize on this benefit more. The reviewer has provided some helpful suggestions for doing so in their recommendations, which we discuss in detail below.

We have added the following discussion on the topic of qualitative versus quantitative analysis to the Introduction:

“We sought to construct a model that is able to recognize words regardless of length, size, typeface and rotation, as well as humans can, so essentially perfectly, whilst producing activity that mimics the type-I, type-II, and N400m components which serve as snapshots of this process unfolding in the brain.

[…]

These variations were first evaluated on their ability to replicate the experimental effects in that study, namely that the type-I response is larger for noise embedded words than all other stimuli, the type-II response is larger for all letter strings than symbols, and that the N400m is larger for real and pseudowords than consonant strings. Once a variation was found that could reproduce these effects satisfactorily, it was further evaluated based on the correlation between the amount of activation of the units in the model and MEG response amplitude.”

And follow this up in the Discussion with a new sub-section entitled “On the importance of experimental contrasts and qualitative analysis of the model”

The experiments only consider a rather outdated vision model (VGG).

VGG was designed to use a minimal number of operations (convolution-and-pooling, fullyconnected linear steps, ReLU activations, and batch normalization) and rely mostly on scale to solve the classification task. This makes VGG a good place to start our explorations and see how far a basic CNN can take us in terms of explaining experimental MEG effects in visual word recognition. However, we agree with the reviewer that it is easy to envision more advanced models that could potentially explain more. In our revision, we expand on the question of where to go from here and outline our vision on what types of models would be worth investigating and how one may go about doing that in a way that provides insights beyond higher correlation values.

We have included the following in our Discussion sub-sections on “Limitations of the current model and the path forward”:

“The VGG-11 architecture was originally designed to achieve high image classification accuracy on the ImageNet challenge (Simonyan & Zisserman, 2015). Although we have introduced some modifications that make the model more biologically plausible, the final model is still incomplete in many ways as a complete model of brain function during reading.

[…]

In this paper we have restricted our simulations to feed-forward processes. Now, the way is open to incorporate convolution-and-pooling principles in models of reading that simulate feed-back processes as well, which should allow the model to capture more nuance in the Type-II and N400m components, as well as extend the simulation to encompass a realistic semantic representation. A promising way forward may be to use a network architecture like CORNet (Kubilius et al., 2019), that performs convolution multiple times in a recurrent fashion, yet simultaneously propagates activity forward after each pass. The introduction of recursion into the model will furthermore align it better with traditional-style models, since it can cause a model to exhibit attractor behavior (McLeod et al., 2000), which will be especially important when extending the model into the semantic domain. Furthermore, convolution-and-pooling has recently been explored in the domain of predictive coding models (Ororbia & Mali, 2023), a type of model that seems particularly well suited to model feed-back processes during reading (Gagl et al., 2020; Heilbron et al., 2020; Nour Eddine et al., 2024).”

Reviewer #3 (Recommendations For The Authors):

(1) The method used to select the experimental conditions under which the behavior of the CNN is the most brain-like is rather qualitative (Figure 4). It would have been nice to have a plot where the noisyness of the activations, the vocab size and the amount of frequency balancing are varied continuously, and show how these three parameters impact the correlation of the model layers with the MEG responses.

We now include this analysis (Figure 6 in the revised manuscript, Supplementary Figures 47) and discuss these factors in the revised Results section:

“Various other aspects of the model architecture were evaluated which ultimately did not lead to any improvements of the model. The response profiles can be found in the supplementary information (Supplementary Figures 4–7) and the correlations between the models and the MEG components are presented in Figure 6. The vocabulary of the final model (10 000) exceeds the number of units in its fullyconnected layers, which means that a bottleneck is created in which a sub-lexical representation is formed. The number of units in the fully-connected layers, i.e. the width of the bottleneck, has some effect on the correlation between model and brain (Figure 6A), and the amount of noise added to the unit activations less so (Figure 6B). We already saw that the size of the vocabulary, i.e. the number of wordforms in the training data and number of units in the output layer of the model, had a large effect on the response profiles (Figure 4). Having a large vocabulary is of course desirable from a functional point of view, but also modestly improves correlation between model and brain (Figure 6C). For large vocabularies, we found it beneficial to apply frequency-balancing of the training data, meaning that the number of times a word-form appears in the training data is scaled according to its frequency in a large text corpus. However, this cannot be a one-to-one scaling, since the most frequent words occur so much more often than other words that the training data would consist of mostly the top-ten most common words, with less common words only occurring once or not at all. Therefore, we decided to scale not by the frequency 𝑓 directly, but by 𝑓𝑠, where 0 < 𝑠 < 1, opting for 𝑠 = 0.2 for the final model (Figure 6D).”

(2) It is not clear which layers exactly correspond to which of the three response components. For this to be clearer, it would have been nice to have a plot with all the layers of VGG on the x-axis and three curves corresponding to the correlation of each layer with each of the three response components.

This is a great suggestion that we were happy to incorporate in the revised version of the manuscript. Every figure comparing the response patterns of the model and brain now includes a panel depicting the correlation between each layer of the model and each of the three MEG components (Figures 4 & 5, Supplementary Figures 2-5). This has given us (and now also the reader) the ability to better benchmark the different models quantitatively, adding to our discussion on qualitative to quantitative analysis.

(3) It is not clear to me why the authors report the correlation of all layers with the MEG responses in Figure 5: why not only report the correlation of the final layers for N400, and that of the first layers for type-I?

We agree with the reviewer that it would have been better to compare the correlation scores for those layers which response profile matches the MEG component. While the old Figure 5 has been merged with Figure 4, and now provides the correlations between all the layers and all MEG components, we have taken the Reviewer’s advice and marked the layers which qualitatively best correspond to each MEG component, so the reader can take that into account when interpreting the correlation scores.

(4) The authors mention that the reason that they did not reproduce the protocol with more advanced vision models is that they needed the minimal setup capable of yielding the desired experiment effect. I am not fully convinced by this and think the paper could be significantly strengthened by reporting results for a vision transformer, in particular to study the role of attention layers which are expected to play an important role in processing higher-level features.

We appreciate and share the Reviewer’s enthusiasm in seeing how other model architectures would fare when it comes to modeling MEG components. However, we regard modifying the core model architecture (i.e., a series of convolution-and-pooling followed by fully-connected layers) to be out of scope for the current paper.

One of the key points of our study is to create a model that reproduces the experimental effects of an existing MEG study, which necessitates modeling the initial feed-forward processing from pixel to word-form. For this purpose, a convolution-and-pooling model was the obvious choice, because these operations play a big role in cognitive models of vision in general. In order to properly capture all experimental contrasts in the MEG study, many variations of the CNN were trained and evaluated. This iterative design process concluded when all experimental contrasts could be faithfully reproduced.

If we were to explore different model architectures, such as a transformer architecture, reproducing the experimental contrasts of the MEG study would no longer be the end goal, and it would be unclear what the end goal should be. Maximizing correlation scores has no end, and there are a nearly endless number of model architectures one could try. We could bring in a second MEG study with experimental contrasts that the CNN cannot explain and a transformer architecture potentially could and set the end goal to explain all experimental effects in both MEG studies. But even if we had access to such a dataset, this would almost double the length of the paper, which is already too long.

Author response:

The following is the authors’ response to the previous reviews.

Reviewer 1:

(1) The results do not support the conclusions. The main "selling point" as summarized in the title is that the apoptotic rate of zebrafish motorneurons during development is strikingly low (~2% ) as compared to the much higher estimate (~50%) by previous studies in other systems. The results used to support the conclusion are that only a small percentage (under 2%) of apoptotic cells were found over a large population at a variety of stages 24-120hpf. This is fundamentally flawed logic, as a short-time window measure of percentage cannot represent the percentage on the long-term. For example, at any year under 1% of human population die, but over 100 years >99% of the starting group will have died. To find the real percentage of motorneurons that died, the motorneurons born at different times must be tracked over long term, or the new motorneuron birth rate must be estimated. Similar argument can be applied to the macrophage results.<br />

In the revised manuscript (revised Figure 4), we extended the observation time window as long as possible, from 24 hpf to 240 hpf. After 240 hpf, the transparency of zebrafish body decreased dramatically, which made optical imaging quite difficult.

We are confident that this 24-240 hpf time window covers the major time window during which motor neurons undergo programmed cell death during zebrafish early development. We chose the observation time window based on the following two reasons: 1) Previous studies showed that although the time windows of motor neuron death vary in chick (E5-E10), mouse (E11.5-E15.5), rat (E15-E18), and human (11-25 weeks of gestation), the common feature of these time windows is that they are all the developmental periods when motor neurons contact with muscle cells. The contact between zebrafish motor neurons and muscle cells occurs before 72 hpf, which is included in our observation time window. 2) Most organs of zebrafish form before 48-72 hpf, and they complete hatching during 48-72 hpf. Food-seeking and active avoidance behaviors also start at 72 hpf, indicating that motor neurons are fully functional at 72 hpf.

Previous studies in zebrafish have shown that the production of spinal cord motor neurons largely ceases before 48 hpf, and then the motor neurons remain largely constant until adulthood (doi: 10.1016/j.celrep.2015.09.050; 10.1016/j.devcel.2013.04.012; 10.1007/BF00304606; 10.3389/fcell.2021.640414). Our observation time window covers the major motor neuron production process. Therefore, we believe that neurogenesis will not affect our findings and conclusions.

Although we are confident that 240 h tracking is long enough to measure the motor neuron death rate, several sentences have been added in the discussion part, “In our manuscript, we tracked the motor neuron death in live zebrafish until 240 hpf, which was the longest time window we could achieve. But there was still a possibility that zebrafish motor neurons might die after 240 hpf.”

We agreed that the “2%” description might not be very accurate. Thus, we have revised our title to “Zebrafish live imaging reveals a surprisingly small percentage of spinal cord motor neurons die during early development.”

(2) The conclusion regarding timing of axon and cell body caspase activation and apoptosis timing also has clear issues. The ~minutes measurement are too long as compared to the transport/diffusion timescale between the cell body and the axon, caspase activity could have been activated in the cell body and either caspase or the cleaved sensor move to the axon in several seconds. The authors' results are not high frequency enough to resolve these dynamics. Many statements suggest oversight of literature, for example, in abstract "however, there is still no real-time observation showing this dying process in live animals.".

Real-time imaging of live animals is quite challenging in the field. Currently, using confocal microscopy, we can only achieve minute-scale tracking. In the future, with more advanced imaging techniques, the sensor fish in the present study may provide us with more detailed information on motor neuron death. We have removed “real-time” from our revised manuscript. We also revised the mentioned sentence in the abstract.

(3) Many statements should use more scholarly terms and descriptions from the spinal cord or motorneuron, neuromuscular development fields, such as line 87 "their axons converged into one bundle to extend into individual somite, which serves as a functional unit for the development and contraction of muscle cells"

We have removed this sentence.

(4) The transgenic line is perhaps the most meaningful contribution to the field as the work stands. However, mnx1 promoter is well known for its non-specific activation - while the images do suggest the authors' line is good, motorneuron markers should be used to validate the line. This is especially important for assessing this population later as mnx1 may be turned off in mature neurons. The author's response regarding mnx1 specificity does not mitigate the original concern.

The mnx1 promoter has been widely used to label motor neurons in transgenic zebrafish. Previous studies have shown that most of the cells labeled in the mnx1 transgenic zebrafish are motor neurons. In this study, we observed that the neuronal cells in our sensor zebrafish formed green cell bodies inside of the spinal cord and extended to the muscle region, which is an important morphological feature of the motor neurons.

Furthermore, a few of those green cell bodies turned into blue apoptotic bodies inside the spinal cord and changed to blue axons in the muscle regions at the same time, which strongly suggests that those apoptotic neurons are not interneurons.

In fact, no matter what method is used, such as using antibodies to stain specific markers to label motor neurons, 100% specificity cannot be achieved. More importantly, although the mnx1 promoter might have labeled some interneurons, this will not affect our major finding that only a small percentage of spinal cord motor neurons die during the early development of zebrafish.

Reviewer 2:

(1) Title: The 50% figure of motor neurons dying through apoptosis during early vertebrate development is not precisely accurate. In papers referenced by the authors, there is a wide distribution of percentages of motor neurons that die depending on the species and the spinal cord region. In addition, the authors did not examine limb-innervating motor neurons, which are the ones best studied in motor neuron programmed cell death in other species. Thus, a better title that reflects what they actually show would be something like "A surprisingly small percentage of early developing zebrafish motor neurons die through apoptosis in non-limb innervating regions of the spinal cord."

In fish, there are no such structures as limbs, although fins may be evolutionarily related to limbs. In our manuscript, we studied the naturally occurring motor neuron death in the whole spinal cord during the early stage of zebrafish development. The death of motor neurons in limb-innervating motor neurons has been extensively studied in chicks and rodents, as it is easy to undergo operations such as amputation. However, previous studies have shown this dramatic motor neuron death occurs not only in limb-innervating motor neurons but also in other spinal cord motor neurons (doi: 10.1006/dbio.1999.9413).

We have revised our title to “Zebrafish live imaging reveals a surprisingly small percentage of spinal cord motor neurons die during early development.”

(2) lines 18-19: "embryonic stage of vertebrates" is very broad, since zebrafish are also vertebrates; it would be better to be more specific

lines 25-26: The authors should be more specific about which animals have widespread neuronal cell death.

We have revised our manuscript accordingly.

(3) lines 98-99; 110-111; 113; 122-123; 140-141: A cell can undergo apoptosis. But an axon, which is only part of a cell, cannot undergo apoptosis. Especially since the axon doesn't have a separate nucleus, and the definition of apoptosis usually includes nuclear fragmentation. A better subheading would describe the result, which is that caspase activation is seen in both the cell body and the axon.

We have revised the subheadings and related words in the manuscript accordingly. In the introduction, we also revised the expression of the third aim from “Which part of a neuron (cell body vs. axon) will die first?” to “Which part of a neuron (cell body vs. axon) will degrade first?”.

(4) lines 159-160; 178-179: This is an oversimplification of the literature. The authors should spell out which populations of motor neuron have been examined and say something about the similarities and difference in motor neuron death.

We have revised it accordingly.

(5) lines 200; 216: The authors did not observe macrophages engulfing motor neurons. But that does not mean that they cannot. Making the conclusion stated in this subheading would require some kind of experiment, not just observations.

We did observe few colocalizations of macrophages and dead motor neurons.  To more accurately express these data, in the revised manuscript, we used “colocalization” to replace “engulfment.” The subheading has been revised to “Most dead motor neurons were not colocalized with macrophages.” Accordingly, panel C of Figure 5 has also been revised.

(6) lines 234-246: The authors seem to have missed the point about VaP motor neuron death, which was two-fold. First, VaP death has been previously described, thus it could serve as a control for the work in this paper, especially since the conditions underlying VaP death and survival have been experimentally tested. Second, they should acknowledge that previous work showed that at least some motor neuron death in zebrafish differs from that described in chick and rodents. This conclusion came from work showing that death of VaP is independent of limitations in muscle innervation area, suggesting it is not coupled to muscle-derived neurotrophic factors.

Figures: The authors should say which level of the spinal cord they examined in each figure.

We have compared our findings with previous findings in the revised manuscript. The death of VaP motor neurons is not related to neurotrophic factors, but the death of other motor neurons may be related to neurotrophic factors, which needs further study and evidence. Our study examined the overall motor neuron apoptosis regardless of the causes and locations. To avoid misunderstanding, in the revised manuscript, we removed the data and words related to neurotrophic factors.

We also extended the observation time window as long as possible, from 24 hpf to 240 hpf (revised Figure 4). After 240 hpf, the transparency of zebrafish body decreased dramatically, which made the optical imaging quite difficult.

Formatering

Ledningskollen

Formatering

Vad är nyttjänstenät?

Vilken är datakällan?

Finns det någon mer information som kan inkluderas i popupen? Nu står det bara "1" när man klickar på punkten.

Beskriv varför? Hänvisa till kartmaterialet som visar att det finns ett borrhålslager där tidigare, vilket betyder att det har utfärdats borrtillstånd i området tidigare.

Skriv att (eller liknande):

Borrningsytor på respektive fastighet har identiferiats med hänsyn till [ledningar, avstånd till grannfastigheter med mera..]. Antalet borrhål som får plats inom dessa ytor varierar beroende på hur geoenergisystemet utformas - om borrhålslagret är välbalanserad med avseende på värmeuttag- och tillförsel över året kan borrhålen borras med tätare mellanrum och vice versa.

Av ovanstående anledning har antalet borrhål bedömts för två olika scenarion - 6 m avstånd som speglar ett fall med balanserat borrhålslager samt 10 m avstånd för ett fall då lagret är relativt obalanserat.

Jättebra med dessa uppdaterade kartor.

Tror det behövs lite förklaringar till lagrena så att man som utomstående förstår vad de betyder. typ

• Identifierad borrningsyta
• Centrumpunkt borrhål

etc. Denna förklaring kan skrivas i texten - men även kartlegenden kan vara lite mer beskrivande

Tror det blir snyggare om man använder centrerade kartor. Att kartan blir mindre gör inte så mycket eftersom man kan zooma.

The book of books.

Vad är nyttjänstenät?

Se kommentar ovan

Bättre att inte använda full width här, allt hamnar så att man inte ser kontrollerna (iaf för mig)

Formatering

Formatering punktlista

Var är figuren?

Beskriv vilka kriterier som vi har tittat på för att ta fram dessa uppskattningar. Kanske mer logiskt att ha avsnittet "Specifika problemområden" först som du inleder med vilka förutsättnignar som krävs för att man ska kunna borra (eller inte kunna borra)

Centrera caption, gäller alla figurer.

Går det att skapa en frame kring kartan?

UX-mässigt känns det som att detta är en bild/printscreen typ och inte en interaktiv karta :) Den är riktigt snygg gjord men läsaren behöver förstå vad man kan göra. Du kanske kan lägga en backround map som är aktiv som default med låg opacitet så att din georefererade pdf sticker ut samtidigt som man förstår att det är en karta.

Man kan skapa en callout note där man förklarar att rapporten är interaktiv. SÅ här har jag skrivit i en annan rapport:

"Denna rapport utförs i form av en s.k. interaktiv rapport, vilket innebär att läsaren kan interagera med data och resultat i figurer och andra media i rapporten. Syftet är att ge läsaren en ökad inblick i och förståelse av de resultat som rapporten förmedlar, t.ex. via möjligheten att se extra information genom att hovra över datapunkter eller att zooma in intressanta områden i figurer."

Också: highlighta på ett bättre sätt vilka fastigheterna är, trots att det finns massa underlag inlagt är det inte helt uppenbart vilka fastigheter som rapporten handlar om.

Beskriv kort vad det är för ritningar och varför de är relevant i detta uppdrag.

Du kan nog vara mindre specifik här i inledningen, skriv istället "bedömnings av tillgängliga borrningsytor"

Skriv "Bengt Dahlgren" har fått i uppdrag

It is weird how they are trying to help while ignoring the people that they are trying to help. Economic profit? The rich always end up trying to take advantage of the desperate and poor situation of people.

wrapper 的概念是 maven 从 gradle 那里借鉴过来的

Reviewer #1 (Public review):

Summary:

This study investigates the relationship between climate variables and malaria incidence from monthly records, for rainfall, temperature, and a measure of ENSO, in a lowland region of Kenya in East Africa. Wavelet analyses show significant variability at the seasonal scale at the 6-month scale with some variation in its signal over time, and some additional variability at the 12-month scale for some variables. As conducted, the analyses show weak (non-significant) signals at the interannual time scales (longer than seasonal). Cross-wavelet analysis also highlights the 6-month scale and the association of malaria and climate variables at that scale, with some signal at 12 months, reflecting the role of climate in seasonality. Evidence is presented for some small changes in the lags of the response of malaria to the seasonal climate drivers over time.

Strengths:

Although there have been many studies of climate drivers of malaria dynamics in East Africa, these analyses have been largely focused on highlands where these drivers are expected to exhibit the strongest signal of association with disease burden at interannual and longer time scales. It is therefore of interest to take advantage of a relatively long time series of cases to examine the role of climate variables in more endemic malaria in lowlands.

Weaknesses:

(1) Major comments:

The work is not sufficiently placed in the context of what is known about climate variability in East Africa, and the role of climate variables in the temporal variation of malaria cases in this region. This context includes the relationship between large (global/regional) drivers of interannual climate variability such as ENSO (and the Indian Ocean Dipole) and local temporal patterns in rainfall and temperature. There is for example literature on the influence of those drivers and the short and long rains in East Africa. That is, phenomena such as ENSO would influence malaria through those local climate variables. This context should be considered when formulating and interpreting the analyses.

There are conceptual problems with the design of the analyses which can limit the findings on association. It is not surprising that rainfall would exhibit a clear association at seasonal scales. It is nevertheless valuable to confirm this as the authors have done and to examine the faster than 12-month scale, given the typical pattern of two rainfall seasons in this area. However, the results on temperature are less clear. If rainfall is the main limiting factor for the transmission season, the temperature variation that would matter can be during the rainy periods. One would then see an association with temperature only in particular windows of time during the year, when rainfall is sufficient (see for example, Rodo et al. Nat. Commun. 2022, for this finding in a highland region of Ethiopia). For this situation, there would be no clear association with temperature when all months are considered, and one would not find a significant relationship (or a lagged one) between peak times in this climate factor and malaria's seasonal cases. It would be difficult for the wavelet analysis to reveal such an effect. Another consideration is whether to use an ENSO variable that includes seasonality or to use an ENSO index computed as an anomaly, to focus on interannual variability. That is, it is most relevant to consider how ENSO influences time scales of variation longer than seasonal (the multiannual variation in seasonal epidemics) and for this purpose, one would typically rely on an anomaly. This choice would better enable one to see whether there is a role of ENSO at interannual time scales. It would also make sense to analyze with cross-wavelets the effect of ENSO on local climate factors, temperature, and rainfall, and not only on malaria. This would allow us to establish evidence for a chain of causality, from a global driver of interannual variability to local climate variability to malaria incidence.

The multiresolution analysis and associated analysis of lag variations were confusing and difficult to follow as presented: (1) the lags chosen by the multiresolution analysis do not match the phase differences of the cross-wavelet analysis if I followed what was presented. On page 8, phase differences are expressed in months. I do not understand then the following statements on page 9: "The phase differences obtained by the cross-wavelet transforms were turned into lags, allowing us to plot the evolution of the lags over time". The resulting lags in Figure 6 are shorter than the phase differences provided in the text on page 8. (2) The phase difference of the cross-wavelet analyses for malaria and temperature is also too long for this climate factor to explain an effect on the vector and then on the disease. (3) In Table 3, the regression results that are highlighted are those for Land Surface Temperatures (LST) and ENSO, with a weak but significant negative linear correlation, and for LST and bednet coverage, and this is considered part of the lag analysis. The previous text and analyses up to that point do not seem to consider the relationship of ENSO and local climate variables, or that between local climate variables and bednets (which would benefit from some context for the causal pathways this would reflect).

The conclusion in the Abstract: "Our study underlines the importance of considering long-term time scales when assessing malaria dynamics. The presented wavelet approach could be applicable to other infectious diseases" needs to be reformulated. The use of "long-term" time scales for those of ENSO and interannual variability is not consistent with the climate literature, where long-term could be interpreted as decadal and longer. The time scales beyond those of seasonality, especially those of climate variability, have been addressed in many malaria studies. It is not compelling to have the significance of this study be the importance of considering those time scales. This is not new. I recommend focusing on what has been done for lowland malaria and endemic regions (for example, in Laneri et al. PNAS 2015) as there has been less work for those regions than for seasonal epidemic ones of low transmission (e.g. altitude fringes and desert ones, e.g. Laneri et al. PloS Comp. Biol. 2010; Roy et al. Mal. J. 2015). Also, wavelet analyses have been used extensively by now to consider the association of climate variables and infectious diseases at multiple time scales. There is here an additional component of the analysis but the decomposition that underlies the linear regressions is also not that new, as decompositions of time series have been used before in this area. In summary, I recommend a more appropriate and compelling conclusion on what was learned about malaria at this location and what it may tell us about other, similar, locations, but not malaria dynamics everywhere.

The conversion from monthly cases to monthly incidence needs a better explanation of the Methods, rather than a referral to another paper. This is a key aspect of the data. It may be useful to plot the monthly time series of both variables in the Supplement, for comparison.

There is plenty of evidence of the seasonal role of rainfall on malaria's seasonality in many regions. The literature cited here to support this well-known association is quite limited. It would be useful to provide a context that better reflects the literature and some context for the environmental conditions of this lowland region that would explain the dominant role of rainfall on malaria seasonality. Two papers (from 2017 and 2019) are cited in the second paragraph of the introduction as showing that "key climatic factors are rainfall and temperatures". This is a misrepresentation of the field. That these factors matter to malaria in general has been known for a very long time given that the vectors are mosquitoes, and the cited studies are particular ones that examine the mechanistic basis of this link for modeling purposes. Either these papers are presented as examples, with a more accurate description of what they add to the earlier literature or earlier literature should be acknowledged. Also, what has been much less studied is the role of these variables at interannual time scales, as potentially mediating the effects of global drivers in teleconnections.

(2) Minor comments:

In relation to the conceptual issues raised above, it would be valuable to consider whether the negative association with temperature persists if one considers mean temperature during the rainy seasons only, against the total cases in the transmission season each year (as in Rodó et al. 2021). This would allow one to disentangle whether the negative association reflects a robust result or an artifact of an interaction between temperature and rainfall so that the former matters when the latter is permissive for transmission.

The conclusion in the Discussion " This suggests that minor climate variations have a limited impact on malaria incidence at shorter time scales, whereas climatic trends may play a more substantial role in shaping long-term malaria dynamics" is unsubstantiated. There is no clear result in the paper on climatic trends that I can see.

The Abstract writes: "The true impact of climate change...". This paper is not about climate change but about climate seasonality and variability. This text needs to be changed to make it consistent with the content of the paper.

Page 2, Introduction: The statement on Pascual et al. 2008 is not completely accurate. This paper shows an interplay of climate variability and disease dynamics, but not cycles that are completely independent of climate.

Page 2, next sentence: "More recently, such cycles have been attributed to global climate drivers such as ENSO (Cazelles et al., 2023)". This writing is also somewhat unclear. Are you referring to the cycles for the same location in Kenya? Or generically, to the interannual variability of malaria?

There are multiple places in the writing that could be edited.

нельзя просто так взять объект из другого потока, и как то его использовать. это как нельзя брать чужую игрушку и с ней играть

один должны по очереди вызывать, иначе может быть блокирвока

This work has been peer reviewed in GigaScience (https://doi.org/10.1093/gigascience/giaf007), which carries out open, named peer-review. These reviews are published under a CC-BY 4.0 license and were as follows:

Reviewer #2: Anuradha Wickramarachchi

Overall comments.

Authors of the manuscript have developed an iterative overlap graph construction algorithm to support genome assembly. This is both an interesting and a demanding area of research due to very recent advancements in sequencing technologies.

Although the text in the manuscript is interesting, grammar must be rechecked and revised. At some point it is difficult to keep track of the content and references to supplementary to make sense out of the content.

Specific comments

Page 1 Line 13: I believe the authors are talking about assembly sizes and not genome sizes. The sentences here could be a bit short to make them easy to understand.

Page 2 Line 19: Theoretical time complexity O(m2n2) is bit of an overstatement due to the heuristics employed by most assemblers. For example, mash distance, minimisers and k-mer bins are there to prevent this explosion of complexity. Either acknowledge such methods or provide a range for the time complexity. I would be interesting to know the time complexities of the methods expressed in sentence starting Line 15.

Page 5 Line 11: Was this performed with overlapping windows of 1gb? Otherwise, simulations may not have reads spanning across such regions.

Page 5 Line 14: It seems you are simulating 9 + 4 + 4 datasets. This is unclear, please make this into bullet points or separate paragraphs and explain clearly. Include simulator information in the table itself by may be making it landscape (in supplementary).

Fig 2: I believe authors should expand their analysis to more recent and popular assemblers. For example, wtdbg2 is designed for noisy reads and not specifically for more accurate R10/ HiFi reads. So please include, HiFi-asm, Flye where appropriate. Flye supports ONT out of the box and in my experience does produce good assemblies.

Although, you are evaluating read overlaps, it is hard to ignore assemblers themselves just because they do not produce intermediate overlaps graphs.

Page 5-9: In the benchmarks section, please include how True Positives and False Positives were labelled. Was this from simulation data?

Page 11: Use of xRead has been evaluated on genome assemblies. This is a very important and it is a bit unfortunate that existing assemblers are not very flexible in terms of plugging in new intermediate steps. It might be worth exploring into creating a new assembler using the wtpoa2 cli command of wtdbg2.

Page 16: What will happen if you only capture reads from a single chromosome due to longer length? I believe the objective is to gather longest reads capturing as much as possible covering the whole genome. Please comment on this.

Page 19: In the Github Readme the download URL was wrong. Please correct it to the latest release

Correct: https://github.com/tcKong47/xRead/releases/download/xRead-v1.0.0.1/xRead-v1.0.0.tar.gz Existing: https://github.com/tcKong47/xRead/releases/download/v1.0.0/xRead-v1.0.0.tar.gz

Make command failed with make: *** No rule to make target main.h', needed bymain.o'. Stop.

It seems the release does not have source code, but rather the compiled version. Please update github instructing how to compile code properly with a git clone.

标记这个页面

Reviewer #3 (Public review):

Summary:

This paper aims to investigate how the human brain represents different forms of value and uncertainty that participate in active inference within a free-energy framework, in a two-stage decision task involving contextual information sampling, and choices between safe and risky rewards, which promotes shifting between exploration and exploitation. They examine neural correlates by recording EEG and comparing activity in the first vs second half of trials and between trials in which subjects did and did not sample contextual information, and perform a regression with free-energy-related regressors against data "mapped to source space."

Strengths:

This two-stage paradigm is cleverly designed to incorporate several important processes of learning, exploration/exploitation and information sampling that pertain to active inference. Although scalp/brain regions showing sensitivity to the active-inference related quantities do not necessarily suggest what role they play, they are illuminating and useful as candidate regions for further investigation. The aims are ambitious, and the methodologies are impressive. The paper lays out an extensive introduction to the free energy principle and active inference to make the findings accessible to a broad readership.

Weaknesses:

It is worth noting that the high lower-cutoff of 1 Hz in the bandpass filter, included to reduce the impact of EEG noise, would remove from the EEG any sustained, iteratively updated representation that evolves with learning across trials, or choice-related processes that unfold slowly over the course of the 2-second task windows. It is thus possible there are additional processes related to the active inference quantities that are missed here. This is not a flaw as one must always try to balance noise removal against signal removal in filter settings - it is just a caveat. As the authors also note, the regions showing up as correlated with model parameters change depending on source modelling method and correction for multiple comparisons, warranting some caution around the localisation aspect.

Author response:

Reviewer #1 (Public review):

Summary:

This manuscript by Guo and Uusisaari describes a series of experiments that employ a novel approach to address long-standing questions on the inferior olive in general and the role of the nucleoolivary projection specifically. For the first time, they optimized the ventral approach to the inferior olive to facilitate imaging in this area that is notoriously difficult to reach. Using this approach, they are able to compare activity in two olivary regions, the PO and DAO, during different types of stimulation. They demonstrate the difference between the two regions, linked to Aldoc-identities of downstream Purkinje cells, and that there is co-activation resulting in larger events when they are clustered. Periocular stimulation also drives larger events, related to co-activation. Using optogenetic stimulation they activate the nucleoolivary (N-O) tract and observe a wide range of responses, from excitation to inhibition. Zooming in on inhibition they test the assumption that N-O activation can be responsible for suppression of sensoryevoked events. Instead, they suggest that the N-O input can function to suppress background activity while preserving the sensory-driven responses.

Strengths:

This is an important study, tackling the long-standing issue of the impossibility to do imaging in the inferior olive and using that novel method to address the most relevant questions. The experiments are technically very challenging, the results are presented clearly and the analysis is quite rigorous. There is quite a lot of room for interpretation, see weaknesses, but the authors make an effort to cover many options.

Weaknesses:

The heavy anesthesia that is required during the experiment could severely impact the findings. Because of the anesthesia, the firing rate of IO neurons is found to be 0.1 Hz, significantly lower than the 1 Hz found in non-anesthetized mice. This is mentioned and discussed, but what the consequences could be cannot be understated and should be addressed more. Although the methods and results are described in sufficient detail, there are a few points that, when addressed, would improve the manuscript.

We sincerely thank the reviewer for their encouraging comments and recognition of our study’s significance. We fully acknowledge the confounding effects of the deep anesthesia used in our experiments, which was necessary to ensure the animals’ welfare while establishing this technically demanding methodology. We elaborate on these effects below and will further clarify them in the revised manuscript.

Ultimately, the full resolution of this issue will require recordings in awake animals, as we consider our approach an advancement from acute slice preparations but not yet a complete representation of in vivo IO function. However, key findings from our study—such as amplitude modulation with co-activation and the potential role of IO refractoriness in complex spike generation—could be further explored in existing cerebellar cortical recordings from awake, behaving animals. We hope our work will motivate re-examination of such datasets to assess whether these mechanisms contribute to overall cerebellar function.

Reviewer #1 (Recommendations for the authors):

On page 10 the authors indicate that 2084 events were included for DAO and 1176 for PO. Is that the total number of events? What was the average and the range per neuron and the average recording duration?

Thank you for pointing out lack of clarity. The sentence should say "in total, 2084 and 1176 detected events from DAO and PO were included in the study". We will add the averages and ranges of events detected per neuron in different categories, as well as the durations of the recordings (ranging from 120s to 270s) to the tables.

On page 10 it is also stated that: "events in PO reached larger values than those in DAO even though the average values did not differ". Please clarify that statement. Which parameter + p-value in the table indicates this difference?

Apologies for omission. Currently the observation is only visible in the longer tail to the right in the PO data in Figure 2B2. We will add the range of values (3.0-75.2 vs 3.1-39.6 for PO and DAO amplitudes, respectively) in text and the tables in the revision.

Abbreviating airpuff to AP is confusing, I would suggest not abbreviating it.

Understood. We will change AP to airpuff in the text. In figure labels, at least in some panels, the abbreviation will be necessary due to space constraints.

What type of pulse was used to drive ChrimsonR? Could it be that the pulse caused a rebound-like phenomenon with the pulse duration that drove the excitation?

As described on line 229 and in the Methods, we used 5-second trains of 5-ms LED light pulses. Importantly, these stimulation parameters were informed by our extensive in vitro examination of various stimulation patterns (Lefler et al., 2014), which consistently produced stable postsynaptic responses without inducing depolarization or rebound effects. Additionally, Loyola et al. (2024) reported no evidence of rebound activity in IO cells following optogenetic activation of N-O axons in the absence of direct neuronal depolarization. We will incorporate these considerations into the discussion, while also acknowledging that unequivocal confirmation of “direct” rebound excitation would require intracellular recordings, such as patch clamp experiments.

The authors indicate that the excitatory activity was indistinguishable in shape from other calcium activity, but can anything be said about the timing (the scale bar in Figure 4A2 has no value, is it the same 2s pulse)?

Apologies for oversight in labeling the scale bar in Figure 4A2 (it is 2s). While we deliberately refrain from making strong claims regarding the origin of the NO-evoked spikes, their timing can be examined in more detail in Figure 4 - Supplement 1, panels C and D. We will make sure this is clearly stated in the revised text.

Did the authors check for accidental sparse transfection with ChrimsonR of olivary neurons in the post-mortem analysis?

Good point! However, we have never seen this AAV9-based viral construct to drive trans-synaptic expression in the IO, nor is this version of AAV known to have the capacity for transsynaptic expression in general.

No sign of retrograde labeling (via the CF collaterals in the cerebellar nuclei) was seen either. Notably, the hSyn promoter used to drive ChrimsonR expression is extremely ineffective in the IO. Thus, we doubt that such accidental labeling could underlie the excitatory events seen upon N-O stimulation. We will add these mentions with relevant references to the discussion of the revised manuscript.

On page 18 the authors state that: "The lower SS rate was attributed to intrinsic factors of PNs, while the reduced frequency of CSs was speculated to result from increased inhibition of the IO via the nucleo-olivary (N-O) pathway targeting the same microzone." I think I understand what you mean to say, but this is a bit confusing.

Agreed. We will rephrase this sentence to clarify that a lower SS rate in a given microzone may lead to increased activation of inhibitory N-O axons that target the region of IO that sends CF to the same microzone.

Is airpuff stimulation not more likely to activate PO dan DAO because of the related modalities (more face vs. more trunk/limbs?), and thereby also more likely to drive event co-activation (as it is stated in the abstract).

We agree that the specific innervation patterns of different IO regions likely explain the discrepancy between previous reports of airpuff-evoked complex spikes in cerebellar cortical regions targeted by DAO and the absence of airpuff responses in the particular region of DAO accessible via our surgical approach. As in the present dataset virtually no airpuff-evoked events were seen in DAO regions, we are unable to directly compare airpuff-evoked event co-activation between PO and DAO. The higher co-activation for PO was observed for "spontaneous" activity.

The Discussion addresses the question of why N-O pathway activation does not remove the airpuff response.

Given the potentially profound effect, I would propose to expand the discussion on the role of aneasthesia, including longer refractory periods but also potential disruption of normal network interactions (even though individually the stimulations work). Briefly indicating what is known about alpha-chloralose would help interpret the results as well.

We fully agree that the anesthetic state introduces confounding factors that must be considered when interpreting our results. We will expand the discussion to address how anesthesia, particularly alphachloralose as well as tissue cooling, may contribute to prolonged refractory periods and potential disruptions in normal network interactions. However, we recognize that certain aspects cannot be fully resolved without recordings in awake animals. For this reason, we characterize our preparation as an "upgraded" in vitro approach rather than a fully representative in vivo model.

Please clearly indicate that the age range of P35-45 is for the moment of virus injection and specify the age range for the imaging experiment.

Apologies for the oversight. We will indicate these age ranges in the results (as they are currently only specified in Methods). The P35-45 range refers to moment of virus injection.

The methods indicate that a low-pass filter of 1Hz was used. I am sure this helps with smoothing, but does it not remove a lot of potentially interesting information. How would a higher low-pass filter affect the analysis and results?

We acknowledge that applying a 1 Hz low-pass filter inevitably removes high-frequency components, including potential IO oscillations and fine details such as spike "doublets." However, given the temporal resolution constraints of our recording approach, we prioritized capturing robust, interpretable events over attempting to extract finer features that might be obscured by both the indicator kinetics and imaging speed.

While a higher cut-off frequency could, in principle, allow more precise measurement of rise times and peak timings, it would also amplify high-frequency noise, complicating automated event detection and reducing confidence in distinguishing genuine neural signals from artifacts. Given these trade-offs, we opted for a conservative filtering approach to ensure stable event detection. Future work, particularly with faster imaging rates and improved sensors (GCaMP8s) will be used to explore the finer temporal structure of IO activity. We will deliberate on these matters more extensively in the revised discussion.

Reviewer #2 (Public review):

The authors developed a strategy to image inferior olive somata via viral GCaMP6s expression, an implanted GRIN lens, and a one-photon head-mounted microscope, providing the first in vivo somatic recordings from these neurons. The main new findings relate to the activation of the nucleoolivary pathway, specifically that: this manipulation does not produce a spiking rebound in the IO; it exerts a larger effect on spontaneous IO spiking than stimulus (airpuff)-evoked spiking. In addition, several findings previously demonstrated in vivo in Purkinje cell complex spikes or inferior olivary axons are confirmed here in olivary somata: differences in event sizes from single cells versus co-activated cells; reduced coactivation when activating the NO pathway; more coactivation within a single zebrin compartment.

The study presents some interesting findings, and for the most part, the analyses are appropriate. My two principal critiques are that the study does not acknowledge major technical limitations and their impact on the claims; and the study does not accurately represent prior work with respect to the current findings.

We thank the reviewer for recognising the value of the findings in our "reduced" in vivo preparation, and apologize for omissions in the work that led to critique. We will elaborate on these matters below and prepare a revised manuscript.

The authors use GCaMP6s, which has a tau1/2 of >1 s for a normal spike, and probably closer to 2 s (10.1038/nature12354) for the unique and long type of olivary spikes that give rise to axonal bursts (10.1016/j.neuron.2009.03.023). Indeed, the authors demonstrate as much (Fig. 2B1). This affects at least several claims:

a. The authors report spontaneous spike rates of 0.1 Hz. They attribute this to anesthesia, yet other studies under anesthesia recording Purkinje complex spikes via either imaging or electrophysiology report spike rates as high as 1.5 Hz (10.1523/JNEUROSCI.2525-10.2011). This discrepancy is not acknowledged and a plausible explanation is not given. Citations are not provided that demonstrate such low anesthetized spike rates, nor are citations provided for the claim that spike rates drop increasingly with increasing levels of anesthesia when compared to awake resting conditions.

We fully acknowledge that anesthesia is a major confounding factor in our study. Given the unusually invasive nature of our surgical preparation, we prioritized deep anesthesia to ensure the animals’ welfare. This, along with potential cooling effects from tissue removal and GRIN lens contact, likely contributed to the observed suppression of IO activity.

We recognize that reported complex spike rates under anesthesia vary considerably across studies, and we will expand our discussion to provide a more comprehensive comparison with prior literature. Notably, different anesthetic protocols, levels of anesthesia, and recording methodologies can lead to widely different estimates of firing rates. While we cannot resolve this issue without recordings in awake animals, we will clarify that our observed rates likely reflect both the effects of anesthesia and specific methodological constraints. We will also incorporate additional references to studies examining cerebellar activity under different anesthetic conditions.

More likely, this discrepancy reflects spikes that are missed due to a combination of the indicator kinetics and low imaging sensitivity (see (2)), neither of which are presented as possible plausible alternative explanations.

We acknowledge that the combination of slow indicator kinetics and limited optical power in our miniature microscope setup constrains the temporal resolution of our recordings. However, we are confident that we can reliably detect events occurring at intervals of 1 second or longer. This confidence is based on data from another preparation using the same viral vector and optical system, where we observed spike rates an order of magnitude higher.

That said, we do not make claims regarding the presence or absence of somatic events occurring at very short intervals (e.g., 100-ms "doublets," as described by Titley et al., 2019), as these would likely fall below our temporal resolution. We will clarify this limitation in the revised manuscript to ensure that the constraints of our approach are fully acknowledged.

While GCaMP6s is not as sensitive as more recent variants (Zhang et al., 2023, PMID 36922596), our previous work (Dorgans et al., 2022) demonstrated that its dynamic range and sensitivity are sufficient to detect both spikes and subthreshold activity in vitro. Although the experimental conditions differ in the current miniscope experiments, we took measures to optimize signal quality, including excluding recordings with a low signal-to-noise ratio (see Methods). This need for high signal fidelity also informed our decision to limit the sampling rate to 20 fps. In future work, we plan to adopt newer GCaMP variants that were not available at the start of this project, which should further improve sensitivity and temporal resolution.

Many claims are made throughout about co-activation ("clustering"), but with the GCaMP6s rise time to peak (0.5 s), there is little technical possibility to resolve co-activation. This limitation is not acknowledged as a caveat and the implications for the claims are not engaged with in the text.

As noted in the manuscript (L492-), "interpreting fluorescence signals relative to underlying voltage changes is challenging, particularly in IO neurons with unusual calcium dynamics." We acknowledge that the slow rise time of GCaMP6s ( 0.5 s) limits our ability to precisely resolve the timing of co-activation at very short intervals. However, given the relatively slow timescales of IO event clustering and the inherent synchrony in olivary network dynamics, we believe that the observed co-activation patterns remain meaningful, even if finer temporal details cannot be fully resolved.

To ensure clarity, we will expand this section to explicitly acknowledge the temporal resolution limitations of our approach and discuss their implications for interpreting co-activation. While the precise timing of individual spikes within a cluster may not be resolvable, the observed increase in event magnitude with coarse co-activation suggests that clustering effects remain functionally relevant even when exact spike synchrony is not detectable at millisecond resolution.

This finding is consistent with the idea that co-activation enhances calcium influx, leading to larger amplitude events — a relationship that does not require perfect temporal resolution to be observed. The fact that this effect persists across a broad range of clustering windows (as shown in Figure 2 Supplement 2) further supports its robustness. While we cannot make strong claims about precise spike timing within these clusters nor about the mechanism underlying enhanced calcium signal, our results demonstrate that co-activation may influence IO activity in a quantifiable way. We will clarify these points in the revised manuscript to ensure that our findings are appropriately framed given the temporal constraints of our imaging approach.

The study reports an ultralong "refractory period" (L422-etc) in the IO, but this again must be tempered by the possibility that spikes are simply being missed due to very slow indicator kinetics and limited sensitivity. Indeed, the headline numeric estimate of 1.5 s (L445) is suspiciously close to the underlying indicator kinetic limitation of 1-2 s.

Our findings suggest a potential refractory period limiting the frequency of events in the inferior olive under our recording conditions. This interpretation is supported by the observed inter-event interval distribution, the inability of N-O stimulation to suppress airpuff-evoked events, and lower bounds reported in earlier literature on complex spike intervals recorded in awake animals under various behavioral contexts. Taking into account the likely cooling of tissue, a refractory period of 1.5s is not unreasonable. Of course, we recognize that the slow decay kinetics of GCaMP6s may cause overlapping fluorescence signals, potentially obscuring closely spaced events. This is in line with data presented in the Chen et al 2013 manuscript describing GCaMp6s (PMID: 36922596; Figure 3b showing events detected with intervals less than 500 ms).

The consideration of refractoriness only arose late in the project while we were investigating the explanations for lack of inhibition of airpuff-evoked spikes. Future experiments, particularly in awake animals, will be instrumental in validating this interpretation. To ensure that the refractory period is understood as one possible mechanism rather than a definitive explanation, we will rephrase the discussion to clarify that while our data are compatible with a refractory period, they do not establish it conclusively.

The study uses endoscopic one-photon miniaturized microscope imaging. Realistically, this is expected to permit an axial point spread function (z-PSF) on the order of 40um, which must substantially reduce resolution and sensitivity. This means that if there *is* local coactivation, the data in this study will very likely have individual ROIs that integrate signals from multiple neighboring cells. The study reports relationships between event magnitude and clustering, etc; but a fluorescence signal that contains photons contributed by multiple neighboring neurons will be larger than a single neuron, regardless of the underlying physiology - the text does not acknowledge this possibility or limitation.

We acknowledge that the use of one-photon endoscopic imaging imposes limitations on axial resolution, potentially leading to signal contributions from neighboring neurons. To mitigate this, we applied CNMFe processing, which allows for the deconvolution of overlapping signals and the differentiation of multiple neuronal sources within shared pixels. However, as the reviewer points out, if two neurons are perfectly overlapping in space, they may be treated as a single unit.

To clarify this limitation, we will expand the discussion to explicitly acknowledge the impact of one-photon imaging on signal separation and to emphasize that, while CNMFe helps resolve some overlaps, perfect separation is not always possible. As already noted in the manuscript (L495-), "the absence of optical sectioning in the whole-field imaging method can lead to confounding artifacts in densely labeled structures such as the IO’s tortuous neuropil." We will further elaborate on how this factor was considered in our analysis and interpretation.

Second, the text makes several claims for the first multicellular in vivo olivary recordings. (L11; L324, etc).

I am aware of at least two studies that have recorded populations of single olivary axons using two-photon Ca2+ imaging up to 6 years ago (10.1016/j.neuron.2019.03.010; 10.7554/eLife.61593). This technique is not acknowledged or discussed, and one of these studies is not cited. No argument is presented for why axonal imaging should not "count" as multicellular in vivo olivary recording: axonal Ca2+ reflects somatic spiking.

We appreciate the reviewer’s point and acknowledge the important prior work using two-photon imaging to record olivary axonal activity in the cerebellar cortex. However, while axonal calcium signals do reflect somatic spiking, these recordings inherently lack information about the local network interactions within the inferior olive itself.

A key motivation for our study was to observe neuronal activity within the IO at the level of its gap-junctioncoupled local circuits, rather than at the level of its divergent axonal outputs. The fan-like spread of climbing fibers across rostrocaudal microzones in the cerebellar cortex makes them relatively easy to record in vivo, but it also means that individual imaging fields contain axons from neurons that may be distributed across different IO microdomains. As a result, while previous work has provided valuable insight into olivary output patterns, it has not allowed for the examination of coordinated somatic activity within localized IO neuron clusters.

With apologies, we recognize that this distinction was not sufficiently emphasized in our introduction. We will clarify this key point and ensure that the important climbing fiber imaging studies are properly cited and contextualized in the revised manuscript.

Reviewer #2 (Recommendations for the authors):

The authors state: "we found no reports that examined coactivation levels between Z+ and Z- microzones in cerebellar complex spike recordings" (L359). Multiple papers (that are not cited) using AldolaceC-tdTomato mice with two photon Purkinje dendritic calcium imaging showed synchronization (at similar levels) within but not across z+/z- bands. (2015 10.1523/JNEUROSCI.2170-14.2015, 2023 https://doi.org/10.7554/eLife.86340).

We apologize for the misleading phrasing. We will rephrase this statement to: "While complex spike coactivation within individual zebrin zones has been extensively studied (references), we found no reports directly comparing the levels of intra-zone co-activation between Z+ and Z microzones."

Additionally, we will ensure that the relevant studies demonstrating synchronization within zebrin zones, as well as (lack of) interactions between neighboring zones, are properly cited and discussed in the revised manuscript.

The figures could use more proofreading, and several decisions should be reconsidered:

Normalizing the amplitude to maximum is not a good strategy, as it can overemphasize noise or extremely small-magnitude signals, and should instead follow standard convention and present in fixed units (3A2, 4B2, and even 2C).

As noted earlier, we have excluded recordings and cells with high noise or a low signal-to-noise ratio for event amplitudes, ensuring that such data do not influence the color-coded panels. Importantly, all quantitative analyses and traces presented in the manuscript are normalized to baseline noise level, not to maximal amplitude, ensuring that noise or low-magnitude signals do not skew the analysis.

The decision to use max-amplitude normalization in color-coded panels was made specifically to aid visualization of temporal structure across recordings. This approach allows for clearer comparisons without the distraction of inter-cell variability in absolute signal strength. However, we recognize the potential for confusion and will revise the Results text to explicitly clarify that the color-coded visualizations use a different scaling method than the quantitative analyses.

x axes with no units: Figures 2B2, 2E1, 3B2, 3C2, 5B2, 5C2, 5D2.

No colorbar units: 5A3 (and should be shown in real not normalized units).

No y axis units: 5D1.

No x axis label or units: 5E1.

5E3 says "stim/baseline" for the y-axis units and then the first-panel title says "absolute frequencies" meaning it’s *not* normalized and needs a separate (accurate) y-axis with units.

Illegibly tiny fonts: 2E1, 3E1, etc.

We will correct all these in the revised manuscript. Thank you for careful reading.

Kernel data structures maintain essential state information about I/O activities, ensuring efficient system operation. UNIX’s open-file table, for example, tracks file descriptors, file system records, and active inodes, streamlining file management. Similar structures exist for network connections and character devices. Object-oriented approaches further enhance modularity, as seen in UNIX’s dispatch tables and Windows’ message-passing system. In Windows, I/O requests are encapsulated as messages, enabling flexible interactions between the kernel, I/O manager, and device drivers. Although this approach introduces additional processing overhead, it simplifies I/O management and enhances system flexibility. These structured methods ensure that operating systems can efficiently track, manage, and process diverse I/O operations, leading to more stable and efficient computing environments.

Spooling is a technique used to manage output devices that cannot handle interleaved data streams, such as printers. Instead of sending data directly to the device, the system stores it in a spool—a designated buffer in secondary storage—ensuring orderly processing. A spooler daemon manages the queue, allowing users to monitor and manipulate print jobs. Additionally, some devices require exclusive access to prevent conflicts. Device reservation mechanisms prevent multiple processes from simultaneously accessing non-shareable devices like tape drives. Operating systems enforce these constraints through explicit allocation requests or file-handle restrictions, ensuring coordinated access. In environments like Windows, system calls allow processes to wait for device availability, preventing deadlocks. Effective spooling and reservation strategies optimize device utilization and maintain orderly, conflict-free access to critical system resources.

Due to distinct performance and addressing characteristics, network I/O employs a different interface than disk I/O. Many operating systems, including UNIX and Windows, implement the socket interface for network communication. The socket system calls enable applications to create, connect, listen, send, and receive data over network connections. The select() function is highlighted for managing multiple sockets efficiently, allowing applications to detect readable or writable sockets without continuous polling. Network communication in UNIX is further explored, showcasing diverse interprocess communication methods such as pipes, FIFOs, message queues, and STREAMS. The section underscores the importance of encapsulating network behaviors through standardized interfaces, simplifying the development of distributed applications that function across different network hardware and protocol stacks.

There are three general approaches to handling deadlocks:

1. Ignoring Deadlocks: Many operating systems, including Linux and Windows, do not implement specific deadlock-handling mechanisms. Instead, they leave the responsibility to kernel and application developers to prevent and resolve deadlocks.

2. Deadlock Prevention or Avoidance: This approach ensures that a system never enters a deadlocked state. Deadlock prevention works by ensuring that at least one of the necessary conditions for deadlocks is eliminated. Deadlock avoidance requires prior knowledge of resource requests to make decisions that prevent circular waits.

3. Deadlock Detection and Recovery: Some systems, such as databases, allow deadlocks to occur and then implement algorithms to detect and recover from them. If no detection and recovery mechanisms exist, the system may deteriorate until manual intervention is required.

Each approach has its own advantages and drawbacks. Ignoring deadlocks is cost-effective but risks system failure. Prevention and avoidance require additional constraints on resource requests, which may reduce system efficiency. Detection and recovery offer flexibility but may introduce significant computational overhead.

Windows threads follow a one-to-one mapping, where each user-level thread corresponds to a kernel thread. Every thread has essential components, including a thread ID, program counter, register set, and user/kernel stacks. Windows threads are represented by key data structures: ETHREAD (executive thread block), KTHREAD (kernel thread block), and TEB (thread environment block). The ETHREAD links the thread to its parent process, while KTHREAD manages scheduling and kernel stack usage. The TEB resides in user space, containing thread-local storage and other user-mode data. Since ETHREAD and KTHREAD exist in kernel space, only the operating system can access them. This structured approach ensures efficient thread management, optimizing performance in multithreaded Windows applications.

Multithreaded programs face several challenges, including handling system calls like fork() and exec(). The behavior of fork() varies—some UNIX implementations duplicate all threads, while others duplicate only the calling thread. If exec() is called immediately after forking, duplicating all threads is unnecessary. Signal handling is another challenge, as signals can be synchronous (e.g., division by zero) or asynchronous (e.g., termination signals). Signals in multithreaded programs can be delivered to a specific thread, all threads, or certain threads based on signal type. POSIX Pthreads provide pthread_kill() to direct signals to a specific thread. Windows uses Asynchronous Procedure Calls (APCs) to handle event-driven notifications, similar to UNIX signals. Developers must ensure proper signal handling to avoid unintended behavior. Thread cancellation is another concern, requiring careful implementation to prevent resource leaks and ensure that terminated threads do not leave operations incomplete.

Intel TBB is a powerful template library for parallel programming in C++. Unlike OpenMP, it does not require compiler support but instead provides a task-based approach for parallel execution. The TBB task scheduler dynamically maps tasks to threads, optimizing load balancing and cache efficiency. The parallel_for template automates parallelization of loops, ensuring efficient distribution of workload. This method abstracts hardware-specific optimizations, making it adaptable across different multicore systems. TBB also provides concurrent data structures such as thread-safe hash maps and vectors, enhancing performance in multithreaded applications. The flexibility of TBB allows developers to scale applications efficiently without modifying code for different hardware configurations. While TBB offers advanced features for parallelism, it requires a solid understanding of lambda functions and task dependencies to maximize performance. Proper use of TBB ensures improved computational efficiency in complex applications.

The Java Virtual Machine (JVM) operates on top of a host OS, abstracting platform-specific threading details. The JVM does not dictate how Java threads map to OS threads, leaving it to individual implementations. Windows, for example, employs a one-to-one threading model, meaning each Java thread corresponds to a kernel thread. This adaptability allows Java applications to run consistently across different OS environments, utilizing native threading libraries such as Windows API or Pthreads on Linux/macOS.

A thread library provides an API for creating and managing threads, allowing developers to handle multithreading efficiently. These libraries can be implemented at either the user level or the kernel level.

A user-level thread library operates in user space, avoiding system calls for better performance. However, if a thread makes a blocking system call, all threads in the process may be blocked.

A kernel-level thread library is supported directly by the operating system. While slightly slower due to system calls, it allows better thread management and parallel execution.

Some of the most widely used thread libraries include POSIX Pthreads (common in UNIX/Linux systems), Windows Threads, and Java Threads, which rely on the underlying OS for execution.

In the one-to-one model, each user thread is directly mapped to a kernel thread. This model improves concurrency because if one thread blocks, others can continue execution. Furthermore, it supports true parallelism, allowing threads to run on multiple processors. Despite these advantages, the model comes with high overhead. Creating a user thread also requires creating a corresponding kernel thread, which can lead to resource exhaustion if too many threads are created.

Remote Procedure Calls (RPCs) abstract the communication process between distributed systems by allowing a client to invoke procedures on a remote machine as if they were local functions. Unlike raw socket communication, which requires applications to structure their own messages, RPCs handle function calls transparently. A key component of RPCs is parameter marshaling, which ensures data compatibility between different architectures (e.g., big-endian vs. little-endian). Additionally, RPC implementations must handle network failures and duplicate execution risks, requiring techniques such as timestamping and acknowledgment messages. The use of a matchmaker service for dynamic binding enhances flexibility, allowing clients to locate available RPC services at runtime rather than relying on fixed port assignments.

Windows systems implement ordinary pipes, called anonymous pipes, with the CreatePipe() function. The difference from UNIX is that Windows requires explicit handling of pipe inheritance for child processes. The parent creates the pipe, and when invoking CreateProcess(), the parent's pipe handles must be passed to the child. This is achieved by manipulating the STARTUPINFO structure to redirect the child’s standard input and output to the pipe’s read and write ends.

Once the child process is created, the parent writes to the pipe using WriteFile(), and the child reads from the pipe using ReadFile(). After communication, both ends of the pipe are closed. The main advantage of Windows anonymous pipes is the ability to explicitly control the inheritance of pipe handles, ensuring that the child only has access to the required pipe ends. This approach facilitates secure and isolated communication between the parent and child processes.

Windows ALPC (Advanced Local Procedure Call) provides a message-passing mechanism for communication between processes, optimized for performance within the Windows environment. Windows uses connection ports for clients to request services from server processes. ALPC supports different message-passing methods, depending on the message size, including smaller messages using a message queue and larger ones via section objects, which leverage shared memory. Notably, ALPC is designed to handle local communication on the same machine efficiently, and its internal complexity allows clients to make decisions about message size to avoid unnecessary copying. While ALPC is a powerful IPC mechanism, it is not directly exposed to application developers but is used behind the scenes by the system's standard remote procedure call (RPC) system.

The section on Pipes focuses on a basic but foundational IPC method in UNIX and Windows systems, where data is transmitted between two processes through a unidirectional or bidirectional conduit. The design of pipes involves considerations like whether they allow bidirectional communication, whether a relationship is required between processes (such as parent–child), and whether communication can occur across a network. Named pipes, an extension of regular pipes, allow for more flexible communication, as they can be used by unrelated processes and even across machines, though they still rely on local IPC principles.

Reviewer #1 (Public review):

Summary:

The manuscript puts forward a statistical method to more accurately report the significance of correlations within data. The motivation for this study is two-fold. First, the publication of biological studies demands the report of p-values, and it is widely accepted that p-values below the arbitrary threshold of 0.05 give the authors of such studies justification to draw conclusions about their data. Second, many biological studies are limited by the number of replicate samples that are feasible, with replicates of less than 5 typical. The authors report a statistical tool that uses a permute-match approach to calculate p-values. Notably, the proposed method reduces p-values from around 0.2 to 0.04 as compared to a standard permutation test with a small sample size. The approach is clearly explained, including detailed mathematical explanations and derivations. The advantage of the approach is also demonstrated through analysis of computer-generated synthetic data with specified correlation and analysis of previously published data related to fish schooling. The authors make a clear case that this method is an improvement over the more standard approach currently used, and also demonstrate the impact of this methodology on the ability to obtain p-values that are the standard for biological research. Overall, this paper is very strong. While the subject matter seems somewhat specialized, I would make the case that this will be an important study that has broad general interest to readers. The findings are very general and applicable to many research contexts. Experimentalists also want to report accurate p-values in their work and better understand how these values are calculated. Although I believe the previous statement is true, I am not sure that many research groups doing biological work are reading specialized statistics journals regularly. Therefore a useful and broadly applicable statistical tool is well placed in this journal.<br /> Strengths:

The proposed method is broadly applicable to many realistic datasets in many experimental contexts.

The power of this method was demonstrated with both real experimental data and "synthetic" data. The advantages of the tool are clearly reported. The zebrafish data is a great example dataset.

The method solves a real-life problem that is frequently encountered by many experimental groups in the biological sciences.

The writing of the paper is surprisingly clear, given the technical nature of the subject matter. I would not at all consider myself a statistician or mathematician, but I found the text easy to follow. The authors did an impressive job guiding the reader through material that would often be difficult to grasp. The introduction was also well-written and clearly motivated the goals of the study.

Weaknesses:

A few changes could be made if the manuscript is revised. I would consider all of these points minor, but the paper could be improved if these points were addressed.

(1) The caption of Figure 2 doesn't seem to mention panel D. Figure A-2 also does not mention C in the caption.

(2) Figure 2D is a little hard to follow. First, the definition of "Power" is not clear, and I couldn't find the precise definition in the text. Second, the legend for the different lines in 2D is only given in Figure A-2. Perhaps a portion of the caption for Figure 2 is missing?

(3) The concept of circular variance for the fish data was heard to understand/visualize. The equation on line 326 did not help much. If there is a very simple picture that could be added near line 326 that helps to explain Ct and theta, that could be a big help for some readers who do not work on related systems. The analysis performed is understandable, the reader just has to accept that circular variance captions the degree of alignment of the fish.

(4) For the data discussed in Figure 3, I wasn't 100% sure how the time windows were selected. In the caption, it says "time series to different lengths starting from the first frame". So the 20 s time window was from t=0 to t= 20 s. Would a different result be obtained if a different 20 s window was chosen (from t = 4 min to t = 4 min 20 s just to give a specific example). I suppose by chance one of the time windows would give a p-value less than the target 0.05, that wouldn't be surprising. Maybe a random time window should be selected (although I am not indicating what was reported was incorrect)? A little more discussion on this aspect of the study may be helpful.

Author response:

The following is the authors’ response to the original reviews.

Public Reviews

Reviewer #1 (Public Review):

Summary:

The authors have created a system for designing and running experimental pipelines to control and coordinate different programs and devices during an experiment, called Heron. Heron is based around a graphical tool for creating a Knowledge Graph made up of nodes connected by edges, with each node representing a separate Python script, and each edge being a communication pathway connecting a specific output from one node to an iput on another. Each node also has parameters that can be set by the user during setup and runtime, and all of this behavior is concisely specified in the code that defines each node. This tool tries to marry the ease of use, clarity, and selfdocumentation of a purely graphical system like Bonsai with the flexibility and power of a purely code-based system like Robot Operating System (ROS).

Strengths:

The underlying idea behind Heron, of combining a graphical design and execution tool with nodes that are made as straightforward Python scripts seems like a great way to get the relative strengths of each approach. The graphical design side is clear, selfexplanatory, and self-documenting, as described in the paper. The underlying code for each node tends to also be relatively simple and straightforward, with a lot of the complex communication architecture successfully abstracted away from the user. This makes it easy to develop new nodes, without needing to understand the underlying communications between them. The authors also provide useful and well-documented templates for each type of node to further facilitate this process. Overall this seems like it could be a great tool for designing and running a wide variety of experiments, without requiring too much advanced technical knowledge from the users.

The system was relatively easy to download and get running, following the directions and already has a significant amount of documentation available to explain how to use it and expand its capabilities. Heron has also been built from the ground up to easily incorporate nodes stored in separate Git repositories and to thus become a large community-driven platform, with different nodes written and shared by different groups. This gives Heron a wide scope for future utility and usefulness, as more groups use it, write new nodes, and share them with the community. With any system of this sort, the overall strength of the system is thus somewhat dependent on how widely it is used and contributed to, but the authors did a good job of making this easy and accessible for people who are interested. I could certainly see Heron growing into a versatile and popular system for designing and running many types of experiments.

Weaknesses:

(1) The number one thing that was missing from the paper was any kind of quantification of the performance of Heron in different circumstances. Several useful and illustrative examples were discussed in depth to show the strengths and flexibility of Heron, but there was no discussion or quantification of performance, timing, or latency for any of these examples. These seem like very important metrics to measure and discuss when creating a new experimental system.

Heron is practically a thin layer of obfuscation of signal passing across processes. Given its design approach it is up to the code of each Node to deal with issues of timing, synching and latency and thus up to each user to make sure the Nodes they author fulfil their experimental requirements. Having said that, Heron provides a large number of tools to allow users to optimise the generated Knowledge Graphs for their use cases. To showcase these tools, we have expanded on the third experimental example in the paper with three extra sections, two of which relate to Heron’s performance and synching capabilities. One is focusing on Heron’s CPU load requirements (and existing Heron tools to keep those at acceptable limits) and another focusing on post experiment synchronisation of all the different data sets a multi Node experiment generates.   

(2) After downloading and running Heron with some basic test Nodes, I noticed that many of the nodes were each using a full CPU core on their own. Given that this basic test experiment was just waiting for a keypress, triggering a random number generator, and displaying the result, I was quite surprised to see over 50% of my 8-core CPU fully utilized. I don’t think that Heron needs to be perfectly efficient to accomplish its intended purpose, but I do think that some level of efficiency is required. Some optimization of the codebase should be done so that basic tests like this can run with minimal CPU utilization. This would then inspire confidence that Heron could deal with a real experiment that was significantly more complex without running out of CPU power and thus slowing down.

The original Heron allowed the OS to choose how to manage resources over the required process. We were aware that this could lead to significant use of CPU time, as well as occasionally significant drop of packets (which was dependent on the OS and its configuration). This drop happened mainly when the Node was running a secondary process (like in the Unity game process in the 3rd example). To mitigate these problems, we have now implemented a feature allowing the user to choose the CPU that each Node’s worker function runs on as well as any extra processes the worker process initialises. This is accessible from the Saving secondary window of the node. This stops the OS from swapping processes between CPUs and eliminates the dropping of packages due to the OS behaviour. It also significantly reduces the utilised CPU time. To showcase this, we initially run the simple example mentioned by the reviewer. The computer running only background services was using 8% of CPU (8 cores). With Heron GUI running but with no active Graph, the CPU usage went to 15%. With the Graph running and Heron’s processes running on OS attributed CPU cores, the total CPU was at 65% (so very close to the reviewer’s 50%). By choosing a different CPU core for each of the three worker processes the CPU went down to 47% and finally when all processes were forced to run on the same CPU core the CPU load dropped to 30%.  So, Heron in its current implementation running its GUI and 3 Nodes takes 22% of CPU load. This is still not ideal but is a consequence of the overhead of running multiple processes vs multiple threads. We believe that, given Heron’s latest optimisation, offering more control of system management to the user, the benefits of multi process applications outweigh this hit in system resources. 

We have also increased the scope of the third example we provide in the paper and there we describe in detail how a full-scale experiment with 15 Nodes (which is the upper limit of number of Nodes usually required in most experiments) impacts CPU load. 

Finally, we have added on Heron’s roadmap projects extra tasks focusing only on optimisation (profiling and using Numba for the time critical parts of the Heron code).

(3) I was also surprised to see that, despite being meant specifically to run on and connect diverse types of computer operating systems and being written purely in Python, the Heron Editor and GUI must be run on Windows. This seems like an unfortunate and unnecessary restriction, and it would be great to see the codebase adjusted to make it fully crossplatform-compatible.

This point was also mentioned by reviewer 2. This was a mistake on our part and has now been corrected in the paper. Heron (GUI and underlying communication functionality) can run on any machine that the underlying python libraries run, which is Windows, Linux (both for x86 and Arm architectures) and MacOS. We have tested it on Windows (10 and 11, both x64), Linux PC (Ubuntu 20.04.6, x64) and Raspberry Pi 4 (Debian GNU/Linux 12 (bookworm), aarch64). The Windows and Linux versions of Heron have undergone extensive debugging and all of the available Nodes (that are not OS specific) run on those two systems. We are in the process of debugging the Nodes’ functionality for RasPi. The MacOS version, although functional requires further work to make sure all of the basic Nodes are functional (which is not the case at the moment). We have also updated our manuscript (Multiple machines, operating systems and environments) to include the above information. 

(4) Lastly, when I was running test experiments, sometimes one of the nodes, or part of the Heron editor itself would throw an exception or otherwise crash. Sometimes this left the Heron editor in a zombie state where some aspects of the GUI were responsive and others were not. It would be good to see a more graceful full shutdown of the program when part of it crashes or throws an exception, especially as this is likely to be common as people learn to use it. More problematically, in some of these cases, after closing or force quitting Heron, the TCP ports were not properly relinquished, and thus restarting Heron would run into an "address in use" error. Finding and killing the processes that were still using the ports is not something that is obvious, especially to a beginner, and it would be great to see Heron deal with this better. Ideally, code would be introduced to carefully avoid leaving ports occupied during a hard shutdown, and furthermore, when the address in use error comes up, it would be great to give the user some idea of what to do about it.

A lot of effort has been put into Heron to achieve graceful shut down of processes, especially when these run on different machines that do not know when the GUI process has closed. The code that is being suggested to avoid leaving ports open has been implemented and this works properly when processes do not crash (Heron is terminated by the user) and almost always when there is a bug in a process that forces it to crash. In the version of Heron available during the reviewing process there were bugs that caused the above behaviour (Node code hanging and leaving zombie processes) on MacOS systems. These have now been fixed. There are very seldom instances though, especially during Node development, that crashing processes will hang and need to be terminated manually. We have taken on board the reviewer’s comments that users should be made more aware of these issues and have also described this situation in the Debugging part of Heron’s documentation. There we explain the logging and other tools Heron provides to help users debug their own Nodes and how to deal with hanging processes.

Heron is still in alpha (usable but with bugs) and the best way to debug it and iron out all the bugs in all use cases is through usage from multiple users and error reporting (we would be grateful if the errors the reviewer mentions could be reported in Heron’s github Issues page). We are always addressing and closing any reported errors, since this is the only way for Heron to transition from alpha to beta and eventually to production code quality.

Overall I think that, with these improvements, this could be the beginning of a powerful and versatile new system that would enable flexible experiment design with a relatively low technical barrier to entry. I could see this system being useful to many different labs and fields. 

We thank the reviewer for positive and supportive words and for the constructive feedbacks. We believe we have now addressed all the raised concerns.  

Reviewer #2 (Public Review):

Summary:

The authors provide an open-source graphic user interface (GUI) called Heron, implemented in Python, that is designed to help experimentalists to

(1) design experimental pipelines and implement them in a way that is closely aligned with their mental schemata of the experiments,

(2) execute and control the experimental pipelines with numerous interconnected hardware and software on a network.

The former is achieved by representing an experimental pipeline using a Knowledge Graph and visually representing this graph in the GUI. The latter is accomplished by using an actor model to govern the interaction among interconnected nodes through messaging, implemented using ZeroMQ. The nodes themselves execute user-supplied code in, but not limited to, Python.

Using three showcases of behavioral experiments on rats, the authors highlighted three benefits of their software design:

(1) the knowledge graph serves as a self-documentation of the logic of the experiment, enhancing the readability and reproducibility of the experiment,

(2) the experiment can be executed in a distributed fashion across multiple machines that each has a different operating system or computing environment, such that the experiment can take advantage of hardware that sometimes can only work on a specific computer/OS, a commonly seen issue nowadays,

(3) he users supply their own Python code for node execution that is supposed to be more friendly to those who do not have a strong programming background.

Strengths:

(1) The software is light-weight and open-source, provides a clean and easy-to-use GUI,

(2) The software answers the need of experimentalists, particularly in the field of behavioral science, to deal with the diversity of hardware that becomes restricted to run on dedicated systems.

(3) The software has a solid design that seems to be functionally reliable and useful under many conditions, demonstrated by a number of sophisticated experimental setups.

(4) The software is well documented. The authors pay special attention to documenting the usage of the software and setting up experiments using this software.

Weaknesses:

(1) While the software implementation is solid and has proven effective in designing the experiment showcased in the paper, the novelty of the design is not made clear in the manuscript. Conceptually, both the use of graphs and visual experimental flow design have been key features in many widely used softwares as suggested in the background section of the manuscript. In particular, contrary to the authors’ claim that only pre-defined elements can be used in Simulink or LabView, Simulink introduced MATLAB Function Block back in 2011, and Python code can be used in LabView since 2018. Such customization of nodes is akin to what the authors presented.

In the Heron manuscript we have provided an extensive literature review of existing systems from which Heron has borrowed ideas. We never wished to say that graphs and visual code is what sets Heron apart since these are technologies predating Heron by many years and implemented by a large number of software. We do not believe also that we have mentioned that LabView or Simulink can utilise only predefined nodes. What we have said is that in such systems (like LabView, Simulink and Bonsai) the focus of the architecture is on prespecified low level elements while the ability for users to author their own is there but only as an afterthought. The difference with Heron is that in the latter the focus is on the users developing their own elements. One could think of LabView style software as node-based languages (with low level visual elements like loops and variables) that also allow extra scripting while Heron is a graphical wrapper around python where nodes are graphical representations of whole processes. To our knowledge there is no other software that allows the very fast generation of graphical elements representing whole processes whose communication can also be defined graphically. Apart from this distinction, Heron also allows a graphical approach to writing code for processes that span different machines which again to our knowledge is a novelty of our approach and one of its strongest points towards ease of experimental pipeline creation (without sacrificing expressivity). 

(2) The authors claim that the knowledge graph can be considered as a self-documentation of an experiment. I found it to be true to some extent. Conceptually it’s a welcoming feature and the fact that the same visualization of the knowledge graph can be used to run and control experiments is highly desirable (but see point 1 about novelty). However, I found it largely inadequate for a person to understand an experiment from the knowledge graph as visualized in the GUI alone. While the information flow is clear, and it seems easier to navigate a codebase for an experiment using this method, the design of the GUI does not make it a one-stop place to understand the experiment. Take the Knowledge Graph in Supplementary Figure 2B as an example, it is associated with the first showcase in the result section highlighting this self-documentation capability. I can see what the basic flow is through the disjoint graph where 1) one needs to press a key to start a trial, and 2) camera frames are saved into an avi file presumably using FFMPEG. Unfortunately, it is not clear what the parameters are and what each block is trying to accomplish without the explanation from the authors in the main text. Neither is it clear about what the experiment protocol is without the help of Supplementary Figure 2A.

In my opinion, text/figures are still key to documenting an experiment, including its goals and protocols, but the authors could take advantage of the fact that they are designing a GUI where this information, with properly designed API, could be easily displayed, perhaps through user interaction. For example, in Local Network -> Edit IPs/ports in the GUI configuration, there is a good tooltip displaying additional information for the "password" entry. The GUI for the knowledge graph nodes can very well utilize these tooltips to show additional information about the meaning of the parameters, what a node does, etc, if the API also enforces users to provide this information in the form of, e.g., Python docstrings in their node template. Similarly, this can be applied to edges to make it clear what messages/data are communicated between the nodes. This could greatly enhance the representation of the experiment from the Knowledge graph.

In the first showcase example in the paper “Probabilistic reversal learning.

Implementation as self-documentation” we go through the steps that one would follow in order to understand the functionality of an experiment through Heron’s Knowledge Graph. The Graph is not just the visual representation of the Nodes in the GUI but also their corresponding code bases. We mention that the way Heron’s API limits the way a Node’s code is constructed (through an Actor based paradigm) allows for experimenters to easily go to the code base of a specific Node and understand its 2 functions (initialisation and worker) without getting bogged down in the code base of the whole Graph (since these two functions never call code from any other Nodes). Newer versions of Heron facilitate this easy access to the appropriate code by also allowing users to attach to Heron their favourite IDE and open in it any Node’s two scripts (worker and com) when they double click on the Node in Heron’s GUI. On top of this, Heron now (in the versions developed as answers to the reviewers’ comments) allows Node creators to add extensive comments on a Node but also separate comments on the Node’s parameters and input and output ports. Those can be seen as tooltips when one hovers over the Node (a feature that can be turned off or on by the Info button on every Node).  

As Heron stands at the moment we have not made the claim that the Heron GUI is the full picture in the self-documentation of a Graph. We take note though the reviewer’s desire to have the GUI be the only tool a user would need to use to understand an experimental implementation. The solution to this is the same as the one described by the reviewer of using the GUI to show the user the parts of the code relevant to a specific Node without the user having to go to a separate IDE or code editor. The reason this has not been implemented yet is the lack of a text editor widget in the underlying gui library (DearPyGUI). This is in their roadmap for their next large release and when this exists we will use it to implement exactly the idea the reviewer is suggesting, but also with the capability to not only read comments and code but also directly edit a Node’s code (see Heron’s roadmap). Heron’s API at the moment is ideal for providing such a text editor straight from the GUI.

(3) The design of Heron was primarily with behavioral experiments in mind, in which highly accurate timing is not a strong requirement. Experiments in some other areas that this software is also hoping to expand to, for example, electrophysiology, may need very strong synchronization between apparatus, for example, the record timing and stimulus delivery should be synced. The communication mechanism implemented in Heron is asynchronous, as I understand it, and the code for each node is executed once upon receiving an event at one or more of its inputs. The paper, however, does not include a discussion, or example, about how Heron could be used to address issues that could arise in this type of communication. There is also a lack of information about, for example, how nodes handle inputs when their ability to execute their work function cannot keep up with the frequency of input events. Does the publication/subscription handle the queue intrinsically? Will it create problems in real-time experiments that make multiple nodes run out of sync? The reader could benefit from a discussion about this if they already exist, and if not, the software could benefit from implementing additional mechanisms such that it can meet the requirements from more types of experiments.

In order to address the above lack of explanation (that also the first reviewer pointed out) we expanded the third experimental example in the paper with three more sections. One focuses solely on explaining how in this example (which acquires and saves large amounts of data from separate Nodes running on different machines) one would be able to time align the different data packets generated in different Nodes to each other. The techniques described there are directly implementable on experiments where the requirements of synching are more stringent than the behavioural experiment we showcase (like in ephys experiments). 

Regarding what happens to packages when the worker function of a Node is too slow to handle its traffic, this is mentioned in the paper (Code architecture paragraph): “Heron is designed to have no message buffering, thus automatically dropping any messages that come into a Node’s inputs while the Node’s worker function is still running.” This is also explained in more detail in Heron’s documentation. The reasoning for a no buffer system (as described in the documentation) is that for the use cases Heron is designed to handle we believe there is no situation where a Node would receive large amounts of data in bursts while very little data during the rest of the time (in which case a buffer would make sense). Nodes in most experiments will either be data intensive but with a constant or near constant data receiving speed (e.g. input from a camera or ephys system) or will have variable data load reception but always with small data loads (e.g. buttons). The second case is not an issue and the first case cannot be dealt with a buffer but with the appropriate code design, since buffering data coming in a Node too slow for its input will just postpone the inevitable crash. Heron’s architecture principle in this case is to allow these ‘mistakes’ (i.e. package dropping) to happen so that the pipeline continues to run and transfer the responsibility of making Nodes fast enough to the author of each Node. At the same time Heron provides tools (see the Debugging section of the documentation and the time alignment paragraph of the “Rats playing computer games”  example in the manuscript) that make it easy to detect package drops and either correct them or allow them but also allow time alignment between incoming and outgoing packets. In the very rare case where a buffer is required Heron’s do-it-yourself logic makes it easy for a Node developer to implement their own Node specific buffer.

(4) The authors mentioned in "Heron GUI’s multiple uses" that the GUI can be used as an experimental control panel where the user can update the parameters of the different Nodes on the fly. This is a very useful feature, but it was not demonstrated in the three showcases. A demonstration could greatly help to support this claim.

As the reviewer mentions, we have found Heron’s GUI double role also as an experimental on-line controller a very useful capability during our experiments. We have expanded the last experimental example to also showcase this by showing how on the “Rats playing computer games” experiment we used the parameters of two Nodes to change the arena’s behaviour while the experiment was running, depending on how the subject was behaving at the time (thus exploring a much larger set of parameter combinations, faster during exploratory periods of our shaping protocols construction). 

(5) The API for node scripts can benefit from having a better structure as well as having additional utilities to help users navigate the requirements, and provide more guidance to users in creating new nodes. A more standard practice in the field is to create three abstract Python classes, Source, Sink, and Transform that dictate the requirements for initialisation, work_function, and on_end_of_life, and provide additional utility methods to help users connect between their code and the communication mechanism. They can be properly docstringed, along with templates. In this way, the com and worker scripts can be merged into a single unified API. A simple example that can cause confusion in the worker script is the "worker_object", which is passed into the initialise function. It is unclear what this object this variable should be, and what attributes are available without looking into the source code. As the software is also targeting those who are less experienced in programming, setting up more guidance in the API can be really helpful. In addition, the self-documentation aspect of the GUI can also benefit from a better structured API as discussed in point 2 above.

The reviewer is right that using abstract classes to expose to users the required API would be a more standard practice. The reason we did not choose to do this was to keep Heron easily accessible to entry level Python programmers who do not have familiarity yet with object oriented programming ideas. So instead of providing abstract classes we expose only the implementation of three functions which are part of the worker classes but the classes themselves are not seen by the users of the API. The point about the users’ accessibility to more information regarding a few objects used in the API (the worker object for example) has been taken on board and we have now addressed this by type hinting all these objects both in the templates and more importantly in the automatically generated code that Heron now creates when a user chooses to create a Node graphically (a feature of Heron not present in the version available in the initial submission of this manuscript).  

(6) The authors should provide more pre-defined elements. Even though the ability for users to run arbitrary code is the main feature, the initial adoption of a codebase by a community, in which many members are not so experienced with programming, is the ability for them to use off-the-shelf components as much as possible. I believe the software could benefit from a suite of commonly used Nodes.

There are currently 12 Node repositories in the Heron-repositories project on Github with more than 30 Nodes, 20 of which are general use (not implementing a specific experiment’ logic). This list will continue to grow but we fully appreciate the truth of the reviewer’s comment that adoption will depend on the existence of a large number of commonly used Nodes (for example Numpy, and OpenCV Nodes) and are working towards this goal.

(7) It is not clear to me if there is any capability or utilities for testing individual nodes without invoking a full system execution. This would be critical when designing new experiments and testing out each component.

There is no capability to run the code of an individual Node outside Heron’s GUI. A user could potentially design and test parts of the Node before they get added into a Node but we have found this to be a highly inefficient way of developing new Nodes. In our hands the best approach for Node development was to quickly generate test inputs and/or outputs using the “User Defined Function 1I 1O” Node where one can quickly write a function and make it accessible from a Node. Those test outputs can then be pushed in the Node under development or its outputs can be pushed in the test function, to allow for incremental development without having to connect it to the Nodes it would be connected in an actual pipeline. For example, one can easily create a small function that if a user presses a key will generate the same output (if run from a “User Defined Function 1I 1O” Node) as an Arduino Node reading some buttons. This output can then be passed into an experiment logic Node under development that needs to do something with this input. In this way during a Node development Heron allows the generation of simulated hardware inputs and outputs without actually running the actual hardware. We have added this way of developing Nodes also in our manuscript (Creating a new Node).

Reviewer #3 (Public Review):

Summary:

The authors present a Python tool, Heron, that provides a framework for defining and running experiments in a lab setting (e.g. in behavioural neuroscience). It consists of a graphical editor for defining the pipeline (interconnected nodes with parameters that can pass data between them), an API for defining the nodes of these pipelines, and a framework based on ZeroMQ, responsible for the overall control and data exchange between nodes. Since nodes run independently and only communicate via network messages, an experiment can make use of nodes running on several machines and in separate environments, including on different operating systems.

Strengths:

As the authors correctly identify, lab experiments often require a hodgepodge of separate hardware and software tools working together. A single, unified interface for defining these connections and running/supervising the experiment, together with flexibility in defining the individual subtasks (nodes) is therefore a very welcome approach. The GUI editor seems fairly intuitive, and Python as an accessible programming environment is a very sensible choice. By basing the communication on the widely used ZeroMQ framework, they have a solid base for the required non-trivial coordination and communication. Potential users reading the paper will have a good idea of how to use the software and whether it would be helpful for their own work. The presented experiments convincingly demonstrate the usefulness of the tool for realistic scientific applications.

Weaknesses:

(1) In my opinion, the authors somewhat oversell the reproducibility and "selfdocumentation" aspect of their solution. While it is certainly true that the graph representation gives a useful high-level overview of an experiment, it can also suffer from the same shortcomings as a "pure code" description of a model - if a user gives their nodes and parameters generic/unhelpful names, reading the graph will not help much. 

This is a problem that to our understanding no software solution can possibly address. Yet having a visual representation of how different inputs and outputs connect to each other we argue would be a substantial benefit in contrast to the case of “pure code” especially when the developer of the experiment has used badly formatted variable names.

(2) Making the link between the nodes and the actual code is also not straightforward, since the code for the nodes is spread out over several directories (or potentially even machines), and not directly accessible from within the GUI. 

This is not accurate. The obligatory code of a Node always exists within a single folder and Heron’s API makes it rather cumbersome to spread scripts relating to a Node across separate folders. The Node folder structure can potentially be copied over different machines but this is why Heron is tightly integrated with git practices (and even politely asks the user with popup windows to create git repositories of any Nodes they create whilst using Heron’s automatic Node generator system). Heron’s documentation is also very clear on the folder structure of a Node which keeps the required code always in the same place across machines and more importantly across experiments and labs. Regarding the direct accessibility of the code from the GUI, we took on board the reviewers’ comments and have taken the first step towards correcting this. Now one can attach to Heron their favourite IDE and then they can double click on any Node to open its two main scripts (com and worker) in that IDE embedded in whatever code project they choose (also set in Heron’s settings windows). On top of this, Heron now allows the addition of notes both for a Node and for all its parameters, inputs and outputs which can be viewed by hovering the mouse over them on the Nodes’ GUIs. The final step towards GUI-code integration will be to have a Heron GUI code editor but this is something that has to wait for further development from Heron’s underlying GUI library DearPyGUI.

(3) The authors state that "[Heron’s approach] confers obvious benefits to the exchange and reproducibility of experiments", but the paper does not discuss how one would actually exchange an experiment and its parameters, given that the graph (and its json representation) contains user-specific absolute filenames, machine IP addresses, etc, and the parameter values that were used are stored in general data frames, potentially separate from the results. Neither does it address how a user could keep track of which versions of files were used (including Heron itself).

Heron’s Graphs, like any experimental implementation, must contain machine specific strings. These are accessible either from Heron’s GUI when a Graph json file is opened or from the json file itself. Heron in this regard does not do anything different to any other software, other than saving the graphs into human readable json files that users can easily manipulate directly.

Heron provides a method for users to save every change of the Node parameters that might happen during an experiment so that it can be fully reproduced. The dataframes generated are done so in the folders specified by the user in each of the Nodes (and all those paths are saved in the json file of the Graph). We understand that Heron offers a certain degree of freedom to the user (Heron’s main reason to exist is exactly this versatility) to generate data files wherever they want but makes sure every file path gets recorded for subsequent reproduction. So, Heron behaves pretty much exactly like any other open source software. What we wanted to focus on as the benefits of Heron on exchange and reproducibility was the ability of experimenters to take a Graph from another lab (with its machine specific file paths and IP addresses) and by examining the graphical interface of it to be able to quickly tweak it to make it run on their own systems. That is achievable through the fact that a Heron experiment will be constructed by a small amount of Nodes (5 to 15 usually) whose file paths can be trivially changed in the GUI or directly in the json file while the LAN setup of the machines used can be easily reconstructed from the information saved in the secondary GUIs.

Where Heron needs to improve (and this is a major point in Heron’s roadmap) is the need to better integrate the different saved experiments with the git versions of Heron and the Nodes that were used for that specific save. This, we appreciate is very important for full reproducibility of the experiment and it is a feature we will soon implement. More specifically users will save together with a graph the versions of all the used repositories and during load the code base utilised will come from the recorded versions and not from the current head of the different repositories. This is a feature that we are currently working on now and as our roadmap suggests will be implemented by the release of Heron 1.0. 

(4) Another limitation that in my opinion is not sufficiently addressed is the communication between the nodes, and the effect of passing all communications via the host machine and SSH. What does this mean for the resulting throughput and latency - in particular in comparison to software such as Bonsai or Autopilot? The paper also states that "Heron is designed to have no message buffering, thus automatically dropping any messages that come into a Node’s inputs while the Node’s worker function is still running."- it seems to be up to the user to debug and handle this manually?

There are a few points raised here that require addressing. The first is Heron’s requirement to pass all communication through the main (GUI) machine. We understand (and also state in the manuscript) that this is a limitation that needs to be addressed. We plan to do this is by adding to Heron the feature of running headless (see our roadmap). This will allow us to run whole Heron pipelines in a second machine which will communicate with the main pipeline (run on the GUI machine) with special Nodes. That will allow experimenters to define whole pipelines on secondary machines where the data between their Nodes stay on the machine running the pipeline. This is an important feature for Heron and it will be one of the first features to be implemented next (after the integration of the saving system with git). 

The second point is regarding Heron’s throughput latency. In our original manuscript we did not have any description of Heron’s capabilities in this respect and both other reviewers mentioned this as a limitation. As mentioned above, we have now addressed this by adding a section to our third experimental example that fully describes how much CPU is required to run a full experimental pipeline running on two machines and utilising also non python code executables (a Unity game). This gives an overview of how heavy pipelines can run on normal computers given adequate optimisation and utilising Heron’s feature of forcing some Nodes to run their Worker processes on a specific core. At the same time, Heron’s use of 0MQ protocol makes sure there are no other delays or speed limitations to message passing. So, message passing within the same machine is just an exchange of memory pointers while messages passing between different machines face the standard speed limitations of the Local Access Network’s ethernet card speeds. 

Finally, regarding the message dropping feature of Heron, as mentioned above this is an architectural decision given the use cases of message passing we expect Heron to come in contact with. For a full explanation of the logic here please see our answer to the 3rd comment by Reviewer 2.

(5) As a final comment, I have to admit that I was a bit confused by the use of the term "Knowledge Graph" in the title and elsewhere. In my opinion, the Heron software describes "pipelines" or "data workflows", not knowledge graphs - I’d understand a knowledge graph to be about entities and their relationships. As the authors state, it is usually meant to make it possible to "test propositions against the knowledge and also create novel propositions" - how would this apply here?

We have described Heron as a Knowledge Graph instead of a pipeline, data workflow or computation graph in order to emphasise Heron’s distinct operation in contrast to what one would consider a standard pipeline and data workflow generated by other visual based software (like LabView and Bonsai). This difference exists on what a user should think of as the base element of a graph, i.e. the Node. In all other visual programming paradigms, the Node is defined as a low-level computation, usually a language keyword, language flow control or some simple function. The logic in this case is generated by composing together the visual elements (Nodes). In Heron the Node is to be thought of as a process which can be of arbitrary complexity and the logic of the graph is composed by the user both within each Node and by the way the Nodes are combined together. This is an important distinction in Heron’s basic operation logic and it is we argue the main way Heron allows flexibility in what can be achieved while retaining ease of graph composition (by users defining their own level of complexity and functionality encompassed within each Node). We have found that calling this approach a computation graph (which it is) or a pipeline or data workflow would not accentuate this difference. The term Knowledge Graph was the most appropriate as it captures the essence of variable information complexity (even in terms of length of shortest string required) defined by a Node.

Recommendations for the authors:  

Reviewer #1 (Recommendations For The Authors):

-  No buffering implies dropped messages when a node is busy. It seems like this could be very problematic for some use cases... 

This is a design principle of Heron. We have now provided a detailed explanation of the reasoning behind it in our answer to Reviewer 2 (Paragraph 3) as well as in the manuscript. 

-  How are ssh passwords stored, and is it secure in some way or just in plain text?  

For now they are plain text in an unencrypted file that is not part of the repo (if one gets Heron from the repo). Eventually, we would like to go to private/public key pairs but this is not a priority due to the local nature of Heron’s use cases (all machines in an experiment are expected to connect in a LAN).  

Minor notes / copyedits:

-  Figure 2A: right and left seem to be reversed in the caption. 

They were. This is now fixed. 

-  Figure 2B: the text says that proof of life messages are sent to each worker process but in the figure, it looks like they are published by the workers? Also true in the online documentation.  

The Figure caption was wrong. This is now fixed.

-  psutil package is not included in the requirements for GitHub

We have now included psutil in the requirements.

-  GitHub readme says Python >=3.7 but Heron will not run as written without python >= 3.9 (which is alluded to in the paper)

The new Heron updates require Python 3.11. We have now updated GitHub and the documentation to reflect this.

-  The paper mentions that the Heron editor must be run on Windows, but this is not mentioned in the Github readme.  

This was an error in the manuscript that we have now corrected.

-  It’s unclear from the readme/manual how to remove a node from the editor once it’s been added.  

We have now added an X button on each Node to complement the Del button on the keyboard (for MacOS users that do not have this button most of the times).

-  The first example experiment is called the Probabilistic Reversal Learning experiment in text, but the uncertainty experiment in the supplemental and on GitHub.  

We have now used the correct name (Probabilistic Reversal Learning) in both the supplemental material and on GitHub

-  Since Python >=3.9 is required, consider using fstrings instead of str.format for clarity in the codebase  

Thank you for the suggestion. Latest Heron development has been using f strings and we will do a refactoring in the near future.

-  Grasshopper cameras can run on linux as well through the spinnaker SDK, not just Windows.  

Fixed in the manuscript. 

-  Figure 4: Square and star indicators are unclear.

Increased the size of the indicators to make them clear.

-  End of page 9: "an of the self" presumably a typo for "off the shelf"?  

Corrected.

-  Page 10 first paragraph. "second root" should be "second route"

Corrected.

-  When running Heron, the terminal constantly spams Blowfish encryption deprecation warnings, making it difficult to see the useful messages.  

The solution to this problem is to either update paramiko or install Heron through pip. This possible issue is mentioned in the documentation.

-  Node input /output hitboxes in the GUI are pretty small. If they could be bigger it would make it easier to connect nodes reliably without mis-clicks.

We have redone the Node GUI, also increasing the size of the In/Out points.

Reviewer #2 (Recommendations For The Authors):

(1) There are quite a few typos in the manuscript, for example: "one can accessess the code", "an of the self", etc.  

Thanks for the comment. We have now screened the manuscript for possible typos.

(2) Heron’s GUI can only run on Windows! This seems to be the opposite of the key argument about the portability of the experimental setup.  

As explained in the answers to Reviewer 1, Heron can run on most machines that the underlying python libraries run, i.e. Windows and Linux (both for x86 and Arm architectures). We have tested it on Windows (10 and 11, both x64), Linux PC (Ubuntu 20.04.6, x64) and Raspberry Pi 4 (Debian GNU/Linux 12 (bookworm), aarch64). We have now revised the manuscript and the GitHub repo to reflect this.

(3) Currently, the output is displayed along the left edge of the node, but the yellow dot connector is on the right. It would make more sense to have the text displayed next to the connectors.  

We have redesigned the Node GUI and have now placed the Out connectors on the right side of the Node.

(4) The edges are often occluded by the nodes in the GUI. Sometimes it leads to some confusion, particularly when the number of nodes is large, e.g., Fig 4.

This is something that is dependent on the capabilities of the DearPyGUI module. At the moment there is no way to control the way the edges are drawn.

Reviewer #3 (Recommendations For The Authors):

A few comments on the software and the documentation itself:

- From a software engineering point of view, the implementation seems to be rather immature. While I get the general appeal of "no installation necessary", I do not think that installing dependencies by hand and cloning a GitHub repository is easier than installing a standard package.

We have now added a pip install capability which also creates a Heron command line command to start Heron with. 

-The generous use of global variables to store state (minor point, given that all nodes run in different processes), boilerplate code that each node needs to repeat, and the absence of any kind of automatic testing do not give the impression of a very mature software (case in point: I had to delete a line from editor.py to be able to start it on a non-Windows system).  

As mentioned, the use of global variables in the worker scripts is fine partly due to the multi process nature of the development and we have found it is a friendly approach to Matlab users who are just starting with Python (a serious consideration for Heron). Also, the parts of the code that would require a singleton (the Editor for example) are treated as scripts with global variables while the parts that require the construction of objects are fully embedded in classes (the Node for example). A future refactoring might make also all the parts of the code not seen by the user fully object oriented but this is a decision with pros and cons needing to be weighted first. 

Absence of testing is an important issue we recognise but Heron is a GUI app and nontrivial unit tests would require some keystroke/mouse movement emulator (like QTest of pytest-qt for QT based GUIs). This will be dealt with in the near future (using more general solutions like PyAutoGUI) but it is something that needs a serious amount of effort (quite a bit more that writing unit tests for non GUI based software) and more importantly it is nowhere as robust as standard unit tests (due to the variable nature of the GUI through development) making automatic test authoring an almost as laborious a process as the one it is supposed to automate.

-  From looking at the examples, I did not quite see why it is necessary to write the ..._com.py scripts as Python files, since they only seem to consist of boilerplate code and variable definitions. Wouldn’t it be more convenient to represent this information in configuration files (e.g. yaml or toml)?  

The com is not a configuration file, it is a script that launches the communication process of the Node. We could remove the variable definitions to a separate toml file (which then the com script would have to read). The pros and cons of such a set up should be considered in a future refactoring.

Minor comments for the paper:

-  p.7 (top left): "through its return statement" - the worker loop is an infinite loop that forwards data with a return statement?  

This is now corrected. The worker loop is an infinite loop and does not return anything but at each iteration pushes data to the Nodes output.

-  p.9 (bottom right): "of the self" → "off-the-shelf"  

Corrected.

-  p.10 (bottom left): "second root" → "second route"  

Corrected.

-  Supplementary Figure 3: Green start and square seem to be swapped (the green star on top is a camera image and the green star on the bottom is value visualization - inversely for the green square).  

The star and square have been swapped around.

-  Caption Supplementary Figure 4 (end): "rashes to receive" → "rushes to receive"  

Corrected.

The reason for the fight is obviously much bigger than just hockey.

"Shattered windows and the sounds of drums" uses visual imagery and suggests that people might have started to turn and rebel against him. "People couldn't believe what I'd become" tells me that there were people still confused and unsure about what had happened.

Reviewer #1 (Public review):

Summary:

This study investigates whether pupil dilation reflects prediction error signals during associative learning, defined formally by Kullback-Leibler (KL) divergence, an information-theoretic measure of information gain. Two independent tasks with different entropy dynamics (decreasing and increasing uncertainty) were analyzed: the cue-target 2AFC task and the letter-color 2AFC task. Results revealed that pupil responses scaled with KL divergence shortly after feedback onset, but the direction of this relationship depended on whether uncertainty (entropy) increased or decreased across trials. Furthermore, signed prediction errors (interaction between frequency and accuracy) emerged at different time windows across tasks, suggesting task-specific temporal components of model updating. Overall, the findings highlight that pupil dilation reflects information-theoretic processes in a complex, context-dependent manner.

Strengths:

This study provides a novel and convincing contribution by linking pupil dilation to information-theoretic measures, such as KL divergence, supporting Zénon's hypothesis that pupil responses reflect information gained during learning. The robust methodology, including two independent datasets with distinct entropy dynamics, enhances the reliability and generalisability of the findings. By carefully analysing early and late time windows, the authors capture the temporal dynamics of prediction error signals, offering new insights into the timing of model updates. The use of an ideal learner model to quantify prediction errors, surprise, and entropy provides a principled framework for understanding the computational processes underlying pupil responses. Furthermore, the study highlights the critical role of task context - specifically increasing versus decreasing entropy - in shaping the directionality and magnitude of these effects, revealing the adaptability of predictive processing mechanisms.

Weaknesses:

While this study offers important insights, several limitations remain. The two tasks differ significantly in design (e.g., sensory modality and learning type), complicating direct comparisons and limiting the interpretation of differences in pupil dynamics. Importantly, the apparent context-dependent reversal between pupil constriction and dilation in response to feedback raises concerns about how these opposing effects might confound the observed correlations with KL divergence. Finally, subjective factors such as participants' confidence and internal belief states were not measured, despite their potential influence on prediction errors and pupil responses.

Not true on Windows

Author response:

The following is the authors’ response to the original reviews.

Reviewer #1 (Public Review): 

Summary: 

In this work, Noorman and colleagues test the predictions of the "four-stage model" of consciousness by combining psychophysics and scalp EEG in humans. The study relies on an elegant experimental design to investigate the respective impact of attentional and perceptual blindness on visual processing. 

The study is very well summarised, the text is clear and the methods seem sound. Overall, a very solid piece of work. I haven't identified any major weaknesses. Below I raise a few questions of interpretation that may possibly be the subject of a revision of the text. 

We thank the reviewer for their positive assessment of our work and for their extremely helpful and constructive comments that helped to significantly improve the quality of our manuscript.

(1) The perceptual performance on Fig1D appears to show huge variation across participants, with some participants at chance levels and others with performance > 90% in the attentional blink and/or masked conditions. This seems to reveal that the procedure to match performance across participants was not very successful. Could this impact the results? The authors highlight the fact that they did not resort to postselection or exclusion of participants, but at the same time do not discuss this equally important point. 

Performance was indeed highly variable between observers, as is commonly found in attentional-blink (AB) and masking studies. For some observers, the AB pushes performance almost to chance level, whereas for others it has almost no effect. A similar effect can be seen in masking. We did our best to match accuracy over participants, while also matching accuracy within participants as well as possible, adjusting mask contrast manually during the experimental session. Naturally, those that are strongly affected by masking need not be the same participants as those that are strongly affected by the AB, given the fact that they rely on different mechanisms (which is also one of the main points of the manuscript). To answer the research question, what mattered most was that at the group-level, performance was well matched between the two key conditions. As all our statistical inferences, both for behavior and EEG decoding, rest on this group level. We do not think that variability at the individualsubject level detracts from this general approach.  

In the Results, we added that our goal was to match performance across participants:

“Importantly, mask contrast in the masked condition was adjusted using a staircasing procedure to match performance in the AB condition, ensuring comparable perceptual performance in the masked and the AB condition across participants (see Methods for more details).”

In the Methods, we added:

“Second, during the experimental session, after every 32 masked trials, mask contrast could be manually updated in accordance with our goal to match accuracy over participants, while also matching accuracy within participants as well as possible.”

(2) In the analysis on collinearity and illusion-specific processing, the authors conclude that the absence of a significant effect of training set demonstrates collinearity-only processing. I don't think that this conclusion is warranted: as the illusory and nonillusory share the same shape, so more elaborate object processing could also be occurring. Please discuss. 

We agree with this qualification of our interpretation, and included the reviewer’s account as an alternative explanation in the Discussion section:  

“It should be noted that not all neurophysiological evidence unequivocally links processing of collinearity and of the Kanizsa illusion to lateral and feedback processing, respectively (Angelucci et al., 2002; Bair et al., 2003; Chen et al., 2014), so that overlap in decoding the illusory and non-illusory triangle may reflect other mechanisms, for example feedback processes representing the triangular shapes as well.”

(3) Discussion, lines 426-429: It is stated that the results align with the notion that processes of perceptual segmentation and organization represent the mechanism of conscious experience. My interpretation of the results is that they show the contrary: for the same visibility level in the attentional blind or masking conditions, these processes can be implicated or not, which suggests a role during unconscious processing instead. 

We agree with the reviewer that the interpretation of this result depends on the definition of consciousness that one adheres to. If one takes report as the leading metric for consciousness (=conscious access), one can indeed conclude that perceptual segmentation/organization can also occur unconsciously. However, if the processing that results in the qualitative nature of an image (rather than whether it is reported) is taken as leading – such as the processing that results in the formation of an illusory percept – (=phenomenal) the conclusion can be quite different. This speaks to the still ongoing debate regarding the existence of phenomenal vs access consciousness, and the literature on no-report paradigms amongst others (see last paragraph of the discussion). Because the current data do not speak directly to this debate, we decided to remove  the sentence about “conscious experience”, and edited this part of the manuscript (also addressing a comment about preserved unconscious processing during masking by Reviewer 2) by limiting the interpretation of unconscious processing to those aspects that are uncontroversial:

“Such deep feedforward processing can be sufficient for unconscious high-level processing, as indicated by a rich literature demonstrating high-level (e.g., semantic) processing during masking (Kouider & Dehaene, 2007; Van den Bussche et al., 2009; van Gaal & Lamme, 2012). Thus, rather than enabling deep unconscious processing, preserved local recurrency during inattention may afford other processing advantages linked to its proposed role in perceptual integration (Lamme, 2020), such as integration of stimulus elements over space or time.”

(4) The two paradigms developed here could be used jointly to highlight nonidiosyncratic NCCs, i.e. EEG markers of visibility or confidence that generalise regardless of the method used. Have the authors attempted to train the classifier on one method and apply it to another (e.g. AB to masking and vice versa)? What perceptual level is assumed to transfer? 

To avoid issues with post-hoc selection of (visible vs. invisible) trials (discussed in the Introduction), we did not divide our trials into conscious and unconscious trials, and thus did not attempt to reveal NCCs, or NCCs generalizing across the two paradigms. Note also that this approach alone would not resolve the debate regarding the ‘true’ NCC as it hinges on the operational definition of consciousness one adheres to; also see our response to the previous point the reviewer raised. Our main analysis revealed that the illusory triangle could be decoded with above-chance accuracy during both masking and the AB over extended periods of time with similar topographies (Fig. 2B), so that significant cross-decoding would be expected over roughly the same extended period of time (except for the heightened 200-250 ms peak). However, as our focus was on differences between the two manipulations and because we did not use post-hoc sorting of trials, we did not add these analyses.

(5) How can the results be integrated with the attentional literature showing that attentional filters can be applied early in the processing hierarchy? 

Compared to certain manipulations of spatial attention, the AB phenomenon is generally considered to represent an instance of  “late” attentional filtering. In the Discussion section we included a paragraph on classic load theory, where early and late filtering depend on perceptual and attentional load. Just preceding this paragraph, we added this:  

“Clearly, these findings do not imply that unconscious high-level (e.g., semantic) processing can only occur during inattention, nor do they necessarily generalize to other forms of inattention. Indeed, while the AB represents a prime example of late attentional filtering, other ways of inducing inattention or distraction (e.g., by manipulating spatial attention) may filter information earlier in the processing hierarchy (e.g., Luck & Hillyard, 1994 vs. Vogel et al., 1998).”

Reviewer #2 (Public Review): 

Summary: 

This is a very elegant and important EEG study that unifies within a single set of behaviorally equated experimental conditions conscious access (and therefore also conscious access failures) during visual masking and attentional blink (AB) paradigms in humans. By a systematic and clever use of multivariate pattern classifiers across conditions, they could dissect, confirm, and extend a key distinction (initially framed within the GNWT framework) between 'subliminal' and 'pre-conscious' unconscious levels of processing. In particular, the authors could provide strong evidence to distinguish here within the same paradigm these two levels of unconscious processing that precede conscious access : (i) an early (< 80ms) bottom-up and local (in brain) stage of perceptual processing ('local contrast processing') that was preserved in both unconscious conditions, (ii) a later stage and more integrated processing (200-250ms) that was impaired by masking but preserved during AB. On the basis of preexisting studies and theoretical arguments, they suggest that this later stage could correspond to lateral and local recurrent feedback processes. Then, the late conscious access stage appeared as a P3b-like event. 

Strengths: 

The methodology and analyses are strong and valid. This work adds an important piece in the current scientific debate about levels of unconscious processing and specificities of conscious access in relation to feed-forward, lateral, and late brain-scale top-down recurrent processing. 

Weaknesses: 

- The authors could improve clarity of the rich set of decoding analyses across conditions. 

- They could also enrich their Introduction and Discussion sections by taking into account the importance of conscious influences on some unconscious cognitive processes (revision of traditional concept of 'automaticity'), that may introduce some complexity in Results interpretation 

- They should discuss the rich literature reporting high-level unconscious processing in masking paradigms (culminating in semantic processing of digits, words or even small group of words, and pictures) in the light of their proposal (deeper unconscious processing during AB than during masking). 

We thank the reviewer for their positive assessment of our study and for their insightful comments and helpful suggestions that helped to significantly strengthen our paper. We provide a more detailed point-by-point response in the “recommendations for the authors” section below. In brief, we followed the reviewer’s suggestions and revised the Results/Discussion to include references to influences on unconscious processes and expanded our discussion of unconscious effects during masking vs. AB.  

Reviewer #3 (Public Review): 

Summary: 

This work aims to investigate how perceptual and attentional processes affect conscious access in humans. By using multivariate decoding analysis of electroencephalography (EEG) data, the authors explored the neural temporal dynamics of visual processing across different levels of complexity (local contrast, collinearity, and illusory perception). This is achieved by comparing the decidability of an illusory percept in matched conditions of perceptual (i.e., degrading the strength of sensory input using visual masking) and attentional impairment (i.e., impairing topdown attention using attentional blink, AB). The decoding results reveal three distinct temporal responses associated with the three levels of visual processing. Interestingly, the early stage of local contrast processing remains unaffected by both masking and AB. However, the later stage of collinearity and illusory percept processing are impaired by the perceptual manipulation but remain unaffected by the attentional manipulation. These findings contribute to the understanding of the unique neural dynamics of perceptual and attentional functions and how they interact with the different stages of conscious access. 

Strengths: 

The study investigates perceptual and attentional impairments across multiple levels of visual processing in a single experiment. Local contrast, collinearity, and illusory perception were manipulated using different configurations of the same visual stimuli. This clever design allows for the investigation of different levels of visual processing under similar low-level conditions. 

Moreover, behavioural performance was matched between perceptual and attentional manipulations. One of the main problems when comparing perceptual and attentional manipulations on conscious access is that they tend to impact performance at different levels, with perceptual manipulations like masking producing larger effects. The study utilizes a staircasing procedure to find the optimal contrast of the mask stimuli to produce a performance impairment to the illusory perception comparable to the attentional condition, both in terms of perceptual performance (i.e., indicating whether the target contained the Kanizsa illusion) and metacognition (i.e., confidence in the response). 

The results show a clear dissociation between the three levels of visual processing in terms of temporal dynamics. Local contrast was represented at an early stage (~80 ms), while collinearity and illusory perception were associated with later stages (~200-250 ms). Furthermore, the results provide clear evidence in support of a dissociation between the effects of perceptual and attentional processes on conscious access: while the former affected both neuronal correlates of collinearity and illusory perception, the latter did not have any effect on the processing of the more complex visual features involved in the illusion perception. 

Weaknesses: 

The design of the study and the results presented are very similar to those in Fahrenfort et al. (2017), reducing its novelty. Similar to the current study, Fahrenfort et al. (2017) tested the idea that if both masking and AB impact perceptual integration, they should affect the neural markers of perceptual integration in a similar way. They found that behavioural performance (hit/false alarm rate) was affected by both masking and AB, even though only the latter was significant in the unmasked condition. An early classification peak was instead only affected by masking. However, a late classification peak showed a pattern similar to the behavioural results, with classification affected by both masking and AB. 

The interpretation of the results mainly centres on the theoretical framework of the recurrent processing theory of consciousness (Lamme, 2020), which lead to the assumption that local contrast, collinearity, and the illusory perception reflect feedforward, local recurrent, and global recurrent connections, respectively. It should be mentioned, however, that this theoretical prediction is not directly tested in the study. Moreover, the evidence for the dissociation between illusion and collinearity in terms of lateral and feedback connections seems at least limited. For instance, Kok et al. (2016) found that, whereas bottom-up stimulation activated all cortical layers, feedback activity induced by illusory figures led to a selective activation of the deep layers. Lee & Nguyen (2001), instead, found that V1 neurons respond to illusory contours of the Kanizsa figures, particularly in the superficial layers. They all mention feedback connections, but none seem to point to lateral connections. 

Moreover, the evidence in favour of primarily lateral connections driving collinearity seems mixed as well. On one hand, Liang et al. (2017) showed that feedback and lateral connections closely interact to mediate image grouping and segmentation. On the other hand, Stettler et al. (2002) showed that, whereas the intrinsic connections link similarly oriented domains in V1, V2 to V1 feedback displays no such specificity. Furthermore, the other studies mentioned in the manuscript did not investigate feedback connections but only lateral ones, making it difficult to draw any clear conclusions. 

We thank the reviewer for their careful review and positive assessment of our study, as well as for their constructive criticism and helpful suggestions. We provide a more detailed point-by-point response in the “recommendations for the authors” section below. In brief, we addressed the reviewer’s comments and suggestions by better relating our study to Fahrenfort et al.’s (2017) paper and by highlighting the limitations inherent in linking our findings to distinct neural mechanisms (in particular, to lateral vs. feedback connections).

Recommendations for the authors:  

Reviewer #1 (Recommendations For The Authors): 

-  Methods: it states that "The distance between the three Pac-Man stimuli as well as between the three aligned two-legged white circles was 2.8 degrees of visual angle". It is unclear what this distance refers to. Is it the shortest distance between the edges of the objects? 

It is indeed the shortest distance between the edges of the objects. This is now included in the Methods.

-  Methods: It's unclear to me if the mask updating procedure during the experimental session was based on detection rate or on the perceptual performance index reported on Fig1D. Please clarify. 

It was based on accuracy calculated over 32 trials. We have included this information in the Methods.

-  Methods and Results: I did not understand why the described procedure used to ensure that confidence ratings are not contaminated by differences in perceptual performance was necessary. To me, it just seems to make the "no manipulations" and "both manipulations" less comparable to the other 2 conditions. 

To calculate accurate estimates of metacognitive sensitivity for the two matched conditions, we wanted participants to make use of the full confidence scale (asking them to distribute their responses evenly over all ratings within a block). By mixing all conditions in the same block, we would have run the risk of participants anchoring their confidence ratings to the unmatched very easy and very difficult conditions (no and both manipulations condition). We made this point explicit in the Results section and in the Methods section:

“To ensure that the distribution of confidence ratings in the performancematched masked and AB condition was not influenced by participants anchoring their confidence ratings to the unmatched very easy and very difficult conditions (no and both manipulations condition, respectively), the masked and AB condition were presented in the same experimental block, while the other block type included the no and both manipulations condition.”

“To ensure that confidence ratings for these matched conditions (masked, long lag and unmasked, short lag) were not influenced by participants anchoring their confidence ratings to the very easy and very difficult unmatched conditions (no and both manipulations, respectively), one type of block only contained the matched conditions, while the other block type contained the two remaining, unmatched conditions (masked, short lag and unmasked, long lag).”

- Methods: what priors were used for Bayesian analyses? 

Bayesian statistics were calculated in JASP (JASP Team, 2024) with default prior scales (Cauchy distribution, scale 0.707). This is now added to the Methods.

- Results, line 162: It states that classifiers were applied on "raw EEG activity" but the Methods specify preprocessing steps. "Preprocessed EEG activity" seems more appropriate. 

We changed the term to “preprocessed EEG activity” in the Methods and to “(minimally) preprocessed EEG activity (see Methods)” in the  Results, respectively.

- Results, line 173: The effect of masking on local contrast decoding is reported as "marginal". If the alpha is set at 0.05, it seems that this effect is significant and should not be reported as marginal. 

We changed the wording from “marginal” to “small but significant.”  

- Fig1: The fixation cross is not displayed. 

Because adding the fixation cross would have made the figure of the trial design look crowded and less clear, we decided to exclude it from this schematic trial representation. We are now stating this also in the legend of figure 1.  

- Fig 3A: In the upper left panel, isn't there a missing significant effect of the "local contrast training and testing" condition in the first window? If not, this condition seems oddly underpowered compared to the other two conditions. 

Thanks for the catch! The highlighting in bold and the significance bar were indeed lacking for this condition in the upper left panel (blue line). We corrected the figure in our revision.

- Supplementary text and Fig S6: It is unclear to me why the two control analyses (the black lines vs. the green and purple lines) are pooled together in the same figure. They seem to test for different, non-comparable contrasts (they share neither training nor testing sets), and I find it confusing to find them on the same figure. 

We agree that this may be confusing, and deleted the results from one control analysis from the figure (black line, i.e., training on contrast, testing on illusion), as the reviewer correctly pointed out that it displayed a non-comparable analysis. Given that this control analysis did not reveal any significant decoding, we now report its results only in the Supplementary text.  

- Fig S6: I think the title of the legend should say testing on the non-illusory triangle instead of testing on the illusory triangle to match the supplementary text. 

This was a typo – thank you! Corrected.  

Reviewer #2 (Recommendations For The Authors): 

Issue #1: One key asymmetry between the three levels of T2 attributes (i.e.: local contrast; non-illusory triangle; illusory Kanisza triangle) is related to the top-down conscious posture driven by the task that was exclusively focusing on the last attribute (illusory Kanisza triangle). Therefore, any difference in EEG decoding performance across these three levels could also depend to this asymmetry. For instance, if participants were engaged to report local contrast or non-illusory triangle, one could wonder if decoding performance could differ from the one used here. This potential confound was addressed by the authors by using decoders trained in different datasets in which the main task was to report one the two other attributes. They could then test how classifiers trained on the task-related attribute behave on the main dataset. However, this part of the study is crucial but not 100% clear, and the links with the results of these control experiments are not fully explicit. Could the author better clarity this important point (see also Issue #1 and #3). 

The reviewer raises an important point, alluding to potential differences between decoded features regarding task relevance. There are two separate sets of analyses where task relevance may have been a factor, our main analyses comparing illusion to contrast decoding, and our comparison of collinearity vs. illusion-specific processing.  

In our main analysis, we are indeed reporting decoding of a task-relevant feature (illusion) and of a task-irrelevant feature (local contrast, i.e., rotation of the Pac-Man inducers). Note, however, that the Pac-Man inducers were always task-relevant, as they needed to be processed to perceive illusory triangles, so that local contrast decoding was based on task-relevant stimulus elements, even though participants did not respond to local contrast differences in the main experiment. However, we also ran control analyses testing the effect of task-relevance on local contrast decoding in our independent training data set and in another (independent) study, where local contrast was, in separate experimental blocks, task-relevant or task-irrelevant. The results are reported in the Supplementary Text and in Figure S5. In brief, task-relevance did not improve early (70–95 ms) decoding of local contrast. We are thus confident that the comparison of local contrast to illusion decoding in our main analysis was not substantially affected by differences in task relevance. In our previous manuscript version, we referred to these control analyses only in the collinearity-vs-illusion section of the Results. In our revision, we added the following in the Results section comparing illusion to contrast decoding:

“In the light of evidence showing that unconscious processing is susceptible to conscious top-down influences (Kentridge et al., 2004; Kiefer & Brendel, 2006; Naccache et al., 2002), we ran control analyses showing that early local contrast decoding was not improved by rendering contrast task-relevant (see Supplementary Information and Fig. S5), indicating that these differences between illusion and contrast decoding did not reflect differences in task-relevance.”

In addition to our main analysis, there is the concern that our comparison of collinearity vs. illusion-specific processing may have been affected by differences in task-relevance between the stimuli inducing the non-illusory triangle (the “two-legged white circles”, collinearity-only) and the stimuli inducing the Kanizsa illusion (the PacMan inducers, collinearity-plus-illusion). We would like to emphasize that in our main analysis classifiers were always used to decode T2 illusion presence vs. absence (collinearity-plus-illusion), and never to decode T2 collinearity-only. To distinguish collinearity-only from collinearity-plus-illusion processing, we only varied the training data (training classifiers on collinearity-only or collinearity-plus-illusion), using the independent training data set, where collinearity-only and collinearity-plus-illusion (and rotation) were task-relevant (in separate blocks). As discussed in the Supplementary Information, for this analysis approach to be valid, collinearity-only processing should be similar for the illusory and the non-illusory triangle, and this is what control analyses demonstrated (Fig. S7). In any case, general task-relevance was equated for the collinearity-only and the collinearity-plus-illusion classifiers.  

Finally, in supplementary Figure 6 we also show that our main results reported in Figure 2 (discussed at the top of this response) were very similar when the classifiers were trained on the independent localizer dataset in which each stimulus feature could be task-relevant.  

Together, for the reasons described above, we believe that differences in EEG decoding performance across these three stimulus levels did  are unlikely to depend also depend on a “task-relevance” asymmetry.

Issue #2: Following on my previous point the authors should better mention the concept of conscious influences on unconscious processing that led to a full revision of the notion of automaticity in cognitive science [1 , 2 , 3 , 4]. For instance, the discovery that conscious endogenous temporal and spatial attention modulate unconscious subliminal processing paved the way to this revision. This concept raises the importance of Issue#1: equating performance on the main task across AB and masking is not enough to guarantee that differences of neural processing of the unattended attributes of T2 (i.e.: task-unrelated attributes) are not, in part, due to this asymmetry rather than to a systematic difference of unconscious processing strengtsh [5 , 6-8]. Obviously, the reported differences for real-triangle decoding between AB and masking cannot be totally explained by such a factor (because this is a task-unrelated attribute for both AB and masking conditions), but still this issue should be better introduced, addressed, clarified (Issue #1 and #3) and discussed. 

We would like to refer to our response to the previous point: Control analyses for local contrast decoding showed that task relevance had no influence on our marker for feedforward processing. Most importantly, as outlined above, we did not perform real-triangle decoding – all our decoding analyses focused on comparing collinearity-only vs. collinearity-plus-illusion were run on the task-relevant T2 illusion (decoding its presence vs. absence). The key difference was solely the training set, where the collinearity-only classifier was trained on the (task-relevant) real triangle and the collinearity-plus-illusion classifier was trained on the (task-relevant) Kanizsa triangle. Thus, overall task relevance was controlled in these analyses.  

In our revision, we are now also citing the studies proposed by the reviewer, when discussing the control analyses testing for an effect of task-relevance on local contrast decoding:

“In the light of evidence showing that unconscious processing is susceptible to conscious top-down influences (Kentridge et al., 2004; Kiefer & Brendel, 2006; Naccache et al., 2002), we ran control analyses showing that early local contrast decoding was not improved by rendering contrast task-relevant (see Supplementary Information and Fig. S5), indicating that these differences between illusion and contrast decoding did not reflect differences in task-relevance.”

Issue #3: In terms of clarity, I would suggest the authors to add a synthetic figure providing an overall view of all pairs of intra and cross-conditions decoding analyses and mentioning main task for training and testing sets for each analysis (see my previous and related points). Indeed, at one point, the reader can get lost and this would not only strengthen accessibility to the detailed picture of results, but also pinpoint the limits of the work (see previous point). 

We understand the point the reviewer is raising and acknowledge that some of our analyses, in particular those using different training and testing sets, may be difficult to grasp. But given the variety of different analyses using different training and testing sets, different temporal windows, as well as different stimulus features, it was not possible to design an intuitive synthetic figure summarizing the key results. We hope that the added text in the Results and Discussion section will be sufficient to guide the reader through our set of analyses.  

In our revision, we are now more clearly highlighting that, in addition to presenting the key results in our main text that were based on training classifiers on the T1 data, “we replicated all key findings when training the classifiers on an independent training set where individual stimuli were presented in isolation (Fig. 3A, results in the Supplementary Information and Fig. S6).” For this, we added a schematic showing the procedure of the independent training set to Figure 3, more clearly pointing the reader to the use of a separate training data set.  

Issue #4: In the light of these findings the authors should discuss more thoroughly the question of unconscious high-level representations in masking versus AB: in particular, a longstanding issue relates to unconscious semantic processing of words, numbers or pictures. According to their findings, they tend to suggest that semantic processing should be more enabled in AB than in masking. However, a rich literature provided a substantial number of results (including results from the last authors Simon Van Gaal) that tend to support the notion of unconscious semantic processing in subliminal processing (see in particular: [9 , 10 , 11 , 12 , 13]). So, and as mentioned by the authors, while there is evidence for semantic processing during AB they should better discuss how they would explain unconscious semantic subliminal processing. While a possibility could be to question the unconscious attribute of several subliminal results, the same argument also holds for AB studies. Another possible track of discussion would be to differentiate AB and subliminal perception in terms of strength and durability of the corresponding unconscious representations, but not necessarily in terms of cognitive richness. Indeed, one may discuss that semantic processing of stimuli that do not need complex spatial integration (e.g.: words or digits as compared to illusory Kanisza tested here) can still be observed under subliminal conditions. 

We thank the reviewer for pointing us to this shortcoming of our previous Discussion. Note that our data does not directly speak to the question of high-level unconscious representations in masking vs AB, because such conclusions would hinge on the operational definition of consciousness one adheres to (also see response to Reviewer 1). Nevertheless, we do follow the reviewer’s suggestions and added the following in the Discussion (also addressing a point about other forms of attention raised by Reviewer 1):

“Clearly, these findings do not imply that unconscious high-level (e.g., semantic) processing can only occur during inattention, nor do they necessarily generalize to other forms of inattention. Indeed, while the AB represents a prime example of late attentional filtering, other ways of inducing inattention or distraction (e.g., by manipulating spatial attention) may filter information earlier in the processing hierarchy (e.g., Luck & Hillyard, 1994 vs. Vogel et al., 1998).”

And, in a following paragraph in the Discussion:

“Such deep feedforward processing can be sufficient for unconscious high-level processing, as indicated by a rich literature demonstrating high-level (e.g., semantic) processing during masking (Kouider & Dehaene, 2007; Van den Bussche et al., 2009; van Gaal & Lamme, 2012). Thus, rather than enabling high-level unconscious processing, preserved local recurrency during inattention may afford other processing advantages linked to its proposed role in perceptual integration (Lamme, 2020), such as integration of stimulus elements over space or time.  

Reviewer #3 (Recommendations For The Authors): 

(1) The objective of Fahrenfort et al., 2017 seems very similar to that of the current study. What are the main differences between the two studies? Moreover, Fahrenfort et al., 2017 conducted similar decoding analyses to those performed in the current study.

Which results were replicated in the current study, and which ones are novel? Highlighting these differences in the manuscript would be beneficial. 

We now provide a more comprehensive coverage of the study by Fahrenfort et al., 2017. In the Introduction, we added a brief summary of the key findings, highlighting that this study’s findings could have reflected differences in task performance rather than differences between masking and AB:

“For example, Fahrenfort and colleagues (2017) found that illusory surfaces could be decoded from electroencephalogram (EEG) data during the AB but not during masking. This was taken as evidence that local recurrent interactions, supporting perceptual integration, were preserved during inattention but fully abolished by masking. However, masking had a much stronger behavioral effect than the AB, effectively reducing task performance to chance level. Indeed, a control experiment using weaker masking, which resulted in behavioral performance well above chance similar to the main experiment’s AB condition, revealed some evidence for preserved local recurrent interactions also during masking. However, these conditions were tested in separate experiments with small samples, precluding a direct comparison of perceptual vs. attentional blindness at matched levels of behavioral performance. To test …”

In the Results , we are now also highlighting this key advancement by directly referencing the previous study:

“Thus, whereas in previous studies task performance was considerably higher during the AB than during masking (e.g., Fahrenfort et al., 2017), in the present study the masked and the AB condition were matched in both measures of conscious access.” When reporting the EEG decoding results in the Results section, we continuously cite the Fahrenfort et al. (2017) study to highlight similarities in the study’s findings. We also added a few sentences explicitly relating the key findings of the two studies:

“This suggests that the AB allowed for greater local recurrent processing than masking, replicating the key finding by Fahrenfort and colleagues (2017). Importantly, the present result demonstrates that this effect reflects the difference between the perceptual vs. attentional manipulation rather than differences in behavior, as the masked and the AB condition were matched for perceptual performance and metacognition.”

“This similarity between behavior and EEG decoding replicates the findings of Fahrenfort and colleagues  (2017) who also found a striking similarity between late Kanizsa decoding (at 406 ms) and behavioral Kanizsa detection. These results indicate that global recurrent processing at these later points in time reflected conscious access to the Kanizsa illusion.”  

We also more clearly highlighted where our study goes beyond Fahrenfort et al.’s (2017), e.g., in the Results:

“The addition of this element of collinearity to our stimuli was a key difference to the study by Fahrenfort and colleagues (2017), allowing us to compare non-illusory triangle decoding to illusory triangle decoding in order to distinguish between collinearity and illusion-specific processing.”

And in the Discussion:

“Furthermore, the addition of line segments forming a non-illusory triangle to the stimulus employed in the present study allowed us to distinguish between collinearity and illusion-specific processing.”

Also, in the Discussion, we added a paragraph “summarizing which results were replicated in the current study, and which ones are novel”, as suggested by the reviewer:

“This pattern of results is consistent with a previous study that used EEG to decode Kanizsa-like illusory surfaces during masking and the AB (Fahrenfort et al., 2017). However, the present study also revealed some effects where Fahrenfort and colleagues (2017) failed to obtain statistical significance, likely reflecting the present study’s considerably larger sample size and greater statistical power. For example, in the present study the marker for feedforward processing was weakly but significantly impaired by masking, and the marker for local recurrency was significantly impaired not only by masking but also by the AB, although to a lesser extent. Most importantly, however, we replicated the key findings that local recurrent processing was more strongly impaired by masking than by the AB, and that global recurrent processing was similarly impaired by masking and the AB and closely linked to task performance, reflecting conscious access. Crucially, having matched the key conditions behaviorally, the present finding of greater local recurrency during the AB can now unequivocally be attributed to the attentional vs. perceptual manipulation of consciousness.”

Finally, we changed the title to “Distinct neural mechanisms underlying perceptual and attentional impairments of conscious access despite equal task performance” to highlight one of the crucial differences between the Fahrenfort et al., study and this study, namely the fact that we equalized task performance between the two critical conditions (AB and masking).

(2) It is not clear from the text the link between the current study and the literature on the role of lateral and feedback connections in consciousness (Lamme, 2020). A better explanation is needed. 

To our knowledge, consciousness theories such as recurrent processing theory by Lamme make currently no distinction between the role of lateral and feedback connections for consciousness. The principled distinction lies between unconscious feedforward processing and phenomenally conscious or “preconscious” local recurrent processing, where local recurrency refers to both lateral (or horizontal) and feedback connections. We added a sentence in the Discussion:

“As current theories do not distinguish between the roles of lateral vs. feedback connections for consciousness, the present findings may enrich empirical and theoretical work on perceptual vs. attentional mechanisms of consciousness …”

(3) When training on T1 and testing on T2, EEG data showed an early peak in local contrast classification at 75-95 ms over posterior electrodes. The authors stated that this modulation was only marginally affected by masking (and not at all by AB); however, the main effect of masking is significant. Why was this effect interpreted as nonrelevant? 

Following this and Reviewer 1’s comment, we changed the wording from “marginal” to “weak but significant.” We considered this effect “weak” and of lesser relevance, because its Bayes factor indicated that the alternative hypothesis was only 1.31 times more likely than the null hypothesis of no effect, representing only “anecdotal” evidence, which is in sharp contrast to the robust effects of the consciousness manipulations on illusion decoding reported later. Furthermore, later ANOVAs comparing the effect of masking on contrast vs. illusion decoding revealed much stronger effects on illusion decoding than on contrast decoding (BFs>3.59×10⁴).

(4) The decoding analysis on the illusory percept yielded two separate peaks of decoding, one from 200 to 250 ms and another from 275 to 475 ms. The early component was localized occipitally and interpreted as local sensory processing, while the late peak was described as a marker for global recurrent processing. This latter peak was localized in the parietal cortex and associated with the P300. Can the authors show the topography of the P300 evoked response obtained from the current study as a comparison? Moreover, source reconstruction analysis would probably provide a better understanding of the cortical localization of the two peaks. 

Figure S4 now shows the P300 from electrode Pz, demonstrating a stronger positivity between 375 and 475 ms when the illusory triangle was present than when it was absent. We did not run a source reconstruction analysis.  

(5) The authors mention that the behavioural results closely resembled the pattern of the second decoding peak results. However, they did not show any evidence for this relationship. For instance, is there a correlation between the two measures across or within participants? Does this relationship differ between the illusion report and the confidence rating? 

This relationship became evident from simply eyeballing the results figures: Both in behavior and EEG decoding performance dropped from the both-manipulations condition to the AB and masked conditions, while these conditions did not differ significantly. Following a similar observation of a close similarity between behavior and the second/late illusion decoding peak in the study by Fahrenfort et al. (2017), we adopted their analysis approach and ran two additional ANOVAs, adding “measure” (behavior vs. EEG) as a factor. For this analysis, we dropped the both-manipulations condition due to scale restrictions (as noted in footnote 1: “We excluded the bothmanipulations condition from this analysis due to scale restrictions: in this condition, EEG decoding at the second peak was at chance, while behavioral performance was above chance, leaving more room for behavior to drop from the masked and AB condition.”). The analysis revealed that there were no interactions with condition:

“The pattern of behavioral results, both for perceptual performance and metacognitive sensitivity, closely resembled the second decoding peak: sensitivity in all three metrics dropped from the no-manipulations condition to the masked and AB conditions, while sensitivity did not differ significantly between these performancematched conditions (Fig. 2C). Two additional rm ANOVAs with the factors measure (behavior, second EEG decoding peak) and condition (no-manipulations, masked, AB)¹ for perceptual performance and metacognitive sensitivity revealed no significant interaction (performance: F</iv>_2,58=0.27, P\=0.762, BF₀₁=8.47; metacognition: F</iv><sub>2,58</sub=0.54, P\=0.586, BF_2,58=6.04). This similarity between behavior and EEG decoding replicates the findings of Fahrenfort and colleagues  (2017) who also found a striking similarity between late Kanizsa decoding (at 406 ms) and behavioral Kanizsa detection. These results indicate that global recurrent processing at these later points in time reflected conscious access to the Kanizsa illusion.”

Reviewer #2 (Public review):

Summary:

This work by den Bakker and Kloosterman contributes to the vast body of research exploring the dynamics governing the communication between the hippocampus (HPC) and the medial prefrontal cortex (mPFC) during spatial learning and navigation. Previous research showed that population activity of mPFC neurons is replayed during HPC sharp-wave ripple events (SWRs), which may therefore correspond to privileged windows for the transfer of learned navigation information from the HPC, where initial learning occurs, to the mPFC, which is thought to store this information long term. Indeed, it was also previously shown that the activity of mPFC neurons contains task-related information that can inform about the location of an animal in a maze, which can predict the animals' navigational choices. Here, the authors aim to show that the mPFC neurons that are modulated by HPC activity (SWRs and theta rhythms) are distinct from those "encoding" spatial information. This result could suggest that the integration of spatial information originating from the HPC within the mPFC may require the cooperation of separate sets of neurons.

This observation may be useful to further extend our understanding of the dynamics regulating the exchange of information between the HPC and mPFC during learning. However, my understanding is that this finding is mainly based upon a negative result, which cannot be statistically proven by the failure to reject the null hypothesis. Moreover, in my reading, the rest of the paper mainly replicates phenomena that have already been described, with the original reports not correctly cited. My opinion is that the novel elements should be precisely identified and discussed, while the current phrasing in the manuscript, in most cases, leads readers to think that these results are new. Detailed comments are provided below.

Major concerns:

(1) The main claim of the manuscript is that the neurons involved in predicting upcoming choices are not the neurons modulated by the HPC. This is based upon the evidence provided in Figure 5, which is a negative result that the authors employ to claim that predictive non-local representations in the mPFC are not linked to hippocampal SWRs and theta phase. However, it is important to remember that in a statistical test, the failure to reject the null hypothesis does not prove that the null hypothesis is true. Since this claim is so central in this work, the authors should use appropriate statistics to demonstrate that the null hypothesis is true. This can be accomplished by showing that there is no effect above some size that is so small that it would make the effect meaningless (see https://doi.org/10.1177/070674370304801108).

(2) The main claim of the work is also based on Figure 3, where the authors show that SWRs-unmodulated mPFC neurons have higher spatial tuning, and higher directional selectivity scores, and a higher percentage of these neurons show theta skipping. This is used to support the claim that SWRs-unmodulated cells encode spatial information. However, it must be noted that in this kind of task, it is not possible to disentangle space and specific task variables involving separate cognitive processes from processing spatial information such as decision-making, attention, motor control, etc., which always happen at specific locations of the maze. Therefore, the results shown in Figure 3 may relate to other specific processes rather than encoding of space and it cannot be unequivocally claimed that mPFC neurons "encode spatial information". This limitation is presented by Mashoori et al (2018), an article that appears to be a major inspiration for this work. Can the authors provide a control analysis/experiment that supports their claim? Otherwise, this claim should be tempered. Also, the authors say that Jadhav et al. (2016) showed that mPFC neurons unmodulated by SWRs are less tuned to space. How do they reconcile it with their results?

(3) My reading is that the rest of the paper mainly consists of replications or incremental observations of already known phenomena with some not necessarily surprising new observations:<br /> a) Figure 2 shows that a subset of mPFC neurons is modulated by HPC SWRs and theta (already known), that vmPFC neurons are more strongly modulated by SWRs (not surprising given anatomy), and that theta phase preference is different between vmPFC and dmPFC (not surprising given the fact that theta is a travelling wave).<br /> b) Figure 4 shows that non-local representations in mPFC are predictive of the animal's choice. This is mostly an increment to the work of Mashoori et al (2018). My understanding is that in addition to what had already been shown by Mashoori et al here it is shown how the upcoming choice can be predicted. The author may want to emphasize this novel aspect.<br /> c) Figure 6 shows that prospective activity in the HPC is linked to SWRs and theta oscillations. This has been described in various forms since at least the works of Johnson and Redish in 2007, Pastalkova et al 2008, and Dragoi and Tonegawa (2011 and 2013), as well as in earlier literature on splitter cells. These foundational papers on this topic are not even cited in the current manuscript.<br /> Although some previous work is cited, the current narrative of the results section may lead the reader to think that these results are new, which I think is unfair. Previous evidence of the same phenomena should be cited all along the results and what is new and/or different from previous results should be clearly stated and discussed. Pure replications of previous works may actually just be supplementary figures. It is not fair that the titles of paragraphs and main figures correspond to notions that are well established in the literature (e.g., Figure 2, 2nd paragraph of results, etc.).<br /> d) My opinion is that, overall, the paper gives the impression of being somewhat rushed and lacking attention to detail. Many figure panels are difficult to understand due to incomplete legends and visualizations with tiny, indistinguishable details. Moreover, some previous works are not correctly cited. I tried to make a list of everything I spotted below.

Reviewer #2 (Public review):

Summary:

Dr. Adam Kim and collaborators study the changes in chromatin structure in monocytes obtained from alcohol-associated hepatitis (AH) when compared to healthy controls (HC). Through the usage of high throughput chromatin conformation capture technology (Hi-C), they collected data on contact frequencies between both contiguous and distal DNA windows (100 kB each); mainly within the same chromosome. From the analyses of those data in the two cohorts under analysis, authors describe frequent pairs of regions subject to significant changes in contact frequency across cohorts. Their accumulation onto specific regions of the genome -referred to as hotspots- motivated authors to narrow down their analyses to these disease-associated regions, in many of which, authors claim, a number of key innate immune genes can be found. Ultimately, the authors try to draw a link between the changes observed in chromatin architecture in some of these hotspots and the differential co-expression of the genes lying within those regions, as ascertained in previous single-cell transcriptomic analyses.

Strengths:

The main strength of this paper lies in the generation of Hi-C data from patients, a valuable asset that, as the authors emphasize, offers critical insights into the role of chromatin architecture dysregulation in the pathogenesis of alcohol-associated hepatitis (AH). If confirmed, the reported findings have the potential to highlight an important, yet overlooked, aspect of cellular dysregulation-chromatin conformation changes - not only in AH but potentially in other immune-related conditions with a component of pathological inflammation.

Weaknesses:

In what I regard as the two most important weaknesses of the work, I feel that they are more methodological than conceptual. The first of these issues concerns the perhaps insufficient level of description provided on the definition of some key types of genomic regions, such as topologically associated domains, DNA hotspots, or even DNA loci showing significant changes in contact frequency between AH and HC. In spite of the importance of these concepts in the paper, no operational, explicit description of how are they defined, from a statistical point of view, is provided in the current version of the manuscript.

Without these definitions, some of the claims that authors make in their work become hard to sustain. Some examples are the claim that randomizing samples does not lead to significant differences between cohorts; the claim that most of the changes in contact frequency happen locally; or the claim that most changes do not alter the structure of TADs, but appear either within, or between TADs. In my viewpoint, specific descriptions and implementation of proper tests to check these hypotheses and back up the mentioned specific claims, along with the inclusion of explicit results on these matters, would contribute very significantly to strengthening the overall message of the paper.

The second notable weakness of the study pertains to the characterization of the changes observed around immune genes in relation to genome-wide expectations. Although the authors suggest that certain hotspots contain a high number of immune-related genes, no enrichment analysis is provided to verify whether these regions indeed harbor a higher concentration of such genes compared to other genomic areas. It would be important for readers to be promptly informed if no such enrichment is observed, for in that case, the presence of some immune genes within these hotspots would carry more limited implications.

Additionally, the criteria used to define a hotspot are not clearly outlined, making it difficult to assess whether the changes in contact frequencies around the immune genes highlighted in figures 5-8 are truly more pronounced than what would be expected genome-wide.

This is a weird way to put it, since the very first note given in this section is that all fields unless otherwise noted are unsigned...

Author response:

The following is the authors’ response to the original reviews.

eLife Assessment

This valuable study combined whole-head magnetoencephalography (MEG) and subthalamic (STN) local field potential (LFP) recordings in patients with Parkinson's disease undergoing deep brain stimulation surgery. The paper provides solid evidence that cortical and STN beta oscillations are sensitive to movement context and may play a role in the coordination of movement redirection.

We are grateful for the expert assessment by the editor and the reviewers. Below we provide pointby-point replies to both public and private reviews. We have tried to keep the answers in the public section short and concise, not citing the changed passages unless the point does not re-appear in the recommendations. There, we did include all of the changes to the manuscript, such that the reviewers need not go back and forth between replies and manuscript.

The reviewer comments have not only led to numerous improvements of the text, but also to new analyses, such as Granger causality analysis, and to methodological improvements e.g. including numerous covariates in the statistical analyses. We believe that the article improved substantially through the feedback, and we thank the reviewers and the editor for their effort.

Public Reviews

Reviewer #1 (Public review):

Summary:

Winkler et al. present brain activity patterns related to complex motor behaviour by combining wholehead magnetoencephalography (MEG) with subthalamic local field potential (LFP) recordings from people with Parkinson's disease. The motor task involved repetitive circular movements with stops or reversals associated with either predictable or unpredictable cues. Beta and gamma frequency oscillations are described, and the authors found complex interactions between recording sites and task conditions. For example, they observed stronger modulation of connectivity in unpredictable conditions. Moreover, STN power varied across patients during reversals, which differed from stopping movements. The authors conclude that cortex-STN beta modulation is sensitive to movement context, with potential relevance for movement redirection.

Strengths:

This study employs a unique methodology, leveraging the rare opportunity to simultaneously record both invasive and non-invasive brain activity to explore oscillatory networks.

Weaknesses:

It is difficult to interpret the role of the STN in the context of reversals because no consistent activity pattern emerged.

We thank the reviewer for the valuable feedback to our study. We agree that the interpretation of the role of the STN during reversals is rather difficult, because reversal-related STN activity was highly variable across patients. Although there seem to be consistent patterns in sub-groups of the current cohort, with some patients showing event-related increases (Fig. 3b) and others showing decreases, the current dataset is not large enough to substantiate or even explain the existence of such clusters. Thus, we limit ourselves to acknowledging this limitation and discussing potential reasons for the high variability, namely variability in electrode placement and insufficient spatial resolution for the separation of specialized cell ensembles within the STN (see Discussion, section Limitations and future directions).

Reviewer #2 (Public review):

Summary:

This study examines the role of beta oscillations in motor control, particularly during rapid changes in movement direction among patients with Parkinson's disease. The researchers utilized magnetoencephalography (MEG) and local field potential (LFP) recordings from the subthalamic nucleus to investigate variations in beta band activity within the cortex and STN during the initiation, cessation, and reversal of movements, as well as the impact of external cue predictability on these dynamics. The primary finding indicates that beta oscillations more effectively signify the start and end of motor sequences than transitions within those sequences. The article is well-written, clear, and concise.

Strengths:

The use of a continuous motion paradigm with rapid reversals extends the understanding of beta oscillations in motor control beyond simple tasks. It offers a comprehensive perspective on subthalamocortical interactions by combining MEG and LFP.

Weaknesses:

(1) The small and clinically diverse sample size may limit the robustness and generalizability of the findings. Additionally, the limited exploration of causal mechanisms reduces the depth of its conclusions and focusing solely on Parkinson's disease patients might restrict the applicability of the results to broader populations.

We thank the reviewer for the insightful feedback. We address these issues one by one in our responses to points 2, 4 and 6, respectively.

(2) The small sample size and variability in clinical characteristics among patients may limit the robustness of the study's conclusions. It would be beneficial for the authors to acknowledge this limitation and propose strategies for addressing it in future research. Additionally, incorporating patient-specific factors as covariates in the ANOVA could help mitigate the confounding effects of heterogeneity.

Thank you for this comment. The challenges associated with recording brain activity peri-operatively can be a limiting factor when it comes to sample size and cohort stratification. We now acknowledge this in the revised discussion (section Limitations and future directions). Furthermore, we suggest using sensing-capable devices in the future as a measure to increase sample sizes (Discussion, section Limitations and future directions). Lastly, we appreciate the idea of adding patient-specific factors as covariates to the ANOVAs and have thus included age, disease duration and pre-surgical UPDRS score into our models. This did not lead to any qualitative changes of statistical effects.

(3) The author may consider using standardized statistics, such as effect size, that would provide a clearer picture of the observed effect magnitude and improve comparability.

Thanks for the suggestion. As measures of effect size, we have added partial eta squared (η_p<sup2</sup>) to the results of all ANOVAs and Cohen’s d to all follow-up t-tests.

(4) Although the study identifies relevance between beta activity and motor events, it lacks causal analysis and discussion of potential causal mechanisms. Given the valuable datasets collected, exploring or discussing causal mechanisms would enhance the depth of the study.

We appreciate this idea and have conducted Granger causality analyses in response to this comment. This new analysis reveals that there is a strong cortical drive to the STN for all movements of interest and predictability conditions in the beta band. The detailed results can be viewed on p. 16 in the section on Granger causality. For statistical testing, we conducted an rmANCOVA, similar to those for power and coherence (see p. 46-48 and 54-56 for the corresponding tables), as well as t-tests assessing directionality (Figure 6-figure supplement 2 on p. 35). In the discussion section, we connect these results with prior findings suggesting that the frontal cortex drives the STN in the beta band, likely through hyperdirect pathway fibers (p. 17).

(5) The study cohort focused on senior adults, who may exhibit age-related cortical responses during movement planning in neural mechanisms. These aspects were not discussed in the study.

We appreciate the comment and agree that age may have impacted neural oscillatory activity of patients in the present study. We now acknowledge this in the limitations section, and point out that our approach to handling these effects was including age as a covariate in the statistical analyses.

(6) Including a control group of patients with other movement disorders who also undergo DBS surgery would be beneficial. Because we cannot exclude the possibility that the observed findings are specific to PD or can be generalized. Additionally, the current title and the article, which are oriented toward understanding human motor control, may not be appropriate.

We thank the reviewer for this comment and fully agree that it cannot be ruled out that the present findings are, in part, specific to PD. We acknowledge this limitation in the Limitations and future directions section (p. 20-21). Indeed, including a control group of patients with other disorders would be ideal, but the scarcity of patients with diseases other than PD who receive STN DBS in our centre makes this an unfeasible option in practical terms. We do suggest that future research may address this issue by extending our approach to different disorders or healthy participants on the cortical level (p. 21). Lastly, we appreciate the idea to adjust the title of the present article. The adjusted title is: “Context-Dependent Modulations of Subthalamo-Cortical Synchronization during Rapid Reversals of Movement Direction in Parkinson’s Disease”.

That being said, we do believe that our findings at least approximate healthy functioning and are not solely related to PD. For one, patients were on their usual dopaminergic medication and dopamine has been found to normalize pathological alterations of beta activity. Further, the general pattern of movement-related beta and gamma oscillations reported here has been observed in numerous diseases and brain structures, including cortical beta oscillations measured non-invasively in healthy participants.

Reviewer #3 (Public review):

Summary:

The study highlights how the initiation, reversal, and cessation of movements are linked to changes in beta synchronization within the basal ganglia-cortex loops. It was observed that different movement phases, such as starting, stopping briefly, and stopping completely, affect beta oscillations in the motor system.

It was found that unpredictable cues lead to stronger changes in STN-cortex beta coherence. Additionally, specific patterns of beta and gamma oscillations related to different movement actions and contexts were observed. Stopping movements was associated with a lack of the expected beta rebound during brief pauses within a movement sequence.

Overall, the results underline the complex and context-dependent nature of motor-control and emphasize the role of beta oscillations in managing movement according to changing external cues.

Strengths:

The paper is very well written, clear, and appears methodologically sound.

Although the use of continuous movement (turning) with reversals is more naturalistic than many previous button push paradigms.

Weaknesses:

The generalizability of the findings is somewhat curtailed by the fact that this was performed perioperatively during the period of the microlesion effect. Given the availability of sensing-enabled DBS devices now and HD-EEG, does MEG offer a significant enough gain in spatial localizability to offset the fact that it has to be done shortly postoperatively with externalized leads, with an attendant stun effect? Specifically, for paradigms that are not asking very spatially localized questions as a primary hypothesis?

We appreciate the reviewer’s feedback and acknowledge the valid point raised on the timing of our measurements. Indeed, sensing-enabled devices offer a valid alternative to peri-operative recordings, circumventing the stun effect. We acknowledge this in the revised discussion, section Limitations and future directions (p. 23): “Additionally, future research could capitalize on sensingcapable devices to circumvent the necessity to record brain activity peri-operatively, facilitating larger sample sizes and circumventing the stun effect, an immediate improvement in motor symptoms arising as a consequence of electrode implantation (Mann et al., 2009).” This alternative strategy, however, was not an option here because we did not have a sufficient number of patients implanted with sensing-enabled devices at the time when the data collection was initialized.

That being said, we would like to highlight that in the present study, our goal was not to study pathology related to Parkinson’s disease. Rather, we aimed to learn about motor control in general. The stun effect may have facilitated motor performance in our patients, which is actually beneficial to the research goals at hand.

Further investigation of the gamma signal seems warranted, even though it has a slightly lower proportional change in amplitude in beta. Given that the changes in gamma here are relatively wide band, this could represent a marker of neural firing that could be interestingly contrasted against the rhythm account presented.

We appreciate the reviewer’s interest and we have extended the investigation of gamma oscillations. We now provide statistics regarding the influence of predictability on gamma power and gamma coherence (no significant effects) and explore Granger causality in the gamma (and beta) band (see comment 4 of reviewer 2). Unfortunately, we cannot measure spiking via the DBS electrode, and therefore we cannot investigate correlations between gamma oscillatory activity and action potentials. We do agree with the reviewer, however, that action potentials rather than oscillations form the basis of motor control in the brain. This view of ours is now reflected in the revised discussion, section Limitations and future directions (p. 21): “Lastly, given the present study’s focus on understanding movement-related rhythms, particularly in the beta range, future research could further explore the role of gamma oscillations in continuous movement and their relation to action potentials in motor areas (Fischer et al., 2020; Igarashi, Isomura, Arai, Harukuni, & Fukai, 2013), which form the basis of movement encoding in the brain.”

Recommendations for the authors:

Reviewer #1 (Recommendations for the authors):

This is a well-conducted study and overall the results are clear. I only have one minor suggestion for improvement of the manuscript. I found the order of appearance of the results somewhat confusing, switching from predictability-related behavioral effects to primarily stopping and reversal-related neurophysiological effects, back to predictability but starting with coherence. I would suggest that the authors try to follow a systematic order focused on the questions at hand. E.g. perhaps readability could be improved if the results section is split into reversal vs. stopping related effects, reporting behavior, power, and coherence in this order, followed by a predictability section, again reporting behavior, power, and coherence. Obviously, this is an optional suggestion. Apart from that, I just missed a more direct message related to the absence of statistical significance related to STN power changes during reversal. I think this could be made more clear in the text.

We thank the reviewer for the feedback to our study. In order to ease reading, we modified the order and further added additional sub-titles to the results section. We start with Behavior (p. 4) and then move on to Power (general movement effects on power – movement effects on STN power – movement effects on cortical power – predictability effects on power). Next, we move on to Connectivity (movement effects on connectivity – predictability effects on connectivity – Granger causality). We hope that these adaptations will help guide the reader.

Additionally, we thank the reviewer for noting that we did not explicitly mention the lack of statistical significance of reversal-related beta power modulations in the STN. We have adapted the section on modulation of STN beta power associated with reversals (p. 8) to: “In the STN, reversals were associated with a brief modulation of beta power, which was weak in the group-average spectrum and did not reach significance (Fig. 3a).”

Reviewer #2 (Recommendations for the authors):

(1) The small sample size and variability in clinical characteristics among patients may limit the robustness of the study's conclusions. It would be beneficial for the authors to acknowledge this limitation and propose strategies for addressing it in future research. Additionally, incorporating patient-specific factors as covariates in the ANOVA could help mitigate the confounding effects of heterogeneity.

Thank you for this comment. The challenges associated with recording brain activity peri-operatively can be a limiting factor when it comes to sample size. We now acknowledge this in the revised discussion, section Limitations and future directions (p. 20):

“Invasive measurements of STN activity are only possible in patients who are undergoing or have undergone brain surgery. Studies drawing from this limited pool of candidate participants are typically limited in terms of sample size and cohort stratification, particularly when carried out in a peri-operative setting. Here, we had a sample size of 20, which is rather high for a peri-operative study, but still low in terms of absolute numbers.”

Furthermore, we suggest using sensing-capable devices in the future as a measure to increase sample sizes (p. 21):

“Additionally, future research could capitalize on sensing-capable devices to circumvent the necessity to record brain activity peri-operatively, facilitating larger sample sizes and circumventing the stun effect, an immediate improvement in motor symptoms arising as a consequence of electrode implantation (Mann et al., 2009).”

Lastly, we appreciate the idea of adding patient-specific factors as covariates to the ANOVAs and have thus included age, disease duration and pre-surgical UPDRS score into our models. This did not lead to any qualitative changes of statistical effects.

Revised article

Methods, Statistical analysis:

“To account for their potential influence on brain activity, we added age, pre-operative UPDRS score, and disease duration as covariates to all ANOVAs. Covariates were standardized by means of zscoring.”

(2) The author may consider using standardized statistics, such as effect size, that would provide a clearer picture of the observed effect magnitude and improve comparability.

Thanks for this useful suggestion. As measures of effect size, we have added partial eta squared (η_p<sup2</sup>) to the results of all ANOVAs and Cohen’s d to all follow-up _t-_tests.

(3) Although the study identifies relevance between beta activity and motor events, it lacks causal analysis and discussion of potential causal mechanisms. Given the valuable datasets collected, exploring or discussing causal mechanisms would enhance the depth of the study.

We appreciate this idea and have conducted Granger causality analyses in response to this comment. This new analysis reveals that there is a strong cortical drive to the STN for all movements of interest and predictability conditions in the beta band, but no directed interactions in the gamma band. For statistical testing, we conducted an rmANCOVA, similar to the analysis of power and coherence (see p. 46-48 and 54-56 for the corresponding tables), as well as t-tests assessing directionality (Figure 6 figure supplement 2 on p. 35). In the discussion section, we connect these results with prior findings suggesting that the frontal cortex drives the STN in the beta band, likely through hyperdirect pathway fibers (p. 17).

Revised article

Methods Section, Granger Causality Analysis

“We computed beta and gamma band non-parametric Granger causality (Dhamala, Rangarajan, & Ding, 2008) between cortical ROIs and the STN in the hemisphere contralateral to movement for the post-event time windows (0 – 2 s with respect to start, reversal, and stop). Because estimates of Granger causality are often biased, we compared the original data to time-reversed data to suppress non-causal interactions. True directional influence is reflected by a higher causality measure in the original data than in its time-reversed version, resulting in a positive difference between the two, the opposite being the case for a signal that is “Granger-caused” by the other. Directionality is thus reflected by the sign of the estimate (Haufe, Nikulin, Müller, & Nolte, 2013). Because rmANCOVA results indicated no significant effects for predictability and movement type, and post-hoc tests did not detect significant differences between hemispheres, we averaged Granger causality estimates over movement types, hemispheres and predictability conditions in Figure 6-figure supplement 2.”

Results, Granger causality

“In general, cortex appeared to drive the STN in the beta band, regardless of the movement type and predictability condition. This was reflected in a main effect of ROI on Granger causality estimates (F_ROI(7,9) = 3.443, p_ROI = 0.044, η_p<sup2</sup> = 0.728; refer to Supplementary File 4 for the full results of the ANOVA). In the hemisphere contralateral to movement, follow-up t-tests revealed significantly higher Granger causality estimates from M1 to the STN (t = 3.609, one-sided p < 0.001, d = 0.807) and from MSMC to the STN (t = 2.051, one-sided p < 0.027, d = 0.459) than the other way around. The same picture emerged in the hemisphere ipsilateral to movement (M1 to STN: t = 3.082, one-sided p = 0.003, d = 0.689; MSMC to STN: t \= 1.833, one-sided p < 0.041, d = 0.410). In the gamma band, we did not detect a significant drive from one area to the other (F_ROI(7,9) = 0.338, p_ROI = 0.917, η_p<sup2</sup> = 0.208, Supplementary File 6). Figure 6-figure supplement 2 demonstrates the differences in Granger causality between original and time-reversed data for the beta and gamma band.”

Discussion, The dynamics of STN-cortex coherence

“Considering the timing of the increase observed here, the STN’s role in movement inhibition (Benis et al., 2014; Ray et al., 2012) and the fact that frontal and prefrontal cortical areas are believed to drive subthalamic beta activity via the hyperdirect pathway (Chen et al., 2020; Oswal et al., 2021) it seems plausible that the increase of beta coherence reflects feedback of sensorimotor cortex to the STN in the course of post-movement processing. In line with this idea, we observed a cortical drive of subthalamic activity in the beta band.”

(4) The study cohort focused on senior adults, who may exhibit age-related cortical responses during movement planning in neural mechanisms. These aspects were not discussed in the study.

We appreciate the comment and agree that age may have impacted neural oscillatory activity of patients in the present study. We now acknowledge this in the limitations section, and point out that our approach to handling these effects was including age as a covariate in the statistical analyses.

Revised article

Discussion, Limitations and Future Directions

“Further, most of our participants were older than 60 years. To diminish any confounding effects of age on movement-related modulations of neural oscillations, such as beta suppression and rebound (Bardouille & Bailey, 2019; Espenhahn et al., 2019), we included age as a covariate in the statistical analyses.”

(5) Including a control group of patients with other movement disorders who also undergo DBS surgery would be beneficial. Because we cannot exclude the possibility that the observed findings are specific to PD or can be generalized. Additionally, the current title and the article, which are oriented toward understanding human motor control, may not be appropriate.

We thank the reviewer for this comment and fully agree that it cannot be ruled out that the present findings are, in part, specific to PD. We acknowledge this limitation in the Limitations and future directions section (p. 20-21). Indeed, including a control group of patients with other disorders would be ideal, but the scarcity of patients with diseases other than PD who receive STN DBS makes this an unfeasible option. We do suggest that future research may address this issue by extending our approach to different disorders or healthy participants on the cortical level (p. 21). Lastly, we appreciate the idea to adjust the title of the present article. The adjusted title is: “Context-Dependent Modulations of Subthalamo-Cortical Synchronization during Rapid Reversals of Movement Direction in Parkinson’s Disease”.

That being said, we do believe that our findings at least approximate healthy functioning and are not solely related to PD. For one, patients were on their usual dopaminergic medication for the study and dopamine has been found to normalize pathological alterations of beta activity. More importantly, the general pattern of movement-related beta and gamma oscillations has been observed in numerous diseases and brain structures, including cortical beta oscillations measured non-invasively in healthy participants. Thus, it is not unlikely that the new aspects discovered here are also general features of motor processing.

Revised article

Discussion, Limitations and future directions

“Furthermore, we cannot be sure to what extent the present study’s findings relate to PD pathology rather than general motor processing. We suggest that our approach at least approximates healthy brain functioning as patients were on their usual dopaminergic medication. Dopaminergic medication has been demonstrated to normalize power within the STN and globus pallidus internus, as well as STN-globus pallidus internus and STN-cortex coherence (Brown et al., 2001; Hirschmann et al., 2013). Additionally, several of our findings match observations made in other patient populations and healthy participants, who exhibit the same beta power dynamics at movement start and stop (Alegre et al., 2004) that we observed here. Notably, our finding of enhanced cortical involvement in face of uncertainty aligns well with established theories of cognitive processing, given the cortex' prominent role in managing higher cognitive functions (Altamura et al., 2010). Yet, transferring our approach and task to patients with different disorders, e.g. obsessive compulsive disorder, or examining young and healthy participants solely at the cortical level, could contribute to elucidating whether the synchronization dynamics reported here are indeed independent of PD and age.”

Reviewer #3 (Recommendations for the authors):

Despite the strengths of the "rhythm" account of cognitive processes, the paper could possibly be improved by making it less skewed to rhythms explaining all of the movement encoding.

Thank you for this comment - the point is well taken. There is a large body of literature relating neural oscillations to spiking in larger neural populations, which itself is likely the most relevant signal with respect to motor control. In our eyes, it is this link that justifies the rhythm account, i.e. we agree with the reviewer that action potentials are the basis of movement encoding in the brain, not oscillations. Unfortunately, we cannot measure spiking with the method at hand.

To better integrate this view into the current manuscript, we make the following suggestion for future research in the Limitations and future directions section (p. 21): “Lastly, given the present study’s focus on understanding movement-related rhythms, particularly in the beta range, future research could further explore the role of gamma oscillations in continuous movement and their relation to action potentials in motor areas (Fischer et al., 2020; Igarashi, Isomura, Arai, Harukuni, & Fukai, 2013), which form the basis of movement encoding in the brain.”

In Figure 5 - is the legend correct? Is it really just a 0.2% change in power only? That would be a very surprisingly small effect size.

We thank the reviewer for noting this. Indeed, the numbers on the scale quantify relative change (post - pre)/pre and should be multiplied by 100 to obtain %-change. We have adjusted the color bars accordingly.

The dissociation between the effects of unpredictable cues in coherence versus raw power is interesting and could potentially be directly contrasted further in the discussion (here they are presented separately with separate discussions, but this seems like a pretty important and novel finding as beta coherence and power usually go in the same direction).

We appreciate the reviewer’s interest in our findings on the predictability of movement instructions. In case of coherence, the difference between pre- and post-event was generally more positive in the unpredictable condition, meaning that suppressions (negative pre-post difference) were diminished whereas increases (positive pre-post difference) were enhanced. With respect to power, we also observed less suppression in the unpredictable condition at movement start. Therefore, the direction of change is in fact the same. We made this clearer in the revised version by adapting the corresponding sections of the abstract, results and discussion (see below).

The only instance of coherence and power diverging (on a qualitative level) was observed during reversals: here, we noted post-event increases in coherence and post-event decreases in M1 power in the group-average spectra. However, when comparing the pre- and post-event epochs statistically by means of permutation testing, the coherence increase did not reach significance. Hence, we did not highlight this aspect.

Revised version

Abstract

“… Event-related increases of STN-cortex beta coherence were generally stronger in the unpredictable than in the predictable condition. … “

Results, Effects of predictability on beta power  

“With respect to the effect of predictability of movement instructions on beta power dynamics (research aim 2), we observed an interaction between movement type and condition (F_cond*mov (2,14) = 4.206, p_cond*mov = 0.037, η_p<sup2</sup> = 0.375), such that the beta power suppression at movement start was generally stronger in the predictable (M = -0.170, SD = 0.065) than in the unpredictable (M \= -0.154, SD = 0.070) condition across ROIs (t = -1.888, one-sided p \= 0.037, d = -0.422). We did not observe any modulation of gamma power by the predictability of movement instructions (F_cond (1,15) = 0.792, p_cond = 0.388, η_p<sup2</sup> = 0.050, Supplementary File 5).”

Effects of predictability on STN-cortex coherence

“With respect to the effect of predictability of movement instructions on beta coherence (research aim 2), we found that the pre-post event differences were generally more positive in the unpredictable condition (main effect of predictability condition; F_cond(1,15) = 8.684, p_cond = 0.010, η_p<sup2</sup> = 0.367; Supplementary File 3), meaning that the suppression following movement start was diminished and the increases following stop and reversal were enhanced in the unpredictable condition (Fig. 6a). This effect was most pronounced in the MSMC (Fig. 6b). When comparing regionaverage TFRs between the unpredictable and the predictable condition, we observed a significant difference only for stopping (t_clustersum = 142.8, p = 0.023), suggesting that the predictability effect was mostly carried by increased beta coherence following stops. When repeating the rmANCOVA for preevent coherence, we did not observe an effect of predictability (F_cond(1,15) = 0.163, p_cond = 0.692, η_p<sup2</sup> = 0.011), i.e. the effect was most likely not due to a shift of baseline levels. The increased tendency for upward modulations and decreased tendency for downward modulations rather suggests that the inability to predict the next cue prompted intensified event-related interaction between STN and cortex. STN-cortex gamma coherence was not modulated by predictability (F_cond(1,15) = 0.005, p_cond = 0.944, η_p<sup2</sup> = 0.000, Supplementary File 5).”

Discussion, Beta coherence and beta power are modulated by predictability

“In the present paradigm, patients were presented with cues that were either temporally predictable or unpredictable. We found that unpredictable movement prompts were associated with stronger upward modulations and weaker downward modulations of STN-cortex beta coherence, likely reflecting the patients adopting a more cautious approach, paying greater attention to instructive cues. Enhanced STN-cortex interactions might thus indicate the recruitment of additional neural resources, which might have allowed patients to maintain the same movement speed in both conditions. […]”

With respect to power, we observed reduced beta suppression in the unpredictable condition at movement start, consistent with the effect on coherence, likely demonstrating a lower level of motor preparation.

Given that you have a nice continuous data task here - the turning of the wheel, it might be interesting to cross-correlate the circular position (and separately - velocity) of the turning with the envelope of the beta signal. This would be a nice finding if you could also show that the beta is modulated continuously by the continuous movements. In the natural world, we rarely do a continuous movement with a sudden reversal, or stop, most of the time we are in continuous movement. Look at this might also be a strength of your dataset.

We could not agree more. In fact, having a continuous behavioral output was a major motivation for choosing this particular task. We are very interested in state space models such as preferential subspace identification (Sani et al., 2021), for example. These models relate continuous brain signals to continuous behavioral target variables and should be of great help for questions such as: do oscillations relate to moment-by-moment adaptations of continuous movement? Which frequency bands and brain areas are important? Is angular position encoded by different brain areas/frequency bands than angular speed? These analyses are in fact ongoing. This project, however, is too large to fit into the current article.

Makes me think of the videos I saw of Philadelphia after the Eagles just won the Superbowl. I saw broken windows and trashed streets.

Emily Dickinson brings out emotion by challenging our views on death. Most would believe death to be some kind of dramatic transformation into whatever comes after, but Dickinson argues how mundane the process of death might be. A process so ordinary that it could be easily interrupted by the buzz of a fly.

Emily Dickinson uses symbols to show death as confusing and unsettling. The "blue – uncertain – stumbling Buzz" of the fly represents the body's breakdown, interrupting the expected peaceful journey to the afterlife. Light may stand for life or understanding, but the fly gets in the way, focusing on the physical side of death instead of a spiritual one. The failing windows show the loss of awareness, and "I could not see to see" emphasizes the finality and mystery of death.

she is confused about death and the sound of the buzzing while her eyes are closing

In these lines, Dickinson uses metaphors to express emotional confusion and disorientation. The "Blue - uncertain - stumbling Buzz" conveys a sense of disarray and unease, while the "Windows failed" suggests a loss of clarity or vision. The phrase "I could not see to see" highlights a moment of emotional blindness, unable to comprehend or process the situation. The imagery portrays the overwhelming and isolating nature of the speaker's emotions.

Author response:

The following is the authors’ response to the original reviews.

Public Reviews:

Reviewer #1 (Public Review):

Thank you for your constructive feedback and recognition of our work. We followed your suggestion and improved the accuracy of the language used to interpret some of our findings. 

Summary:

The present study by Mikati et al demonstrates an improved method for in-vivo detection of enkephalin release and studies the impact of stress on the activation of enkephalin neurons and enkephalin release in the nucleus accumbens (NAc). The authors refine their pipeline to measure met and leu enkephalin using liquid chromatography and mass spectrometry. The authors subsequently measured met and leu enkephalin in the NAc during stress induced by handling, and fox urine, in addition to calcium activity of enkephalinergic cells using fiber photometry. The authors conclude that this improved tool for measuring enkephalin reveals experimenter handling stress-induced enkephalin release in the NAc that habituates and is dissociable from the calcium activity of these cells, whose activity doesn't habituate. The authors subsequently show that NAc enkephalin neuron calcium activity does habituate to fox urine exposure, is activated by a novel weigh boat, and that fox urine acutely causes increases in met-enk levels, in some animals, as assessed by microdialysis.

Strengths:

A new approach to monitoring two distinct enkephalins and a more robust analytical approach for more sensitive detection of neuropeptides. A pipeline that potentially could help for the detection of other neuropeptides.

Weaknesses:

Some of the interpretations are not fully supported by the existing data or would require further testing to draw those conclusions. This can be addressed by appropriately tampering down interpretations and acknowledging other limitations the authors did not cover brought by procedural differences between experiments.

We have taken time to go through the manuscript ensuring we are more detailed and precise with our interpretations as well as appropriately acknowledging limitations. 

Reviewer #2 (Public Review):

Thank you for your constructive and thorough assessment of our work. In our revised manuscript, we adjusted the text to reflect the references you mentioned regarding the methionine oxidation procedure. Additionally, we expanded the methods section to include the key details of the statistical tests and procedures that you outlined. 

Summary:

The authors aimed to improve the detection of enkephalins, opioid peptides involved in pain modulation, reward, and stress. They used optogenetics, microdialysis, and mass spectrometry to measure enkephalin release during acute stress in freely moving rodents. Their study provided better detection of enkephalins due to the implementation of previously reported derivatization reaction combined with improved sample collection and offered insights into the dynamics and relationship between Met- and Leu-Enkephalin in the Nucleus Accumbens shell during stress.

Strengths:

A strength of this work is the enhanced opioid peptide detection resulting from an improved microdialysis technique coupled with an established derivatization approach and sensitive and quantitative nLC-MS measurements. These improvements allowed basal and stimulated peptide release with higher temporal resolution, lower detection thresholds, and native-state endogenous peptide measurement.

Weaknesses:

The draft incorrectly credits itself for the development of an oxidation method for the stabilization of Met- and Leu-Enk peptides. The use of hydrogen peroxide reaction for the oxidation of Met-Enk in various biological samples, including brain regions, has been reported previously, although the protocols may slightly vary. Specifically, the manuscript writes about "a critical discovery in the stabilization of enkephalin detection" and that they have "developed a method of methionine stabilization." Those statements are incorrect and the preceding papers that relied on hydrogen peroxide reaction for oxidation of Met-Enk and HPLC for quantification of oxidized Enk forms should be cited. One suggested example is Finn A, Agren G, Bjellerup P, Vedin I, Lundeberg T. Production and characterization of antibodies for the specific determination of the opioid peptide Met5-Enkephalin-Arg6-Phe7. Scand J Clin Lab Invest. 2004;64(1):49-56. doi: 10.1080/00365510410004119. PMID: 15025428.

Thank you for highlighting this. It was not our intention to imply that we developed the oxidation method, rather that we were able improve the detection of metenkephalin by oxidation of the methionine without compromising the detection resolution of leu-enkephalin, enabling the simultaneous detection of both peptides. We have addressed this is the manuscript and included the suggested citation. 

Another suggestion for this draft is to make the method section more comprehensive by adding information on specific tools and parameters used for statistical analysis:

(1) Need to define "proteomics data" and explain whether calculations were performed on EIC for each m/z corresponding to specific peptides or as a batch processing for all detected peptides, from which only select findings are reported here. What type of data normalization was used, and other relevant details of data handling? Explain how Met- and Leu-Enk were identified from DIA data, and what tools were used.

Thank you for pointing out this source of confusion. We believe it is because we use a different DIA method than is typically used in other literature. Briefly, we use a DIA method with the targeted inclusion list to ensure MS2 triggering as opposed to using large isolation widths to capture all precursors for fragmentation, as is typically done with MS1 features. For our method, MS2 is triggered based on the 4 selected m/z values (heavy and light versions of Leu and Met-Enkephalin peptides) at specific retention time windows with isolation width of 2 Da; regardless of the intensity of MS1 of the peptides. 

(2) Simple Linear Regression Analysis: The text mentions that simple linear regression analysis was performed on forward and reverse curves, and line equations were reported, but it lacks details such as the specific variables being regressed (although figures have labels) and any associated statistical parameters (e.g., R-squared values). 

Additional detail about the linear regression process was added to the methods section, please see lines 614-618. The R squared values are also now shown on the figure. 

‘For the forward curves, the regression was applied to the measured concentration of the light standard as the theoretical concentration was increased. For plotting purposes, we show the measured peak area ratios for the light standards in the forward curves. For the reverse curves, the regression was applied to the measured concentration of the heavy standard, as the theoretical concentration was varied.’

(3) Violin Plots: The proteomics data is represented as violin plots with quartiles and median lines. This visual representation is mentioned, but there is no detail regarding the software/tools used for creating these plots.

We used Graphpad Prism to create these plots. This detail has been added to the statistical analysis section. See line 630.

(4) Log Transformation: The text states that the data was log-transformed to reduce skewness, which is a common data preprocessing step. However, it does not specify the base of the logarithm used or any information about the distribution before and after transformation.

We have added the requested details about the log transformation, and how the data looked before and after, into the statistical analysis section. We followed convention that the use of log is generally base 10 unless otherwise specified as natural log (base 2) or a different base. See lines 622-625

‘The data was log10 transformed to reduce the skewness of the dataset caused by the variable range of concentrations measured across experiments/animals. Prior to log transformation, the measurements failed normality testing for a Gaussian distribution. After the log transformation, the data passed normality testing, which provided the rationale for the use of statistical analyses that assume normality.’

(5) Two-Way ANOVA: Two-way ANOVA was conducted with peptide and treatment as independent variables. This analysis is described, but there is no information regarding the software or statistical tests used, p-values, post-hoc tests, or any results of this analysis.

Information about the two-way ANOVA analysis has been added to the statistical analysis section. Additionally, more detailed information has been added to the figure legends about the statistical results. Please see lines 625-628.

‘Two-way ANOVA testing with peptide (Met-Enk or Leu-Enk) and treatment (buffer or stress for example) as the two independent variables. Post-hoc testing was done using Šídák's multiple comparisons test and the p values for each of these analyses are shown in the figures (Figs. 1F, 2A).’ 

(6) Paired T-Test: A paired t-test was performed on predator odor proteomic data before and after treatment. This step is mentioned, but specific details like sample sizes, and the hypothesis being tested are not provided.

The sample size is included in the figure legend to which we have included a reference. We have also included the following text to highlight the purpose of this test. See lines 628-630

A paired t-test was performed on the predator odor proteomic data before and after odor exposure to test that hypothesis that Met-Enk increases following exposure to predator odor  (Fig. 3F). These analyses were conducted using Graphpad Prism.

(7) Correlation Analysis: The text mentions a simple linear regression analysis to correlate the levels of Met-Enk and Leu-Enk and reports the slopes. However, details such as correlation coefficients, and p-values are missing.

We apologize for the use of the word correlation as we think it may have caused some confusion and have adjusted the language accordingly. Since this was a linear regression analysis, there is no correlation coefficient. The slope of the fitted line is reported on the figures to show the fitted values of Met-Enk to Leu-Enk. 

(8) Fiber Photometry Data: Z-scores were calculated for fiber photometry data, and a reference to a cited source is provided. This section lacks details about the calculation of zscores, and their use in the analysis. 

These details have been added to the statistical analysis section. See lines 634-637

‘For the fiber photometry data, the z-scores were calculated as described in using GuPPy which is an open-source python toolbox for fiber photometry analysis. The z-score equation used in GuPPy is z=(DF/F-(mean of DF/F)/standard deviation of DF/F) where F refers to fluorescence of the GCaMP6s signal.’

(9) Averaged Plots: Z-scores from individual animals were averaged and represented with SEM. It is briefly described, but more details about the number of animals, the purpose of averaging, and the significance of SEM are needed.

We have added additional information about the averaging process in the statistical analysis section. See lines 639-643.

‘The purpose of the averaged traces is to show the extent of concordance of the response to experimenter handling and predator odor stress among animals with the SEM demonstrating that variability. The heatmaps depict the individual responses of each animal. The heatmaps were plotted using Seaborn in Python and mean traces were plotted using Matplotlib in Python.’

A more comprehensive and objective interpretation of results could enhance the overall quality of the paper.

We have taken this opportunity to improve our manuscript following comments from all the reviewers that we hope has resulted in a manuscript with a more objective interpretation of results. 

Reviewer #3 (Public Review):

Thank you for your thoughtful review of our work. To clarify some of the points you raised, we revised the manuscript to include more detail on how we distinguish between the oxidized endogenous and standard signal, as well as refine the language concerning the spatial resolution. We also edited the manuscript regarding the concentration measurements. We conducted technical replicates, so we appreciate you raising this point and clarify that in the main text. 

Summary:

This important paper describes improvements to the measurement of enkephalins in vivo using microdialysis and LC-MS. The key improvement is the oxidation of met- to prevent having a mix of reduced and oxidized methionine in the sample which makes quantification more difficult. It then shows measurements of enkephalins in the nucleus accumbens in two different stress situations - handling and exposure to predator odor. It also reports the ratio of released met- and leu-enkephalin matching what is expected from the digestion of proenkephalin. Measurements are also made by photometry of Ca2+ changes for the fox odor stressor. Some key takeaways are the reliable measurement of met-enkephalin, the significance of directly measuring peptides as opposed to proxy measurements, and the opening of a new avenue into the research of enkephalins due to stress based on these direct measurements.

Strengths:

-Improved methods for measurement of enkephalins in vivo.

-Compelling examples of using this method.

-Opening a new area of looking at stress responses through the lens of enkephalin concentrations.

Weaknesses:

(1) It is not clear if oxidized met-enk is endogenous or not and this method eliminates being able to discern that.

We clarified our wording in the text copied below to provide an explanation on how we distinguish between the two. Even after oxidation, the standard signal has a higher m/z ratio due to the presence of the Carbon and Nitrogen isotopes as described in the Chemicals section of the methods ‘For Met Enkephalin, a fully labeled L-Phenylalanine (¹³C₉, ¹⁵N) was added (YGGFM). The resulting mass shift between the endogenous (light) and heavy isotope-labeled peptide are 7Da and 10Da, respectively.’, so they can still be differentiated from the endogenous signal. We have clarified the language in the results section. See lines 82-87. 

‘After each sample collection, we add a consistent known concentration of isotopically labeled internal standard of Met-Enk and Leu-Enk of 40 amol/sample to the collected ISF for the accurate identification and quantification of endogenous peptide. These internal standards have a different mass/charge (m/z) ratio than endogenous Met- and Leu-Enk. Thus, we can identify true endogenous signal for Met-Enk and Leu-Enk (Suppl Fig. 1A,C) versus noise, interfering signals, and standard signal (Suppl. Fig. 1B,D).’

(2) It is not clear if the spatial resolution is really better as claimed since other probes of similar dimensions have been used.

Apologies for any confusion here. To clarify we primarily state that our approach improves temporal resolution and in a few cases refer to improved spatiotemporal resolution, which we believe we show. The dimensions of the microdialysis probe used in these experiments allow us to target the nucleus accumbens shell and as well as being smaller – especially at the membrane level - than a fiber photometry probe. 

(3) Claims of having the first concentration measurement are not quite accurate.

Thank you for your feedback. To clarify, we do not claim that we have the first concentration measurements, rather we are the first to quantify the ratio of Met-Enk to Leu-Enk in vivo in freely behaving animals in the NAcSh. 

(4) Without a report of technical replicates, the reliability of the method is not as wellevaluated as might be expected.

We have added these details in the methods section, please see lines 521-530. 

‘Each sample was run in two technical replicates and the peak area ratio was averaged before concentration calculations of the peptides were conducted. Several quality control steps were conducted prior to running the in vivo samples. 1) Two technical replicates of a known concentration were injected and analyzed – an example table from 4 random experiments included in this manuscript is shown below. 2) The buffers used on the day of the experiment (aCSF and high K+ buffer) were also tested for any contaminating Met-Enk or Leu-Enk signals by injecting two technical replicates for each buffer. Once these two criteria were met, the experiment was analyzed through the system. If either step failed, which happened a few times, the samples were frozen and the machines were cleaned and restarted until the quality control measures were met.’

Recommendations For The Authors:

Reviewer #1 (Recommendations For The Authors):

• The authors should provide appropriate citations of a study that has validated the Enkephalin-Cre mouse line in the nucleus accumbens or provide verification experiments if they have any available.

Thank you for your comment. We have added a reference validating the Enk-Cre mouse line in the nucleus accumbens to the methods section and is copied here. 

D.C. Castro, C.S. Oswell, E.T. Zhang, C.E. Pedersen, S.C. Piantadosi, M.A. Rossi, A.C. Hunker, A. Guglin, J.A. Morón, L.S. Zweifel, G.D. Stuber, M.R. Bruchas, An endogenous opioid circuit determines state-dependent reward consumption, Nature 2021 598:7882 598 (2021) 646–651. https://doi.org/10.1038/s41586-02104013-0.

• Better definition of the labels y1,y2,b3 in Figures 1 and S1 would be useful. I may have missed it but it wasn't described in methods, results, or legends.

Thank you for this comment. We have added this information to Fig.1 legend ‘Y1, y2, b3 refer to the different elution fragments resulting from Met-Enk during LC-MS.

• It is interesting that the ratio of KCl-evoked release is what changes differentially for Met- vs Leu. Leu enk increases to the range of met-enk. There is non-detectable or approaching being non-detectable leu-enk (below the 40 amol / sample limit of quantification) in most of the subjects that become apparent and approach basal levels of met-enkephalin. This suggests that the K+ evoked response may be more pronounced for leu-enk. This is something that should be considered for further analysis and should be discussed.

Thank you for this astute observation, and you make a great point. We have added some discussion of this finding in the results and discussion sections see lines 111112 and lines 253-257. 

‘Interestingly, Leu-Enk showed a greater fold change compared to baseline than did Met-Enk with the fold changes being 28 and 7 respectively based on the data in Fig.1F.’

‘We also noted that Leu-Enk showed a greater fold increase relative to baseline after depolarization with high K+ buffer as compared to Met-Enk. This may be due to increased Leu-Enk packaging in dense core vesicles compared to Met-Enk or due to the fact that there are two distinct precursor sources for Leu-Enk, namely both proenkephalin and prodynorphin while Met-Enk is mostly cleaved from proenkephalin (see Table 1 [48]).’

• For example in 2E, it would be helpful to label in the graph axis what samples correspond to the manipulation and also in the text provide the reader with the sample numbers. The authors interpret the relationship between the last two samples of baseline and posthandling stress as the following in the figure legend "the concentration released in later samples is affected; such influence suggests that there is regulation of the maximum amount of peptide to be released in NAcSh. E. The negative correlation in panel d is reversed by using a high K+ buffer to evoke Met-Enk release, suggesting that the limited release observed in D is due to modulation of peptide release rather than depletion of reserves." However, the correlations are similar between 2D and E and it appears that two mice are mediating the difference between the two groups. The appropriate statistical analysis would be to compare the regressions of the two groups. Statistics for the high K+ (and all other graphs where appropriate) need to be reported, including the r2 and p-value.

Thank you for your constructive critique. To elucidate the effect of high K+, we have plotted the regression line and reported the slope for Fig. 2E. Notably, the slope is reduced by a factor of 2 and appears to be driven by a large subset of the animals. The statistics for the high K+ graph are shown on the figure (Fig 1F) which test the hypothesis of whether high K+ leads to the release of Leu-Enk and Met-Enk respectively compared to baseline with aCSF. We have added the test statistics to the figure legend for additional clarity. Fig. 1G has no statistics because it is only there to elucidate the ratio between Met-Enk and Leu-Enk in the same samples. We did not test any hypotheses related to whether there are differences between their levels as that is not relevant to our question. The correlation on the same data is depicted in Fig. 1H, and we have added the R² value per your request. 

• The interpretation that handling stress induces enkephalin release from microdialysis experiments is also confounded by other factors. For instance, from the methods, it appears that mice were connected and sample collection started 30 min after surgery, therefore recovery from anesthesia is also a confounding variable, among other technical aspects, such as equilibration of the interstitial fluid to the aCSF running through the probe that is acting as a transmitter and extracellular molecule "sink". Did the authors try to handle the mice post hookup similar to what was done with photometry to have a more direct comparison to photometry experiments? This procedural difference, recording from recently surgerized animals (microdialysis) vs well-recovered animals with photometry should be mentioned in addition to the other caveats the authors mention.

Thank you for your comment. We are aware of this technical limitation, and it is largely why we sought to conduct the fiber photometry experiments to get at the same question. As you requested, we have included additional language in the discussion to acknowledge this limitation and how we chose to address it by measuring calcium activity in the enkephalinergic neurons, which would presumably be the same cell population whose release we are quantifying using microdialysis. See lines 262-273.  

‘Our findings showed a robust increase in peptide release at the beginning of experiments, which we interpreted as due to experimenter handling stress that directly precedes microdialysis collections. However, there are other technical limitations to consider such as the fact that we were collecting samples from mice that were recently operated on. Another consideration is that the circulation of aCSF through the probe may cause a sudden shift in oncotic and hydrostatic forces, leading to increased peptide release to the extracellular space. As such, we wanted to examine our findings using a different technique, so we chose to record calcium activity from enkephalinergic neurons - the same cell population leading to peptide release. Using fiber photometry, we showed that enkephalinergic neurons are activated by stress exposure, both experimenter handling and fox odor, thereby adding more evidence to suggest that enkephalinergic neurons are activated by stress exposure which could explain the heightened peptide levels at the beginning of microdialysis experiments.’

• The authors should provide more details on handling stress manipulation during photometry. For photometry what was the duration of the handling bout, what was the interval between handling events, and can the authors provide a description of what handling entailed? Were mice habituated to handling days before doing photometry recording experiments?

Thank you for your suggestion. We have addressed all of your points in the methods section. See lines 564-570. 

‘The handling bout which mimicked traditional scruffing lasted about 3-5 seconds. The mouse was then let go and the handling was repeated another two times in a single session with a minimum of 1-2 minutes between handling bouts. Mice were habituated to this manipulation by being attached to the fiber photometry rig, for 3-5 consecutive days prior to the experimental recording. Additionally, the same maneuver was employed when attaching/detaching the fiber photometry cord, so the mice were subjected to the same process several times.’

• For the novel weigh boat experiments, the authors should explicitly state when these experiments were done in relation to the fox urine, was it a different session or the same session? Were they the same animals? Statements like the following (line 251) imply it was done in the same animals in the same session but it should be clarified in the methods "We also showed using fiber photometry that the novelty of the introduction of a foreign object to the cage, before adding fox odor, was sufficient to activate enkephalinergic neurons."

As shown in supplementary figure 4, individual animal data is shown for both water and fox urine exposure (overlaid) to depict whether there were differences in their responses to each manipulation – in the same animal. And yes, you are correct, the animals were first exposed to water 3 times in the recording session and then exposed to fox urine 3 times in the same session. We have added that to the methods section describing in vivo fiber photometry. See lines 575-576.  

• Statistical testing would be needed to affirm the conclusions the authors draw from the fox urine and novel weigh boat experiments. For example, it shows stats that the response attenuates, that it is not different between fox urine and novel (it looks like the response is stronger to the fox urine when looking at the individual animals), etc. These data look clear but stats are formally needed. Formal statistics are also missing in other parts of the manuscript where conclusions are drawn from the data but direct statistical comparisons are not included (e.g. Fig 2.G-I).

The photometry data is shown as z-scores which is a formal statistical analysis. ANOVA would be inappropriate to run to compare z-scores. We understand that this is erroneously done in fiber photometry literature, however, it remains incorrect. The z-scores alone provide all the information needed about the deviation from baseline. We understand that this is not immediately clear to readers, and we thank you for allowing us to explain why this is the case. We have added test statistics to figure legends where hypothesis testing was done and p-values were reported. 

• Did the authors try to present the animals with repeated fox urine exposure to see if this habituates like the photometry?

No, we did not do that experiment due to the constrained timing within which we had to run our microdialysis/LC-MS timeline, but it is a great point for future exploration. 

• It would be useful to present the time course of the odor experiment for the microdialysis experiment.

The timeline is shown in Fig.1a and Fig.3e. To reiterate, each sample is 13 minutes long.

• Can the authors determine if differences in behavior (e.g. excessive avoidance in animals with with one type of response) or microdialysis probe location dictate whether animals fall into categories of increased release, no release, or no-detection? From the breakdown, it looks like it is almost equally split into three parts but the authors' descriptions of this split are somewhat misleading (line 210). " The response to predator odor varies appreciably: although most animals show increased Met-Enk release after fox odor exposure, some show continued release with no elevation in Met-Enk levels, and a minority show no detectable release".

Thank you for your constructive feedback. We do not believe the difference in behavior is correlated with probe placement. The hit map can be found in suppl. Fig 3 and shows that all mice included in the manuscript had probes in the NAcSh. We purposely did not distinguish between dorsal and ventral because of our 1 mm membrane would make it hard to presume exclusive sampling from one subregion. That is a great point though, and we have thought about it extensively for future studies. We have edited the language to reflect the almost even split of responses for Met-Enk and appreciate you pointing that out. 

• Overall, given the inconsistencies in experimental design and overall caveats associated, I think the authors are unable to draw reasonable conclusions from the repeated stressor experiments and something they should either consider is not trying to draw strong conclusions from these observations or perform additional experiments that provide the grounds to derive those conclusions.

We have included additional language on the caveats of our study, and our use of a dual approach using fiber photometry and microdialysis was largely driven by a

desire to offer additional support of our conclusions. We expected pushback about our conclusions, so we wanted to offer a secondary analysis using a different technique to test our hypothesis. To be honest the tone of this comment and content is not particularly constructive (especially for trainees) nor does it offer a space to realistically address anything. This work took multiple years to optimize, it was led by a graduate student, and required a multidisciplinary team. As highlighted, we believe it offers an important contribution to the literature and pushes the field of peptide detection forward.  

Reviewer #2 (Recommendations For The Authors):

A more comprehensive and objective interpretation of results could enhance the overall quality of the paper. The manuscript contains statements like "we are the first to confirm," which can be challenging to substantiate and may not significantly enhance the paper. It's essential to ensure that novelty statements are well-founded. For example, the release of enkephalins from other brain regions after stress exposure is well-documented but not addressed in the paper. Similarly, the role of the NA shell in stress has been extensively studied but lacks coverage in this manuscript.

We have edited the language to reflect your feedback. We have also included relevant literature expanding on the demonstrated roles of enkephalins in the literature. We would like to note that most studies have focused on chronic stress, and we were particularly interested in acute stress. See lines 129-134.

‘These studies have included regions such as the locus coeruleus, the ventral medulla, the basolateral nucleus of the amygdala, and the nucleus accumbens core and shell. Studies using global knockout of enkephalins have shown varying responses to chronic stress interventions where male knockout mice showed resistance to chronic mild stress in one study, while another study showed that enkephalin-knockout mice showed delayed termination of corticosteroid release. [33,34]’ 

Finally, not a weakness but a clarification suggestion: the method description mentions the use of 1% FA in the sample reconstitution solution and LC solvents, which is an unusually high concentration of acid. If this concentration is intentional for maintaining the peptides' oxidation state, it would be beneficial to mention this in the text to assist readers who might want to replicate the method.

This is correct and has been clarified in the methods section

Reviewer #3 (Recommendations For The Authors):

-The Abstract should state the critical improvements that are made. Also, quantify the improvements in spatiotemporal resolution.

Thank you for your comment. We have edited the abstract to reflect this. 

- The use of "amol/sample" as concentration is less informative than an SI units (e.g., pM concentration) and should be changed. Especially since the volume used was the same for in vivo sampling experiments.

Thank you for your comment. We chose to report amol/sample because we are measuring such a small concentration and wanted to account for any slight errors in volume that can make drastic differences on reported concentrations especially since samples are dried and resuspended.  

-Please check this sentence: "After each collection, the samples were spiked with 2 µL of 12.5 fM isotopically labeled Met-Enkephalin and Leu-Enkephalin" This dilution would yield a concentration of ~2 fM. In a 12 uL sample, that would be ~0.02 amol, well below the detection limit. (note that fM would femtomolar concentration and fmol would be femtomoles added).

-"liquid chromatography/mass spectrometry (LC-MS) [9-12]"... Reference 9 is a RIA analysis paper, not LC-MS as stated.

Thank you for catching these. We have corrected the unit and citation. 

-Given that improvements in temporal resolution are claimed, the lack of time course data with a time axis is surprising. Rather, data for baseline and during treatment appear to be combined in different plots. Time course plots of individuals and group averages would be informative.

Due to the expected variability between individual animal time course data, where for example, we measure detectable levels in one sample followed by no detection, it was very difficult to combine data across time. Therefore, to maximize data inclusion from all animals that showed baseline measurements and responses to individual manipulations, we opted to report snapshot data. Our improvement in temporal resolution refers to the duration of each sample rather than continuous sampling, so those two are unrelated. Thank you for your feedback and allowing us to clarify this.

- I do not understand this claim "We use custom-made microdialysis probes, intentionally modified so they are similar in size to commonly used fiber photometry probes to avoid extensive tissue damage caused by traditional microdialysis probes (Fig. 1B)." The probes used are 320 um OD and 1 mm long. This is not an uncommon size of microdialysis probes and indeed many are smaller, so is their probe really causing less damage than traditional probes?

Thank you for your comment. We are only trying to make the point that the tissue damage from these probes is comparable to commonly used fiber photometry probes. We only point that out because tissue damage is used as a point to dissuade the usage of microdialysis in some literature, and we just wanted to disambiguate that. We have clarified the statement you pointed out.  

-The oxidation procedure is a good idea, as mentioned above. It would be interesting to compare met-enk with and without the oxidation procedure to see how much it affects the result (I would not say this is necessary though). It is not uncommon to add antioxidants to avoid losses like this. Also, it should be acknowledged that the treatment does prevent the detection of any in vivo oxidation, perhaps that is important in met-enk metabolism?

The comparison between oxidized and unoxidized Met-Enk detection is in figure 1C. 

-It would be a best practice to report the standard deviation of signal for technical replicates (say near in vivo concentrations) of standards and repeated analysis of a dialysate sample to be able to understand the variability associated with this method. Similarly, an averaged basal concentration from all rats.

Thank you for your comment. We have included a table showing example quality control standard injections from 4 randomly selected experiments included in the manuscript that were run before and after each experiment and descriptive statistics associated with these technical replicates. We also added some detail to the methods section to describe how quality control is done. See lines 521-530. 

‘Each sample was run in two technical replicates and the peak area ratio was averaged before concentration calculations of the peptides were conducted. Several quality control steps were conducted prior to running the in vivo samples. 1) Two technical replicates of a known concentration were injected and analyzed – an example table from 4 random experiments included in this manuscript is shown below. 2) The buffers used on the day of the experiment (aCSF and high K+ buffer) were also tested for any contaminating Met-Enk or Leu-Enk signals by injecting two technical replicates for each buffer. Once these two criteria were met, the experiment was analyzed through the system. If either step failed, which happened a few times, the samples were frozen and the machines were cleaned and restarted until the quality control measures were met.’

EDITORS NOTE

Should you choose to revise your manuscript, please include full statistical reporting including exact p-values wherever possible alongside the summary statistics (test statistic and df) and 95% confidence intervals. These should be reported for all key questions and not only when the p-value is less than 0.05.

Thank you for your suggestion. We have included more detail about statistical analysis in the figure legends per this comment and reviewer comments.

Listing multiple security strategies, and noting that the article intentionally provides several solutions rather than a single quick fix, aligns with my opinion that a multi-step security approach is required.

Reviewer #1 (Public review):

In this manuscript, the authors report that GPR55 activation in presynaptic terminals of Purkinje cells decrease GABA release at the PC-DCN synapse. The authors use an impressive array of techniques (including highly challenging presynaptic recordings) to show that GPR55 activation reduces the readily releasable pool of vesicle without affecting presynaptic AP waveform and presynaptic Ca2+ influx. This is an interesting study, which is seemingly well-executed and proposes a novel mechanism for the control of neurotransmitter release. However, the authors' main conclusions are heavily, if not solely, based on pharmacological agents that most often than not demonstrate affinity at multiple targets. Below are points that the authors should consider in a revised version.

Major points:

(1) There is no clear evidence that GPR55 is specifically expressed in presynaptic terminals at the PC-DCN synapse. The authors cited Ryberg 2007 and Wu 2013 in the introduction, mentioning that GPR55 is potentially expressed in PCs. Ryberg (2007) offers no such evidence, and the expression in PC suggested by Wu (2013) does not necessarily correlate with presynaptic expression. The authors should perform additional experiments to demonstrate the presynaptic expression of GPR55 at PC-DCN synapse.

(2) The authors' conclusions rest heavily on pharmacological experiments, with compounds that are sometimes not selective for single targets. Genetic deletion of GPR55 would be a more appropriate control. The authors should also expand their experiments with occlusion experiments, showing if the effects of LPI are absent after AM251 or O-1602 treatment. In addition, the authors may want to consider AM281 as a CB1R antagonist without reported effects at GPR55.

(3) It is not clear how long the different drugs were applied, and at what time the recordings were performed during or following drug application. It appears that GPR55 agonists can have transient effects (Sylantyev, 2013; Rosenberg, 2023), possibly due to receptor internalization. The timeline of drug application should be reported, where IPSC amplitude is shown as a function of time and drug application windows are illustrated.

(4) A previous investigation on the role of GPR55 in the control of neurotransmitter release is not cited nor discussed Sylantyev et al., (2013, PNAS, Cannabinoid- and lysophosphatidylinositol-sensitive receptor GPR55 boosts neurotransmitter release at central synapses). Similarities and differences should be discussed.

Minor point:

(1) What is the source of LPI? What isoform was used? The multiple isoforms of LPI have different affinities for GPR55.

Author response:

Public Reviews:

Reviewer #1 (Public review):

In this manuscript, the authors report that GPR55 activation in presynaptic terminals of Purkinje cells decrease GABA release at the PC-DCN synapse. The authors use an impressive array of techniques (including highly challenging presynaptic recordings) to show that GPR55 activation reduces the readily releasable pool of vesicle without affecting presynaptic AP waveform and presynaptic Ca2+ influx. This is an interesting study, which is seemingly well-executed and proposes a novel mechanism for the control of neurotransmitter release. However, the authors' main conclusions are heavily, if not solely, based on pharmacological agents that most often than not demonstrate affinity at multiple targets. Below are points that the authors should consider in a revised version.

We thank the reviewer for the encouraging comments, and will fully address the reviewer’s concerns as detailed below.

Major points:

(1) There is no clear evidence that GPR55 is specifically expressed in presynaptic terminals at the PC-DCN synapse. The authors cited Ryberg 2007 and Wu 2013 in the introduction, mentioning that GPR55 is potentially expressed in PCs. Ryberg (2007) offers no such evidence, and the expression in PC suggested by Wu (2013) does not necessarily correlate with presynaptic expression. The authors should perform additional experiments to demonstrate the presynaptic expression of GPR55 at PC-DCN synapse.

We agree with the reviewer’s concern that the present manuscript lacks the evidence for localization of GPR55 at PC axon terminals. Honestly, our previous attempt to immune-label GPR55 did not work well. Now, we realize that different antibodies are commercially available, and are going to test them. Hopefully, in the revised manuscript, we will demonstrate immunocytochemical images showing GPR55 at terminals of PCs.

(2) The authors' conclusions rest heavily on pharmacological experiments, with compounds that are sometimes not selective for single targets. Genetic deletion of GPR55 would be a more appropriate control. The authors should also expand their experiments with occlusion experiments, showing if the effects of LPI are absent after AM251 or O-1602 treatment. In addition, the authors may want to consider AM281 as a CB1R antagonist without reported effects at GPR55.

We appreciate the reviewer for pointing out the essential issue regarding the specificity of activation of GPR55 in our study. Regarding the direct manipulation of GPR55, such as genetic deletion, we will try acute knock-down of its expression, considering the possibility of compensation which sometimes occur when the complete knock-out is performed. In addition, according to the reviewer’s suggestion, we will examine whether the effects of LPI and AM251 occlude each other, and also perform control experiments showing the lack of CB1R involvement.

(3) It is not clear how long the different drugs were applied, and at what time the recordings were performed during or following drug application. It appears that GPR55 agonists can have transient effects (Sylantyev, 2013; Rosenberg, 2023), possibly due to receptor internalization. The timeline of drug application should be reported, where IPSC amplitude is shown as a function of time and drug application windows are illustrated.

As suggested, the timing and duration of drug application will be indicated together with the time course of changes of IPSC amplitudes. This change will make things much clearer. Thank you for the suggestion.

(4) A previous investigation on the role of GPR55 in the control of neurotransmitter release is not cited nor discussed Sylantyev et al., (2013, PNAS, Cannabinoid- and lysophosphatidylinositol-sensitive receptor GPR55 boosts neurotransmitter release at central synapses). Similarities and differences should be discussed.

We are really sorry for missing this important study in discussion and citation. In the revised version, of course, we will discuss their findings and our data.

Minor point:

(1) What is the source of LPI? What isoform was used? The multiple isoforms of LPI have different affinities for GPR55.

We are sorry for insufficient explanation about the LPI used in our study. We used LPI derived from soy (Merck, catalog #L7635) that was estimated to contain 58% C16:0 and 42% C18:0 or C18:2 LPI. This information will be added to the Materials and Methods in the revised manuscript.

Reviewer #2 (Public review):

Summary:

This paper investigates the mode of action of GPR55, a relatively understudied type of cannabinoid receptor, in presynaptic terminals of Purkinje cells. The authors use demanding techniques of patch clamp recording of the terminals, sometimes coupled with another recording of the postsynaptic cell. They find a lower release probability of synaptic vesicles after activation of GPR55 receptors, while presynaptic voltage-dependent calcium currents are unaffected. They propose that the size of a specific pool of synaptic vesicles supplying release sites is decreased upon activation of GPR55 receptors.

Strengths:

The paper uses cutting-edge techniques to shed light on a little-studied, potentially important type of cannabinoid receptor. The results are clearly presented, and the conclusions are for the most part sound.

We are really happy to hear the encouraging comments from the reviewer.

Weaknesses:

The nature of the vesicular pool that is modified following activation of GPR55 is not definitively characterized.

During revision, we will perform further analysis and additional experiments to obtain deeper insights into the vesicle pools affected by GPR55 as much as possible.

Reviewer #3 (Public review):

Summary:

Inoshita and Kawaguchi investigated the effects of GPR55 activation on synaptic transmission in vitro. To address this question, they performed direct patch-clamp recordings from axon terminals of cerebellar Purkinje cells and fluorescent imaging of vesicular exocytosis utilizing synapto-pHluorin. They found that exogenous activation of GPR55 suppresses GABA release at Purkinje cell to deep cerebellar nuclei (PC-DCN) synapses by reducing the readily releasable pool (RRP) of vesicles. This mechanism may also operate at other synapses.

Strengths:

The main strength of this study lies in combining patch-clamp recordings from axon terminals with imaging of presynaptic vesicular exocytosis to reveal a novel mechanism by which activation of GPR55 suppresses inhibitory synaptic strength. The results strongly suggest that GPR55 activation reduces the RRP size without altering presynaptic calcium influx.

We thank the reviewer for the positive evaluation on our conclusions.

Weaknesses:

The study relies on the exogenous application of GPR55 agonists. It remains unclear whether endogenous ligands released due to physiological or pathological activities would have similar effects. There is no information regarding the time course of the agonist-induced suppression. There is also little evidence that GPR55 is expressed in Purkinje cells. This study would benefit from using GPR55 knockout (KO) mice. The downstream mechanism by which GPR55 mediates the suppression of GABA release remains unknown.

We agree with the reviewer in all respects suggested as weaknesses. Most issues will be made much clearer by the additional experiments and analysis described above to respond to respective issues raised by other reviewers. The situation of endogenous ligands for GPR55 causing the synaptic depression and its downstream mechanism are very important issues, and we are going to discuss these points in the revised manuscript, and like to work on these in the future study.

Author response:

The following is the authors’ response to the original reviews.

Reviewer #1 (Public review):

Summary:

The paper proposes that the placement of criteria for determining whether a stimulus is 'seen' or 'unseen' can significantly impact the validity of neural measures of consciousness. The authors found that conservative criteria, which require stronger evidence to classify a stimulus as 'seen,' tend to inflate effect sizes in neural measures, making conscious processing appear more pronounced than it is. Conversely, liberal criteria, which require less evidence, reduce these effect sizes, potentially underestimating conscious processing. This variability in effect sizes due to criterion placement can lead to misleading conclusions about the nature of conscious and unconscious processing.

Furthermore, the study highlights that the Perceptual Awareness Scale (PAS), a commonly used tool in consciousness research, does not effectively mitigate these criterion-related confounds. This means that even with PAS, the validity of neural measures can still be compromised by how criteria are set. The authors emphasize the need for careful consideration and standardization of criterion placement in experimental designs to ensure that neural measures accurately reflect the underlying cognitive processes. By addressing this issue, the paper aims to improve the reliability and validity of findings in the field of consciousness research.

Strengths:

(1) This research provides a fresh perspective on how criterion placement can significantly impact the validity of neural measures in consciousness research.

(2) The study employs robust simulations and EEG experiments to demonstrate the effects of criterion placement, ensuring that the findings are well-supported by empirical evidence.

(3) By highlighting the limitations of the PAS and the impact of criterion placement, the study offers practical recommendations for improving experimental designs in consciousness research.

Weaknesses:

The primary focused criterion of PAS is a commonly used tool, but there are other measures of consciousness that were not evaluated, which might also be subject to similar or different criterion limitations. A simulation could applied to these metrics to show how generalizable the conclusion of the study is.

We would like to thank reviewer 1 for their positive words and for taking the time to evaluate our manuscript. We agree that it would be important to gauge generalization to other metrics of consciousness. Note however, that the most commonly used alternative methods are postdecision wagering and confidence, both of which are known to behave quite similarly to the PAS (Sandberg, Timmermans , Overgaard & Cleeremans, 2010). Indeed, we have confirmed in other work that confidence is also sensitive to criterion shifts (see https://osf.io/preprints/psyarxiv/xa4fj). Although it has been claimed that confidence-derived aggregate metrics like meta-d’ or metacognitive efficiency may overcome criterion shifts, it would require empirical data rather than simulation to settle whether this is true or not (also see the discussion in https://osf.io/preprints/psyarxiv/xa4fj). Furthermore, out of these metrics, the PAS seems to be the preferred one amongst consciouness researchers (see figure 4 in Francken, Beerendonk, Molenaar, Fahrenfort, Kiverstein, Seth, Gaal S van, 2022; as well as https://osf.io/preprints/psyarxiv/bkxzh). Thus, given the fact that other metrics are either expected to behave in similar ways and/or because it would require more empirical work to determine along which dimension(s) criterion shifts would operate in alternative metrics, we see no clear path to implement the suggested simulations. We anticipate that aiming to do this would require a considerable amount of additional work, figuring out many things which we believe would better suit a future project. We would of course be open to doing this if the reviewer would have more specific suggestions for how to go about the proposed simulations.

Reviewer #2 (Public review):

Summary:

The study investigates the potential influence of the response criterion on neural decoding accuracy in consciousness and unconsciousness, utilizing either simulated data or reanalyzing experimental data with post-hoc sorting data.

Strengths:

When comparing the neural decoding performance of Target versus NonTarget with or without post-hoc sorting based on subject reports, it is evident that response criterion can influence the results. This was observed in simulated data as well as in two experiments that manipulated the subject response criterion to be either more liberal or more conservative. One experiment involved a two-level response (seen vs unseen), while the other included a more detailed four-level response (ranging from 0 for no experience to 3 for a clear experience). The findings consistently indicated that adopting a more conservative response criterion could enhance neural decoding performance, whether in conscious or unconscious states, depending on the sensitivity or overall response threshold.

Weaknesses:

(1) The response criterion plays a crucial role in influencing neural decoding because a subject's report may not always align with the actual stimulus presented. This discrepancy can occur in cases of false alarms, where a subject reports seeing a target that was not actually there, or in cases where a target is present but not reported. Some may argue that only using data from consistent trials (those with correct responses) would not be affected by the response criterion. However, the authors' analysis suggests that a conservative response criterion not only reduces false alarms but also impacts hit rates. It is important for the authors to further investigate how the response criterion affects neural decoding even when considering only correct trials.

We would like to thank reviewer 2 for taking the time to evaluate our manuscript. We appreciate the suggestion to investigate neural decoding on only correct trials. What we in fact did is consider target trials that are 'correct' (hits = seen target present trials) and 'incorrect' (misses = unseen target present trials) separately, see figure 4A and figure 4B. This shows that the response criterion also affects the neural measure of consciousness when only considering correct target present trials. Note however, that one cannot decode 'unseen' (target present) trials if one only aims to decode 'correct' trials, because those are all incorrect by definition. We did not analyze false alarms (these would be the 'seen' trials on the noise distribution of Figure 1A), as there were not enough trials of those, especially in the conservative condition (see Figure 2C and 2D), making comparisons between conservative and liberal impossible. However, the predictions for false alarms are pretty straightforward, and follow directly from the framework in Figure 1.

(2) The author has utilized decoding target vs. nontarget as the neural measures of unconscious and/or conscious processing. However, it is important to note that this is just one of the many neural measures used in the field. There are an increasing number of studies that focus on decoding the conscious content, such as target location or target category. If the author were to include results on decoding target orientation and how it may be influenced by response criterion, the field would greatly benefit from this paper.

We thank the reviewer for the suggestion to decode orientation of the target. In our experiments, the target itself does not have an orientation, but the texture of which it is composed does. We used four orientations, which were balanced out within and across conditions such that presence-absence decoding is never driven by orientation, but rather by texture based figure-ground segregation (for similar logic, see for example Fahrenfort et al, 2007; 2008 etc). There are a couple of things to consider when wanting to apply a decoding analysis on the orientation of these textures:

(1) Our behavioral task was only on the presence or absence of the target, not on the orientation of the textures. This makes it impossible to draw any conclusions about the visibility of the orientation of the textures. Put differently: based on behavior there is no way of identifying seen or unseen orientations, correctly or incorrectly identified orientations etc. For examply, it is easy to envision that an observer detects a target without knowing the orientation that defines it, or vice versa a situation in which an observer does not detect the target while still being aware of the orientation of a texture in the image (either of the figure, or of the background). The fact that we have no behavioral response to the orientation of the textures severely limits the usefulness of a hypothetical decoding effect on these orientations, as such results would be uninterpretable with respect to the relevant dimension in this experiment, which is visibility.

(2) This problem is further excarbated by the fact that the orientation of the background is always orthogonal to the orientation of the target. Therefore, one would not only be decoding the orientation of the texture that constitutes the target itself, but also the texture that constitutes the background. Given that we also have no behavioral metric of how/whether the orientation of the background is perceived, it is similarly unclear how one would interpret any observed effect.

(3) Finally, it is important to note that – even when categorization/content is sometimes used as an auxiliary measure in consciousness research (often as a way to assay objective performance) - consciousness is most commonly conceptualized on the presence-absence dimension. A clear illustration of this is the concept of blindsight. Blindsight is the ability of observers to discriminate stimuli (i.e. identify content) without being able to detect them. Blindsight is often considered the bedrock of the cognitive neuroscience of consciousness as it acts as proof that one can dissociate between unconscious processing (the categorization of a stimulus, i.e. the content) and conscious processing of that stimulus (i.e. the ability to detect it).

Given the above, we do not see how the suggested analysis could contribute to the conclusions that the manuscript already establishes. We hope that – given the above - the reviewer agrees with this assessment.

Reviewer #3 (Public review):

Summary:

Fahrenfort et al. investigate how liberal or conservative criterion placement in a detection task affects the construct validity of neural measures of unconscious cognition and conscious processing. Participants identified instances of "seen" or "unseen" in a detection task, a method known as post hoc sorting. Simulation data convincingly demonstrate that, counterintuitively, a conservative criterion inflates effect sizes of neural measures compared to a liberal criterion. While the impact of criterion shifts on effect size is suggested by signal detection theory, this study is the first to address this explicitly within the consciousness literature. Decoding analysis of data from two EEG experiments further shows that different criteria lead to differential effects on classifier performance in post hoc sorting. The findings underscore the pervasive influence of experimental design and participants report on neural measures of consciousness, revealing that criterion placement poses a critical challenge for researchers.

Strengths and Weaknesses:

One of the strengths of this study is the inclusion of the Perceptual Awareness Scale (PAS), which allows participants to provide more nuanced responses regarding their perceptual experiences. This approach ensures that responses at the lowest awareness level (selection 0) are made only when trials are genuinely unseen. This methodological choice is important as it helps prevent the overestimation of unconscious processing, enhancing the validity of the findings.

A potential area for improvement in this study is the use of single time-points from peak decoding accuracy to generate current source density topography maps. While we recognize that the decoding analysis employed here differs from traditional ERP approaches, the robustness of the findings could be enhanced by exploring current source density over relevant time windows. Event-related peaks, both in terms of timing and amplitude, can sometimes be influenced by noise or variability in trial-averaged EEG data, and a time-window analysis might provide a more comprehensive and stable representation of the underlying neural dynamics.

We thank reviewer 3 for their positive words and for taking the time to evaluate our manuscript. If we understand the reviewer correctly, he/she suggests that the signal-to-noise ratio could be improved by averaging over time windows rather than taking the values at singular peaks in time. Before addressing this suggestion, we would like to point out that we plotted the relevant effects across time in Supplementary Figure S1A and S1B. These show that the observed effects were not somehow limited in time, i.e. only occuring around the peaks, but that they consistenly occured throughout the time course of the trial. In line with this observation one might argue that the results could be improved further by averaging across windows of interest rather than taking the peak moments alone, as the reviewer suggests. Although this might be true, there are many analysis choices that one can make, each of which could have a positive (or negative) effect on the signal to noise ratio. For example, when taking a window of interest, one is faced with a new choice to make, this time regarding the number of consecutive samples to average across (i.e. the size of the window), etc. More generally there is a long list of choices that may affect the precise outcome of analyses, either positively or negatively. Having analyzed the data in one way, the problem with adding new analysis approaches is that there is no objective criterion for deciding which analysis would be ‘best’, other than looking at the outcome of the statistical analyses themselves. Doing this would constitute an explorative double-dipping-like approach to analyzing the results, which – aside from potentially increasing the signal to noise ratio – is likely to also result in an increase of the type I error rate. In the past, when the first author of this manuscript has attempted to minimize the number of statistical tests, he has lowered the number of EEG time points by simply taking the peaks (for example see https://doi.org/10.1073/pnas.1617268114), and that is the approach that was taken here as well. Given the above, we prefer not to further ‘try out’ additional analytical approaches on this dataset, simply to improve the results. We hope the reviewer sympathizes with our position that it is methodologically most sound to stick to the analyses we have already performed and reported, without further exploration.

It is helpful that the authors show the standard error of the mean for the classifier performance over time. A similar indication of a measure of variance in other figures could improve clarity and transparency.

That said, the paper appears solid regarding technical issues overall. The authors also do a commendable job in the discussion by addressing alternative paradigms, such as wagering paradigms, as a possible remedy to the criterion problem (Peters & Lau, 2015; Dienes & Seth, 2010). Their consideration of these alternatives provides a balanced view and strengthens the overall discussion.

We thank the reviewer for this suggestion. Note that we already have a measure of variance in the other figures too, namely showing the connected data points of individual participants. Indvidual data points as a visualization of variance is preferred by many journals (e.g., see https://www.nature.com/documents/cr-gta.pdf), and also shows the spread of relevant differences when paired points are connected. For example, in Figure 2, 3 and 4, the relevant difference is between the liberal and conservative condition. When wanting to show the spread of the differences between these conditions, one option would be to first subtract the two measures in a pairwise fashion (e.g., liberal-conservative), and then plot the spread of those differences using some metric (e.g. standard error/CI of the mean difference). However, this has the disadvantage of no longer separately showing the raw scores on the conditions that are being compared. Showing conditions separately provides clarity to the reader about what is being compared to what. The most common approach to visualizing the variance of the relevant difference in such cases, is to plot the connected individual data points of all participants in the same plot. The uniformity of the slope of these lines in such a visualization provides direct insight into the spread of the relevant difference. Plotting the standard error of the mean on the raw scores of the conditions in these plots would not help, because this would not visualize the spread of the relevant difference (liberal-conservative). We therefore opted in the manuscript to show the mean scores on the conditions that we compare, while also showing the connected raw data points of individual participants in the same plot. One might argue that we should then use that same visualization in figure 3A, but note that this figure is merely intended to identify the peaks, i.e. it does not compare liberal to conservative. Furthermore, plotting the decoding time lines of individual participants would greatly diminish the clarity of this figure. Given our explanation, we hope the reviewer agrees with the approach that we chose, although we are of course open to modifying the figures if the reviewer has a suggestion for doing so while taking into account the points we raise here in our response.

Impact of the Work:

This study effectively demonstrates a phenomenon that has been largely unexplored within the consciousness literature. Subjective measures may not reliably capture the construct they aim to measure due to criterion confounds. Future research on neural measures of consciousness should account for this issue, and no-report measures may be necessary until the criterion problem is resolved.

Recommendations for the authors:

Reviewer #2 (Recommendations for the authors):

The authors could further elaborate on the results of the PAS to provide a clearer insight into the impact of response criteria, which is notably more complex than in other experiments. Specifically, the results demonstrate that conservative response criterion condition displays a considerably higher sensitivity compared to those with a liberal response criterion. It would be interesting to explore whether this shift in sensitivity suggests a correlation between changes in response criteria and conscious experiences, and how the interaction between sensitivity and response criteria can affect the neural measure of consciousness.

We thank the reviewer for this suggestion. Note that the change in sensitivity that we observed is minor compared to the change we observed in response criterion (hedges g criterion in exp 2 = 2.02, compared to hedges g sensitivity/d’ in exp 2 = 0.42). However, we do investigate the effect of sensitivity (disregarding response criterion) on decoding accuracy. To this end we devised Figure 3C (for the full decoding time course see Supplementary Figure S1B). These figures show that the small behavioral sensitivity effects observed in both experiments (hedges g sensitivity in exp 1 = 0.30, exp 2 = 0.42) did not translate into significant decoding differences between conservative and liberal in either experiment. This comes as no surprise given the small corresponding behavioral effects. Note that small sensitivity differences between liberal and conservative conditions are commonplace, plausibly driven by the fact that being liberal also involves being more noisy in one’s response tendencies (i.e. sometimes randomly indicating presence). Further, the reviewer suggests that we might correlate changes in response criteria to changes in conscious experience. The only relevant metric of conscious experience for which we have data in this manuscript is the Perceptual Awareness Scale (PAS), so we assume the reviewer asks for a correlation between experimentally induced changes in response criterion with the equivalent changes in d’. To this end we computed the difference in the PAS-based d’ metric between conservative and liberal, as well as the difference in the PAS-based criterion metric between conservative and liberal, and correlated these across subjects (N=26) using a Spearman rank correlation. The result shows that these metrics do not correlate r(24)=0.04, p=0.85. Note however that small-N correlations like these are only somewhat reliable for large effect sizes. An N of 26 and a mere power of 80% requires an effect size of at least r=0.5 to be detectable, so even if a correlation were to exist we may not have had enough power to detect it. Due to these caveats we opted to not report this null-correlation in the manuscript, but we are of course willing to do so if the reviewer and/or editor disagrees with this assessment.

Unnecessary since you already established this in a previous paragraph.

в документации производителя стороннего по?

"перейдите"

"параметр"!

в справке прям навзание таблицы и столбца такое

Reviewer #3 (Public review):

Summary

The manuscript investigates the role of norepinephrine (NE) release in the rodent hippocampus during event boundaries, such as transitions between spatial contexts and the introduction of novel objects. It also explores how NE release is altered by experience and how novelty drives the amplitude and decay times of extracellular NE. By utilizing the GRABNE sensor for sub-second resolution measurement of NE, the authors demonstrate that NE release is driven primarily by the time elapsed since an event boundary and is independent of behaviors like movement or reward. The study further explores how hippocampal neural representations are altered over time, showing that these representations stabilize shortly after event transitions, potentially linking NE release to episodic memory encoding.

Strengths

Overall, the work provides novel insights into the interplay between NE signaling and hippocampal activity and presents an intriguing hypothesis on how NE release may help push hippocampal activity into unique attractor states to encode novel experiences. The experiments are well-controlled, and the analysis is well-presented, with a detailed and engaging discussion that points towards several new and exciting research directions. The use of several behavioral paradigms to demonstrate the strongest predictor of NE release is a strength, as well as the regression analysis to disambiguate the contribution of other correlated variables. The suggestion that NE does not select ensembles for subsequent replay is also an interesting result.

Weaknesses

The authors have not convincingly established a link between hippocampal neural activity and NE release, showing qualitative rather than quantitative correlations. Therefore, at this stage, the role of NE on hippocampal function remains speculative.

Another general concern is that the smoothing/ kinetics of the sensor impacts the regression analyses. Most of the other variables, such as speed, acceleration, and even reward time points are highly dynamic and it is possible that the limitations of the sensor decorrelate the signal from (potentially) causal variables, therefore resulting in the time since the event start having the most explanatory power for most of the analyses.

More broadly, the figure legends should be expanded to better describe error bounds, mean vs median, sample sizes, and averaging choices for plots.

There are also some concerns regarding the nearest neighbor analysis and the reported differences in the rate of reactivations after familiar and novel environments, as outlined below.

(1) Lines 657-658. How far away in time can the top three nearest neighbor time points be? Must they lie in different trials, or can they also be within the same trial? Is there a systematic difference in the average time lags for the nearest neighbors over the course of the session?

The authors should only allow nearest neighbors to be in a different lap because systematic changes in behavior (running fast initially) might force earlier time bins in a certain location to match with a different trial, while the later time bins can be from within the same trial if the mice are moving slower and stay in the same spatial bin location longer. The authors should also provide information on how the averaging is performed because there are several axes of variability - spatial bin locations, sessions, different environments, and animals.

(2) Figure 8: These results are very interesting. However, I am confused by the differences between Figure 8B and D because the significant reactivations in A and C are very similar. The 1-minute and 10-minute windows seem somewhat arbitrary and prone to noise and variability. Perhaps the authors should fit a slope for the curves on A and C and compare whether the slope/ intercept are significantly different between the novel and familiar environments.

Accordion ToC?

This section explores operating-system debugging, covering failure analysis, performance monitoring, and advanced tracing tools. Debugging involves identifying and fixing errors in software and hardware, with performance tuning aiming to eliminate processing bottlenecks. When a process fails, operating systems log errors and may generate core dumps for analysis, while kernel failures result in crash dumps. Debugging kernel issues is complex due to hardware control and limited debugging tools. Performance monitoring relies on counters and tracing methods. Linux provides tools like ps, top, vmstat, and /proc for tracking resource usage, while Windows uses Task Manager. Tracing tools, such as strace, gdb, and tcpdump, capture event-based data for in-depth analysis. The BCC toolkit, built on eBPF, enables secure and low-impact debugging of live systems by tracing interactions between user and kernel code. BCC tools, such as disksnoop for disk I/O and opensnoop for system calls, provide real-time insights into system performance and security without disrupting critical applications.

Operating system structure is crucial for managing complexity and ensuring functionality. Modern operating systems are typically organized into modular components, each with well-defined interfaces, similar to how programs are divided into functions. The monolithic structure, used in UNIX and Linux, combines all kernel functionalities into a single address space, offering high performance due to minimal overhead but making the system difficult to extend and maintain. In contrast, the layered approach divides the OS into distinct layers, each relying on the services of lower layers, simplifying debugging and verification but often suffering from performance overhead due to inter-layer communication. The microkernel approach, exemplified by Mach (used in macOS and iOS), minimizes the kernel by moving nonessential services to user space, communicating via message passing. This enhances security, reliability, and portability but incurs performance penalties due to message-passing overhead. For example, Windows NT initially used a microkernel but shifted to a more monolithic design to improve performance. A modern compromise is the use of loadable kernel modules (LKMs), as seen in Linux, macOS, and Windows. This approach combines the efficiency of monolithic kernels with the flexibility of modular design. Core services remain in the kernel, while additional functionalities, like device drivers or file systems, are dynamically loaded at runtime. This allows for easier updates and customization without recompiling the entire kernel, maintaining performance while enhancing modularity. Overall, the choice of structure depends on balancing performance, flexibility, and ease of maintenance. Hybrid operating systems combine different structures to balance performance, security, and usability. Linux, primarily monolithic, is modular for dynamic functionality. Windows shares monolithic traits but supports microkernel features like user-mode subsystems. Apple's macOS and iOS share the Darwin kernel, featuring Mach and BSD UNIX. macOS targets desktops with Intel chips, while iOS is optimized for mobile ARM architectures with strict security controls. Both provide Cocoa frameworks for application development. Darwin’s hybrid nature integrates Mach’s memory management and IPC with BSD’s POSIX functions. Despite Darwin’s open-source status, Apple’s proprietary frameworks, such as Cocoa, remain closed to developers. Android is an open-source mobile OS developed by Google, supporting various hardware platforms. It features a layered architecture, including the Linux kernel, Bionic C library, and Android Runtime (ART), which performs ahead-of-time (AOT) compilation for efficiency. Developers use Java with Android’s custom API, accessing hardware via the hardware abstraction layer (HAL).

Operating system design and implementation involve defining clear goals and balancing user and system requirements. User goals focus on convenience, reliability, and speed, while system goals emphasize ease of design, flexibility, and efficiency. A key principle is separating mechanisms (how to do something) from policies (what to do), enabling flexibility and adaptability. For example, microkernel systems use minimal, policy-free mechanisms, allowing customization, while systems like Windows integrate both for consistency. Modern operating systems are typically written in higher-level languages like C or C++, with some assembly for critical parts, improving portability, maintainability, and performance. Compiler optimizations and efficient algorithms often outweigh the benefits of assembly language, making higher-level languages preferable for most OS development.

Applications are often operating-system specific due to differences in system calls, binary formats, and CPU instruction sets. System calls, which enable applications to interact with the OS, vary across platforms, making cross-platform execution challenging. Three approaches enable multi-OS compatibility: 1) Interpreted languages (e.g., Python) use interpreters to execute code on different OSes, though performance and feature sets may be limited. 2) Virtual machines (e.g., Java) run applications within a portable runtime environment, but with similar limitations. 3) Porting applications to each OS using standard APIs (e.g., POSIX) is time-consuming. Binary formats (e.g., ELF, PE) and ABIs further complicate cross-platform compatibility, as they are tied to specific architectures and OSes. These factors make developing cross-platform applications, like Firefox, a complex task requiring significant adaptation for each OS and CPU architecture.

Linkers and loaders play a crucial role in transforming a program from a disk-based binary executable (e.g., a.out or prog.exe) into a memory-resident process ready for CPU execution. Source files are compiled into relocatable object files, which the linker combines into a single executable, incorporating libraries like the standard C library. The loader then loads this executable into memory, adjusting addresses through relocation to enable proper function calls and variable access. On UNIX systems, the shell uses fork() and exec() system calls to create a process and invoke the loader. Dynamic linking, as seen with Windows DLLs, allows libraries to be linked and loaded only when needed, saving memory. Executable files follow standard formats like ELF (Linux), PE (Windows), or Mach-O (macOS), containing machine code, symbol tables, and entry points. Tools like the file and readelf commands help analyze ELF files, distinguishing between relocatable and executable formats.

System services, or utilities, enhance program development and execution by offering a structured environment. They include file management tools for manipulating files, status information programs for system details, and file modification utilities like text editors. Programming support features compilers and debuggers, while program loading tools manage execution. Communication programs enable virtual connections, and background services, or daemons, run essential processes continuously. These services, alongside application programs, shape the user’s perception of the operating system, often masking the underlying system calls. Different interfaces, like macOS GUI or UNIX shell, can coexist on the same hardware, offering varied user experiences despite shared resources.

System calls act as a bridge between user applications and the operating system, enabling access to essential system services. The section elaborates on their role through an example of copying data from one file to another, illustrating how system calls facilitate file operations, error handling, and user interaction. It highlights how even simple tasks require multiple system calls, such as reading input, opening files, writing data, and handling errors. The discussion extends to APIs, which simplify system call usage for developers, ensuring portability and ease of development. The role of the runtime environment (RTE) in managing system calls is also explored. Various system call types; process control, file management, device management, information maintenance, communications, and protection; are categorized with examples from Windows and UNIX. File management involves a set of essential system calls that enable users to create, delete, open, read, write, reposition, and close files. These operations also extend to directories, facilitating structured organization within a file system. File attributes, such as name, type, and protection codes, can be retrieved or modified using get_file_attributes() and set_file_attributes() system calls. Additionally, some operating systems provide built-in move() and copy() calls, while others integrate these functionalities through APIs. The close relationship between files and devices is evident in UNIX-like systems, where both share a unified interface. Device management encompasses requesting and releasing devices, which prevents conflicts such as deadlocks. System calls also support debugging, retrieving system information, and process monitoring. Communication between processes occurs through message-passing or shared memory, each with distinct advantages. Finally, system protection mechanisms, including permission controls, ensure secure resource access, safeguarding data in multiprogrammed and networked environments.

The user interface of an operating system determines how users interact with it. There are three primary methods: the command-line interface (CLI), the graphical user interface (GUI), and the touch-screen interface. Each has its unique advantages and use cases. The command interpreter, or shell, allows users to enter commands directly. Operating systems like Linux, UNIX, and Windows provide different shells, such as the Bash shell, which execute user commands. These commands can be built-in or executed via system programs. UNIX, for instance, runs external programs to process commands, making it easier to add new commands without modifying the shell. The graphical user interface (GUI), on the other hand, offers a more intuitive approach by using windows, icons, and menus. It was first developed at Xerox PARC in the 1970s and later popularized by Apple Macintosh and Microsoft Windows. GUIs allow users to interact with applications by clicking icons and navigating menus, significantly simplifying the user experience compared to text-based commands. For mobile devices, the touch-screen interface replaces both CLI and GUI with direct touch-based interactions. Users tap, swipe, or use gestures to navigate through applications. The iPhone and iPad, for example, use the Springboard interface to manage apps and settings, minimizing the need for external input devices like keyboards or mice. Ultimately, the choice between these interfaces depends on user preference and system functionality. Power users and administrators prefer CLI for efficiency and automation, while everyday users favor GUIs and touch interfaces for their ease of use.

The Overview of Cloud Computing Using virtualization to its advantage, cloud computing provides on-demand online access to computers, storage, and software. The resources used determine how much users pay. Public, private, hybrid, and SaaS, PaaS, and IaaS cloud types are among them. These settings can integrate several kinds of cloud services.

Integrated Systems For certain tasks, embedded systems are customized computing devices with constrained interfaces. In order to satisfy stringent timing constraints, they frequently use real-time operating systems. These systems need to process sensor data quickly and respond to it. They are utilized in a variety of industries, including industrial automation, medical devices, and automobile control.

Operating Systems That Are Free and Open-Source GNU/Linux is one example of an open-source operating system that makes its source code available for free alteration and dissemination. The push for free software encourages user liberties,

Cultural anthropology and ethnography offers one of the more fascinating windows into the human psyche and it's "trainability." We generally make the mistake of thinking the way we are is the only way we could be. For example, I heard recently of a professed Christian who refuses to believe he'd be a Muslim if he had been born in a Muslim country. Silliness, of course. Foreign kids adopted by American families grow up full American. While there are certainly genetic predispositions (i.e., evidenced by studies of separated twins), I think we are basically born as an unprogrammed computer waiting for our parents and society to load an operating program.

Reviewer #1 (Public review):

Summary:

This study presents convincing findings that oligodendrocytes play a regulatory role in spontaneous neural activity synchronisation during early postnatal development, with implications for adult brain function. Utilising targeted genetic approaches, the authors demonstrate how oligodendrocyte depletion impacts Purkinje cell activity and behaviours dependent on cerebellar function. Delayed myelination during critical developmental windows is linked to persistent alterations in neural circuit function, underscoring the lasting impact of oligodendrocyte activity.

Strengths:

(1) The research leverages the anatomically distinct olivocerebellar circuit, a well-characterized system with known developmental timelines and inputs, strengthening the link between oligodendrocyte function and neural synchronization.

(2) Functional assessments, supported by behavioral tests, validate the findings of in vivo calcium imaging, enhancing the study's credibility.

(3) Extending the study to assess the long-term effects of early-life myelination disruptions adds depth to the implications for both circuit function and behavior.

Weaknesses:

(1) The study would benefit from a closer analysis of myelination during the periods when synchrony is recorded. Direct correlations between myelination and synchronized activity would substantiate the mechanistic link and clarify if observed behavioral deficits stem from altered myelination timing.

(2) Although the study focuses on Purkinje cells in the cerebellum, neural synchrony typically involves cross-regional interactions. Expanding the discussion on how localized Purkinje synchrony affects broader behaviors - such as anxiety, motor function, and sociality - would enhance the findings' functional significance.

(3) The authors discuss the possibility of oligodendrocyte-mediated synapse elimination as a possible mechanism behind their findings, drawing from relevant recent literature on oligodendrocyte precursor cells. However, there are no data presented supporting this assumption. The authors should explain why they think the mechanism behind their observation extends beyond the contribution of myelination or remove this point from the discussion entirely.

(4) It would be valuable to investigate the secondary effects of oligodendrocyte depletion on other glial cells, particularly astrocytes or microglia, which could influence long-term behavioral outcomes. Identifying whether the lasting effects stem from developmental oligodendrocyte function alone or also involve myelination could deepen the study's insights.

(5) The authors should explore the use of different methods to disturb myelin production for a longer time, in order to further determine if the observed effects are transient or if they could have longer-lasting effects.

(6) Throughout the paper, there are concerns about statistical analyses, particularly on the use of the Mann-Whitney test or using fields of view as biological replicates.

Reviewer #2 (Public review):

Summary:

In this manuscript, the authors use genetic tools to ablate oligodendrocytes in the cerebellum during postnatal development. They show that the oligodendrocyte numbers return to normal post-weaning. Yet, the loss of oligodendrocytes during development seems to result in decreased synchrony of calcium transients in Purkinje neurons across the cerebellum. Further, there were deficits in social behaviors and motor coordination. Finally, they suppress activity in a subset of climbing fibers to show that it results in similar phenotypes in the calcium signaling and behavioral assays. They conclude that the behavioral deficits in the oligodendrocyte ablation experiments must result from loss of synchrony.

Strengths:

Use of genetic tools to induce perturbations in a spatiotemporally specific manner.

Weaknesses:

The main weakness in this manuscript is the lack of a cohesive causal connection between the experimental manipulation performed and the phenotypes observed. Though they have taken great care to induce oligodendrocyte loss specifically in the cerebellum and at specific time windows, the subsequent experiments do not address specific questions regarding the effect of this manipulation. Calcium transients in Purkinje neurons are caused to a large extent by climbing fibers, but there is evidence for simple spikes to also underlie the dF/F signatures (Ramirez and Stell, Cell Reports, 2016). Also, it is erroneous to categorize these calcium signals as signatures of "spontaneous activity" of Purkinje neurons as they can have dual origins. Further, the effect of developmental oligodendrocyte ablation on the cerebellum has been previously reported by Mathis et al., Development, 2003. They report very severe effects such as the loss of molecular layer interneurons, stunted Purkinje neuron dendritic arbors, abnormal foliations, etc. In this context, it is hardly surprising that one would observe a reduction of synchrony in Purkinje neurons (perhaps due to loss of synaptic contacts, not only from CFs but also from granule cells). The last experiment with the expression of Kir2.1 in the inferior olive is hardly convincing. In summary, while the authors used a specific tool to probe the role of developmental oligodendrocytes in cerebellar physiology and function, they failed to answer specific questions regarding this role, which they could have done with more fine-grained experimental analysis.

I think to a point, these programs can be considered history. Some shows accurately tell the stories of their time period, while also highlighting fictional characters and stories surrounding them. They can indirectly represent the ideologies of the time period and address sensitive topics.

Multi-head Latent Attention

Compress the key-value store of tokens, which decreases memory usage during inferencing.

I enjoy the symbolism of this statement. You can really imagine what the speaker is talking about.

Author response:

The following is the authors’ response to the current reviews.

Public Reviews:

Reviewer #1 (Public review):

The hypothesis is based on the idea that inversions capture genetic variants that have antagonistic effects on male sexual success (via some display traits) and survival of females (or both sexes) until reproduction. Furthermore, a sufficiently skewed distribution of male sexual success will tend to generate synergistic epistasis for male fitness even if the individual loci contribute to sexually selected traits in an additive way. This should favor inversions that keep these male-beneficial alleles at different loci together at a cis-LD. A series of simulations are presented and show that the scenario works at least under some conditions. While a polymorphism at a single locus with large antagonistic effects can be maintained for a certain range of parameters, a second such variant with somewhat smaller effects tends to be lost unless closely linked. It becomes much more likely for genomically distant variants that add to the antagonism to spread if they get trapped in an inversion; the model predicts this should drive accumulation of sexually antagonistic variants on the inversion versus standard haplotype, leading to the evolution of haplotypes with very strong cumulative antagonistic pleiotropic effects. This idea has some analogies with one of predominant hypotheses for the evolution of sex chromosomes, and the authors discuss these similarities. The model is quite specific, but the basic idea is intuitive and thus should be robust to the details of model assumption. It makes perfect sense in the context of the geographic pattern of inversion frequencies. One prediction of the models (notably that leads to the evolution of nearly homozygously lethal haplotypes) does not seem to reflect the reality of chromosomal inversions in Drosophila, as the authors carefully discuss, but it is the case of some other "supergenes", notably in ants. So the theoretical part is a strong novel contribution.

We appreciate the detailed and accurate summary of our main theoretic results.

To provide empirical support for this idea, the authors study the dynamics of inversions in population cages over one generation, tracking their frequencies through amplicon sequencing at three time points: (young adults), embryos and very old adult offspring of either sex (>2 months from adult emergence). Out of four inversions included in the experiment, two show patterns consistent with antagonistic effects on male sexual success (competitive paternity) and the survival of offspring, especially females, until an old age, which the authors interpret as consistent with their theory.

As I have argued in my comments on previous versions, the experiment only addresses one of the elements of the theoretical hypothesis, namely antagonistic effects of inversions on male reproductive success and other fitness components, in particular of females. Furthermore, the design of this experiment is not ideal from the viewpoint of the biological hypothesis it is aiming to test. This is in part because, rather than testing for the effects of inversion on male reproductive success versus the key fitness components of survival to maturity and female reproductive output, it looks at the effects on male reproductive success versus survival to a rather old age of 2 months. The relevance of survival until old age to fitness under natural conditions is unclear, as the authors now acknowledge. Furthermore, up to 15% of males that may have contributed to the next generation did not survive until genotyping, and thus the difference between these males' inversion frequency and that in their offspring may be confounded by this potential survival-based sampling bias. The experiment does not test for two other key elements of the proposed theory: the assumption of frequency-dependence of selection on male sexual success, and the prediction of synergistic epistasis for male fitness among genetic variants in the inversion. To be fair, particularly testing for synergistic epistasis would be exceedingly difficult, and the authors have now included a discussion of the above caveats and limitations, making their conclusions more tentative. This is good but of course does not make these limitations of the experiment go away. These limitations mean that the paper is stronger as a theoretical than as an empirical contribution.

We discuss the choice to focus on exploring the potential antagonistic effects of the inversion karyotype on male reproductive success and survival in our general response above. Primarily, this prediction seemed to be the most specific to the proposed model as compared to other alternate models. Still, further studies are clearly needed to elucidate the potential frequency dependence and genetic architecture of the inversions.

Regarding the choice of age at collection, it is unknown to what degree our selected collection age of 10 weeks correlates with survival in the wild, but we feel confident that there will be some positive correlation.

We now further clarify that across our experiments, a minimum of 5% and a mean of 9% of the males used in the parental generation died before collection. These proportions do not appear sufficient to explain the differences between paternal and embryo inversion frequencies shown in Figure 9.

Reviewer #2 (Public review):

Summary:

In their manuscript the authors address the question whether the inversion polymorphism in D. melanogaster can be explained by sexually antagonistic selection. They designed a new simulation tool to perform computer simulations, which confirmed their hypothesis. They also show a tradeoff between male reproduction and survival. Furthermore, some inversions display sex-specific survival.

Strengths:

It is an interesting idea on how chromosomal inversions may be maintained

Weaknesses:

The authors motivate their study by the observation that inversions are maintained in D. melanogaster and because inversions are more frequent closer to the equator, the authors conclude that it is unlikely that the inversion contributes to adaptation in more stressful environments. Rather the inversion seems to be more common in habitats that are closer to the native environment of ancestral Drosophila populations.

While I do agree with the authors that this observation is interesting, I do not think that it rules out a role in local adaptation. After all, the inversion is common in Africa, so it is perfectly conceivable that the non-inverted chromosome may have acquired a mutation contributing to the novel environment.

Based on their hypothesis, the authors propose an alternative strategy, which could maintain the inversion in a population. They perform some computer simulations, which are in line with the predicted behavior. Finally, the authors perform experiments and interpret the results as empirical evidence for their hypothesis. While the reviewer is not fully convinced about the empirical support, the key problem is that the proposed model does not explain the patterns of clinal variation observed for inversions in D. melanogaster. According to the proposed model, the inversions should have a similar frequency along latitudinal clines. So in essence, the authors develop a complicated theory because they felt that the current models do not explain the patterns of clinal variation, but this model also fails to explain the pattern of clinal variation.

To the contrary – in the Discussion paragraph beginning on Line 671, we explain why we would predict that a tradeoff between survival and reproduction should lead to clinal inversion frequencies. We suggest that a karyotype associated with a survival penalty should be increasingly disadvantageous in more challenging environments (such as high altitudes and latitudes for this species). Furthermore, an advantage in male reproductive competition conferred by that same haplotype may be reduced by the lower population densities that we would expect in more challenging environments (meaning that each female should encounter fewer males). Individually or jointly, these two factors predict that the equilibrium frequency of a balanced inversion frequency polymorphism should depend on a local population’s environmental harshness and population density, with the ensuing prediction that inversion frequency should correlate with certain environmental variables.

Reviewer #3 (Public review):

Summary:

In this study, McAllester and Pool develop a new model to explain the maintenance of balanced inversion polymorphism, based on (sexually) antagonistic alleles and a trade-off between male reproduction and survival (in females or both sexes). Simulations of this model support the plausibility of this mechanism. In addition, the authors use experiments on four naturally occurring inversion polymorphisms in D. melanogaster and find tentative evidence for one aspect of their theoretical model, namely the existence of the above-mentioned trade-off in two out of the four inversions.

Strengths:

(1) The study develops and analyzes a new (Drosophila melanogaster-inspired) model for the maintenance of balanced inversion polymorphism, combining elements of (sexually) antagonistically (pleiotropic) alleles, negative frequency-dependent selection and synergistic epistasis. Simulations of the model suggest that the hypothesized mechanism might be plausible.

(2) The above-mentioned model assumes, as a specific example, a trade-off between male reproductive display and survival; in the second part of their study, the authors perform laboratory experiments on four common D. melanogaster inversions to study whether these polymorphisms may be subject to such a trade-off. The authors observe that two of the four inversions show suggestive evidence that is consistent with a trade-off between male reproduction and survival.

Open issues:

(1) A gap in the current modeling is that, while a diploid situation is being studied, the model does not investigate the effects of varying degrees of dominance. It would thus be important and interesting, as the authors mention, to fill this gap in future work.

(2) It will also be important to further explore and corroborate the potential importance and generality of trade-offs between different fitness components in maintaining inversion polymorphisms in future work.

We appreciate the work put in to evaluating, improving, and summarizing our study. We agree that further work studying the effects of dominance and of the fitness components of the inversions is important.

Recommendations for the authors:

Reviewer #1 (Recommendations for the authors):

l. 354 : I don't understand what the authors mean by "an antagonistic and non-antagonistic allele". If there is a antagonistic polymorphism at a locus, then both alleles have antagonistic effects; i.e., allele B increases trait 1 and reduced trait 2 relative to allele A and vice versa.

Edited, agreed that the terminology used here was sub-optimal.

Reviewer #2 (Recommendations for the authors):

The motivation for their model is their claim that the clinal inversion frequencies are not compatible with local adaptation. The reviewer doubts this strong statement. Furthermore, the proposed model also fails to explain the inversion frequencies in natural populations.

Hence, rather than building a straw man, it would be better if the authors first show their experiments and then present their model as an explanation for the empirical results. Nevertheless, it is also clear that the empirical data are not very strong and cannot be fully explained by the proposed model.

This claim that we reject any role of local adaptation in clinal variation and selection upon inversion polymorphism does not hold up in a reading of our manuscript. We even suggest that locally varying selective pressures must be playing some role, although that does not imply that local adaptation is the ultimate driver of inversion frequencies. Indeed, we suggest that local adaptation alone is an insufficient explanation for inversion frequency clines in D. melanogaster, including because (1) these frequency clines do not approach the alternate fixed genotypes predicted by local directional selection, (2) these derived inversions tend to be more frequent in more ancestral environments (l.113-158).

In our public review response above, and in the Discussion section of our paper, we explain why our model can predict both the clinal frequencies of many Drosophila inversions and their intermediate maximal frequencies. Of course, we do not predict that most inversions in this species should follow the specific tradeoff investigated here. In fact, we were surprised to find even two inversions that experimentally supported our predicted tradeoff. Still, it remains possible that other inversions in this species are subject to other balanced tradeoffs not investigated here, which could help explain why they rarely reach high local frequencies.

Reviewer #3 (Recommendations for the authors):

My previous comments have been adequately addressed.

The following is the authors’ response to the original reviews.

Reviewer #1 (Public Review):

[…]

To provide empirical support for this idea, the authors study the dynamics of inversions in population cages over one generation, tracking their frequencies through amplicon sequencing at three time points: (young adults), embryos and very old adult offspring of either sex (>2 months from adult emergence). Out of four inversions included in the experiment, two show patterns consistent with antagonistic effects on male sexual success (competitive paternity) and the survival of offspring, especially females, until an old age, which the authors interpret as consistent with their theory.

There are several reasons why the support from these data for the proposed theory is not waterproof.

(1) As I have already pointed out in my previous review, survival until 2 months (in fact, it is 10 weeks and so 2.3 months) of age is of little direct relevance to fitness, whether under natural conditions or under typical lab conditions.

The authors argue this objection away with two arguments

First, citing Pool (2015) they claim that the average generation time (i.e. the average age at which flies reproduce) in nature is 24 days. That paper made an estimate of 14.7 generations per year under the North Carolina climate. As also stated in Pool (2015), the conditions in that locality for Drosophila reproduction and development are not suitable during three months of the year. This yields an average generation length of about 19.5 days during the 9 months during which the flies can reproduce. On the highly nutritional food used in the lab and at the optimal temperature of 25 C, Drosophila need about 11-12 days to develop from egg to adult. Even assuming these perfect conditions, the average age (counted from adult eclosion) would be about 8 days. In practice, larval development in nature is likely longer for nutritional and temperature reasons, and thus the genomic data analyzed by Pool imply that the average adult age of reproducing flies in nature would be about 5 days, and not 24 days, and even less 10 weeks. This corresponds neatly to the 2-6 days median life expectancy of Drosophila adults in the field based on capture-recapture (e.g., Rosewell and Shorrocks 1987).

Second, the authors also claim that survival over a period of 2 month is highly relevant because flies have to survive long periods where reproduction is not possible. However, to survive the winter flies enter a reproductive diapause, which involves profound physiological changes that indeed allow them to survive for months, remaining mostly inactive, stress resistant and hidden from predators. Flies in the authors' experiment were not diapausing, given that they were given plentiful food and kept warm. It is still possible that survival to the ripe old age of 10 weeks under these conditions still correlates well with surviving diapause under harsh conditions, but if so, the authors should cite relevant data. Even then, I do not think this allows the authors to conclude that longevity is "the main selective pressure" on Drosophila (l. 936).

This is overall a thoughtfully presented critique and we have endeavored to improve our discussion of Pool (2015) and to clarify some of the language used about survival elsewhere. While we agree that challenges other than survival to 10 weeks are very relevant to Drosophila melanogaster, collection at 10 weeks does encompass some of these other challenges. Egg to adult viability still contributes to the frequencies of the inversions at collection and is not separable from longevity in this data. Collection at longevity was chosen in part to encompass all lifetime fitness challenges that might influence the inversion frequency at collection, albeit still within permissive laboratory conditions. Future experiments exploring specific stressors independently and beyond permissive lab conditions would generate a clearer picture.

In addition to general edits, the specific phrase mentioned at 1. 936 [now line 1003] has been revised from “In many such cases females are in reproductive diapause, and so longevity is the main selective pressure.” to “While longevity is a key selective pressure underlying overwintering, the relationship between longevity in permissive lab conditions without diapause and in natural conditions under diapause is unclear (Schmidt et al. 2005; Flatt 2020), and our experiment represents just one of many possible ways to examine tradeoffs involving survival.”

(2) It appears that the "parental" (in fact, paternal) inversion frequency was estimated by sequencing sires that survived until the end of the two-week mating period. No information is provided on male mortality during the mating period, but substantial mortality is likely given constant courtship and mating opportunities. If so, the difference between the parental and embryo inversion frequency could reflect the differential survival of males until the point of sampling rather than / in addition to sexual selection.

We have further clarified that when referenced as parental frequency, the frequency presented is ½ the paternal frequency as the mothers were homokaryotypic for the standard arrangement. We chose to present both due to considerations in representing the frequency change from paternal to embryo frequencies, where a hypothetical change from 0.20 frequency in fathers to 0.15 frequency in embryos represents a selective benefit (a frequency increase in the population), despite the reality that this is a decrease in allele frequency between paternal and embryo cohorts.

We mentioned a maximum 15% paternal mortality at line 827 [now l.1056], but have now added complete data on the counts of flies in the experiment as a supplemental table (Table S1) and have added or corrected further references to this in the results and methods [lines 555, 638, 975]. It is true that this may influence the observed frequency changes to some degree, and while we adjusted our sampling method to account for the effects of this mortality on statistical power [l.1056ff], we have now edited the manuscript to better highlight potential effects of this phenomenon on the recorded frequency changes.

It is also worth noting that, if mortality among fathers over the mating period is codirectional with mortality among aged offspring, this would bias the results against detecting an opposing antagonistic selective effect of the inversions on paternity share. This is now also mentioned in the manuscript, l.639ff.

(3) Finally, irrespective of the above caveats, the experimental data only address one of the elements of the theoretical hypothesis, namely antagonistic effects of inversions on reproduction and survival, notably that of females. It does not test for two other key elements of the proposed theory: the assumption of frequency-dependence of selection on male sexual success, and the prediction of synergistic epistasis for male fitness among genetic variants in the inversion. To be fair, particularly testing the latter prediction would be exceedingly difficult. Nonetheless, these limitations of the experiment mean that the paper is much stronger theoretical than empirical contribution.

This is a fair criticism of the limitations of our results, and we now summarize such caveats more directly in the discussion summary, lines 876ff.

Reviewer #2 (Public Review): 

[…]

Comments on the latest version:

I would like to give an example of the confusing terminology of the authors:

"Additionally, fitness conveyed by an allele favoring display quality is also frequency-dependent: since mating success depends on the display qualities of other males, the relative advantage of a display trait will be diminished as more males carry it..."

I do not understand the difference to an advantageous allele, as it increases in frequency the frequency increase of this allele decreases, but this has nothing to do with frequency dependent selection. In my opinion, the authors re-define frequency dependent selection, as for frequency dependent selection needs to change with frequency, but from their verbal description this is not clear.

We have edited this text for greater clarity, now line 232ff. We did not seek to redefine frequency dependence, and did mean by “the relative advantage of a display trait will be diminished” that an equivalent s would diminish with frequency. We have now remedied terminological issues introduced in the prior revision with regard to frequency dependent selection.

One example of how challenging the style of the manuscript is comes from their description of the DNA extraction procedure. In principle a straightforward method, but even here the authors provide a convoluted uninformative description of the procedure.

We have edited for clarity the text on lines 1016-1020. Citing a published protocol and mentioning our modifications seems an appropriate trade-off between representing what was done accurately, citing the sources we relied on in doing it, and limiting the volume of information in the main text for such a straightforward and common method. 

It is not apparent to the reviewer why the authors have not invested more effort to make their manuscript digestible.

We have invested a great deal of effort in making this manuscript as clear as we are able to.  We regret that our writing has not been to this reviewer’s liking. We believe we have been highly responsive to all specific criticisms, including revising all passages cited as unclear. In this round, we have again scrutinized the entire manuscript for any opportunity to clarify it, and we have made further changes throughout.  Although our subject matter is conceptually nuanced, we nevertheless remain optimistic that a careful, fresh reading of our revised manuscript would yield a more favorable impression.

Reviewer #3 (Public Review):

[…]

Weaknesses:

A gap in the current modeling is that, while a diploid situation is being studied, the model does not investigate the effects of varying degrees of dominance. It would be important and interesting to fill this gap in future work.

Agreed, and now reinforced at lines 892ff.

Comments on the latest version:

Most of the comments which I have made in my public review have been adequately addressed.

Some of the writing still seems somewhat verbose and perhaps not yet maximally succinct; some additional line-by-line polishing might still be helpful at this stage in terms of further improving clarity and flow (for the authors to consider and decide).

We have made further changes and some polishing in this draft, and greatly appreciate the guidance provided in improving the draft so far. 

Reviewer #1 (Recommendations For The Authors):

(1) While the model results are convincing, some of the verbal interpretation is confusing. In particular, the authors state that in their model the allele favoring male display quality shows a negative frequency dependence whereas the alternative allele has a positive frequency dependence. This does not make sense to me in the context of population genetics theory. For a one-locus, two-allele model the change of allele frequency under selection depends on the fitness of the genotypes concerned relative to each other. Thus, at least under no dominance assumed in this model, if the relative fitness of AA decreases with the frequency of allele A, the relative fitness of aa must decrease with the frequency of allele a. I.e., if selection is negatively frequency dependent, then it is so for both alleles.

This phrasing was wrong, and we have edited the relevant section.

(2) I am still not entirely sure that the synergistic epistasis assumed in the verbal model is actually generated in the simulations; this would be easy enough to check by extracting the mating success of males with different genotypes from the simulation output should be reported, e.g., as a figure supplement.

Our new Figure S2, which depicts haplotype frequencies for a set of the simulations presented in Figure 4, should demonstrate a necessary presence of synergistic epistasis. These results further clarify that the weaker allele B is only kept when linked to A. The same fitness classes of genotype are present in the simulations with and without the inversion, so the only mechanical difference is the rate of recombination, and the only way this might change selection on the alleles is if a variant has a different fitness in one haplotype background than another – i.e. epistasis. The maintenance of haplotypes AB and ab to the exclusion of Ab and aB relies on the lesser relative fitness of Ab and aB. And since survival values are multiplicative, this additional contribution must come from the mate success of AB being disproportionately larger than Ab or aB, indicating the emergent synergistic epistasis posited by our model. We have clarified this point in the text at line 363ff.

(3) l. 318ff: What was this set number of males? I could not find this information anywhere. Also, this model of the mating system is commonly referred to as "best of N", so the authors may want to include this label in the description.

We indicate this detail just after the referenced line, now reworded and on l. 338-340 as “For each female’s mating competition, 100 males were sampled, though see Figure S1 for plots with varying encounter number.”  Among these edits, “one hundred” has been changed to a numeral for easier skimming, and Figure S1 is now referenced here earlier in the text. Several edits have also been made in the caption of Figures 2 and 3, and in the relevant methods section to clarify the number of encountered males simulated, mention best of N terminology, and clarify how the quality score is used in the mate competition.

(4) The description of the experiment is still confusing. The number of individuals of each sex entered in each mating cage is missing from the Methods (l. 914); although I did finally find it in the Results. These flies were laying over 2 weeks - does this mean that offspring from the entire period were used to obtain the embryo and aged offspring frequencies, or only from a particular egg collection? If the former, does this mean that the offspring obtained from different egg batches were aged separately? Were the offspring aged in cages or bottles, at what density? Given that only those males that survived until the end of the two-week mating period were sequenced, it is important to know what % of the initial number of males these survivors were. A substantial mortality of the parental males could bias the estimate of parental frequencies. How many parental males, embryos and aged offspring were sequenced? Were all individuals of a given cage and stage extracted and sequenced as a single pool or were there multiple pools? The description could also be structured better. For example, the food and grape agar recipes and cage construction are inserted at random points of the description of the crossing design, which does not help.

We have now reorganized and edited these portions of the Methods text. Portions of this comment overlap with edits responding to (2) of the Public Review and below for l. 921 in Details. Offspring from different laying periods were aged in different bottles, further separated by the time at which they eclosed. They were then pooled for DNA extraction and library preparation by sex and a binary early or late eclosion time. This data was present in the “D. mel. Sample Size” column of supplemental tables S6 and S7 (now S7 and S8), but we have added and referenced a new table to specifically collate the sample sizes of different experimental stages, table S1. Now referenced at lines 555, 638, 975, 1057.

(5) The caption of figure 9 and the discussion of its results should be clear and explicit about the fact that "adult offspring" in Fig 9A and "female" and "male" refers to adults surviving to old age (whereas "parental" in Fig 9A refers to young adults in their reproductive prime. This has consequences for the interpretation of the difference between "parental" and "adult offspring", as it combines one generation of usual selection as it occurs under the conditions of the lab culture (young adult at generation t -> young adult in generation t+1) with an additional step of selection for longevity. Thus, a marked change in allele frequency does not imply that the "parental" frequency does not represent an equilibrium frequency of the inversions under the lab culture conditions. Furthermore, it would be useful to state explicitly that Figure 9B represents the same results as figure 9A, but with the aged offspring split by sex.

Figure caption edited to provide further clarity on the age of cohorts and presented data, along with the relevant results section (2.3) referencing this figure.

We avoid making any statements about the equilibrium frequencies of inversions under lab conditions, and whether or not any step of our experiment reflects such equilibria, because our investigation does not rely upon or test for such conditions. Instead, our analysis focuses on whether inversions have contrasting effects (as indicated by frequency changes that are incompatible with neutral sampling) between different life history components.  Under our model, such frequency reversals might be detectable both at equilibrium balanced inversion frequencies and also at frequencies some distance away from equilibria. We have now clarified this point at l. 970-972.

Details:

l. 211: this should be modified as male-only costs are now included.

Edited. “survival likelihood (of either or both sexes).”

l. 343: misplaced period

Edited.

l. 814: "We confirmed model predictions...": This sounds like it refers to an empirical confirmation of a theory prediction, but I think the authors just want to say that their simulations predicted antagonistic variants can be maintained at an intermediate equilibrium frequency. So the wording should be changed to avoid ambiguity.

Edited. Now line 869.

l. 853: How can a genome be "empty"? Do the authors mean an absence of any polymorphism?

Edited to: “In SAIsim, a population is instantiated as a python object, and populated with individuals which are also represented by python objects. These individuals may be instantiated using genomes specified by the user, or by default carry no genomic variation.” Lines 913ff.

l. 853: I do not see this diagramed in Figure 5

Apologies, fixed to Fig. 2

l. 864: is crossing-over in the model limited to female gametogenesis (reflecting the Drosophila case) or does it occur in both sexes?

There is a variable in the simulator to make crossover female-specific. All simulations were performed with female-only crossover. Edited for clarity. “While the simulator can allow recombination in both sexes, all simulations presented only generate crossovers and gene conversion events for female gametes, in accordance with the biology of D. melanogaster.” Lines 928-929.

l. 906: "F2" is ambiguous; does this mean that the mix of lines was allowed to breed for two generations? Also, in other places in the manuscript these flies appear to be referred to are "parental". So do not use F2.

Edited, F2 language removed and replaced with being allowed to breed for two generations. Now lines 967ff.

l. 910: this is incorrect/imprecise; what can be inferred is the frequency of the inversions in male gametes that contributed to fertilization. This would correspond to the frequency in successful males only if each successful male genotype had the same paternity share.

Edited, now “Since no inversions could be inherited through the mothers, inversion frequencies among successful male gametes could be inferred from their pooled offspring.” Now line 994.

l. 912: "without a controlled day/night cycle" meaning what? Constant light? Constant darkness? Daylight falling through the windows?

Edited to “Unless otherwise noted, all flies were kept in a lab space of 23°C with around a degree of temperature fluctuation and without a controlled day/night cycle. Light exposure was dependent on the varying use of the space by laboratory workers but amounted to near constant exposure to at least a minimal level of lighting, with some variable light due to indirect lighting from adjacent rooms with exterior windows.” Now lines 1007-1010.

l. 921: I cannot parse this sentence. Were the offspring isolated as virgins?

No, the logistics of collecting virgins would have been prohibitive, and it did not seem essential for our experiment. Hopefully the edits to this section are clearer, now lines 978ff.

browser

to

http://brewster.kahle.org/files/2015/05/13-bkblog.png

making your PC the deepest node in the network infrastructure, such that your computer serves as both client and server to present web pages to third parties

from

Why Everything is really fast on Windows than explorer index, MFT, NoSQL

Author response:

The following is the authors’ response to the previous reviews.

Public Reviews:

Reviewer #1 (Public Review):

(1) As VRMate (a component of behaviorMate) is written using Unity, what is the main advantage of using behaviorMate/VRMate compared to using Unity alone paired with Arduinos (e.g. Campbell et al. 2018), or compared to using an existing toolbox to interface with Unity (e.g. Alsbury-Nealy et al. 2022, DOI: 10.3758/s13428-021-01664-9)? For instance, one disadvantage of using Unity alone is that it requires programming in C# to code the task logic. It was not entirely clear whether VRMate circumvents this disadvantage somehow -- does it allow customization of task logic and scenery in the GUI? Does VRMate add other features and/or usability compared to Unity alone? It would be helpful if the authors could expand on this topic briefly.

We have updated the manuscript (lines 412-422) to clarify the benefits of separating the VR system as an isolated program and a UI that can be run independently. We argue that “…the recommended behaviorMate architecture has several important advantages. Firstly, by rendering each viewing angle of a scene on a dedicated device, performance is improved by splitting the computational costs across several inexpensive devices rather than requiring specialized or expensive graphics cards in order to run…, the overall system becomes more modular and easier to debug [and] implementing task logic in Unity would require understanding Object-Oriented Programming and C# … which is not always accessible to researchers that are typically more familiar with scripting in Python and Matlab.”

VRMate receives detailed configuration info from behaviorMate at runtime as to which VR objects to display and receives position updates during experiments. Any other necessary information about triggering rewards or presenting non-VR cues is still handled by the UI so no editing of Unity is necessary. Scene configuration information is in the same JSON format as the settings files for behaviorMate, additionally there are Unity Editor scripts which are provided in the VRmate repository which permit customizing scenes through a “drag and drop” interface and then writing the scene configuration files programmatically. Users interested in these features should see our github page to find example scene.vr files and download the VRMate repository (including the editor scripts).  We provided 4 vr contexts, as well as a settings file that uses one of them which can be found on the behaviorMate github page (https://github.com/losonczylab/behaviorMate) in the “vr_contexts” and “example_settigs_files” directories. These examples are provided to assist VRMate users in getting set up and could provide a more detailed example of how VRMate and behaviorMate interact.

(2) The section on "context lists", lines 163-186, seemed to describe an important component of the system, but this section was challenging to follow and readers may find the terminology confusing. Perhaps this section could benefit from an accompanying figure or flow chart, if these terms are important to understand.

We maintain the use of the term context and context list in order to maintain a degree of parity with the java code. However, we have updated lines 173-175 to define the term context for the behaviorMate system: “... a context is grouping of one or more stimuli that get activated concurrently. For many experiments it is desirable to have multiple contexts that are triggered at various locations and times in order to construct distinct or novel environments.”

a. Relatedly, "context" is used to refer to both when the animal enters a particular state in the task like a reward zone ("reward context", line 447) and also to describe a set of characteristics of an environment (Figure 3G), akin to how "context" is often used in the navigation literature. To avoid confusion, one possibility would be to use "environment" instead of "context" in Figure 3G, and/or consider using a word like "state" instead of "context" when referring to the activation of different stimuli.

Thank you for the suggestion. We have updated Figure 3G to say “Environment” in order to avoid confusion.

(3) Given the authors' goal of providing a system that is easily synchronizable with neural data acquisition, especially with 2-photon imaging, I wonder if they could expand on the following features:

a. The authors mention that behaviorMate can send a TTL to trigger scanning on the 2P scope (line 202), which is a very useful feature. Can it also easily generate a TTL for each frame of the VR display and/or each sample of the animal's movement? Such TTLs can be critical for synchronizing the imaging with behavior and accounting for variability in the VR frame rate or sampling rate.

Different experimental demands require varying levels of precision in this kind of synchronization signals. For this reason, we have opted against a “one-size fits all” for synchronization with physiology data in behaviorMate. Importantly this keeps the individual rig costs low which can be useful when constructing setups specifically for use when training animals. behaviorMate will log TTL pulses sent to GPIO pins setup as sensors, and can be configured to generate TTL pulses at regular intervals. Additionally all UDP packets received by the UI are time stamped and logged. We also include the output of the arduino millis() function in all UDP packets which can be used for further investigation of clock drift between system components. Importantly, since the system is event driven there cannot be accumulating drift across running experiments between the behaviorMate UI and networked components such as the VR system.

For these reasons, we have not needed to implement a VR frame synchronization TTL for any of our experiments, however, one could extend VRMate to send "sync" packets back to behaviorMate to log when each frame was displayed precisely or TTL pulses (if using the same ODROID hardware we recommend in the standard setup for rendering scenes). This would be useful if it is important to account for slight changes in the frame rate at which the scenes are displayed. However, splitting rendering of large scenes between several devices results in fast update times and our testing and benchmarks indicate that display updates are smooth and continuous enough to appear coupled to movement updates from the behavioral apparatus and sufficient for engaging navigational circuits in the brain.

b. Is there a limit to the number of I/O ports on the system? This might be worth explicitly mentioning.

We have updated lines 219-220 in the manuscript to provide this information: Sensors and actuators can be connected to the controller using one of the 13 digital or 5 analog input/output connectors.

c. In the VR version, if each display is run by a separate Android computer, is there any risk of clock drift between displays? Or is this circumvented by centralized control of the rendering onset via the "real-time computer"?

This risk is mitigated by the real-time computer/UI sending position updates to the VR displays. The maximum amount scenes can be out of sync is limited because they will all recalibrate on every position update – which occurs multiple times per second as the animal is moving. Moreover, because position updates are constantly being sent by behaviorMate to VRMate and VRMate is immediately updating the scene according to this position, the most the scene can become out of sync with the mouse's position is proportional to the maximum latency multiplied by the running speed of the mouse. For experiments focusing on eliciting an experience of navigation, such a degree of asynchrony is almost always negligible. For other experimental demands it could be possible to incorporate more precise frame timing information but this was not necessary for our use case and likely for most other use cases. Additionally, refer to the response to comment 3a.

Reviewer #2 (Public review):

(1) The central controlling logic is coupled with GUI and an event loop, without a documented plugin system. It's not clear whether arbitrary code can be executed together with the GUI, hence it's not clear how much the functionality of the GUI can be easily extended without substantial change to the source code of the GUI. For example, if the user wants to perform custom real-time analysis on the behavior data (potentially for closed-loop stimulation), it's not clear how to easily incorporate the analysis into the main GUI/control program.

Without any edits to the existing source code behaviorMate is highly customizable through the settings files, which allow users to combine the existing contexts and decorators in arbitrary combinations. Therefore, users have been able to perform a wide variety of 1D navigation tasks, well beyond our anticipated use cases by generating novel settings files. The typical method for providing closed-loop stimulation would be to set up a context which is triggered by animal behavior using decorators (e.g. based on position, lap number and time) and then trigger the stimulation with a TTL pulse. Rarely, if users require a behavioral condition not currently implemented or composable out of existing decorators, it would require generating custom code in Java to extend the UI. Performing such edits requires only knowledge of basic object-oriented programming in Java and generating a single subclass of either the BasicContextList or ContextListDecorator classes. In addition, the JavaFX (under development) version of behaviorMate incorporates a plugin which doesn't require recompiling the code in order to make these changes. However, since the JavaFX software is currently under development, documentation does not yet exist. All software is open-sourced and available on github.com for users interested in generating plugins or altering the source code.

We have added the additional caveat to the manuscript in order to clarify this point (Line 197-202): “However, if the available set of decorators is not enough to implement the required task logic, some modifications to the source code may be necessary. These modifications, in most cases, would be very simple and only a basic understanding of object-oriented programming is required. A case where this might be needed would be performing novel customized real-time analysis on behavior data and activating a stimulus based on the result”

(2) The JSON messaging protocol lacks API documentation. It's not clear what the exact syntax is, supported key/value pairs, and expected response/behavior of the JSON messages. Hence, it's not clear how to develop new hardware that can communicate with the behaviorMate system.

The most common approach for adding novel hardware is to use TTL pulses (or accept an emitted TTL pulse to read sensor states). This type of hardware addition  is possible through the existing GPIO without the need to interact with the software or JSON API. Users looking to take advantage of the ability to set up and configure novel behavioral paradigms without the need to write any software would be limited to adding hardware which could be triggered with and report to the UI with a TTL pulse (however fairly complex actions could be triggered this way).

For users looking to develop more customized hardware solutions that interact closely with the UI or GPIO board, additional documentation on the JSON messaging protocol has been added to the behaviormate-utils repository (https://github.com/losonczylab/behaviormate_utils). Additionally, we have added a link to this repository in the Supplemental Materials section (line 971) and referenced this in the manuscript (line 217) to make it easier for readers to find this information.

Furthermore, developers looking to add completely novel components to the UI  can implement the interface described by Context.java in order to exchange custom messages with hardware. (described  in the JavaDoc: https://www.losonczylab.org/behaviorMate-1.0.0/)  These messages would be defined within the custom context and interact with the custom hardware (meaning the interested developer would make a novel addition to the messaging API). Additionally, it should be noted that without editing any software, any UDP packets sent to behaviorMate from an IP address specified in the settings will get time stamped and logged in the stored behavioral data file meaning that are a large variety of hardware implementation solutions using both standard UDP messaging and through TTL pulses that can work with behaviorMate with minimal effort. Finally, see response to R2.1 for a discussion of the JavaFX version of the behaviorMatee UI including plugin support.

(3) It seems the existing control hardware and the JSON messaging only support GPIO/TTL types of input/output, which limits the applicability of the system to more complicated sensor/controller hardware. The authors mentioned that hardware like Arduino natively supports serial protocols like I2C or SPI, but it's not clear how they are handled and translated to JSON messages.

We provide an implementation for an I2C-based capacitance lick detector which interested developers may wish to copy if support for novel I2C or SPI. Users with less development experience wishing to expand the hardware capabilities of  behaviorMatecould also develop adapters which can be triggered  on a TTL input/output. Additionally, more information about the JSON API and how messages are transmitted to the PC by the arduino is described in point (2) and the expanded online documentation.

a. Additionally, because it's unclear how easy to incorporate arbitrary hardware with behaviorMate, the "Intranet of things" approach seems to lose attraction. Since currently, the manuscript focuses mainly on a specific set of hardware designed for a specific type of experiment, it's not clear what are the advantages of implementing communication over a local network as opposed to the typical connections using USB.

As opposed to serial communication protocols as typical with USB, networking protocols seamlessly function based on asynchronous message passing. Messages may be routed internally (e.g. to a PCs localhost address, i.e. 0.0.0..0) or to a variety of external hardware (e.g. using IP addresses such as those in the range 192.168.1.2 - 192.168.1.254). Furthermore, network-based communication allows modules, such as VR, to be added easily. behavoirMate systems can be easily expanded using low-cost Ethernet switches and consume only a single network adapter on the PC (e.g. not limited by the number of physical USB ports). Furthermore, UDP message passing is implemented in almost all modern programming languages in a platform independent manner (meaning that the same software can run on OSX, Windows, and Linux). Lastly, as we have pointed out (Line 117) a variety of tools exist for inspecting network packets and debugging; meaning that it is possible to run behaviorMate with simulated hardware for testing and debugging.

The IOT nature of behaviorMate means there is no requirement for novel hardware to be implemented  using an arduino,  since any system capable of  UDP communication can  be configured. For example, VRMate is usually run on Odroid C4s, however one could easily create a system using Raspberry Pis or even additional PCs. behaviorMate is agnostic to the format of the UDP messages, but packaging any data in the JSON format for consistency would be encouraged. If a new hardware is a sensor that has input requiring it to be time stamped and logged then all that is needed is to add the IP address and port information to the ‘controllers’ list in a behaviorMate settings file. If more complex interactions are needed with novel hardware than a custom implementation of ContextList.java may be required (see response to R2.2). However, the provided UdpComms.java class could be used to easily send/receive messages from custom Context.java subclasses.

Solutions for highly customized hardware do require basic familiarity with object-oriented programming using the Java programming language. However, in our experience most behavioral experiments do not require these kinds of modifications. The majority of 1D navigation tasks, which behaviorMate is currently best suited to control, require touch/motion sensors, LEDs, speakers, or solenoid valves,  which are easily controlled by the existing GPIO implementation. It is unlikely that custom subclasses would even be needed.

Reviewer #3 (Public review):

(1) While using UDP for data transmission can enhance speed, it is thought that it lacks reliability. Are there error-checking mechanisms in place to ensure reliable communication, given its criticality alongside speed?

The provided GPIO/behavior controller implementation sends acknowledgement packets in response to all incoming messages as well as start and stop messages for contexts and “valves”. In this way the UI can update to reflect both requested state changes as well as when they actually happen (although there is rarely a perceptible gap between these two states unless something is unplugged or not functioning). See Line 85 in the revised manuscript “acknowledgement packets are used to ensure reliable message delivery to and from connected hardware”.

(2) Considering this year's price policy changes in Unity, could this impact the system's operations?

VRMate is not affected by the recent changes in pricing structure of the Unity project.

The existing compiled VRMate software does not need to be regenerated to update VR scenes, or implement new task logic (since this is handled by the behaviorMate GUI). Therefore, the VRMate program is robust to any future pricing changes or other restructuring of the Unity program and does not rely on continued support of Unity. Additionally, while the solution presented in VRMate has many benefits, a developer could easily adapt any open-source VR Maze project to receive the UDP-based position updates from behaviorMate or develop their own novel VR solutions.

(3) Also, does the Arduino offer sufficient precision for ephys recording, particularly with a 10ms check?

Electrophysiology recording hardware typically has additional I/O channels which can provide assistance with tracking behavior/synchronization at a high resolution. While behaviorMate could still be used to trigger reward valves, either the ephys hardware or some additional high-speed DAQ would be recommended to maintain accurately with high-speed physiology data. behaviorMate could still be set up as normal to provide closed and open-loop task control at behaviorally relevant timescales alongside a DAQ circuit recording events at a consistent temporal resolution. While this would increase the relative cost of the individual recording setup, identical rigs for training animals could still be configured without the DAQ circuit avoiding unnecessary cost and complexity.

(4) Could you clarify the purpose of the Sync Pulse? In line 291, it suggests additional cues (potentially represented by the Sync Pulse) are needed to align the treadmill screens, which appear to be directed towards the Real-Time computer. Given that event alignment occurs in the GPIO, the connection of the Sync Pulse to the Real-Time Controller in Figure 1 seems confusing.

A number of methods exist for synchronizing recording devices like microscopes or electrophysiology recordings with behaviorMate’s time-stamped logs of actuators and sensors. For example, the GPIO circuit can be configured to send sync triggers, or receive timing signals as input. Alternatively a dedicated circuit could record frame start signals and relay them to the PC to be logged independently of the GPIO (enabling a high-resolution post-hoc alignment of the time stamps). The optimal method to use varies based on the needs of the experiment. Our setups have a dedicated BNC output and specification in the settings file that sends a TTL pulse at the start of an experiment in order to trigger 2p imaging setups (see line 224, specifically that this is a detail of “our” 2p imaging setup). We provide this information as it might be useful suggesting how to have both behavior and physiology data start recording at the same time. We do not intend this to be the only solution for alignment. Figure 1 indicates an “optional” circuit for capturing a high speed sync pulse and providing time stamps back to the real time PC. This is another option that might be useful for certain setups (or especially for establishing benchmarks between behavior and physiology recordings). In our setup event alignment does not exclusively occur on the GPIO.

a. Additionally, why is there a separate circuit for the treadmill that connects to the UI computer instead of the GPIO? It might be beneficial to elaborate on the rationale behind this decision in line 260.

Event alignment does not occur on the GPIO, separating concerns between position tracking and more general input/output features which improves performance and simplifies debugging.  In this sense we maintain a single event loop on the Arduino, avoiding the need to either run multithreaded operations or rely extensively on interrupts which can cause unpredictable code execution (e.g. when multiple interrupts occur at the same time). Our position tracking circuit is therefore coupled to a separate,low-cost arduino mini which has the singular responsibility of position-tracking.

b. Moreover, should scenarios involving pupil and body camera recordings connect to the Analog input in the PCB or the real-time computer for optimal data handling and processing?

Pupil and body camera recordings would be independent data streams which can be recorded separately from behaviorMate. Aligning these forms of full motion video could require frame triggers which could be configured on the GPIO board using single TTL like outputs or by configuring a valve to be “pulsed” which is a provided type customization.

We also note that a more advanced developer could easily leverage camera signals to provide closed loop control by writing an independent module that sends UDP packets to behavoirMate. For example a separate computer vision based position tracking module could be written in any preferred language and use UDP messaging to send body tracking updates to the UI without editing any of the behaviorMate source code (and even used for updating 1D location).

(5) Given that all references, as far as I can see, come from the same lab, are there other labs capable of implementing this system at a similar optimal level?

To date two additional labs have published using behaviorMate, the Soltez and Henn labs (see revised lines 341-342). Since behaviorMate has only recently been published and made available open source, only external collaborators of the Losonczy lab have had access to the software and design files needed to do this. These collaborators did, however, set up their own behavioral setups in separate locations with minimal direct support from the authors–similar to what would be available to anyone seeking to set a behaviorMate system would find online on our github page or by posting to the message board.

Recommendations for the authors:

Reviewer #1 (Recommendations For The Authors):

(4) To provide additional context for the significance of this work, additional citations would be helpful to demonstrate a ubiquitous need for a system like behaviorMate. This was most needed in the paragraph from lines 46-65, specifically for each sentence after line 55, where the authors discuss existing variants on head-fixed behavioral paradigms. For instance, for the clause "but olfactory and auditory stimuli have also been utilized at regular virtual distance intervals to enrich the experience with more salient cues", suggested citations include Radvansky & Dombeck 2018 (DOI: 10.1038/s41467-018-03262-4), Fischler-Ruiz et al. 2021 (DOI: 10.1016/j.neuron.2021.09.055).

We thank the reviewer for the suggested missing citations and have updated the manuscript accordingly (see line 58).

(5) In addition, it would also be helpful to clarify behaviorMate's implementation in other laboratories. On line 304 the authors mention "other labs" but the following list of citations is almost exclusively from the Losonczy lab. Perhaps the citations just need to be split across the sentence for clarity? E.g. "has been validated by our experimental paradigms" (citation set 1) "and successfully implemented in other labs as well" (citation set 2).

We have split the citation set as suggested (see lines 338-342).

Minor Comments:

(6) In the paragraph starting line 153 and in Fig. 2, please clarify what is meant by "trial" vs. "experiment". In many navigational tasks, "trial" refers to an individual lap in the environment, but here "trial" seems to refer to the whole behavioral session (i.e. synonymous with "experiment"?).

In our software implementation we had originally used “trial” to refer to an imaging session rather than experiment (and have made updates to start moving to the more conventional lexicon). To avoid confusion we have remove this use of “trial” throughout the manuscript and replaced with “experiment” whenever possible

(7) This is very minor, but in Figure 3 and 4, I don't believe the gavage needle is actually shown in the image. This is likely to avoid clutter but might be confusing to some readers, so it may be helpful to have a small inset diagram showing how the needle would be mounted.

We assessed the image both with and without the gavage needle and found the version in the original (without) to be easier to read and less cluttered and therefore maintained that version in the manuscript.

(8) In Figure 5 legend, please list n for mice and cells.

We have updated the Figure 5 legend to indicate that for panels C-G, n=6 mice (all mice were recorded in both VR and TM systems), 3253 cells in VR classified as significantly tuned place cells VR, and 6101 tuned cells in TM,

(9) Line 414: It is not necessary to tilt the entire animal and running wheel as long as the head-bar clamp and objective can rotate to align the imaging window with the objective's plane of focus. Perhaps the authors can just clarify the availability of this option if users have a microscope with a rotatable objective/scan head.

We have added the suggested caveat to the manuscript in order to clarify when the goniometers might be useful (see lines 281-288).

(10) Figure S1 and S2 could be referenced explicitly in the main text with their related main figures.

We have added explicit references to figures S1 and S2 in the relevant sections (see lines 443, 460  and 570)

(11) On line 532-533, is there a citation for "proximal visual cues and tactile cues (which are speculated to be more salient than visual cues)"?

We have added citations to both Knierim & Rao 2003 and Renaudineau et al. 2007 which discuss the differential impact of proximal vs distal cues during navigation as well as Sofroniew et al. 2014 which describe how mice navigate more naturally in a tactile VR setup as opposed to purely visual ones.

(12) There is a typo at the end of the Figure 2 legend, where it should say "Arduino Mini."

This typo has been fixed.

Reviewer #2 (Recommendations For The Authors):

(4) As mentioned in the public review: what is the major advantage of taking the IoT approaches as opposed to USB connections to the host computer, especially when behaviorMate relies on a central master computer regardless? The authors mentioned the readability of the JSON messages, making the system easier to debug. However, the flip side of that is the efficiency of data transmission. Although the bandwidth/latency is usually more than enough for transmitting data and commands for behavior devices, the efficiency may become a problem when neural recording devices (imaging or electrophysiology) need to be included in the system.

behaviorMate is not intended to do everything, and is limited to mainly controlling behavior and providing some synchronizing TTL style triggers. In this way the system can easily and inexpensively be replicated across multiple recording setups; particularly this is useful for constructing additional animal training setups. The system is very much sufficient for capturing behavioral inputs at relevant timescales (see the benchmarks in Figures 3 and 4 as well as the position correlated neural activity in Figures 5 and 6 for demonstration of this). Additional hardware might be needed to align the behaviorMate output with neural data for example a high-speed DAQ or input channels on electrophysiology recording setups could be utilized (if provided). As all recording setups are different the ideal solution would depend on details which are hard to anticipate. We do not mean to convey that the full neural data would be transmitted to the behaviorMate system (especially using the JSON/UDP communications that behaviorMate relies on).

(5) The author mentioned labView. A popular open-source alternative is bonsai (https://github.com/bonsai-rx/bonsai). Both include a graphical-based programming interface that allows the users to easily reconfigure the hardware system, which behaviorMate seems to lack. Additionally, autopilot (https://github.com/auto-pi-lot/autopilot) is a very relevant project that utilizes a local network for multiple behavior devices but focuses more on P2P communication and rigorously defines the API/schema/communication protocols for devices to be compatible. I think it's important to include a discussion on how behaviorMate compares to previous works like these, especially what new features behaviorMate introduces.

We believe that behaviorMate provides a more opinionated and complete solution than the projects mentioned. A wide variety of 1D navigational paradigms can be constructed in behaviorMate without the need to write any novel software. For example, bonsai is a “visual programming language” and would require experimenters to construct a custom implementation of each of their experiments. We have opted to use Java for the UI with distributed computations across modules in various languages. Given the IOT methodology it would be possible to use any number of programming languages or APIs; a large number of design decisions were made  when building the project and we have opted to not include this level of detail in the manuscript in order to maintain readability. We strongly believe in using non-proprietary and open source projects, when possible, which is why the comparison with LabView based solutions was included in the introduction. Also, we have added a reference to the autopilot reference to the section of the introduction where this is discussed.

(6) One of the reasons labView/bonsai are popular is they are inherently parallel and can simultaneously respond to events from different hardware sources. While the JSON events in behaviorMate are asynchronous in nature, the handling of those events seems to happen only in a main event loop coupled with GUI, which is sequential by nature. Is there any multi-threading/multi-processing capability of behaviorMate? If so it's an important feature to highlight. If not I think it's important to discuss the potential limitation of the current implementation.

IOT solutions are inherently concurrent since the computation is distributed. Additional parallelism could be added by further distributing concerns between additional independent modules running on independent hardware. The UI has an eventloop which aggregates inputs and then updates contexts based on the current state of those inputs sequentially. This sort of a “snapshot” of the current state is necessary to reason about when the start certain contexts based on their settings and applied decorators. While the behaviorMate UI uses multithreading libraries in Java to be more performant in certain cases, the degree to which this represents true vs “virtual” concurrency would depend on the individual PC architecture it is run on and how the operating system allocates resources. For this reason, we have argued in the manuscript that behaviorMate is sufficient for controlling experiments at behaviorally relevant timescales, and have presented both benchmarks and discussed different synchronization approaches and permit users to determine if this is sufficient for their needs.

(7) The context list is an interesting and innovative approach to abstract behavior contingencies into a data structure, but it's not currently discussed in depth. I think it's worth highlighting how the context list can be used to cover a wide range of common behavior experimental contingencies with detailed examples (line 185 might be a good example to give). It's also important to discuss the limitation, as currently the context lists seem to only support contingencies based purely on space and time, without support for more complicated behavior metrics (e.g. deliver reward only after X% correct).

To access more complex behavior metrics during runtime, custom context list decorators would need to be implemented. While this is less common in the sort of 1D navigational behaviors the project was originally designed to control, adding novel decorators is a simple process that only requires basic object oriented programming knowledge. As discussed we are also implementing a plugin-architecture in the JavaFX update to streamline these types of additions.

Minor Comments:

(8) In line 202, the author suggests that a single TTL pulse is sent to mark the start of a recording session, and this is used to synchronize behavior data with imaging data later. In other words, there are no synchronization signals for every single sample/frame. This approach either assumes the behavior recording and imaging are running on the same clock or assumes evenly distributed recording samples over the whole recording period. Is this the case? If so, please include a discussion on limitations and alternative approaches supported by behaviorMate. If not, please clarify how exactly synchronization is done with one TTL pulse.

While the TTL pulse triggers the start of neural data in our setups, various options exist for controlling for the described clock drift across experiments and the appropriate one depends on the type of recordings made, frame rate duration of recording etc. Therefore behaviorMate leaves open many options for synchronization at different time scales (e.g. the adding a frame-sync circuit as shown in Figure 1 or sending TTL pulses to the same DAQ recording electrophysiology data).  Expanded consideration of different synchronization methods has been included in the manuscript (see lines 224-238).

(9) Is the computer vision-based calibration included as part of the GUI functionality? Please clarify. If it is part of the GUI, it's worth highlighting as a very useful feature.

The computer vision-based benchmarking is not included in the GUI. It is in the form of a script made specifically for this paper. However for treadmill-based experiments behaviorMate has other calibration tools built into it (see line 301-303).

(10) I went through the source code of the Arduino firmware, and it seems most "open X for Y duration" functions are implemented using the delay function. If this is indeed the case, it's generally a bad idea since delay completely pauses the execution and any events happening during the delay period may be missed. As an alternative, please consider approaches comparing timestamps or using interrupts.

We have avoided the use of interrupts on the GPIO due to the potential for unpredictable code execution. There is a delay which is only just executed if the duration is 10 ms or less as we cannot guarantee precision of the arduino eventloop cycling faster than this. Durations longer than 10 ms would be time stamped and non-blocking. We have adjusted this MAX_WAIT to be specified as a macro so it can be more easily adjusted (or set to 0).

(11) Figure 3 B, C, D, and Figure 4 D, E suffer from noticeable low resolution.

We have converted Figure 3B, C, D and 4C, D, E to vector graphics in order to improve the resolution.

(12) Figure 4C is missing, which is an important figure.

This figure appeared when we rendered and submitted the manuscript. We apologize if the figure was generated such that it did not load properly in all pdf viewers. The panel appears correctly in the online eLife version of the manuscript. Additionally, we have checked the revision in Preview on Mac OS as well as Adobe Acrobat and the built-in viewer in Chrome and all figure panels appear in each so we hope this issue has been resolved.

(13) There are thin white grid lines on all heatmaps. I don't think they are necessary.

The grid lines have been removed from the heatmaps  as suggested.

(14) Line 562 "sometimes devices directly communicate with each other for performance reasons", I didn't find any elaboration on the P2P communication in the main text. This is potentially worth highlighting as it's one of the advantages of taking the IoT approaches.

In our implementation it was not necessary to rely on P2P communication beyond what is indicated in Figure 1. The direct communication referred to in line 562 is meant only to refer to the examples expanded on in the rest of the paragraph i.e. the behavior controller may signal the microscope directly using a TTL signal without looping back to the UI. As necessary users could implement UDP message passing between devices, but this is outside the scope of what we present in the manuscript.

(15) Line 147 "Notably, due to the systems modular architecture, different UIs could be implemented in any programming language and swapped in without impacting the rest of the system.", this claim feels unsupported without a detailed discussion of how new code can be incorporated in the GUI (plugin system).

This comment refers to the idea of implementing “different UIs”. This would entail users desiring to take advantage of the JSON messaging API and the proposed electronics while fully implementing their own interface. In order to facilitate this option we have improved documentation of the messaging API posted in the README file accompanying the arduino source code. We have added reference to the supplemental materials where readers can find a link to the JSON API implementation to clarify this point.

Additionally, while a plugin system is available in the JavaFX version of behaviorMate, this project is currently under development and will update the online documentation as this project matures, but is unrelated to the intended claim about completely swapping out the UI.

Reviewer #3 (Recommendations For The Authors):

(6) Figure 1 - the terminology for each item is slightly different in the text and the figure. I think making the exact match can make it easier for the reader.

- Real-time computer (figure) vs real-time controller (ln88).

The manuscript was adjusted to match figure terminology.

- The position controller (ln565) - position tracking (Figure).

We have updated Figure 1 to highlight that the position controller does the position tracking.

- Maybe add a Behavior Controller next to the GPIO box in Figure 1.

We updated Figure 1 to highlight that the Behavior Controller performs the GPIO responsibility such that "Behavior Controller" and "GPIO circuit" may be used interchangeably.

- Position tracking (fig) and position controller (subtitle - ln209).

We updated Figure 1 to highlight that the position controller does the position tracking.

- Sync Pulse is not explained in the text.

The caption for Figure 1 has been updated to better explain the Sync pulse and additional systems boxes

(7) For Figure 3B/C: What is the number of data points? It would be nice to see the real population, possibly using a swarm plot instead of box plots. How likely are these outliers to occur?

In order to better characterize the distributions presented in our benchmarking data we have added mean and standard deviation information the plots 3 and 4. For Figure 3B: 0.0025 +/- 0.1128, Figure 3C: 12.9749 +/- 7.6581, Figure 4C: 66.0500 +/- 15.6994, Figure 4E: 4.1258 +/- 3.2558.

Author response:

The following is the authors’ response to the original reviews.

Reviewer 1:

We thank the reviewer for the time and effort in providing very useful comments and suggestions for our manuscript.

(1) The results do not support the conclusions. The main "selling point" as summarized in the title is that the apoptotic rate of zebrafish motorneurons during development is strikingly low (~2% ) as compared to the much higher estimate (~50%) by previous studies in other systems. The results used to support the conclusion are that only a small percentage (under 2%) of apoptotic cells were found over a large population at a variety of stages 24-120hpf. This is fundamentally flawed logic, as a short-time window measure of percentage cannot represent the percentage in the long term. For example, at any year under 1% of the human population dies, but over 100 years >99% of the starting group will have died. To find the real percentage of motorneurons that died, the motorneurons born at different times must be tracked over the long term or the new motorneuron birth rate must be estimated. A similar argument can be applied to the macrophage results. Here the authors probably want to discuss well-established mechanisms of apoptotic neuron clearance such as by glia and microglia cells.

We chose the time window of 24-120 hpf based on the following two reasons: 1) Previous studies showed that although the time windows of motor neuron death vary in chick (E5-E10), mouse (E11.5-E15.5), rat (E15-E18), and human (11-25 weeks of gestation), the common feature of these time windows is that they are all the developmental periods when motor neurons contact with muscle cells. The contact between zebrafish motor neurons and muscle cells occurs before 72 hpf, which is included in our observation time window of 24-120 hpf. 2) Zebrafish complete hatching during 48-72 hpf, and most organs form before 72 hpf. More importantly, zebrafish start swimming around 72 hpf, indicating that motor neurons are fully functional at 72 hpf. Thus, we are confident that this 24-120 hpf time window covers the time window during which motor neurons undergo programmed cell death during zebrafish early development. We have added this information to the revised manuscript.

We frequently used “early development” in this manuscript to describe our observation. However, we missed “early” in our title. We therefore have added this ket word of “early” in the title in the revised manuscript.

Previous studies in zebrafish have shown that the production of spinal cord motor neurons largely ceases before 48 hpf, and then the motor neurons remain largely constant until adulthood (doi: 10.1016/j.celrep.2015.09.050; 10.1016/j.devcel.2013.04.012; 10.1007/BF00304606; 10.3389/fcell.2021.640414). Our observation time window covers the major motor neuron production process. Therefore, we believe that neurogenesis will not affect our findings and conclusions.

We discussed the engulfment of dead motor neurons by other types of cells in the discussion section.

(2) The transgenic line is perhaps the most meaningful contribution to the field as the work stands. However, the mnx1 promoter is well known for its non-specific activation - while the images suggest the authors' line is good, motor neuron markers should be used to validate the line. This is especially important for assessing this population later as mnx1 may be turned off in mature neurons.

The mnx1 promoter has been widely used to label motor neurons in transgenic zebrafish. Previous studies have shown that most of the cells labeled in the mnx1 transgenic zebrafish are motor neurons. In this study, we observed that the neuronal cells in our sensor zebrafish formed green cell bodies inside of the spinal cord and extended to the muscle region, which is an important morphological feature of the motor neurons.

Reviewer 2:

We thank the reviewer for the time and effort in making very useful comments and suggestions for our manuscript.

The FRET-based programmed cell death biosensor described in this manuscript could be very useful. However, the authors have not considered what is already known about the development and programmed cell death of zebrafish spinal motor neurons, and potential differences between motor neuron populations innervating different types of muscles in different vertebrate models. Without this context, the application of their new biosensor tool does not provide new insights into zebrafish motor neuron programmed cell death. In addition, the authors have not carried out controls to show the efficacy and specificity of their morpholinos. Nor have they described how they counted dying motor neurons, or why they chose the specific developmental time points they addressed. These issues are addressed more specifically below.

(1) Lines 12-13: Previous studies in zebrafish showed death of identified spinal motor neurons.

Line 103: In Figure 2A the cell body in the middle is that of identified motor neuron VaP. VaP death has previously been described in several publications. The cell body on the right of the same panel appears to belong to an interneuron whose axon can be seen extending off to the left in one of the rostrocaudal axon bundles that traverse the spinal cord. Higher-resolution imaging would clarify this.

Lines 163-164: Is this the absolute number of motor neurons that died? How were the counts done? Were all the motor neurons in every segment counted? There are approximately 30 identifiable VaP motor neurons in each embryo and they have previously been reported to die between 24-36 hpf. So this analysis is likely capturing those cells.

Our study examined the overall motor neuron apoptosis rather than a specific type of motor neuron death, so we did not emphasize the death of VaP motor neurons. We agree that the dead motor neurons observed in our manuscript contain VaP motor neurons. However, there were also other types of dead motor neurons observed in our study. The reasons are as follows: 1) VaP primary motor neurons die before 36 hpf, but our study found motor neuron cells died after 36 hpf and even at 84 hpf (revised Figure 4A). 2) The position of the VaP motor neuron is together with that of the CaP motor neuron, that is, at the caudal region of the motor neuron cluster. Although it’s rare, we did observe the death of motor neurons in the rostral region of the motor neuron cluster (revised Figure 2C). 3) There is only one or zero VaP motor neuron in each motor neuron cluster. Although our data showed that usually one motor neuron died in each motor neuron cluster, we did observe that sometimes more than one motor neuron died in the motor neuron cluster (revised Figure 2C). We included this information in the revised discussion.

(2) Lines 82-83: It is published that mnx1 is expressed in at least one type of spinal interneuron derived from the same embryonic domain as motor neurons.

The mnx1 promoter has been widely used to label motor neurons in transgenic zebrafish. Previous studies have shown that most of the cells labeled in the mnx1 transgenic zebrafish are motor neurons. In this study, we observed that the neuronal cells in our sensor zebrafish formed green cell bodies inside of the spinal cord and extended to the muscle region, which is an important morphological feature of the motor neurons.

Furthermore, a few of those green cell bodies turned into blue apoptotic bodies inside the spinal cord and changed to blue axons in the muscle regions at the same time, which strongly suggests that those apoptotic neurons are not interneurons. Although the mnx1 promoter might have labeled some interneurons, this will not affect our major finding that only a small portion of motor neurons died during zebrafish early development.

(3) Lines 161-162: Although this may be the major time window of neurogenesis, there are many more motor neurons in adults than in larvae. Neither of these references describes the increase in motor neuron numbers over this particular time span, so the rationale for this choice is unclear.

Lines 168-171: It is known that later developing motor neurons are still being generated in the spinal cord at this time, suggesting that if there is a period of programmed cell death similar to that described in chick and mouse, it would likely occur later. In addition, most of the chick and mouse studies were performed on limb-innervating motor neurons, rather than the body wall muscle-innervating motor neurons examined here.

Lines 237-238: Especially since new motor neurons are still being generated at this time.

Previous studies have shown that the production of spinal cord motor neurons largely ceases before 48 hpf in zebrafish, and then the motor neurons remain largely constant until the adulthood (doi: 10.1016/j.celrep.2015.09.050; 10.1016/j.devcel.2013.04.012; 10.1007/BF00304606; 10.3389/fcell.2021.640414). Our observation time window covers the major motor neuron production process. Therefore, we believe that neurogenesis will not affect our data and conclusions.

The death of motor neurons in limb-innervating motor neurons has been extensively studied in chicks and rodents, as it is easy to undergo operations such as amputation. However, previous studies have shown this dramatic motor neuron death does not only occur in limb-innervating motor neurons but also occurs in other spinal cord motor neurons (doi: 10.1006/dbio.1999.9413). In our manuscript, we studied the naturally occurring motor neuron death in the whole spinal cord during the early stage of zebrafish development.

(4) Lines 184-187: Previous publications showed that death of VaP is independent of limitations in muscle innervation area, suggesting it is not coupled to muscle-derived neurotrophic factors.

Lines 328-334: There have been many publications describing appropriate morpholino controls. The authors need to describe their controls and show that they know that the genes they were targeting were downregulated.

For the morpholinos, we did not confirm the downregulation of the target genes. These morpholino-related data are a minor part of our manuscript and shall not affect our major findings. We have removed the neurotrophic factors and morpholino-related data in the revised manuscript.

Reviewer #1 (Public review):

Summary:

The authors of this study set out to find RNA binding proteins in the CNS in cell-type specific sequencing data and discover that the cardiomyopathy-associated protein RBM20 is selectively expressed in olfactory bulb glutamatergic neurons and PV+ GABAergic neurons. They make an HA-tagged RBM20 allele to perform CLIP-seq to identify RBM20 binding sites and find direct targets of RBM20 in olfactory bulb glutmatergic neurons. In these neurons, RBM20 binds intronic regions. RBM20 has previously been implicated in splicing, but when they selectively knockout RBM20 in glutamatergic neurons they do not see changes in splicing, but they do see changes in RNA abundance, especially of long genes with many introns, which are enriched for synapse-associated functions. These data show that RBM20 has important functions in gene regulation in neurons, which was previously unknown, and they suggest it acts through a mechanism distinct from what has been studied before in cardiomyocytes.

Strengths:

The study finds expression of the cardiomyopathy-associated RNA binding protein RBM20 in specific neurons in the brain, opening new windows into its potential functions there.

The study uses CLIP-seq to identify RBM20 binding RNAs in olfactory bulb neurons.

Conditional knockout of RBM20 in glutamatergic or PV neurons allows the authors to detect mRNA expression that is regulated by RBM20.

The data include substantial controls and quality control information to support the rigor of the findings.

Weaknesses:

The authors do not fully identify the mechanism by which RBM20 acts to regulate RNA expression in neurons, though they do provide data suggesting that neuronal RBM20 does not regulate alternate splicing in neurons, which is an interesting contrast to its proposed mechanism of function in cardiomyocytes. Discovery of the RNA regulatory functions of RBM20 in neurons is left as a question for future studies.

The study does not identify functional consequences of the RNA changes in the conditional knockout cells, so this is also a question for the future.

do.how - turn on blank screen saver

• turn on screen saver blank

c:\Windows\System32>scrnsave.scr /s

Reviewer #3 (Public review):

Summary:

The study demonstrates the effectiveness of a cost-effective closed-loop feedback system for modulating brain activity and behavior in head-fixed mice. Authors have tested real-time closed-loop feedback system in head-fixed mice two types of graded feedback: 1) Closed-loop neurofeedback (CLNF), where feedback is derived from neuronal activity (calcium imaging), and 2) Closed-loop movement feedback (CLMF), where feedback is based on observed body movement. It is a python based opensource system, and authors call it CLoPy. The authors also claim to provide all software, hardware schematics, and protocols to adapt it to various experimental scenarios. This system is capable and can be adapted for a wide use case scenario.

Authors have shown that their system can control both positive (water drop) and negative reinforcement (buzzer-vibrator). This study also shows that using the close loop system mice have shown better performance, learnt arbitrary task and can adapt to change in the rule as well. By integrating real-time feedback based on cortical GCaMP imaging and behavior tracking authors have provided strong evidence that such closed-loop systems can be instrumental in exploring the dynamic interplay between brain activity and behavior.

Strengths:

Simplicity of feedback systems designed. Simplicity of implementation and potential adoption.

Weaknesses:

Long latencies, due to slow Ca2+ dynamics and slow imaging (15 FPS), may limit the application of the system.

Major comments:

(1) Page 5 paragraph 1: "We tested our CLNF system on Raspberry Pi for its compactness, general-purpose input/output (GPIO) programmability, and wide community support, while the CLMF system was tested on an Nvidia Jetson GPU device." Can these programs and hardware be integrated with windows-based system and a microcontroller (Arduino/ Tency). As for the broad adaptability that's what a lot of labs would already have (please comment/discuss)?

(2) Hardware Constraints: The reliance on Raspberry Pi and Nvidia Jetson (is expensive) for real-time processing could introduce latency issues (~63 ms for CLNF and ~67 ms for CLMF). This latency might limit precision for faster or more complex behaviors, which authors should discuss in the discussion section.

(3) Neurofeedback Specificity: The task focuses on mesoscale imaging and ignores finer spatiotemporal details. Sub-second events might be significant in more nuanced behaviors. Can this be discussed in the discussion section?

(4) The activity over 6s is being averaged to determine if the threshold is being crossed before the reward is delivered. This is a rather long duration of time during which the mice may be exhibiting stereotyped behaviors that may result in the changes in DFF that are being observed. It would be interesting for the authors to compare (if data is available) the behavior of the mice in trials where they successfully crossed the threshold for reward delivery and in those trials where the threshold was not breached. How is this different from spontaneous behavior and behaviors exhibited when they are performing the test with CLNF?