Despite the clear advantages of using a pre-trained Generalist agent over naive online learning, we still sought to determine if it was approaching the maximum possible performance of our control task. To this end, we compared the Generalist against two heuristic controls (Protocol 2, Fig. 4a), each deliberately depriving a key aspect of the agent’s decision-making. The “Without Best” policy restricted stimulation to all electrodes except the one empirically identified as highest yielding for that network, removing the agent’s ability to exploit its single most successful action. The “Random” policy applied spatially randomized stimulation at the same overall stimulation rate (stims/min) as the Generalist, removing any learned spatial selectivity while preserving the total stimulation budget. Together, these controls probe the Generalist from two directions: whether its advantage stems from knowing which electrode to stimulate, or from the overall rate of stimulation.

Wouldn't you wanna show: - Only Best, as the ceiling for any static policy. - UCB or Thompson sampling online from scratch, as a fair online-learning baseline. - Sliding-window or discounted UCB, which is built for nonstationary rewards and is the natural online method if adaptation is the issue. - Generalist plus online fine-tuning, which uses the pretrained prior and adapts to each culture.

Before concluding one approach is better over the other/proceeding in a direction/ choosing to test the max perf. Likely a generalist + OFT would do better than Generalist. Though all speculation because you need a proper baseline pre-decision / report the results. Alternatively rationale but like "We still still sought to determine" isn't really a logical reason significantly divergent from "We felt like testing this," which is un-principled. Then again.... The controls can't answer the question they're framed around. The stated goal is to find out whether the Generalist approaches maximum possible performance. That's a ceiling question. "Without Best" and "Random," though, are ablations: they take capabilities away and show that performance drops. That tells you which components matter (spatial selectivity, access to the best electrode). It says nothing about the gap between the Generalist and optimal. You can beat every degraded baseline and still be far from the ceiling. To estimate headroom you need stronger comparators. The authors just show "outperforms heuristic controls" which isn't saying much.

To resolve this conflict between theoretical framing and practical latency, we used an adapted version of the Proximal Policy Optimization (PPO) algorithm. This choice offers two main advantages. First, PPO learns a reactive policy capable of low-latency inference (single forward pass), effectively solving the contextual bandit challenge within the strict timing constraints of the biological loop. Second, unlike specialized bandit algorithms, PPO provides a unified framework that naturally generalizes beyond the greedy setting: it allows us to solve the immediate bandit problem (discount factor γ = 0) while retaining the architectural flexibility to incorporate long-horizon credit assignment (γ > 0) should the control objectives evolve.

Wouldn't TRPO give you better results?? TRPO enforces a hard KL-divergence constraint using a natural-gradient step with a line search. PPO's clipping is a heuristic that doesn't strictly bound the KL, and Engstrom showed much of PPO's stability comes from implementation details rather than the clip itself. That could matter when a bad policy update means a bad stimulation on a live biological sample.

LitL proceeds as an active learning loop over multiple design rounds described above. In the first round, minimal labeled data is available. In each successive round, we perform multi-objective Bayesian optimization: labeled data is collected and used to retrain the generative models and the predictive multi-task model,

I'd be curious about your retraining procedure. How exactly are you measuring divergence from prior model and quantifying expected improvement, and does that improvement genuinely translate into superior wet lab outcome.

maximizing the Noisy Expected Hypervolume Improvement (NEHVI) acquisition function

Why this choice of acq function? Wouldn't qParEGO or TSEMO give you similar results with less overhead cost?

2.1 Lab-in-the-loop is a general machine learning system for autonomous drug discovery LitL is a new approach that orchestrates loosely coupled ensembles of generative models, property prediction oracles, a ranking and selection algorithm, and experimental assays, all encompassed in an active learning loop to achieve iterative optimization of antibody lead candidates

Interesting idea

Therapeutic antibody design is a complex multi-property optimization problem with substantial promise for improvement with the application of machine-learning methods. Towards realizing that promise, we introduce “Lab-in-the-loop,”

Interesting premise

1. Beyond cell annotation and spatial niche characterization, recovering unmeasured456proteins represents another important downstream application in spatial proteomics.

How exactly would the authors go about tackling this given the approach presented in the manuscript?

1. vely, these analyses identify a reproducible collagen-rich ECM-remodeling449niche that can be robustly transferred across independent cohorts and is preferentially450enriched in recurrent colorectal cancer

Important application

1. The resulting embedding space exhibited well-organized continuous manifolds,

Surprising to see a continuous manifold.

How exactly would the authors go about tackling this given the approach presented in the manuscript?

Important application

Surprising to see a continuous manifold.

1. dding variants, colored by cancer type. b, Image-to-DNAmretrieval across GBM, LGG, HNSC, and BRCA using matched slide and methylation embeddings,shown as Recall@1, Recall@5, and Recall@10. These analyses use pre-trained embeddings beforesupervised downstream task adaptation, providing a direct

In figure 2(a) why are the Gastric points so disparate/spread out for HistoMethylGigaPath? Same for Lung ADC?

I think either this is a t-SNE artifact or implies something abt the model.

Then again for 2(a) one could argue with the four panels => four separate times get embeddings => four independent optimizations with their own initialization, seed, and convergence. I wouldn't expect there to be a shared coordinate system. Even the same embedding re-run at a different seed will redistribute cluster compactness.... If so though the figure doesn't imply much at all / is lowkey useless.

1. toMethyl overview. a, Cancer cohorts from TCGA [14] used for HistoMethyl pre-training.b, Pre-training inputs, including WSIs divided into tiles and paired DNAm beta values correspondingto each case. c, Pre-training with contrastive alignment. WSI and DNAm embeddings are extractedby their respective encoders, projected by MLPs, and aligned with a contrastive loss. d, Downstreamtask adaptation. HistoMethyl uses task-specific supervised models to adapt aligned slide embeddingsto downstream tasks, including gene mutation, morphology, survival analysis, and DNAm pr

Methodologically its really smart to split the WSI branch from the DNAm branch and then use a contrastive alignment on the embeddings. What does the joint space look like? Any structure that falls out? Have the authors investigated any patterns in the joint embedding space? I would expect some rather clear boundaries given the inputs.

In figure 2(a) why are the Gastric points so disparate/spread out for HistoMethylGigaPath? Same for Lung ADC?

I think either this is a t-SNE artifact or implies something abt the model.

Then again for 2(a) one could argue with the four panels => four separate times get embeddings => four independent optimizations with their own initialization, seed, and convergence. I wouldn't expect there to be a shared coordinate system. Even the same embedding re-run at a different seed will redistribute cluster compactness.... If so though the figure doesn't imply much at all / is lowkey useless.

Important contribution!

Methodologically its really smart to split the WSI branch from the DNAm branch and then use a contrastive alignment on the embeddings. What does the joint space look like? Any structure that falls out? Have the authors investigated any patterns in the joint embedding space? I would expect some rather clear boundaries given the inputs.

1. As the taxonomy progresses from broad, coarse-grained catalytic classes (EC .-.-.-)down to highly specific, four-digit reaction profiles (EC x.x.x.x) and targeted families, the challenge of filtering out false positives among closely related sequences scales exponentially. Throughout this entire taxonomic gradient, Boltz2ESI consistently exhibited superior discriminative resolution compared to baseline approaches. Notably, the sustained precision at the deep EC x.x.x.x level indicates that by capturing the adaptive biophysical microenvironment within the active site, our framework effectively untangles tight sub-family specificity that remains hidden to one-dimensional sequencemetrics or rigid-scaffold modeling

Performance claims here are uninterpretable without the evaluation design. How are negatives sampled at each EC depth?

1. first localizes the active-sitepocket using Multiple sequence alignment (MSA) guided co-folding, and subsequentlyre-folds the active site and substrate in an MSA-free regime to capture ligand-inducedside-chain and backbone adaptations

Stage 2 being MSA-free is presumably to avoid the consensus/apo bias that deep MSAs impose on side-chain placement. But does stage 2 condition on stage-1 coordinates? If so, evolutionary information persists as a geometric prior, and the ablation that matters is MSA-free stage 2 without the stage-1 pose, otherwise you can't attribute the induced-fit gains to the MSA-free regime.

1. From this simulated complex, the framework extracts interaction repre-sentations and a predicted structure to capture local biophysical constraints. Toreconcile this local structure with global biological context, Boltz2ESI fuses these geo-metric descriptors with residue-level evolutionary context derived from ESM3 [28],alongside geometry-aware molecular embeddings [29] and topological fingerprints

Four feature streams, no ablation. Do the topological fingerprints add anything over the molecular embeddings; both substrate-side, presumably overlapping Id guess? And do the geometric descriptors actually survive/lead to signal gain once ESM3 features are present? If ESM3's structure track is used, those two streams overlap by construction, since both descend from the stage-1 co-folded pose; that's also a leakage concern. Separately: stage 1 is MSA-guided, so the pipeline draws on evolutionary information twice. What does the PLM contribute then?

lol, four feature streams, no ablation. Do the topological fingerprints add anything over the molecular embeddings; both substrate-side, presumably overlapping Id guess? And do the geometric descriptors actually survive/lead to signal gain once ESM3 features are present? If ESM3's structure track is used, those two streams overlap by construction, since both descend from the stage-1 co-folded pose; that's also a leakage concern. Separately: stage 1 is MSA-guided, so the pipeline draws on evolutionary information twice. What does the PLM contribute then?

Performance claims here are uninterpretable without the evaluation design. How are negatives sampled at each EC depth?

Beautiful figure!

This is true. Good context!

Stage 2 being MSA-free is presumably to avoid the consensus/apo bias that deep MSAs impose on side-chain placement. But does stage 2 condition on stage-1 coordinates? If so, evolutionary information persists as a geometric prior, and the ablation that matters is MSA-free stage 2 without the stage-1 pose, otherwise you can't attribute the induced-fit gains to the MSA-free regime.

Interesting idea / choice of FM.

Well supported. Good job!

With what certainty/confidence? I see the blue/red shadows in 3B but no numerical bound.

Good background tbh!

Classic problem!

bro. In Dataset 1, they have 16 panels but the leave-one-panel-out drop is <8%. That's the better evidence for robustness.

Then you claim pretraining drives "robust representations despite marker inconsistency" based on the KO task, where Dataset 2 has a completely consistent panel.

Those two claims aren't using the same evidence base and shouldn't be merged into one conclusion.

If you're concluding this based on per-class AUC and markers are inconsistent isn't this claim sketchy at best?

Oh just claiming potential. ok nvm!

So you have 14k samples for dataset 1, a slight imbalance in male/female, but only 72 samples for dataset 2 because of selecting only 5 knockout lines? Also "most knockouts in this dataset were found to lack discernible cellular phenotypes"? Is that not concerning if you want to claim general ability/can build on for flow cytometry?

Pre-training distribution has a significant impact on downstream utility.

What exactly do the pre-training distributions look like? Whats the exact mix? Is dataset 1 sufficiently different from dataset 2, specifically as it relates to sample quality and number of samples?

Interesting concept to make them cell level. Why not clusters of cells?

Why make k=7?

Is within-method exact agreement really the best metric? Recommending to not run against a single stochastic agent is fine but what is the delta? Running many costs more for what benefit?

How does temperature/nucleus sampling/effort affect these results? Did you control for potential variation in these parameters?

Great experiment! I wonder what metrics are reported and how representative/relevant those metrics are given real life tasks.

Absolutely!! A challenge of great importance.

Downsampling/excluding a minority class is an interesting decision. Why not include some of them, use some flavor of stratified sampling/curriculum learning/train a specialized subnetwork/set of heads on the larger sequences with a reasonable split? How does the model generalize/perform on larger sequences?

This is interesting. Evolutionary context can be really informative?

Nice breakdown tbh!

To what extent is this accurate / aligned with biological reality? Does generating natural language answers introduce a source of error/confounding? What happens as answers become shorter vs longer vs less/more complex?

To what extent has this been solved by better molecular representations? Proteins and ligands are still molecules, and wouldn't atom-level representations ensure consistency across these data types? Boltz/BoltzGen does leverage atom-level information...

Why the Llama3 tokenizer among all other choices? Seems odd methodologically? Why not something designed for this kind of purpose? https://arxiv.org/html/2409.15370v1

Other objectives could have been chosen beyond ease of implementation, why in particular this objective? Why not a hybrid objective? what abt retrieved neighbor training?

This is an interesting idea. I wonder what the scaling looks like and what the efficacy of the augmentation is with respect to context window size and the quality of retrieval.

This is an interesting idea. I wonder what the scaling looks like and what the efficacy of the augmentation is with respect to context window size and the quality of retrieval.