The metaweb embeds ecological hypotheses and practices
At the end of the introduction, we mention that the family of machine learning tools - specifying graph embeddings and transfer learning. However, we do not really explicitly talk about transfer learning. I think that maybe when we talk about phylogenetic relatedness, we can potentially talk about the parallel idea of transfer learning in ML which is common practice.
sstructural
structural
Poisot et al. (2021)
I'd add Dalla Riva & Stouffer (2016) too here
posess
possess
Mora et al., 2018
Bramon Mora et al., 2018 (Bernat has 2 last names)
it is possible to establish a list of species in a region of interest; but establishing the interactions between these species is difficult.
by "establish" - do you mean extract a list of species? and by 'establishing the interactions', do you mean estimating/inferring the potential interactions? If so, I think using the specific verbs might make it a bit clearer to the reader.
Here, we hypothesise the presence, in the cells of patients with ME/CFS, of a large number of slightly differentially methylated genes (DMGs) potentially missed by studies with smaller sample sizes.
weird phrasing - not sure I understood what you meant exactly.
This is especially true of the investigations of cytokines, which often have conflicting results (Blundell et al., 2015; Corbitt et al., 2019)—cytokines function in an autocrine and paracrine manner, so their levels in the blood do not necessarily represent levels of inflammation elsewhere (Vanelzakker et al., 2019)
I guess that depending on where you are planning to submit this paper, I think it might be worth trying to walk the non-specialist reader through. Not having the domain knowledge it made it hard for me to necessarily follow. For instance, what does the term "autocrine" and "paracrine" mean? Do you mean that the cytokines indicates inflammation at specific locations only rather than being a global measure?
pathophysiology
it might be helpful to define this term
(ME/CFS)
Minor - but either change to (CFS/MF) based on the order that you wrote
inherit
inherent?
sharp and soft local constraints
This appears a bit abstract to me - what are the constraints?
configuration model
It is the first time this term appears, so I would suggest to move the definition that you provided later in the paragraph here.
–
run on sentence. I would suggest to remove the dash and split this into a new sentence.
Determinine
Determining?
Because the confidence intervals on the inferred trait space are probably over-estimates, we decided to apply a thresholding step to the interactions after the data inflation fig. 5. Cirtwill and Hambäck (2021) highlight a number of strategies to threshold probabilistic networks. Their methods assume the underlying data to be tag-based sequencing, which represents interactions as co-occurrences of predator and prey within the same tags; this is conceptually identical to our Bernoulli-trial based reconstruction of a probabilistic network. We performed a full analysis of the effect of various cutoffs, and as they either resulted in removing too few interactions, or removing enough interactions that species started to be disconnected from the network, we set this threshold for a probability equivalent to 0 to the largest possible value that still allowed all species to have at least one interaction with a non-zero probability.
This part is a bit hard to follow. I would suggest to first explain more generally how does the 'thresholding step' work - as in how does this step reduce overestimation? Then proceed by explaining the ins and outs of the thresholding step. Did you specifically use the same method as in Cirtwill and Hambäch (2021)? This part was unclear to me.
use-case
case study
and most importantly, retain this signal even when it is not detectable at the community scale
I would suggest to break it into another sentence as it reads as a run on sentence. This might make it harder for the reader to grasp the sense of the sentence.
signal
signals? or add 'a' prior to phylogenetic signal - if you meant one signal.
encompasses
encompass
Dormann et al. (2010) previously gave very similar conclusions.
If both Gerhold et al. (2015) and Dormann et al. (2010)'s studies have similar conclusions, why not merge them into one sentence.
nevertheless,
"Nevertheless" might not be the best connector here. You could replace it with either "in particular" or "more specifically"
are
should be "is"
such
remove
Alternatively, REF showed that the profile likelihood (assuming a known distribution of Σ, or a normal distribution otherwise) peaks at the rank maximizing the signal-to-noise ratio.
Do you both methods to find the optimal rank and pick the Zhu and Ghodsi profile likeiihood method - i.e. the sharp drop in the scree plot as the one? If you only used one - I would suggest to remove this bit to avoid confusion, or maybe mention that there are other methods to find the optimal rank as a side note only.
truncated Singular Value Decomposition
Perhaps rather than using the term 'truncated Singular Value Decomposition', we could use the SVD term here? Since the following sentences give further explanation about the rank. So maybe we could talk about the truncated SVD after? Unless it was the truncated SVD which was used rather than the SVD?
Specifically, we use Random Dot Product Graphs (RDPG; Young and Scheinerman 2007) to create a number of latent variables that can be combined into an approximation of the network adjacency matrix.
I think that for people not familiar to RDPGs, they might get a bit confused here with the RDPG and the SVD. Maybe you could give a generic definition of the model - it could be something like/ something along those lines "following the RDPG model, the probability of observing interactions between nodes is a function of the nodes' latent features...". By doing so, it might make the next pqrqgrqphs easier to follow.
Specifically, our approach can be summarized as follows: from the known network in Europe, we use a truncated Singular Values Decomposition (t-SVD; Halko, Martinsson, and Tropp 2011) to generate latent traits representing a low-dimensional embedding of the network; these traits give an unbiased estimate of the node’s position in the latent feature spaces. Then, we map these latent traits onto a reference phylogeny (other distance-based measures of species proximity that allow for the inference of features in the latent space can be used, such as for example dissimilarity in functional traits). Based on the reconstructed latent traits for species in the destination species pool, a Random Dot Product Graph (RDPG) model predicts the interaction between species through a function of the nodes’ features. Thus, from latent traits and nodes position, we can infer interactions.
Similar to how you describe the RDPG, then the truncated SVD in the paragraphs below, I would suggest to keep the order the same to make it easier to follow.
(other distance-based measures of species proximity that allow for the inference of features in the latent space can be used, such as for example dissimilarity in functional traits)
Since you already mention above in the data description section having a reference phylogeny, which implies that you are describing the method section in that particular context. One suggestion could be to have a paragraph where you put the emphasis on the fact that this is generalisable framework - either towards the end of your methods section or in the discussion. In that section you could present other alternative measures/method which could be used if other data were to be available.
This being said, the methods at each step may be substituted by others when the data support it.
I would suggest to move this sentence towards the end of the method section (mentioning that you'll get back to that in the discussion) so that it does not distract readers from to follow the different steps of your method section.
metaweb
I think it might be useful to define metaweb to make it easier for the readers to follow