Suggestions for improving the network chapters 18 and 19

[nuorenarra]

The network chapters present a nicely done framework for analysis on microbial covariation networks. However, I have some critique about the construction and presentation of these networks as well as the breadth of these chapters. First of all, my main critique is about presenting correlations among taxa as evidence for ecological interactions. Secondly, I think that abundance -weighted covariation based on relative abundances, which is currently used as the edge definition in the core example here, is biologically hard to interpret and might also introduce some technical biases in the data. This latter I think might both create a biased measure (as correlation between abundances will always be highest in taxa with most average abundances) and biologically meaningless (as two taxa having similar relative abundances across samples tells very little about their relationship with each other). Lastly, I think the current network chapters are much more detailed and require a higher understanding than the rest of the book.

Taken these points, I suggest the chapters could be greatly improved by
a) Providing some text in the beginning to explain what covariation means in terms of ecological associations between taxa, and having caution with terms such as “ecological interactions” or “ecological networks”. Currently the method in the chapter can be interpreted in a way where positive covariance meaning positive ecological interactions (such as cross-feeding) and negative covariance meaning negative ecological interactions (such as competition). Importantly, the cases where covariance implies ecological interactions are rare and most often covariance may imply that two taxa are selected by shared environmental preferences/filters. For example, two taxa that have very different ecological niche (and are thus unlikely to ever compete of same resources) will generally occur in different kind of communities, which will result in negative covariance, not indicative of competition.

b) Revising the main method for microbial network edge construction to be a presence-absence based co-occurrence, rather than abundance weighted covariation. This will make the edges biologically interpretable and these co-occurrence networks are also more widely used in the field and will be more familiar for the average student. Importantly, using presence-absence based edges would also make the network centrality metrics more interpretable. For example, high degree or betweenness of a node in a presence-absence based microbial co-occurrence network is easy to interpret as a taxon that is observed present in diverse communities or many different types of microbial communities, which can be a nice proxy for a generalist lifestyle and a broad ecological niche. Similarly, transitivity of a node in co-occurrence network can be used as a proxy for specialist lifestyle and narrow ecological niche. However, high degree/strength in an abundance-weighted covariation network is harder to interpret: It can be seen as a taxon that is often observed with many others in similar (relative) abundances across samples. This is biologically less informative/meaningful, especially because having a similar relative abundance with many others is generally highest among taxa that have the most average abundance. Furthermore, in the rare cases where people have absolute abundance data for their microbes, these networks and any network metrics derived from them would drastically change biological meaning.

c) Shortening them and making the overall more simple (such that an average experimental biology grad student can easily understand) to fit the book’s style and aims better.

Lastly, the method of "signed" dissimilarity in the network construction presented here in very non-standard in the field. The paper by Dongen & Enright that is cited for this method is from 2012 but has not been peer-reviewed or published. Taken these, I suggest the chapters would start with a standard "signed " network – where edges have a sign (positive or negative) in addition to weights that denote association strength (e.g., frequency of co-occurrence). Then any abundance-based edge definitions and different dissimilarity methods could added in the end as extra material.

[TuomasBorman]

Thanks @nuorenarra for this throughout review. I'm not an expert in this area, but I also feel that these chapters currently cover more advanced topics than the rest of the book (which shouldn't be the case). At least, there should be easier start for readers, starting from simpler approaches.

The material on networks is already very nice, but these suggestions would help make it more approachable for readers who aren't experts in the field. What do you think @stefpeschel ?

[TuomasBorman]

Bumping again. @stefpeschel

[stefpeschel]

Hi @nuorenarra!

Please excuse the late reply. I appreciate your comments but would like to clarify a few points.

First, regarding the terminology of ecological interactions: I fully agree that it’s important to be precise about what network edges represent. In both the book chapters and the NetCoMi package, we are very careful not to refer to these edges as interactions. Instead, we consistently speak of associations, as the methods estimate statistical relationships between taxa - not direct ecological interactions in the mechanistic or causal sense. If any misleading use of the term "interaction" slipped through, please let me know so I can correct it.

Second, I'd like to clarify that our methods are not based on relative abundances, but on clr-transformed data or other appropriate transformations, depending on the chosen association method. The compositional nature of microbiome data is explicitly taken into account through these transformations and normalization steps, which are explained both in the NetCoMi paper and the book chapters. The concern you raised - that correlation structures could be driven by the most abundant taxa - is exactly the kind of issue we address with these preprocessing steps. This is a key motivation for using methods tailored to compositional data.

Regarding the suggestion to prefer presence-absence based networks: while such approaches have their place, the network estimation methods we present are designed to capture quantitative associations, resulting in weighted networks. Discarding edge weights from the start would mean losing potentially important information. Users who prefer simpler metrics or are concerned about interpretability can still analyze these networks using unweighted network measures, which are supported in NetCoMi.

Concerning your comment about the signed dissimilarity approach, you're right that this method is not yet widely established in the field, and we are aware that the referenced paper by Dongen & Enright has not been peer-reviewed. However, we carefully evaluated the proposed transformations and found them to be well-suited to our goals, offering exactly the flexibility we were looking for. Specifically, they allow users to decide whether positive and negative associations should receive similar weight, or whether to emphasize positive associations while downweighting negative ones - an important modeling decision that is often overlooked or not sufficiently addressed in the literature. In many cases, negative associations are either ignored, converted to positive values via absolute value, or not discussed at all. We believe it is essential to make this step explicit and transparent, rather than relying on implicit defaults. For this reason, we present and explain the available transformation options in both the NetCoMi package and the OMA chapters, so users can make informed choices depending on their specific research questions and data.

Finally, about the chapter length: we are aware that the topic is complex and the chapters are more extensive than some others in the book. This is intentional. Microbiome network analysis involves many methodological choices, and we want to encourage users to engage critically with these options rather than applying a one-size-fits-all pipeline. Reducing the content to a minimal example would risk giving a misleading impression of simplicity, which we would like to avoid. Instead, we aim to provide accessible explanations and practical code that highlight the flexibility and considerations required when applying these methods to real data.

Thanks again for your constructive input. I'm always happy to discuss further suggestions on how to improve the clarity and usability of the chapters.

Best,
Stefanie