Leveraging genetic correlation structure to target discrete signaling mechanisms across metabolic tissues

Reviewer #1 (Public Review):

Zhou et al. have set up a study to examine how metabolism is regulated across the organism by taking a combined approach looking at gene expression in multiple tissues, as well as analysis of the blood. Specifically, they have created a tool for easily analyzing data from GTEx across 18 tissues in 310 people. In principle, this approach should be expandable to any dataset where multiple tissues of data were collected from the same individuals. While not necessary, it would also raise my interest to see the "Mouse(coming soon)" selection functional, given that the authors have good access to multi-tissue transcriptomics done in similarly large mouse cohorts.

Summary:

The authors have assembled a web tool that helps analyze multiple tissues' datasets together, with the aim of identifying how metabolic pathways and gene regulation are connected across tissues. This makes sense conceptually and the web tool is easy to use and runs reasonably quickly, considering the size of the data. I like the tool and I think the approach is necessary and surprisingly under-served; there is a lot of focus on multi-omics recently, but much less on doing a good job of integrating multi-tissue datasets even within a single omics layer.

What I am less convinced about is the "Research Article" aspect of this paper. Studying circadian rhythm in GTEx data seems risky to me, given the huge range in circadian clock in the sample collection. I also wonder (although this is not even remotely in my expertise) whether the circadian rhythm also gets rather desynchronized in people dying of natural causes - although I suppose this could be said for any gene expression pathway. Similarly for looking at secreted proteins in Figure 4 looking at muscle-hippocampus transcript levels for ADAMTS17 doesn't make sense to me - of all tissue pairs to make a vignette about to demonstrate the method, this is not an intuitive choice to me. The "within muscle" results look fine but panels C-E-G look like noise to me...especially panel C and G are almost certainly noise, since those are pathways with gene counts of 2 and 1 respectively.

I think this is an important effort and a good basis but a significant revision is necessary. This can devote more time and space to explaining the methodology and for ensuring that the results shown are actually significant. This could be done by checking a mix of negative controls (e.g. by shuffling gene labels and data) and a more comprehensive look at "positive" genes, so that it can be clearly shown that the genes shown in Fig 1 and 2 are not cherry-picked. For Figure 3, I suspect you would get almost an identical figure if instead of showing pan-tissue circadian clock correlations, you instead selected the electron transport chain, or the ribosome, or any other pathway that has genes that are expressed across all tissues. You show that colon and heart have relatively high connectivity to other tissues, but this may be common to other pathways as well.

Reviewer #2 (Public Review):

Summary:

Zhou et al. use publicly available GTEx data of 18 metabolic tissues from 310 individuals to explore gene expression correlation patterns within-tissue and across-tissues. They detect signatures of known metabolic signaling biology, such as ADIPOQ's role in fatty acid metabolism in adipose tissue. They also emphasize that their approach can help generate new hypotheses, such as the colon playing an important role in circadian clock maintenance. To aid researchers in querying their own genes of interest in metabolic tissues, they have developed an easy-to-use webtool (GD-CAT).

This study makes reasonable conclusions from its data, and the webtool would be useful to researchers focused on metabolic signaling. However, some misconceptions need to be corrected, as well as greater clarification of the methodology used.

Strengths:

GTEx is a very powerful resource for many areas of biomedicine, and this study represents a valid use of gene co-expression network methodology. The authors do a good job of providing examples confirming known signaling biology as well as the potential to discover promising signatures of novel biology for follow-up and future studies. The webtool, GD-CAT, is easy to use and allows researchers with genes and tissues of interest to perform the same analyses in the same GTEx data.

Weaknesses:

A key weakness of the paper is that this study does not involve genetic correlations, which is used in the title and throughout the manuscript, but rather gene co-expression networks. The authors do mention the classic limitation that correlation does not imply causation, but this caveat is even more important given that these are not genetic correlations. Given that the goal of their study aligns closely with multi-tissue WGCNA, which is not a new idea (e.g., Talukdar et al. 2016; https://doi.org/10.1016/j.cels.2016.02.002), it is surprising that the authors only use WGCNA for its robust correlation estimation (bicor), but not its latent factor/module estimation, which could potentially capture cross-tissue signaling patterns. It is possible that the biological signals of interest would be drowned out by all the other variation in the data but given that this is a conventional step in WGCNA, it is a weakness that the authors do not use it or discuss it.

Reviewer #3 (Public Review):

Summary:

A useful and potentially powerful analysis of gene expression correlations across major organ and tissue systems that exploits a subset of 310 humans from the GTEx collection (subjects for whom there are uniformly processed postmortem RNA-seq data for 18 tissues or organs). The analysis is complemented by a Shiny R application web service.

The need for more multisystems analysis of transcript correlation is very well motivated by the authors. Their work should be contrasted with more simple comparisons of correlation structure within different organs and tissues, rather than actual correlations across organs and tissues.

Strengths and Weaknesses:

The strengths and limitations of this work trace back to the nature of the GTEx data set itself. The authors refer to the correlations of transcripts as "gene" and "genetic" correlations throughout. In fact, they name their web service "Genetically-Derived Correlations Across Tissues". But all GTEx subjects had strong exposure to unique environments and all correlations will be driven by developmental and environmental factors, age, sex differences, and shared and unshared pre- and postmortem technical artifacts. In fact we know that the heritability of transcript levels is generally low, often well under 25%, even studies of animals with tight environmental control.

This criticism does not comment materially detract for the importance and utility of the correlations-whether genetic, GXE, or purely environmental-but it does mean that the authors should ideally restructure and reword text so as to NOT claim so much for "genetics". It may be possible to incorporate estimates of chip heritability of transcripts into this work if the genetic component of correlations is regarded as critical (all GTEx cases have genotypes).

Appraisal of Work on the Field:

There are two parts to this paper: 1. "case studies" of cross-tissue/organ correlations and 2. the creation of an R/Shiny application to make this type of analysis much more practical for any biologist. Both parts of the work are of high potential value, but neither is fully developed. My own opinion is that the R/Shiny component is the more important immediate contribution and that the "case studies" could be placed in the context of a more complete primer. Or Alternatively, the case studies could be their own independent contributions with more validation.