Dear Karin

The strategy for the analysis will depend on you research goal. What do you want study by comparing metabolites of different tissues?

The tools, though, still can be the same: software provided by instrument's company, metaboanalyst, kegg, r, cytoscape, gephi. For most of the tools it is not important what effect size or fold-change you input, you just will need to interpret the provided results properly in light of what you have used. But I would say again - think about your goals, I am personally not sure that performing all pair-wise comparisons between 7 groups is the best strategy. Although, it can be only the one if this is your ultimate goal.

Thank you for your response, Albert. The mouse model studied, contains a global KO, however, phenotypically some tissues are severely affacted while others are seemingly unaffected. We want to compare the metabolites of the different tissues to gain insight into the tisssue-specific reponses to the disease. We have done univariate statistics on metabolites within each tissue type, however, as I am a newcomer to network analysis, I am not sure what the best strategy is.

Dear Karin

So basically you have two biological groups - WT and KO? And you compare metabolites from KO vs WT for each tissues? And do not compare tissue1 of WT vs tissue2 of WT? That are very important details.

If this is the case, I would start from the easiest steps:

- go to metaboanalyst pathway analysis

- input the metabolites that significantly more abundant in KO for tissue1

- perform analysis - export results

- input the metabolites that significantly more abundant in KO for tissue2

- perform analysis - export results

- repeat for all tissues

- repeat for metabolites that statistically less abundant in KO

Then you can see how the tissues are differently affected in respect to metabolic pathways.

Network analysis is a non-trivial thing. What kind of network you would like to build? There are multiple types of them. May be this review will help you?

Article Network based strategies in metabolomics data analysis and i...

I also would not exclude PCA for all tissues in WT and KO. I think it can be really interesting to see the plot.

Dear Albert, thank you for your detailed advice. Yes, that is the case. I am considering MetaboAnalyst's Debiased Sparse Partial Correlation (DSPC) network analysis, but was wondering whether there are tools available with which to evaluate inter-tissue interactions with an integrated multi-tissue type network?

Hi Karin Terburgh A great example to look up to for a very relevant study, would be https://www.nature.com/articles/s41586-020-2047-9#Fig1

They did comparison and stats for 29 mice organs!

See specifically, page 8 : https://static-content.springer.com/esm/art%3A10.1038%2Fs41586-020-2047-9/MediaObjects/41586_2020_2047_MOESM1_ESM.pdf in their supplementary materials!

[ To assess the overall similarities between the metabolome of murine organs the Bray-Curtis dissimilarity was calculated between all paired samples (compared for the same subsection location for the same organ) between the GF and SPF states for all mice and these dissimilarities were averaged per organ and plotted with notch plots. }

Hope it helps with some directions with

Dear Karin

There are multiple strategies to analyze biological networks, and they strongly depend on a nature of the network (what do vertices and edges represent). Since you mentioned you would like to use partial correlation, this directs us to correlation-based networks. In this case your vertices will be metabolites, and edges will be correlation coefficients (weighted or unweighted). Interpretation of the correlation-based network of metabolites itself is not a trivial task, so what is commonly used in this case - comparison between two or more networks (conditions). In your case these conditions are tissues. So, and please consider this only as suggestion, you can build 7 tissue-specific networks and perform your analysis by comparing them. However, if you decide to go this way, please keep in mind that it still can be difficult, as you likely have different amount of metabolites detected in each tissue, and perhaps different sample size too. In this case I recommend to unify the data-sets beforehand. Doing this you will lose some information, but will minimize artificial differences in the networks parameters. Another potential challenge - I suspect that you have many hundreds of metabolic features. To build a correlation matrix for these large data-sets you may need to use a high computational cluster, as I am not sure what is the size limit of MetaboAnalyst online tool, and you may need to use something local, like R or Python. This also can cause challenges if you have limited experience with them. If this is the case, I would recommend to consult a local bioinformatician or statistician.

My question is do you really need to go in this direction? If you can show qualitative (for example, Venn diagram of the detected metabolites) and quantitative (for example, heat-map) differences in so broad metabolic profiles between the tissues, and general (dis)similarity by PCA or PCoA, I think these will be already great results.

I agree with @ Albert Batushansky Albert has provided you, and all of us here with loads of very helpful, sound advice.

"If you can show qualitative (for example, Venn diagram of the detected metabolites) and quantitative (for example, heat-map) differences in so broad metabolic profiles between the tissues, and general (dis)similarity by PCA or PCoA, I think these will be already great results." - Indeed very true. With his advice, you can opt for a combination of biostatistics and bioinformatics software which you are comfortable with and can use with a high level of confidence. Many of the software have built-in statistical tools so you can apply the filters according to your need, be it in fold-change, Cohen's D or whatsoever. This software can also build the network-network interaction map and systematic enrichment pathways for you, based on your raw data input.

My experience is that you must recognise what is your desired outcome so that you can instruct the software to come up with the interactions and/or pathways with a minimum of variations. Just like what Albert has advised.

Thank you very much Albert, Biswapriya and Alexis for the great advice.