Dear all,
I am setting up a weighted correlation network analysis (WCNA) in order to identify modules of genes that are co-expressed and correlated or anti-correlated with several quantitative traits. For some traits I have quantifications for the same samples that were profiled (RNAseq and proteomics) but for some other traits we did not have enough tissue to perform both molecular profiling and biochemical determinations from the same animal. Instead, we have repeated measurements in animals of the same genotype than the ones that were profiled.
Would it be possible to assign the mean value per each phenotypic group to the corresponding individual samples? I understand that we would have reduced power to detect significant correlations because we would be unable to model interindividual variability... but I really think it is worth trying. I think that, if present, the correlations should be informative.
An alternative would be averaging replicates and assigning each group the mean value for the different phenotypic traits... but we would be reducing a lot the sample size and we would be then underpowered to detect co-expression modules...
I know it is not the perfect scenario but still, do you think the results would be meaningful?
@Peter Langfelder I would really appreciate your feedback.
Thank you!!
Lídia
Your description is not very clear. Is this part
correct? Do you really mean traits or do you mean samples? Or do you have a separate set of samples for each trait?
I will assume you meant samples, i.e., you have a set of mRNA samples and some of those also have protein data.
I would not combine any samples together since all RNA-seq samples come from different animals (if I understand your design correctly). I would start by creating two separate networks, one for RNA-seq and one for proteomics, get modules and correlate module eigengenes on common samples.
Alternatively, if you want to analyze proteins and mRNA together in a single network, you can pad the protein data with missing data (NA) for the samples for which you only have mRNA data, and run a single network analysis.
Thank you very much for your response and sorry for not being clear enough.
I meant quantitative phenotypic traits that are measured using biochemical assays. The experimental design is very complex, I will try to clarify my question.
For some samples I have both the molecular profile (proteomics or transcriptomics) and a set of biochemical measurements that are relevant for the phenotype (quantitative traits). In this case it is possible to correlate the eigengenes of the modules identified with the phenotypic traits. However, sometimes I don't have the quantitative trait measurements of exactly the same samples that were profiled. For instance, I have three technical replicates for a given animal but I only have one value for the phenotypic trait, which is itself the average value of the quantifications performed on different aliquots of the same animal. Intuitively, I would say it is safe to assign this same value to all technical replicates, even if they showed slight intra-individual variations in their transcriptional profile. A more extreme case would be to use some traits that were not directly measured in the same animals that were profiled but in biological replicates with the same characteristics (age, genotype, sex...).
Let me put an example: imagine we have an experiment with 3 age groups, 2 genotypes, and 6 biological replicates per group. Imagine that 3 animals were used for molecular profiling (RNASeq / proteomics) and the other 3 were used for biochemical profiling (the amount of tissue needed didn't allow us to acquire both measurements from the same animal). Would it be safe to assign to all the samples in a given group (i.e 6 months old trangenic mice 1, 2, 3) the average value of a trait that was measured in a different set of animals with the same characteristics (6 months old trangenic mice 4, 5, 6)? I would loose inter-individual variability but I would capture the variation across age and genotype.
I hope it is a bit more clear now. Reality is indeed more complex than that because I have mRNA and proteomics from two different tissues. In this regard, I was planning to do consensus module analysis, as you suggested in your response.
Thank you very much for your time!
Lídia