Hi everyone!
Would it be possible to use WGCNA to find biologically meaningful networks from RNA-seq data generated by fewer than 7 samples?
I have read in WGCNA Package FAQ that it is not recommended attempting WGCNA on a data set consisting of fewer than 15 samples, but even so I would like to know if there is some parameter to be set that can compensate for this low number of samples. If I set up the pickSoftThreshold to a power = 6 may I overcome this low number of samples? Moreover, as I am considering the utilization of RNA-seq data, may I simply use goodSamplesGenes to filter out missing values and outliers? Should I make anything else to prepare the data for WGCNA, e.g. to exclude transcritps with RPKM > 0.3?
Steve Horvath, the author of WGCNA, admitted that he had once gone as low as 12 samples. But honestly, I would not use WGCNA on less than 20 samples. In my experience, with too few samples you start seeing very strong correlations driven by just a single sample, which look like pure sampling noise to me.
Hi Qi Zhao!
What do you mean for sufficient information in the expression table? Could you provide us some paper describing that is safe to use less than 6 or 7 samples? Can anyone provide concrete examples to illustrate this potential WGCNA limitation?
I am sorry for misleading you, I did not mean that using WGCNA is a good way to construct a network with few samples, at least I did not read such a paper before(As i know, the smallest sample size which perform WGCNA is http://www.ncbi.nlm.nih.gov/pubmed/22556371 (mouse,21 samples) ).
BTW, The enough information can be explained as follows: given a sample set that consisted of several identical expression profile, expresstion value of each gene across all samples is at the same level. In this case, even you have a huge sample set, you can still not get a meanful network, since almost all genes may be clustered into one module and there will be no hub gene at all.
For you have only fewer than 7 samples, you can try getting the correlation coefficient (pearson correlation, etc ) between the genes based on their expression across conditions and construct a network based on a threshold(say 0.9).
hope it helps :)
Thank you Qi Zhao! The paper you cited (Iancu et al. Bioinformatics. 2012 Jun 15; 28(12): 1592–1597) is interesting since authors provide a way of validating the modules. So, I can try using 7 samples and then I can use the permutation procedure suggested by Iancu and colleagues to check whether modules are in fact artifacts of the clustering procedure.
Mikhail V Mat, what do you think of these permutation test in this case?
Could you define a sample?
Is each cDNA library considered a sample?
Or
Each treatment is considered a sample?
* If I have 3 treatments (control, cold shock, and heat hock), and each treatment was repeated 5 times. Therefore I have 15 different cDNA libraries, and RNA-seq was made for each one of these 15 libraries seperately.
Do I have 15 samples? Or 3 samples?
I have a similar question to Arye. My hunch is that the real issue is how many libraries go into building the network structure, but clearly multiple samples from the same individual are non-independent? Anyone else care to weigh in? (In my case I have ~30 + libraries, but drawn from only 4 individuals)
Iancu et al provided some hint at an answer to this question. See Figure 4 in Iancu, O.D. et al. Genes, Brain and Behavior 11:29-37 (2012). By selecting subsets of the total number of samples and seeing how many samples were needed before network correlations converged on the entire large sample set, they derived that 30-35 samples gave ≥ 0.9 correlation with total dataset.
- Badges
- Science topic
- Similar topics
- Heat