I am working on a medication switch-control comparison (RNA sequence data, human tissue) and the results have confused me somewhat. I wonder if anyone can help me make sense of my findings.
I have data for 30 people (half of them have been on the same medication they have taken for years, half of them have had their medication switched to a new one). For each one of them, I have RNA seq data at baseline and 6 months after medication switch/continued medication. I have performed analysis with limma-voom and find that there is a 30% overlap (significantly different genes) between people who have stayed on the same medication and people who have switched medication. Top 20 genes (sorted by p-value) have 80% overlap – fold changes (LFC more than 1) that are quite similar, in the same direction. The clinician has advised me that the results are not possible to explain in any biological sense - nor with any time effect. We started thinking that there could be plate effect (technical effect) in the sequencing platform and found that the genes that are differentially expressed are also correlating with the plate position (correlations as high as 0.8). I can’t adjust for plate position in my model because my phenotype is also correlating with plate position; both sample type (medication switch, control) and visit (baseline, 6 months). My question is on how to remove this plate effect when I have such a strong correlation to phenotype. I fear I now have data I can’t work with - any advice is appreciated.
Thank you in advance,
Mahes Muniandy (University of Helsinki)
design <- model.matrix(~ 0 + TS + Age + Sex + Smoking + BMI ) v <- voom(dge_AT, design) corfit <- duplicateCorrelation(v,design,block=ID) v <- voom(dge_AT,design,block= ID,correlation=corfit$consensus) fit <- lmFit(v,design,block=ID,correlation=corfit$consensus) cm <- makeContrasts( ctrl = TScontrol_V2-TScontrol_V1, treat = TStreatment_change _V2- TStreatment_change _V1, difference = ((TStreatment_change _V2- TStreatment_change _V1)-( TScontrol_V2-TScontrol_V1 , levels=design)