Hi Everyone,
I have a simple query. When we estimate differential gene expression between two data sets or two conditions, we generally observed log(base 2)fold change. However, the fold change may not represent true difference. Suppose gene X has TPM of 10 in condition 1 and TPM of 100 in condition 2. While gene Y has TPM of 1000 in condition 1 and TPM of 10000 in condition 2. Both genes X and Y have 10 fold change between condition 1 and 2 but when we observe closely, the change in gene Y is more significant (9000) as compare to gene X (90). Is there a better way to represent differential expression data? Can we give some weight to more change ( for example to gene Y in above example) instead of simple fold change ?
I am sure there must be some statistical method to address such problems but I could not find anything like that so seeking your suggestions.
Waiting for your responses!
Best,
Ankur
What do you mean with "is more significant"? I think you mean that you find one more biologically relevant? This is not easy to answer. It depends on the gene, the role of the gene product in the cell metabolism. A small change in A may be more relevant than a large change in B. Or the other way around. It depends on the role of te products of A and B in the particular cell.
Statistical significance is a different thing! Here is is a very usual observation that the signal (mean expression) is positively correlated with noise (the variance of the expression), and that this correlation is almost absent on the log scale. As that. significance is somethng like a signal-to-noise ratio, it is simple to calculate on the log scale, but not so on the "original" scale (I delibretaely put "original" in quotes to emphasize that this selection is arbirtaty! I could also have chosen the "log scale" as original and "original scale" as exponentiated scale; this is actually like I would like doing it, but it's uncommon and people would likely misunderstand it).
Definition of "Statistical significance" is not equivalent with "True biological significance". At the same time, it is very hard to define "True biological significance". In this case, "True biological significance" should be determined by the consent of the majority of academic researchers. In my opinion, I also don't agree with judging significance by fold-change alone. However, after the statistical test (after considering the dispersion within and between the groups), I agree to use conservative estimation using additional fold-change criteria. The reason for using both criteria at the same time is that it is based on a lot of researches that there is a high probability of false positives in genes with low expression.
Agreed with all of the above answers! Biological systems/ experiments are generally noisy and you generally find a considerable variability among replicates. After correcting for all potential inherent artifacts in RNAseq data (read edgeR or anyother paper), statistical signifncance generally tells whether intra sample variation is larger than intersample variation. Further FDR correction etc. makes irt more robust with multiple corrections. However, as explained by others there could be genes with expression values like the one you mentioned in you question.
To get the biological sense it is also a good idea to further analyze you differentially expressed genes. Clustering and GO-term enrichment etc. to see whether they convey any common message. Then you may choose individual interesting genes based on these analyses and look at their normalized counts (e.g. cpm in edgeR) to see whether it makes sense.
Cheers,
Shajahan
- Badges
- Science topic
- Similar topics
- Mathematics
- Statistics