I want to kindly ask a question regarding DESeq2. I am getting large log2fold changes (absolute log2fold changes greater than 0.5), but the padj values are not significant (padj>0.05). However, some genes in my data have low log2fold changes (absolute log2fold changes lower than 0.5), but the padj values are significant (padj<0.05).
Does this make sense? Can someone please explain to me how this works?
I don't have a definitive answer to your question, but I will say that low p-values and high fold changes are correlated most of the time, but not always.
Separately, absolute log2FC >0.5 does not on its own mean large. A log2FC of 0.6 may be considered by some to be significant - not by me - but it is definitely not large. I'd be looking for a log2FC of at least 1 before considering it significant, though opinions on this vary. Most people stand firmer on log2FC threshold of 1 than they do when it comes to p-value thresholds.
For p-values some people go for the 0.05 threshold, others for 0.02 or even 0.01 (I would be in this camp). The point is that often p-value thresholds are adjusted based on data: if there are 200 genes with p-value <0.01, most people would not try to raise a threshold to 0.05 because that would give them 500 proteins. In my experience, most people are happier with 200 than 500 proteins to look at. But if they get only 10 proteins with p-value <0.01, then they will go for the 0.05 threshold.
You already have a better explanation below by Istvan Albert, but I will give it a try. Very high absolute log2FC values will tend to be associated with low p-values - big fold changes tend to be significant. So I would be surprised if you ever get log2FC>5 that doesn't have low p-value. Yet there is no perfect correlation between the two, especially for borderline cases where it will depend on the nature of your data. Sometimes log2FC=0.5 will have p-value <0.05 and other times >0.05. In my book they would most likely be insignificant either way because of more stringent p-value thresholds I apply.
A log2 fold change of 0.5 is modest but can be biologically meaningful if it occurs in a set of genes within a similar pathway. My postdoc mentor once pointed out to me that Down syndrome is triploidy (log2(1.5FC) = 0.585) for all of chr21 ~200-300genes; so having a large number of genes with a modest change can have a big effect.
For RNA-seq look at the read counts (RPKM) for the genes with modest fold changes but are not significant. Do they have low read counts? Is there a high degree of variability between samples within each group? is there an outlier that is driving the fold-change? Usually its one of those 3 that is taken into account in keeping the padj from being significant.
unequal variance in your data, i.e., heteroscedasticity
As mentioned, a log [base 2] fold-change of 0.5, while not large, can indeed be statistically significant with a large sample size that is sufficiently powered such that one can detect fold-changes this low as statistically significant.
If we additionally bring 'life' into the equation, then it can be easily understood that 0.5 may have a meaningful biological effect for the mRNA from some genes, while not others. Many genes' expression profiles operate on sort of 'buffers', i.e., one could continually express a gene up to a certain point, with no alteration on the cellular phenotype, but, beyond that point, new pathways would be triggered due to the extra mRNA that, ultimately, bring about an alteration to the phenotype..
I don't have a definitive answer to your question, but I will say that low p-values and high fold changes are correlated most of the time, but not always.
Separately, absolute log2FC >0.5 does not on its own mean large. A log2FC of 0.6 may be considered by some to be significant - not by me - but it is definitely not large. I'd be looking for a log2FC of at least 1 before considering it significant, though opinions on this vary. Most people stand firmer on log2FC threshold of 1 than they do when it comes to p-value thresholds.
For p-values some people go for the 0.05 threshold, others for 0.02 or even 0.01 (I would be in this camp). The point is that often p-value thresholds are adjusted based on data: if there are 200 genes with p-value <0.01, most people would not try to raise a threshold to 0.05 because that would give them 500 proteins. In my experience, most people are happier with 200 than 500 proteins to look at. But if they get only 10 proteins with p-value <0.01, then they will go for the 0.05 threshold.
I see, thank you for your helpful response!
Could you please explain to me how low p-values and high fold changes are correlated? What is the implication of it?
You already have a better explanation below by Istvan Albert, but I will give it a try. Very high absolute log2FC values will tend to be associated with low p-values - big fold changes tend to be significant. So I would be surprised if you ever get log2FC>5 that doesn't have low p-value. Yet there is no perfect correlation between the two, especially for borderline cases where it will depend on the nature of your data. Sometimes log2FC=0.5 will have p-value <0.05 and other times >0.05. In my book they would most likely be insignificant either way because of more stringent p-value thresholds I apply.