DESeq raw variance

[narges]

Variance in DESeq is the sum of shot noise and raw variance. The raw variance is the smooth function of estimated read counts and experimental condition. In the corresponding paper, it is mentioned that this function aim to pool the genes with the same strength to estimate the variance(due to the more probable few number of replicates).
I am not sure if I have understand it correctly. Does this means that DESeq groups n genes (i1, i2, ...in) from a common condition with similar expression levels and then calculate the variance for them? So, instead of calculating the variance for a specific gene, it calculates the variance for a group of similar gene and associate that value for all of them?
And how does it decide which genes have similar strength (? similar expression levels)?
Then can we conclude that DESeq does not calculate the gene-specific variance but group of genes-specific variance?

[dpryan]

Well, variance is the sum of technical (shot noise) and biological sources, though perhaps you mean the latter by "raw".

The process of fitting the smooth curve to the data is essentially the "pooling information" step, since variances are then shrunken toward it (generally you use distance from this line as a penalty and perform maximum-likelihood estimation of the variance with that, though I recall DESeq2 also treats genes with variance >3sigma (or something like that, it's in their paper) from the expected differently). So no, DESeq2 doesn't group genes from a common condition with similar expression levels and then calculates variance from that. It calculates per-gene variance and then shrinks that toward the expected value.

[narges]

Well, variance is the sum of technical (shot noise) and biological sources, though perhaps you mean the latter by "raw".

Yes, I meant the biological oriented variance.
Thanks!
Is it true that point "Px" in smooth curve is the mean of count values for gene gene x in the condition p? So, it forms the curve based on the means for each gene of a condition and next shrunken the variances to their corresponding point on the mean-specific smooth curve?

[dpryan]

No, that's not strictly true. It will typically be the case that there's only 1 gene with a mean expression of a certain value, so that would mean that the curve must go through that mean, which will typically not be the case. The curve is fit to the means, as you mentioned (at least that's my recollection), so the remainder of what you wrote looks correct.