Differential Coverage

[rskr]

I have been trying to come up with a test for differential coverage, that given a mapping and classes for the samples ranks the genes such that genes with coverage that is as extreme or more extreme than would be predicted by a two parameter beta binomial distribution are ranked at the top. Now I am using a weighted sum over the coverages at each position in a gene, this may be less than ideal(since it doesn't take into account the probability that multiple bases could be extreme simultaneously), however it is ranking the genes in a manner which is intuitively what I expect to see.

In the following image, you can see that my test set has a gene with where the magenta, light blue, and brown colored sample have much more coverage on one end, while the red, blue, and green, don't have a jump in coverage.

Will something like this be useful? I think it will help researchers doing RNA-seq get more out of their data, than just gene level expression. Right now, I think the draw back is that the p-value if you can call it that is, essentially testing the hypothesis "Are there any genes that are more extreme", which is essentially an OR relationship, so the probabilities aren't quite as significant as one might expect, though the ranking seems to be good.

[dpryan]

How is the weighting being done? Are you thinking that this might be useful to detect differences in 3' bias between samples or is this geared more toward seeing if a given gene in one sample seems to have a bunch of PCR duplicates being aligned to it?

[rskr]

How is the weighting being done? Are you thinking that this might be useful to detect differences in 3' bias between samples or is this geared more toward seeing if a given gene in one sample seems to have a bunch of PCR duplicates being aligned to it?

I am hoping that bias/duplicates between samples isn't a factor, though I could imagine a situation where samples were run at slightly different temperatures causing there to be bias due to differences in GC vs AT melting temperatures for example.

I am hoping that it will find truncations or differential exon coverage, though I suppose it could also be used to determine the quality of the sequencing in conjunction with the differential expression tests.

[dpryan]

I expect that most truncations would be picked up by DEXseq or similar tools, though I guess if there's some sort of systematic truncation of exons in a group of genes, that would likely be missed by the current methods (I can't think of a case when this would biologically occur, but I guess no one has actually looked!).

By "differential exon coverage", do you mean coverage within a given exon or more what's done in DEXSeq (I assume the former). I suppose someone will find a case where that's changed due to some sort of treatment, though I wonder how the effect of that is separated from more general differences in library quality (though perhaps you've thought about that).

[rskr]

I expect that most truncations would be picked up by DEXseq or similar tools, though I guess if there's some sort of systematic truncation of exons in a group of genes, that would likely be missed by the current methods (I can't think of a case when this would biologically occur, but I guess no one has actually looked!).

By "differential exon coverage", do you mean coverage within a given exon or more what's done in DEXSeq (I assume the former). I suppose someone will find a case where that's changed due to some sort of treatment, though I wonder how the effect of that is separated from more general differences in library quality (though perhaps you've thought about that).

I think DEXseq doesn't use base level information.

[dpryan]

True, I guess it'll depend on how your weighting is done and how big exons are. If all the weighting is done within an exon then you're right, DEXSeq won't see any difference there.