Coverage

[swaraj]

Dear All,
I am curious to know for a RNAseq experiment what would be the ideal statistics to measure the coverage of reads across known features. The options I am confused with are whether I should go with coverage of complete transcripts or simply work on the coverage of exons. Also there are some transcripts which have high coverage on only one of the exons hence how to interpret such results.

[swaraj]

Also I would like to know whether I can use a coverage cut-off to select for the expression of a transcript.

[westerman]

Transcripts. FPKM, not coverage. Even if you are not using the program, see the cufflinks manual for a good explanation and a reference. http://cufflinks.cbcb.umd.edu/howitworks.html

[swaraj]

I am working with non-coding transcripts for whom the FPKM as reported by cufflinks is very low. Many times the FPKM value reported is zero whilst I see coverage of the built transcripts. This behaviour may be because some classes of non-coding transcripts are be expressed in a small spatio-temporal window. Hence my question, that whether I can use some score like coverage or read count instead of FPKM to monitor such transcripts. My apologies for not elaborating on the context of the question before.

[westerman]

I am working with non-coding transcripts for whom the FPKM as reported by cufflinks is very low. Many times the FPKM value reported is zero whilst I see coverage of the built transcripts. This behaviour may be because some classes of non-coding transcripts are be expressed in a small spatio-temporal window. Hence my question, that whether I can use some score like coverage or read count instead of FPKM to monitor such transcripts. My apologies for not elaborating on the context of the question before.

I guess I am at a loss as to what you are asking. Certainly you can use read counts. I have done so many times. Statistically you will probably run into problems proving anything from them (especially if they are low) unless you are working in a binary on-off situation.

Assuming that you are not in a binary situation then with zero FPKMs and low counts then you are in the "noise level". Using the counts as a way to point towards other experiments is acceptable. Using the counts in a paper would be tricky since I would expect the reviewers to question any conclusion made from noisy data.

[swaraj]

Thanks a lot for your valuable suggestions. I will keep them in mind.