RSEM expected counts question

[tboothby]

I want to check that I understand the output of RSEM correctly. As I understand it the "expected_count" output for each gene is the number of fragment reads that are predicted to map to that "gene."

Read fragments that map to multiple "genes" are not thrown away but their mapping is divided among those "genes." Thus, you would expect most expected_count estimations to be non-integers.

So if a particular "genes" expected_count value was 1215.54 that would mean that RSEM estimates 1215.54 reads mapped to it. Is this correct?

Another question:
Is there any benchmark for what a "good" expected_count would be in terms of assessing whether or not that gene is expressed?

Does this mean that a transcript with an expected_count of say 8 only had 8 fragments map to it? If this is the case it probably isn't expressed, expressed at exceedingly low levels, or the sequencing was data is bad... right?

Thanks for your input.

[arvid]

So if a particular "genes" expected_count value was 1215.54 that would mean that RSEM estimates 1215.54 reads mapped to it. Is this correct?

Yes.

Is there any benchmark for what a "good" expected_count would be in terms of assessing whether or not that gene is expressed?

Does this mean that a transcript with an expected_count of say 8 only had 8 fragments map to it? If this is the case it probably isn't expressed, expressed at exceedingly low levels, or the sequencing was data is bad... right?

There is no way of specifying an absolute threshold for a call expressed/not expressed, since this depends on the amount you sampled and the depth of sequencing... Thus it is virtually impossible to tell whether your gene is lowly expressed or not at all - it is simply below or at your detection threshold.

Do replicates, then you can tell whether a weak measurement is robust.

[tboothby]

Another few questions:
I have used Bowtie to map read fragments to a reference transcriptome. When I look at a sample of individual transcripts in IGV I see they have thousand fold coverage.

When I look at the same transcripts in my RSEM output their "expected_count" numbers are low (at least thats my initial impression), with values of 15 or so.

I am trying to understand why this is happening, are those thousands of fragments I see in IGV also mapping to other transcripts?

Is there an easy way to check this?

Would this cause such low expected_count values?

Any other ideas of why this would be happening?

Thanks for your input.