TopHat with and without GTF

[id0]

From what I understand, running TopHat with the GTF will assist with the mappings and make them a little cleaner, but shouldn't make a huge difference. I tried running TopHat with and without GTF on some mouse data.

Alignment summary with GTF:

Left reads:
               Input:  76922617
              Mapped:  37116408 (48.3% of input)
            of these:   1976558 ( 5.3%) have multiple alignments (515208 have >20)
Right reads:
               Input:  76672086
              Mapped:  35646606 (46.5% of input)
            of these:   1395593 ( 3.9%) have multiple alignments (748805 have >20)
47.4% overall read alignment rate.

Aligned pairs:  32699739
     of these:    745858 ( 2.3%) have multiple alignments
          and:    353721 ( 1.1%) are discordant alignments
42.2% concordant pair alignment rate.

Alignment summary without GTF:

Left reads:
               Input:  76922617
              Mapped:   4455809 ( 5.8% of input)
            of these:    212949 ( 4.8%) have multiple alignments (481436 have >20)
Right reads:
               Input:  76672086
              Mapped:   3360721 ( 4.4% of input)
            of these:    128254 ( 3.8%) have multiple alignments (732174 have >20)
 5.1% overall read alignment rate.

Aligned pairs:    817083
     of these:      5615 ( 0.7%) have multiple alignments
          and:       720 ( 0.1%) are discordant alignments
 1.1% concordant pair alignment rate.

The difference between alignment with and without GTF is huge. Is this normal? What would explain such a big discrepancy?

If this is normal, then the conclusion is that providing a GTF is important. By that logic, can TopHat really be trusted to detect novel transcripts if it has so much trouble working with transcripts not described by a GTF file?

[dpryan]

Including a GTF file can make a large difference (see "tophat2" vs. "tophat2 ann" at the bottom):



I recommend reading the whole paper, it's quite useful.

[dpryan]

I should add that in either case your alignment rate is exceedingly low. What sort of organism is this? Also, did you do any adapter trimming?

[id0]

To answer your question, this is mouse without adapter trimming.

Thanks for that informative paper. However, the difference between annotated and non-annotated TopHat there is a few percentage points. For me it's ~5% versus ~50%.

For comparison, I am getting over 80% with just regular genomic alignment with Bowtie, so the reads themselves are of reasonable quality.

[dpryan]

80% with mouse RNAseq is more what one would expect (I get >95% alignment with mouse RNAseq, though only ~85-90% map uniquely).

Are you using local alignment with bowtie? Also, keep in mind that tophat is less tolerant (by default) of mismatches than bowtie, so if you have a number of those (due to using a quite divergent strain, for example), then that might also cause these sorts of problems.

Maybe give STAR a try and see if that produces better results for you. I've been quite happy with it.

[id0]

Based on what I've heard from other people, STAR will be much faster, but only marginally more accurate (if at all).

Regarding mismatches, that should not be affected by adding or removing a GTF. That variable is yielding ~5% versus ~50% alignment rate for me. I don't see how I can find any novel genes based off TopHat alignment if it is having so much difficulty finding known ones.

[dpryan]

True, though if the low alignment rate is due in part to the ends of many reads not mapping then using an aligner that can do soft-clipping (e.g., STAR) might produce better results. Aside from that, I'd have to actually see and play around with your data a bit to be of any more help. I've never had these sorts of issues with mouse RNA.

[id0]

I ran the same sample with STAR. I generated two genomes, one with GTF and one without. I ran the sample against both. I got more than twice the number of splices with GTF, which makes sense to me. For uniquely mapped reads, I got 64% alignment rate with GTF and 63% without. Essentially identical, which is what I would expect from a good aligner.

I will have to evaluate STAR more thoroughly. Based on the literature and this forum, it's main advantage is speed, which is not a concern for me, so I never bothered to test it for myself. At least for this one example, it seems to be far superior than TopHat in terms of alignment. I would also be far more confident in any novel genes detected from this alignment.