novoalign <-> bowtie comparison

[rgregor]

Hello bioinformaticians

I am mapping 50.000 sequences (stored in .fastq file) to a reference genome. I use default parameters (only single hits).

novoalign -Q 20 -g 1000 -x 1000 -d ref_gen -f data_small.fq -o SAM > novoalign.sam

bowtie ref_gen -n 3 -a -y --nomaqround -B 1 -p 3 -m 1 --best --strata data_small.fq --solexa-quals -S --un bowtie_unmapped.fastq > bowtie.sam

Results of reported alignments:

bowtie: 38030
novoalign: 38225
shared: 37344

How come the number of shared reported alignments is so low? (~600 alignments difference)?

Did anybody do some studies on which alignments are reported by bowtie/novoalign?

Any idea on how to make this difference smaller, changing the bowtie parameters?

tnx,
Gregor

[sparks]

Hi Gregor,

I don't think you'll ever get them to be the same. They just work differently.

With the large gap penalties Novoalign needs to align the full length of the read while Bowtie only needs to align the first 28bp of the seed. This might account for some of Bowtie extra mapped reads.

There's also a big difference in what would be considered a unique alignment and there's no options to make them behave in the same way. A few of Novoaligns extra alignments may fall into this category.

Colin

[rgregor]

Thank you Colin. I am thinking: if a unique hit (alignment) is not well defined because there are many parameters regarding quality and mismatches + other parameters, i would say that a unique hit is very well defined if we only look at the sequence.

Is this correct:

* a unique alignment of a read on a reference sequence happens when the read can be aligned on the reference (without mismatches) at exactly 1 position.

Do you perhaps know if bowtie/novoalign/etc. would be expected to perform in the same way if working only with FASTA data (no qualities)? Can you point to some studies on how exact these aligners are? (do they find all possible unique alignments, miss some, how many etc?).

Thank you,
Gregor

[sparks]

Hi Gregor,

If we have a base with low base quality such as 2 (Perr = 0.63) or 4 (Perr = 0.4) and we have one alignment that matches the read sequence and we also have several alignments that differ only at this low quality base, do we have a unique alignment or not?

and we always look at base qualities.

For fasta files bases default to a quality of 30 and so in above scenario the alignment that matches would have a quality of 30 or thereabouts depending on how many alignments had 1 mismatch.

Colin

[rgregor]

From the bowtie manual:

In -v mode, alignments may have no more than V mismatches, where V may be a number from 0 through 3 set using the -v option. Quality values are ignored.

The question remains: when looking at mismatches ignoring qualities, should bowtie vs. other aligners give pretty much the same result? Is the alignment space searched in an exact (optimal) way?

tnx,
Gregor

[colindaven]

Bowtie is very quick, but certainly not the most optimal for some use cases. It can't align reads containing indels, which may be the source of the discrepancy. If you can get them to work on your dataset rapidly enough I would try Shrimp2, Novoalign or Stampy.