can somene explain how BWA do its trimming

[foxyg]

In BWA manual, there is a -q option.

Parameter for read trimming. BWA trims a read down to argmax_x{\sum_{i=x+1}^l(INT-q_i)} if q_l<INT where l is the original read length.


I am not really sure what this means. If I say -q 15, what does it really mean?

[dawe]

In BWA manual, there is a -q option.

Parameter for read trimming. BWA trims a read down to argmax_x{\sum_{i=x+1}^l(INT-q_i)} if q_l<INT where l is the original read length.


I am not really sure what this means. If I say -q 15, what does it really mean?

I believe it tries to find local maxima in the (INT-q_i) function, and chooses the rightmost maximum. That means that it trims when the quality starts to decrease monotonically below your threshold. This is pretty smart: suppose you have two/three bad qualities at the beginning of your read (say from 5 to 8 bp): hard trimming below a certain threshold results in a 5 bp long read. bwa method actually checks if you have better qualities after that and trims later.

d

[foxyg]

So does BWA start scanning from left or right?

Also what is the common trim parameter here I should use if I have offset 64 FASTQ data?

[lh3]

bwa learned from phred.

@foxyg

Use Sanger fastq. That is the standard. Use -q15 or -q20. Usually the threshold does not matter too much.

[dawe]

Hi again!

So does BWA start scanning from left or right?

Well... How else? :-)

Also what is the common trim parameter here I should use if I have offset 64 FASTQ data?

As pointed by lh3 you should always have your scores in Sanger format and then you may apply a filter to 15-20 (which corresponds to a ~0.03-0.01 probability).
BTW, if you have your fastq in Illumina (Pipieline 1.3+) you may try this patch I've written. It enables a '-I' option to bwa aln so that you can use Illumina reads and trim (and output) as they were in Sanger scale.

d

[ElMichael]

bwa learned from phred.
Use Sanger fastq. That is the standard. Use -q15 or -q20. Usually the threshold does not matter too much.

Hi,
I wonder when does the threshold matter too much? And could anybody explain why it usually doesn't matter?
Thanks!

[Orr Shomroni]

As pointed by lh3 you should always have your scores in Sanger format and then you may apply a filter to 15-20 (which corresponds to a ~0.03-0.01 probability).

lh3 also pointed out that

Usually the threshold does not matter too much.

So on the same tone of ElMichael, why would the threshold not matter? I mean, if the quality threshold is higher, you would select fewer bases, as the overall quality would decrease, right?

Dawe, what did you base your choice of read trimming threshold (15-20) upon? Is there a specific paper saying "this is a commonly used threshold value", like the use of p-value=0.05 for hypothesis testing? I just want to have some confirmation of the threshold selection.

[carmeyeii]

I suppose if the probability of a base call being wrong is less than .01, you'd still want to keep it.