Bowtie2: minimum score threshold to percent similarity

[DamianMichael]

Hello,

I am new to using Bowtie2 and am trying to understand the 'minimum score threshold' as it relates to sequence similarity. I will be using the 'local alignment mode' where the default minimum score threshold is 20 + 8.0 * ln(L). If I understand correctly by increasing the values in the formula, i.e. making it 20 + 10.0 * ln(L), I will get a higher alignment score which makes the minimum threshold more stringent. This value will change based on the read length of my sequences.

My question is, how do I adjust the formula to test for different sequence similarities? I.e I would like to know which sequences only hit at 99% sequence similarity or above, or 95 or 90, etc. I can't tell if the values given by the formula are on a scale, if they relate to something other than read length, or something else.

Thank you in advance

Damian

[dpryan]

This becomes a question of how you define similarity in the case of insertions and deletions. Also, do you want to incorporate the fact that mismatches at low quality base calls are more likely?

[DamianMichael]

Way to throw a wrench in my question!

So there isn't a direct correlation of alignment score to sequence similarity and it then becomes a question of adjusting penalties and accounting for sequence quality to determine the quality of the alignment.

If this is true, I'm curious how users adjust penalties and minimum score thresholds to account for these changes. For example, I am running an eDNA set through CLC Genomics, but would like to move it to Bowtie2. At 99% sequence similarity on the CLC I get three species whereas 80% I get 20+ species (I didn't adjust the penalties). Is it just a matter of playing with the penalties and minimum quality score in Bowtie2 until I can match the output of CLC?

[dpryan]

[Devon quickly googles "eDNA"]
Ah, you're working on metagenomics. You might want to have a look at Kraken rather than bowtie2. The goal shouldn't be to match whatever (possibly meaningless thing) CLC is doing, but rather to get the right answer.

[DamianMichael]

Good point. I will do that. Thank you.