forward_only flag

[mepster]

Hey guys, thanks for this great tool, nice clear code too!

I wonder if you could explain more about the forward_only flag.

I understand that when it's False (the default), then for each given sequence you predict from both 1) the forward sequence, and 2) the reverse complement of the sequence. Then you average the pairs of profiles (after re-reversing the latter profile), and the counts.

However, I know that chrombpnet models are trained with artificial reverse complement data, so wouldn't a single prediction from the forward version of a sequence, and the reverse complement version of the sequence be about the same anyway? (therefore not much difference if you average the two)

Related question, in what cases might you want to use forward_only = True or False ?

Thanks!

[soumyakundu]

Hi,

Yes, the predictions from the forward and reverse complement input sequences should indeed generate nearly identical outputs. Averaging the forward predictions with the reverse complement ones simply makes the final output more robust, but should be very highly correlated with just using forward_only. The difference should be minimal. Therefore, if you are running this for a large number of variants, then you should be fine with just using forward_only = True, which should take about half the time to run.

Averaging the forward and reverse complement predictions may provide some small benefits, since taking the mean of the two will provide a more robust and accurate prediction of the variant effect, but the differences, if any, should be minimal for most cases.

I hope this is helpful. Please let me know if you have any more questions!

Best,
Soumya