Why Bidirectional LLR Calculation and Averaging

songlab-cal / gpn

Genomic Pre-trained Network

https://doi.org/10.1073/pnas.2311219120

MIT License

182 stars 22 forks source link

Why Bidirectional LLR Calculation and Averaging #21

Closed yangzhao1230 closed 7 months ago

yangzhao1230 commented 7 months ago

Hello,

I've been examining the implementation of the Log Likelihood Ratio (LLR) calculation in the MLMforVEPModel, and I noticed that the LLR is computed separately for both forward (fwd) and reverse (rev) sequences. Any clarification or insights into the methodology would be greatly appreciated

gonzalobenegas commented 7 months ago

Hi! This serves two purposes:

Ensures reverse-complement invariance (the choice of strand is arbitrary so we don't want the variant score to vary between them)
Averaging the two predictions is a form of "ensembling" and could improve performance (we haven't checked though)

For more discussions about the topic, I recommend Towards a Better Understanding of Reverse-Complement Equivariance for Deep Learning Models in Genomics.