No actual masking of the residue in predicting the likelihood of targeted residue in the code?

[lejecrs]

Hi thanks for the great package!

I am looking at the code to understand the logic.

I looked at the evolocity.tl.velocity_graph, the understanding from paper is that here is the actual code that is predicting substituted residue likelihood from original sequence context (the sequence context does not contain the residue, but the sequences surrounding it).

I just wonder that in predict_sequence_prob , it calculate the prob based on predict_sequence_prob_fb, I didn't see any masking task in the code of predict_sequence_prob_fb.

I think in the paper it states predicting the residue probability only given the surrounding context. Please let me know if I was wrong.

Thanks for looking at this!