May introduce mutations that are below the pssm threshold.
For example, when probs[i,j] =~ 1 (close to 1) and probs[i,k] = 0 (for k!=j) but the pssm_log_odds_mask[i,j]=0 , the forbbiden aa may still be introduced since now:
probs_masked[i,j] =~ 0.001 and probs_masked[i,k] = 0 for k!=j
Then after normalization occurs:
probs = probs_masked/torch.sum(probs_masked, dim=-1, keepdim=True) #[B, 21]
probs[i,j] = 1 now , although it doesn't cross the PSSM threshold.
Is that a bug or a feature? :-D
Meaning, if pssm_log_odds_mask[i,j] = 0 then probs_masked[i,j] = 0 too right?
The following line:
probs_masked+=probs*0.001May introduce mutations that are below the pssm threshold.
For example, when
probs[i,j] =~ 1(close to 1) andprobs[i,k] = 0(for k!=j) but thepssm_log_odds_mask[i,j]=0, the forbbiden aa may still be introduced since now:probs_masked[i,j] =~ 0.001andprobs_masked[i,k] = 0for k!=jThen after normalization occurs:
probs = probs_masked/torch.sum(probs_masked, dim=-1, keepdim=True) #[B, 21]probs[i,j] = 1 now , although it doesn't cross the PSSM threshold.
Is that a bug or a feature? :-D Meaning, if
pssm_log_odds_mask[i,j] = 0thenprobs_masked[i,j] = 0too right?