Clarification on how use_sequence=True affects single_aa_score #73
Description
Hello, thank you very much to your team for this work and for open-sourcing the code. I am a beginner, and while reading the implementation, I encountered the following confusion.
In the function single_aa_score, residues along the backbone are decoded independently and sequentially, one residue at a time. When use_sequence is enabled, the order_mask is constructed such that only the currently scored residue has value 0, while all other residues have value 1. As a result, in the resulting decoding_order, the current residue is always placed at the first decoding position.
Consequently, in both mask_attend and mask_bw, all neighboring residues of the currently decoded residue have mask value 0. Therefore, in the update step
h_ESV = mask_bw * h_ESV + h_EXV_encoder_fw
even if h_ESV contains the true sequence features of neighboring residues, these features still seem to be completely masked out.
My confusion is: how does the model access real sequence information in this case?
It appears to me that the logic might be reversed.
I understand that this question most likely comes from my incomplete understanding of the code, but it has genuinely confused me for quite a long time. I sincerely apologize if this is a naive question, and I would greatly appreciate your clarification.