Section A.17, formula 17.

[1zzt]

I was studying your paper/code and had a question regarding Section A.17, formula 17. When I try to apply the formula, the resulting attention score matrix does not have rows or columns that sum to 1, which is different from standard attention matrices. Could you please clarify if I am misunderstanding something, or if there is an additional normalization step required to ensure each row/column sums to 1?

[hyukjunlim]

Thanks so much for taking the time to study our work!

You are absolutely correct in your observation that the resulting attention score matrix A does not have rows or columns that sum to 1. Actually, the matrix A is not a standard attention probability matrix but rather an aggregated interaction score map between ligand and protein atoms.

That is to say, A is a atom-level back-mapped heatmap using cluster assignment scores and attention scores, where each element A_ij​ represents the interaction strength between ligand atom i and protein atom j. The higher the score, the more significant the interaction, which helps us identify the critical binding regions.

We appreciate you bringing it to our attention! Thank you again for your interest.