ESM_highAttention

This repository contains code and resources for finding high attention sites from the ESM attention matrices. Associated Zenodo: https://zenodo.org/records/15509699

0. environment.yml file contains the package requirements
1. ESM_matrices.py
   • get the attention matrices and representation matrix from the ESM model
2. ESM_ha_loc.py
   • Load the attention matrices (from ESM) and calculate the column sum and normalized heatmaps (provided in Zenodo)
   • Use the heatmaps to identify the convergence layer and the High Attention (HA) sites. Save the protein to convergence layer and HA site dictionaries to data folder (provided in Github). **.pkl files should be read in as binary
3. ESM_ks.py
   • Perform the KS test per family. For each family, calculate the true family distance for each distance measure (ha_based, mean, max, cls) and 1000 random iterations, and calculates the KS statistics (provided in Github and Zenodo).
4. ESM_partial_dist.py
   • Calculates the pairwise distance matrices for each distance measure to perform the silhouette score analysis (results in Github).
5. ESM_spatial_dist.py
   • Takes in the pdbs files and the high attention sites and active sites to calculate the spatial distances between the HA sites and active sites (results in Github).
6. ESM_pfam_msa.py Separates the full human protein fasta file into fasta files per family and performs multiple sequence alignment (provided in Zenodo). Then finds the percent consensus at each HA site (results provided in Github).