ESMPair

Requirement

• Installing the lastest version of AlphaFold.

Installation

• pip install fair-esm
• cd msa_pair; pip install -e .

ESMPair pipeline

• For saving GPU memory, you need to replace the esm/model/msa_transformer.py in the directory of {your python path}/site-packages/ with the provided msa_transformer.py. In the new msa_transformer.py, we remove redundant operations.
• Run the ColAttn MSA Pairing, take the 2p01A in the dataset for example, let's build a paired MSA on heterodimers.
  • Making the subdirectory A and B in the directory of 2p01A;
  • Renaming the 2p01A_domain_0_start_21_end_91.a3m with uniref90.a3m and moving it to the subdirectory A. Similarly, renaming 2p01A_domain_1_start_111_end_215.a3m with uniref90.a3m and moving it to B;
  • Running: python colattn_pair.py ./dataset/ {device_id} to get the scoring output: col_scores_512.json and the final paired output col_pr_512.json

Output format

• Scoring output: col_scores_512.json { 'A':{ "{msa_index}":{ "blocknum": xxx, "description": msa description, "score": colattn score. } } 'B':{ "{msa_index}":{ "blocknum": xxx, "description": msa description, "score": colattn score. } } }

• Paired output: col_pr_512.json { "A":[ 0, # 0 is the index of the primary sequence. 3, # index of other msas. ... ], "B":[ 0,
  5, ... ] }

Notaly, msas with the same rank from the two chain lists should be paired, such as the (3+1)th sequence from chainA and the (5+1)th sequence from chainB should be paired.