MsaPair

Installation

git clone https://github.com/zw2x/msa_pair.git
cd msa_pair
pip install -e .

Requirments

Install the latest versions of alphafold and esm.

Dataset

Download the dataset here. It includes 176 subdirectories; each corresponds to a PDB target. The structure of each of the subdirectory looks like this:

- A: .a3ms for chain A
- B: .a3ms for chain B
- esm_scores.json: the esm scores for sequences in uniprot.a3m
- esm_pr.json: the paired rows based on "esm_scores.json"
- multimer.npz: the input feature to AlphaFold-Multimer; this is derived from 
    "esm_pr.json"

Multimer-pipeline

AlphaFold-Multimer default pipeline

python scripts/build_alphafold_features.py below_medium

MSA-Transformer ColAttn pipeline

After modifing esm/model.py using the model.py file provided by Bo Chen, run

python scripts/build_colattn_features.py below_medium

You can also replace the compute_scores function in the script with your own scoring function, and then use the pair_rows and process in the script to generate features for AlphaFold-Multimer

Genetic locus pipeline

This pipeline needs fast internet connections and a lot of memory and disk space, because we need to download and process whole genome sequence (WGS) data from ENA.

Download required WGS data from ENA

python scripts/export_ena_requests below_medium

Pair sequences

python scripts/build_ena_features.py below_medium

Run AlphaFold-Multimer

python msa_pair/scripts/run_model.py --data-dir database \
    --input-npz multimer.npz --output-dir models

This script can run all AlphaFold-Multimer models but need only compile the code once.

Assess

To assess the quality of prediction.pdb against the true pdb ground_truth.pdb, run:

python scripts/assess_models.py --query prediction.pdb --ground-truth \
    ground_truth.pdb --dst-dir assess_results --tmalign-path ${TMALIGN_BINARY} \
    --dockq-path ${DOCKQ_PY}

All outputs will be in assess_results. The most important output file is assess.json, which shows the DockQ scores.