AFsample3

Modelling protein conformational ensembles with Alphafold3

Setup

Clone repo

git clone https://github.com/wallnerlab/afsample3.git
git submodule update --init --recursive

Install Alphafold3 as a package using instructions their GitHub repo

[OPTIONAL]

Setup for the reference-free state determination system.

conda env update --file pathfinder/environment.yml

Usage

Generating protein ensembles

Activate the correct environment and run the following command (prferably as an executable).

## TEST RUN
chmod +x run.sh
./run.sh <PATH_TO_AF3_WEIGHTS> <DB_DIR>

Update run.sh script to run prediction for your protein.

Key flags to run AFsample3

1. --msa_rand_fraction : Determines the fraction of MSA columns to mask
2. --num_seeds : Number of unique combination of masked MSAs to use
3. --num_diffusion : Number of models to generate per seed (Each seed uses the same version of MSA)
4. --db_dir : Path to sequecne databases if MSAs not provided

#!/bin/bash

if [ -z "$1" ]; then
  echo "Usage: bash run_model.sh <PATH_TO_AF3_WEIGHTS>"
  exit 1
fi

PATH_TO_AF3_WEIGHTS=$1

python run_afsample3.py \
    --json_path=demo/fold_input.json \
    --model_dir=$PATH_TO_AF3_WEIGHTS \
    --db_dir=$DB_DIR \
    --flash_attention_implementation=xla  \
    --output_dir=demo/output \
    --num_seeds 10 \
    --num_diffusion_samples 5 \
    --msa_rand_fraction 0.4 \
    --shuffle_msa=True \
    --num_msa 1024 \
    --tar_output=False

Refererence-free state selection

Once an ensemble is generated, you can used the pathfinder utility to identify unique conformations.

chmod +x pathfinder/run_test.sh
./pathfinder/run_test.sh

Happy sampling !!

Cite

1. Alphafold3
2. Alphafold2
3. AFsample2