Trainable AlphaFold 2 ported to Pytorch

Our version of AlphaFold 2 written using Pytorch library.

The repository provides a training script alphafold/train.py, which contains a similar to AlphaFold 2 training procedure.

We also provide a script for protein folding alphafold/inference.py, which can use AlphaFold 2 original parameters as well as our trained parameters.

Inference with AlphaFold 2 original parameters

Go to download_af2_parameters/ and follow the steps. This will download the parameters and convert them to pth format, which can be used with alphafold/inference.py to fold proteins from MSAs. Check examples/inference_1ezi_with_af2_params/ for an example of how to use AF2 parameters with our code.

Training example

Follow the steps in examples/training_toy_example/ to run training on a small debug dataset (3 proteins).

Deployment

The project was developed and deployed on the Summit Supercomputer inside a conda enviroment, which can be recreated on Summit using bootstrap_summit.sh. The base for the enviroment is https://github.com/open-ce/open-ce, version 1.4.1. Some of the main packages are listed below

• python == 3.8
• pytorch == 1.9.0
• prody == 2.0.1
• numpy == 1.19.2
• mpi4py == 3.1.1
• nccl == 2.8.3
• horovod == 0.22.1
• cudatoolkit == 11.0.221
• biopython == 1.79

For installation instructions refer to https://osuosl.org/services/powerdev/opence/