AF2_resTrain: Injecting distance information into Alphafold-like structure predictors

On Google Colab

A simplified version of resTrain is available to run as notebook on Google Colab. Try it by clicking on the button below:

Installing resTrain

With Docker (recommended)

1. Pull the image from Docker Hub.

docker image pull clami66/restrain

Alternatively, use the provided Dockerfile to build the image from scratch.

2. Run from the docker image while mapping the necessary directories with -v and the GPU devices with --gpus, e.g.:

docker run -v /home:/home --gpus device=0  clami66/restrain run_alphafold.py ...

Stand-alone installation

If you don't wish to use a Docker image, the installation and setup procedure is as follows:

1. Install Anaconda/Miniconda

2. Set up conda environment, install dependencies:

# clone this repository
git clone https://github.com/clami66/resTrain.git
cd resTrain

conda env create --file=environment.yaml

conda activate resTrain

python -m pip install -r requirements.txt

Setting up the databases

[optional] Download and set up the AF parameters and sequence databases. Reduced databases (reduced_dbs option) should suffice as evolutionary inputs are not as important when a good template is provided. If the full databases are needed, run the following by omitting reduced_dbs:

cd scripts
chmod +x download_all_data.sh

./download_all_data.sh ../AF_data/ reduced_dbs

If you have databases and parameters from a precedent AlphaFold installation, it is not necessary to repeat this step, just make sure that the paths inside databases.flag point to the right directories.

Quick start

Restraint file format

Restraints can be written as .tsv files, each line representing a separate AA pair restraint:

chain_id1  res_number1    chain_id2  res_number2    d
...

For example, the file restraints.tsv:

A   5   A   15  5.0
A   10  B   20  8.0

defines two restraints:

• intra-chain restraint at 5Å for residues 5 and 15 in chain A;
• inter-chain restraint at 8Å between residue 10 in chain A and residue 20 in chain B.

Restraints can also be distogram-like binned distance probabilities. In this case, the distance field is a comma-separated list of 64 float numbers, representing probabilities for each distance bin in AlphaFold's distogram (from [2.31, 2.62) Å in the first bin to [22, ∞) in the last bin):

A   5   A   15  0.0,0.0,0.1,0.2,0.3,0.2,0.1,0.1,0.0, ... ,0.0
...

Running resTrain

Restraint files are passed with the flag --restraints:

run_alphafold.py --fasta_paths examples/H1142/H1142.fasta --restraints example/H1142/restraints.tsv

Other flags

• --approximate: treats all restraints as maximum-distance restraints rather than exact distances (default: False)
• --kl: use KL divergence as loss, for example when restraining against distance distributions (default: False)
• --optimization_steps: number of gradient descent steps to optimize distogram loss (default: 10)
• --increase_seed_during_optimization: changes AlphaFold's random seed after each gradient descent steps (default: False)
• --pae_w: include pAE as loss term during search to encourage better quality predictions (default: 0.0)

OF3_resTrain

The OpenFold3 implementation of resTrain will be available soon.