PeSTo: parameter-free geometric deep learning for accurate prediction of protein binding interfaces

PeSTo (Protein Structure Transformer) is a parameter-free geometric deep learning method to predict protein interaction interfaces from a protein structure. It is available for free without registration as an online tool (pesto.epfl.ch).

Installation

Download the source code and examples by cloning the repository.

git clone https://github.com/LBM-EPFL/PeSTo.git
cd PeSTo

The primary requirements for PeSTo are GEMMI to parse PDB files and PyTorch for the deep learning framework. During training, h5py is used to store the processed data in an optimized format. The predicted interfaces can be visualized using PyMOL or ChimeraX.

Using Anaconda

All the specific dependencies are listed in pesto.yml. The specific dependencies can be easily installed using Anaconda. Create and activate the environement with:

conda env create -f pesto.yml
conda activate pesto

Or installing manually de dependencies

conda create -n pesto python=3.9
conda activate pesto
conda install pytorch torchvision torchaudio pytorch-cuda=11.8 -c pytorch -c nvidia
conda install numpy scipy pandas matplotlib scikit-learn h5py tqdm
conda install gemmi tensorboard -c conda-forge

Using virtualenv

Alternatively, it is possible to create a local environment using virtualenv and install all the dependencies.

virtualenv pesto
source pesto/bin/activate
pip install -r requirements.txt

Application

A set a Jupyter notebooks and python scripts are available to apply our trained model. Start a JupyterLab session with:

jupyter-lab

Interfaces predictions

The PeSTo model can be applied to PDB files using the apply_model.ipynb notebook. Specify the path to the folder containing the PDB files using the data_path variable. We recommand using the latest model (i_v4_1) but other pre-trained variants are available. The predictions can be run on CPU or GPU.

The predictions for the interfaces are stored in the b-factor field of the PDB files using a value from 0 (no interface) to 1 (interface). The predicted interfaces can be visualized with a color gradient per residue. This can be done in PyMOL with,

spectrum b, blue_white_red, all, 0, 1

Or in ChimeraX with

color bfactor palette "#2B59C3:#D1D1D1:#D7263D" range 0,1

Reproducibility

We provide all the Jupyter notebooks and scripts used to obtain and process the data, train and evaluate the model. The latest model (i_v4_1) is used for the benchmarks and results shown in the paper.

Interfaces prediction

All bioassemblies used are downloaded from RCSB PDB. The subunits are split into training, testing and validation dataset according to 30% sequence similarity clusters (processing/split_dataset.ipynb). Finaly, we preprocess the structure, detect the interfaces within complexes and store the features and labels into an optimized HDF5 format (processing/build_dataset.py).

The model folder contains the scripts to train the model as well as the selected pre-trained models in model/save. The benchmark and comparison can be reproduced with the interface_*.ipynb notebooks.

MD analysis

Scripts and functions to perform predictions and analysis on MD are found in the md_analysis folder. Molecular dynamics are loaded using MDTraj. An utility tool called data_manager was developed to easily locate simulations within a defined tree-folder structure. We also developed analysis tools based on MDTraj (mdtraj_utils).

Interfaceome

The interfaceome folder contains the Jupyter notebooks and python scripts used to download, process and analyse the data. All the AlphaFold-predicted structures used can be downloaded freely from the AlphaFold Protein Structure Database. Only the corresponding UniProt data is downloaded (interfaceome/download_uniprot.py). We also download the PAE from the AlphaFold Protein Structure Database (interfaceome/download_af_pae.py).

Other available pre-trained models

We provide 4 variants of the trained PeSTo models:

1. i_v3_0 is composed of 16 geometric transformers and uses both atom element and residue type information
2. i_v3_1 is composed of 16 geometric transformers and uses both atom element and residue type information but only predicts protein-protein interfaces
3. i_v4_0 is composed of 16 geometric transformers and uses only atom element
4. i_v4_1 is composed of 32 geometric transformers and uses only atom element

Web server

It is possible to use PeSTo without requiring the user to install it using our web server freely available at pesto.epfl.ch. PDB ID, UniProt ID or PDB files are accepted. The predictions are fast and can be visualized directely in the browser or downloaded as PDB files.

License

This work is licensed under a Creative Commons Attribution-NonCommercial-ShareAlike 4.0 International License.

Reference

Krapp, L.F., Abriata, L.A., Cortés Rodriguez, F. et al. PeSTo: parameter-free geometric deep learning for accurate prediction of protein binding interfaces. Nat Commun 14, 2175 (2023). https://doi.org/10.1038/s41467-023-37701-8