egurapha/igvae

A variational autoencoder that directly generates the 3D coordinates of immunoglobulin protein backbones.

Ig-VAE

Author: Raphael R. Eguchi

The Ig-VAE is a variational autoencoder that directly generates the 3D coordinates of immunoglobulin backbones using a torsion- and distance-based loss function that is rotationally and translationally invariant. The model is trained on structures from AbDb/AbYBank. IgVAE is intended for use with existing protein design suites such as Rosetta.

Requirements

Python 3.7
See environment.yml

Usage

python generate.py -n 100 -device cpu -seed 14 -outdir outputs/ 

Interpolation Example

Notes

The environment uses CPU Pytorch, but the script is GPU-compatible via:

python generate.py -device cuda

Citing and Licensing

This software is distributed under the BSD-3 license. The license file is available at license.txt. This work was published in PLoS Computational Biology in 2022. Please reference it using the following citation:

@article{10.1371/journal.pcbi.1010271,
    doi = {10.1371/journal.pcbi.1010271},
    author = {Eguchi, Raphael R. AND Choe, Christian A. AND Huang, Po-Ssu},
    journal = {PLOS Computational Biology},
    publisher = {Public Library of Science},
    title = {Ig-VAE: Generative modeling of protein structure by direct 3D coordinate generation},
    year = {2022},
    month = {06},
    volume = {18},
    url = {https://doi.org/10.1371/journal.pcbi.1010271},
    pages = {1-18},
    abstract = {While deep learning models have seen increasing applications in protein science, few have been implemented for protein backbone generation—an important task in structure-based problems such as active site and interface design. We present a new approach to building class-specific backbones, using a variational auto-encoder to directly generate the 3D coordinates of immunoglobulins. Our model is torsion- and distance-aware, learns a high-resolution embedding of the dataset, and generates novel, high-quality structures compatible with existing design tools. We show that the Ig-VAE can be used with Rosetta to create a computational model of a SARS-CoV2-RBD binder via latent space sampling. We further demonstrate that the model’s generative prior is a powerful tool for guiding computational protein design, motivating a new paradigm under which backbone design is solved as constrained optimization problem in the latent space of a generative model.},
    number = {6},
}

igvae's Issues

How to generate heavy chain backbones?

Hello, I run generate.py as README but only get length ~108aa, How to generate heavy chain backbones?

Loss functions and training script?

Hi,

Thanks for awesome publication. I couldn't find your loss functions and training script? Can you please provide more details thanks.

source code not finished?

Can you at least please fill up the details: loss functions, codes for setting up torsion angles? thanks

How to retrain the network

Thank you for releasing this wonderful work. Is there a plan to release the code for training? Particularly, the losses as described in the paper.