Learning Protein Family Manifolds with Smoothed Energy-Based Models
Abstract
We resolve difficulties in training and sampling from discrete energy-based models (EBMs) by learning a smoothed energy landscape, sampling the smoothed data manifold with Langevin Markov chain Monte Carlo, and projecting back to the true data manifold with one-step denoising. Our formalism combines the attractive properties of EBMs and improved sample quality of score-based models, while simplifying training and sampling by requiring only a single noise scale. We demonstrate the robustness of our approach on generative modeling of antibody proteins.
Cite
Text
Frey et al. "Learning Protein Family Manifolds with Smoothed Energy-Based Models." ICLR 2023 Workshops: Physics4ML, 2023.Markdown
[Frey et al. "Learning Protein Family Manifolds with Smoothed Energy-Based Models." ICLR 2023 Workshops: Physics4ML, 2023.](https://mlanthology.org/iclrw/2023/frey2023iclrw-learning/)BibTeX
@inproceedings{frey2023iclrw-learning,
title = {{Learning Protein Family Manifolds with Smoothed Energy-Based Models}},
author = {Frey, Nathan C. and Berenberg, Dan and Kleinhenz, Joseph and Hotzel, Isidro and Lafrance-Vanasse, Julien and Kelly, Ryan Lewis and Wu, Yan and Rajpal, Arvind and Ra, Stephen and Bonneau, Richard and Cho, Kyunghyun and Loukas, Andreas and Gligorijevic, Vladimir and Saremi, Saeed},
booktitle = {ICLR 2023 Workshops: Physics4ML},
year = {2023},
url = {https://mlanthology.org/iclrw/2023/frey2023iclrw-learning/}
}