ProxelGen: Generating Proteins as 3D Densities

TL;DR

ProxelGen introduces a novel 3D density-based generative model for proteins, enabling more flexible shape conditioning and achieving high-quality, novel protein structure samples compared to existing point cloud methods.

Contribution

It is the first to operate on 3D densities (proxels) for protein generation, combining a 3D CNN VAE with diffusion models for improved sample quality and conditioning capabilities.

Findings

ProxelGen achieves higher novelty and better FID scores than state-of-the-art models.

Samples demonstrate comparable designability to training data.

The model enables flexible shape conditioning in protein design.

Abstract

We develop ProxelGen, a protein structure generative model that operates on 3D densities as opposed to the prevailing 3D point cloud representations. Representing proteins as voxelized densities, or proxels, enables new tasks and conditioning capabilities. We generate proteins encoded as proxels via a 3D CNN-based VAE in conjunction with a diffusion model operating on its latent space. Compared to state-of-the-art models, ProxelGen's samples achieve higher novelty, better FID scores, and the same level of designability as the training set. ProxelGen's advantages are demonstrated in a standard motif scaffolding benchmark, and we show how 3D density-based generation allows for more flexible shape conditioning.