Learning the Language of Protein Structure

TL;DR

This paper introduces a vector-quantized autoencoder for protein structures, converting complex 3D data into discrete tokens, enabling effective structure modeling and generation with high fidelity, advancing computational protein design.

Contribution

The paper presents a novel vector-quantized autoencoder that tokenizes protein structures into discrete representations, facilitating improved modeling and generation of protein structures.

Findings

High-fidelity reconstructions with RMSD of 1-5 Å

A GPT model trained on codebooks can generate diverse protein structures

The method bridges the gap between sequence and structure modeling

Abstract

Representation learning and \emph{de novo} generation of proteins are pivotal computational biology tasks. Whilst natural language processing (NLP) techniques have proven highly effective for protein sequence modelling, structure modelling presents a complex challenge, primarily due to its continuous and three-dimensional nature. Motivated by this discrepancy, we introduce an approach using a vector-quantized autoencoder that effectively tokenizes protein structures into discrete representations. This method transforms the continuous, complex space of protein structures into a manageable, discrete format with a codebook ranging from 4096 to 64000 tokens, achieving high-fidelity reconstructions with backbone root mean square deviations (RMSD) of approximately 1-5 \AA. To demonstrate the efficacy of our learned representations, we show that a simple GPT model trained on our codebooks can generate novel, diverse, and designable protein structures. Our approach not only provides representations of protein structure, but also mitigates the challenges of disparate modal representations and sets a foundation for seamless, multi-modal integration, enhancing the capabilities of computational methods in protein design.