Atom-level Protein Representation Learning Improves Protein Structure Prediction

2026-05-21

AI Summary

This work addresses the limitation of existing protein representation methods, which predominantly rely on sequence or functional annotations and inadequately capture structural information. The authors propose TriProRep, a novel framework that jointly models three residue-level structural views—amino acid identity, backbone geometry, and local all-atom geometry—for the first time. By leveraging vector-quantized variational autoencoders (VQ-VAE) to generate discrete structural codes, TriProRep performs structure-aware generative pretraining through a masked token reconstruction task to learn high-quality representations. The approach further incorporates multi-view alignment and enhanced discriminative learning. A new benchmark, RepSP, is introduced to systematically evaluate structural prediction capabilities. Experiments demonstrate that TriProRep significantly outperforms current methods on RepSP across homodimer co-folding, interaction property prediction, and monomer structure prediction tasks, while remaining competitive on established benchmarks.

Abstract

Recent advances in generative modeling show that pretrained representations can improve generation as conditioning features or alignment targets. Motivated by this, we study protein representations for predicting structures beyond conventional function annotation. We propose TriProRep, a structure-aware pretraining method that jointly models three aligned residue-level views: amino-acid identity, backbone geometry, and local full-atom geometry, discretely encoded via VQ-VAE tokenizers. By pretraining to recover original tokens from generator-corrupted views, TriProRep learns to distinguish plausible but incorrect cross-view augmentations from the original protein. We further introduce RepSP, a benchmark for evaluating protein representations in structure-predictive settings. RepSP tests three uses of representations: homodimer co-folding from apo-chain representations, residue-level prediction of homodimer-derived interaction properties, and representation-aligned monomer structure prediction. Across these tasks, TriProRep improves over sequence-only and prior structure-aware representation models, while maintaining competitive performance on conventional benchmarks.

Problem

• protein representation
• structure prediction
• atom-level modeling
• homodimer co-folding
• residue-level prediction

Innovation

• TriProRep
• structure-aware pretraining
• VQ-VAE tokenization
• atom-level representation
• RepSP benchmark