Research map / Probabilistic and scientific modeling / Molecules and proteins

Atom-level Protein Representation Learning Improves Protein Structure Prediction

NeurIPS 2026 · Probabilistic and scientific modeling › Molecules and proteins · approach: Generative protein structure design · 0 citations (OpenAlex)

Read the paperCheck your idea against 108,967 papers

AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Proposes TriProRep, a three-view (amino-acid, backbone VQ tokens, full-atom VQ tokens) protein representation model pretrained with ELECTRA-style generator-corrupted token recovery, and RepSP, a homodimer-based benchmark for structure-predictive use of representations.

Problem

Existing protein representation benchmarks (EC/GO) do not test whether structure-aware representations expose residue geometry useful for structure prediction such as homodimer co-folding, interface prediction, and representation-aligned monomer folding.

Approach

Jointly model amino-acid, backbone-geometry, and full-atom-geometry residue tokens (the latter newly introduced via an atom-level VQ-VAE) and pretrain with generator-corrupted token recovery across views, so the model learns cross-view consistency; evaluate on a new homodimer-based benchmark RepSP.

Key results

• On RepSP homodimer flexible docking, TriProRep 650M reaches DockQ 0.477 vs 0.443 for SaProt-650M, and 2.8B reaches DockQ 0.499 and TM-score 0.724.
• Per-residue probing at 650M improves binding-site AUPRC to 0.8524 vs 0.8273 for SaProt; 2.8B reaches 0.8650.
• In REPA distillation into SimpleFold-100M, TriProRep 650M targets give TM-score 0.5822 vs 0.5659 for SaProt and baseline 0.5422.

Surprising findings

• Sequence-only ESM2-3B underperforms several smaller structure-aware models on most homodimer docking metrics.
• At huge scale, TriProRep and ProstT5 achieve comparable distillation results, so performance saturates there.

Limitations

• Evaluation relies on predicted AFDB structures; validation on experimentally resolved structures is limited to a small RCSB subset.
• RepSP uses homodimers only, which are simpler than general heteromeric complexes.
• On EC/GO, TriProRep is only competitive rather than best; it does not clearly beat the strongest baselines.

Experimental setup

Datasets: AlphaFold Protein Structure Database ESMAtlas AFDB multimer release (RepSP) RCSB PDB structures deposited after June 1, 2023 Enzyme Commission Gene Ontology

Benchmarks: RepSP EC GO RCSB real-world homodimers

Metrics: DockQ iRMSD LRMSD Fnat TM-score LDDT RMSD AUPRC AUROC Pearson

Baselines: ESM2 SaProt S-PLM MIF-ST ESM3 ProstT5

Builds on: ELECTRA AminoAseed SimpleFold ESM2 Boltz

Closest papers by meaning

• Pre-Training Protein Bi-level Representation Through Span Mask Strategy On 3D Protein Chains (ICML 2024 · similarity 0.89)
• Evaluating Representation Learning on the Protein Structure Universe (ICLR 2024 · similarity 0.88)
• Protein Structure Tokenization: Benchmarking and New Recipe (ICML 2025 · similarity 0.87)
• FoldToken: Learning Protein Language via Vector Quantization and Beyond (AAAI 2025 · similarity 0.86)
• Learning Complete Protein Representation by Dynamically Coupling of Sequence and Structure (NeurIPS 2024 · similarity 0.86)
• Protein 3D Graph Structure Learning for Robust Structure-Based Protein Property Prediction (AAAI 2024 · similarity 0.86)