Research map / Probabilistic and scientific modeling / Molecules and proteins
An All-Atom Generative Model for Designing Protein Complexes
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.
Introduces APM, an all-atom flow-matching generative model (sequence+backbone, sidechain torsion, and refinement modules) that natively designs multi-chain protein complexes and supports folding, inverse folding, and SFT/zero-shot antibody and peptide design.
Problem
Generative modeling of multi-chain protein complexes at all-atom resolution, where inter-chain atom-level interactions matter for binding, and existing methods mostly handle single chains or connect chains with poly-G linkers.
Approach
Three-module iterative architecture: a flow-matching Seq&BB module co-generates sequence and backbone with decoupled noise levels, a Sidechain module predicts torsion angles (four per residue) for all-atom completion, and a Refine module corrects the generated protein with all-atom information at late timesteps (t>=0.8); trained on mixed single- and multi-chain data with native inter-chain attenti
Key results
- On single-chain folding/inverse folding, APM reaches 50.44% AAR and 0.86 TM vs ESM3's 49.50% AAR and 0.83 TM, but has worse median RMSD (2.64 vs 2.27).
- On RAbD antibody CDR-H3 design, APM SFT achieves 41.20% AAR and 2.08 RMSD, outperforming dyMEAN, DiffAb, and AbDPO.
- On peptide design, APM SFT reaches 69.34% negative binding energy (%<0) and 11.29% DockQ>=0.8, while RFDiffusion has the highest foldability (46.28% success).
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use only Seq&BB, disabling Sidechain and Refine (ablation) | APM · Multi-chain generation, chain lengths 50-100 | ΔG_RSC: -72.44/-71/91 -> -64.30/-67.30 | Hurt |
| Use only Seq&BB, disabling Sidechain and Refine (ablation) | APM · Multi-chain generation, chain lengths 100-100 | ΔG_RSC: -91.61/-94.54 -> -36.74/-69.30 | Hurt |
| Use only Seq&BB, disabling Sidechain and Refine (ablation) | APM · Multi-chain generation, chain lengths 100-200 | ΔG_RSC: -44.02/-39.42 -> -3.42/-33.71 | Hurt |
| Switch from simultaneous chain generation to chain-by-chain generation (ablation) | APM · Multi-chain generation, chain lengths 50-100, 100-100, and 100-200 | ΔG_RSC: 312.66/-14.05; 235.52/-13.77; 1073.51/-11.01 | Hurt |
| Use zero-shot sampling instead of supervised fine-tuning | APM peptide design · LNR | DockQ: 0.40 -> 0.24 | Hurt |
| Use zero-shot sampling instead of supervised fine-tuning | APM antibody design · RAbD | ΔG: 91.64 -> 81.12 | Mixed |
| Use zero-shot sampling instead of supervised fine-tuning | APM antibody design · RAbD | AAR (%): 41.20 -> 28.35 | Hurt |
| Use APM instead of ESM3(1.4B) | Protein folding · PDB date split | TM: 0.828±0.002 -> 0.856±0.002 | Helped |
| Use APM instead of ESM3(1.4B) | Protein folding · PDB date split | RMSD: 4.708±0.094 -> 4.828±0.077 | Hurt |
| Use APM instead of Boltz-1 without MSA | Multi-chain folding · 273-protein test set | TM: 0.44/0.45 -> 0.64/0.62 | Helped |
| Use supervised fine-tuning | APM peptide design · LNR | DockQ: 0.24 -> 0.40 | Helped |
| Use APM instead of the strongest listed baseline | Unconditional protein generation · Length 100 | scTM: 0.91 -> 0.96 | Helped |
Surprising findings
- Zero-shot APM generated antibodies show strong binding energy (-81.12 average dG) despite poor AAR (28.35%) and RMSD (5.81), suggesting different binding patterns than natives.
- Performance degrades for complexes with more than two chains, with increased unstructured backbones or over-idealized secondary structure.
Limitations
- Folding performance for multi-chain and single-chain tasks still lags, with worse median RMSD than ESM3.
- Performance declines for complexes with more than two chains.
- Refine module functionality is relatively restricted and few downstream tasks are evaluated.
Experimental setup
Datasets: PDB Swiss-Prot AFDB PDB Biological Assemblies SAbDab RAbD PepBench LNR
Benchmarks: PDB date split (MultiFlow) RAbD LNR PepBench
Metrics: RMSD TM-score scTM AAR perplexity scRMSD ΔG DockQ pLDDT ipTM
Baselines: ESMFold ProteinMPNN ESM3 MultiFlow ProteinGenerator ProtPardelle Boltz-1 Chroma PPFlow DiffPP PepGLAD RFDiffusion
Builds on: AlphaFold2 MultiFlow Discrete Flow Matching ESMFold ProteinMPNN
Closest papers by meaning
- All-atom inverse protein folding through discrete flow matching (ICML 2025 · similarity 0.89)
- ESM All-Atom: Multi-Scale Protein Language Model for Unified Molecular Modeling (ICML 2024 · similarity 0.89)
- P(all-atom) Is Unlocking New Path For Protein Design (ICML 2025 · similarity 0.87)
- ProteinBench: A Holistic Evaluation of Protein Foundation Models (ICLR 2025 · similarity 0.87)
- Generative Modeling of Full-Atom Protein Conformations using Latent Diffusion on Graph Embeddings (NeurIPS 2025 · similarity 0.87)
- Structure Language Models for Protein Conformation Generation (ICLR 2025 · similarity 0.87)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.