Research map / Probabilistic and scientific modeling / Molecules and proteins
All-atom inverse protein folding through discrete flow matching
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 ADFLIP, a discrete flow-matching model for all-atom inverse protein folding that incorporates predicted side chains, ensemble sampling over NMR conformers, and training-free classifier guidance.
Problem
Inverse protein folding for biomolecular complexes containing non-protein components (ligands, nucleotides, metal ions) and for dynamic complexes with multiple structural states, where existing backbone-only or all-atom methods perform poorly.
Approach
Discrete flow matching with masked sequence states, a multi-scale atom/residue GNN denoiser, progressive side-chain packing (PIPPack) during sampling, ensemble averaging of predictions across conformers, and training-free guidance that plugs pre-trained regressors into the denoiser's expected output.
Key results
- On interaction-residue recovery, ADFLIP reaches 62.19% (ligand), 50.21% (nucleotide), and 75.73% (metal ion), versus 59.21%, 46.14%, and 69.31% for LigandMPNN.
- On NMR ensembles, using the full ensemble raises recovery over single-structure input by 8.6 points for ligands and 5.8 points for nucleotides.
- In binding-affinity guidance with DSMBind, affinity gain rises from 41.9% (unguided) to 58.1% (guided), with foldability 91.4%.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use ADFLIP instead of LigandMPNN for all-atom ligand-complex inverse folding | ADFLIP · Ligand complexes | Recovery Rate (%): 59.21 -> 62.19 | Helped |
| Use ADFLIP instead of LigandMPNN for all-atom nucleotide-complex inverse folding | ADFLIP · Nucleotide complexes | Recovery Rate (%): 46.14 -> 50.21 | Helped |
| Use ADFLIP instead of LigandMPNN for all-atom metal-ion-complex inverse folding | ADFLIP · Metal ion complexes | Recovery Rate (%): 69.31 -> 75.73 | Helped |
| Use multiple NMR conformations instead of a single conformation (ablation) | ADFLIP · NMR ligand ensembles | Recovery Rate (%): 41.48 -> 50.08 | Helped |
| Use multiple NMR conformations instead of a single conformation (ablation) | ADFLIP · NMR nucleotide ensembles | Recovery Rate (%): 34.48 -> 40.33 | Helped |
| Use multiple NMR conformations instead of a single conformation (ablation) | ADFLIP · NMR metal ion ensembles | Recovery Rate (%): 65.44 -> 68.01 | Helped |
| Use multiple NMR conformations instead of a single conformation (ablation) | ADFLIP · NMR structure ensembles | RMSD (Å): 9.11 to 7.61 Å | Helped |
| Guide generation with DSMBind affinity predictions | ADFLIP · Protein-ligand complexes | Affinity gain: 41.9% to 58.1% | Helped |
| Guide generation with DSMBind affinity predictions | ADFLIP · Protein-ligand complexes | Foldability: 100% -> 91.4% | Hurt |
Limitations
- Lower sequence recovery for nucleotide complexes, possibly because a single conformational state does not capture nucleotide dynamics.
- Foldability on NMR ensembles remains low (RMSD above 7 Å), and guidance by DSMBind is limited by the fixed backbone and limited side-chain sampling.
- Sequences have not been validated experimentally.
Experimental setup
Datasets: PDB all-atom complexes (LigandMPNN curation) NMR structure ensembles (219 ensembles, curated by authors)
Benchmarks: LigandMPNN PDB test set (ligand, nucleotide, metal ion) NMR ensemble test set
Metrics: perplexity sequence recovery rate RMSD TM-score pLDDT foldability
Baselines: LigandMPNN ProteinMPNN PiFold
Builds on: LigandMPNN AlphaFold3 discrete flow matching PIPPack
Closest papers by meaning
- An All-Atom Generative Model for Designing Protein Complexes (ICML 2025 · similarity 0.89)
- AlphaFold Meets Flow Matching for Generating Protein Ensembles (ICML 2024 · similarity 0.89)
- Few-Step Cofolding with All-Atom Flow Maps (NeurIPS 2026 · similarity 0.88)
- Bridge-IF: Learning Inverse Protein Folding with Markov Bridges (NeurIPS 2024 · similarity 0.88)
- Sequence-Augmented SE(3)-Flow Matching For Conditional Protein Generation (NeurIPS 2024 · similarity 0.88)
- Protein Inverse Folding From Structure Feedback (NeurIPS 2025 · similarity 0.88)
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.