Research map / Probabilistic and scientific modeling / Molecules and proteins
3D Structure Prediction of Atomic Systems with Flow-based Direct Preference Optimization
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 FlowDPO, which applies Direct Preference Optimization to flow-matching models over arbitrary Gaussian paths, using RMSD-ranked self-generated candidates to suppress hallucinated structures in antibody CDR and crystal structure prediction.
Problem
Generative 3D structure prediction for atomic systems (antibody CDRs, crystals) relies mostly on diffusion models with limited exploration of probability paths, and maximum-likelihood training yields hallucinated low-quality samples.
Approach
Generate multiple candidate structures from a pretrained flow-matching model, rank them by distance to ground truth (RMSD) to build an automatic preference dataset, and apply DPO; the authors derive a DPO objective expressed via MSE terms that holds for arbitrary Gaussian paths (VE, VP, OT).
Key results
- On antibody CDRs, DPO lowers worst-case Cα RMSD for OT path on H3 from 4.81 to 4.28 and for VP path from 5.01 to 4.42.
- On crystal structure prediction, OT+VE+DPO raises MP-20 match rate from 58.94% to 62.47% and MPTS-52 from 18.91% to 20.27%.
- DPO consistently improves all six CDRs and all three crystal datasets in the reported tables, with the largest gain on CDR-H3.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| apply DPO to the VP path (ablation) | Antibody structure prediction · CDR-H3 | Cα-w RMSD: 5.01 -> 4.42 | Helped |
| apply DPO to the VP path (ablation) | Antibody structure prediction · CDR-H3 | Cα RMSD: 3.77 -> 3.44 | Helped |
| apply DPO to the OT path (ablation) | Antibody structure prediction · CDR-H3 | bb-w RMSD: 4.83 -> 4.23 | Helped |
| apply DPO to the OT path (ablation) | Antibody structure prediction · CDR-H3 | bb RMSD: 3.70 -> 3.32 | Helped |
| apply DPO to VP + VE paths (ablation) | Crystal structure prediction · Perov-5 | MR (%): 52.02 -> 53.47 | Helped |
| apply DPO to VP + VE paths (ablation) | Crystal structure prediction · MP-20 | MR (%): 51.49 -> 59.98 | Helped |
| apply DPO to VP + VE paths (ablation) | Crystal structure prediction · MPTS-52 | MR (%): 12.19 -> 14.75 | Helped |
| apply DPO to OT + OT paths (ablation) | Crystal structure prediction · Perov-5 | MR (%): 53.95 -> 55.56 | Helped |
| apply DPO to OT + OT paths (ablation) | Crystal structure prediction · MP-20 | MR (%): 57.40 -> 59.62 | Helped |
| apply DPO to OT + OT paths (ablation) | Crystal structure prediction · MPTS-52 | MR (%): 17.40 -> 22.36 | Helped |
| apply DPO to OT + VE paths (ablation) | Crystal structure prediction · Perov-5 | MR (%): 52.29 -> 53.94 | Helped |
| apply DPO to OT + VE paths (ablation) | Crystal structure prediction · MP-20 | MR (%): 58.94 -> 62.47 | Helped |
Surprising findings
- On Perov-5 and MPTS-52, DPO sometimes lowers match rate for the VP+VE path (e.g., MPTS-52 from 12.19% to 14.75% actually rises; the VP+VE Perov-5 match rate rises but RMSE is nearly unchanged), showing gains are not uniform across settings.
Limitations
- Theoretical derivation of DPO is restricted to Gaussian probability paths.
- Evaluation is purely computational; no wet-lab validation.
- No error bars or repeated runs due to computational cost.
Experimental setup
Datasets: SAbDab DiffAb test set (19 complexes) Perov-5 MP-20 MPTS-52
Benchmarks: DiffAb antibody CDR test set (19 complexes) Perov-5 MP-20 MPTS-52
Metrics: RMSD Cα RMSD backbone worst-of-20 RMSD match rate RMSE normalized
Baselines: DiffAb (VP path) OT path P-cG-SchNet CDVAE VP + VE path (DiffCSP) MEAN regression
Builds on: Direct Preference Optimization Diffusion-DPO Flow Matching DiffAb DiffCSP
Closest papers by meaning
- EP-Flow: Disordered Crystal Structure Prediction without Site-level Annotation (NeurIPS 2026 · similarity 0.88)
- Energy-Based Flow Matching for Generating 3D Molecular Structure (ICML 2025 · similarity 0.87)
- Prior-Guided Flow Matching for Target-Aware Molecule Design with Learnable Atom Number (NeurIPS 2025 · similarity 0.86)
- AlphaFold Meets Flow Matching for Generating Protein Ensembles (ICML 2024 · similarity 0.86)
- A Periodic Bayesian Flow for Material Generation (ICLR 2025 · similarity 0.85)
- MOFFlow: Flow Matching for Structure Prediction of Metal-Organic Frameworks (ICLR 2025 · similarity 0.85)
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.