Research map / Probabilistic and scientific modeling / Molecules and proteins
Energy-Based Flow Matching for Generating 3D Molecular Structure
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 IDFlow, an energy-based flow matching method that trains an idempotent flow map (predictor-refiner) for molecular docking and protein backbone generation, outperforming HarmonicFlow and FrameFlow at matched compute.
Problem
Flow matching and diffusion models for 3D molecular structure generation (docking, protein backbone generation) produce suboptimal samples and lack explicit energy-landscape shaping; a computational-budget-matched improvement is sought.
Approach
Treat the x1-parameterized flow map as an energy-based model with reconstruction-error energy, and train it to be idempotent (f(f(x)) ≈ f(x)) via an extra refinement loss on the model's own detached predictions, then sample with a predictor-refiner Euler scheme using two NFEs per step.
Key results
- Single-ligand docking on time-split radius pocket: IDFlow reaches 34.7% RMSD<2Å vs 28.3% for HarmonicFlow at matched 20 NFEs.
- Multi-ligand docking on Binding MOAD: IDFlow gets 83.1% RMSD<5Å vs 75.0% for HarmonicFlow, with RMSD<2Å on par (43.8% vs 44.4%).
- Protein backbone generation on PDB with 200 NFEs: IDFlow designability 0.927 vs 0.824 for FrameFlow (200 NFEs).
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add IDFlow's idempotent refinement training | HarmonicFlow · PDBBind sequence similarity split, Distance-Pocket | %<2: 30.1 -> 35.6 | Helped |
| add IDFlow's idempotent refinement training | HarmonicFlow · PDBBind sequence similarity split, Radius-Pocket | Med. RMSD: 3.4 -> 3.7 | Hurt |
| add IDFlow's idempotent refinement training | HarmonicFlow · PDBBind time split, Radius-Pocket | %<2: 28.3 -> 34.7 | Helped |
| add IDFlow's idempotent refinement training | HarmonicFlow · Binding MOAD multi-ligand docking | %<5: 75.0 -> 83.1 | Helped |
| add IDFlow's idempotent refinement training | FrameFlow · PDB protein backbone generation | Designability: 0.824 ± 0.037 -> 0.871 ± 0.013 | Helped |
| increase inference from 100 to 200 NFEs | IDFlow · PDB protein backbone generation | Designability: 0.871 ± 0.013 -> 0.927 ± 0.020 | Helped |
| train idempotency with K_max = 1 rather than K_max = 0 (ablation) | IDFlow · PDBBind time split, Radius-Pocket | %<2: 28.3 -> 34.2 ± 2.42 | Helped |
| train idempotency with K_max = 2 rather than K_max = 1 (ablation) | IDFlow · PDBBind time split, Radius-Pocket | %<2: 34.2 ± 2.42 -> 34.7 ± 0.17 | No clear effect |
| use more than one refinement during test-time sampling (ablation) | IDFlow predictor-refiner sampler · PDBBind time split, Radius-Pocket | docking performance | Hurt |
| set the refiner timestep to 1 (ablation) | IDFlow · structure generation | performance | No clear effect |
Surprising findings
- Increasing refinement iterations beyond one at test time degrades performance, attributed to ODE discretization error.
- Increasing K_max beyond 2 during training gives diminishing returns, and the timestep used for the refiner training was found not to matter in practice.
Limitations
- Increased training cost from multiple forward passes.
- Extra test-time refinement increases discretization error of ODE sampling.
- Energy is only the reconstruction error; no biophysical or chemically plausible energy is used.
Experimental setup
Datasets: PDBBind v2020 Binding MOAD SCOPe PDB
Benchmarks: PDBBind v2020 time split PDBBind v2020 30% sequence similarity split Binding MOAD SCOPe PDB protein backbone
Metrics: RMSD <2Å fraction RMSD <5Å fraction median RMSD designability diversity novelty helix content strand content sampling time
Baselines: HarmonicFlow Product Space Diffusion EigenFold Diffusion FrameFlow FoldFlow2 GAFL FoldFlow-SFM FoldFlow-OT FrameDiff RFdiffusion
Builds on: flow matching HarmonicFlow FrameFlow tensor field network self-conditioning
Closest papers by meaning
- SE(3)-Stochastic Flow Matching for Protein Backbone Generation (ICLR 2024 · similarity 0.92)
- AlphaFold Meets Flow Matching for Generating Protein Ensembles (ICML 2024 · similarity 0.91)
- Learning conformational ensembles of proteins based on backbone geometry (NeurIPS 2025 · similarity 0.90)
- Sequence-Augmented SE(3)-Flow Matching For Conditional Protein Generation (NeurIPS 2024 · similarity 0.90)
- Swallowing the Bitter Pill: Simplified Scalable Conformer Generation (ICML 2024 · similarity 0.89)
- Energy Matching: Unifying Flow Matching and Energy-Based Models for Generative Modeling (NeurIPS 2025 · similarity 0.89)
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.