Research map / Probabilistic and scientific modeling / Molecules and proteins

ET-Flow: Equivariant Flow-Matching for Molecular Conformer Generation

NeurIPS 2024 · Probabilistic and scientific modeling › Molecules and proteins · approach: Diffusion for molecular conformations · 3 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 ET-Flow, an SE(3)-style equivariant transformer (TorchMD-NET based) trained with conditional flow matching from a harmonic prior with Kabsch alignment, generating all-atom molecular conformers on GEOM-DRUGS and GEOM-QM9.

Problem

Generating low-energy 3D molecular conformers from molecular graphs accurately, efficiently and with physical validity, without relying on RDKit-generated initial structures or large transformer models.

Approach

Combine an equivariant transformer operating on all-atom coordinates with flow matching, a harmonic prior encoding bonded-atom proximity, and RMSD (Kabsch) alignment of prior samples to data to shorten transport paths; add stochastic sampling and a post-hoc chirality correction.

Key results

• On GEOM-DRUGS, ET-Flow (8.3M params) achieves Precision Coverage mean 74.38 and AMR 0.541, beating MCF-L (242M) on precision metrics (66.8 coverage, 0.618 AMR).
• On GEOM-QM9, ET-Flow reaches Recall Coverage mean 96.47 and AMR 0.073, outperforming MCF (95.0, 0.103).
• Ensemble property median errors on GEOM-DRUGS are lower than MCF and Torsional Diffusion across E, mu, HOMO-LUMO gap and E_min (e.g., E_min 0.02 vs 0.04 for MCF).

Measured effects reported in the paper

Surprising findings

• ET-Flow with only 5 sampling steps still matches or exceeds previous methods on precision metrics on GEOM-DRUGS.
• The post-hoc chirality correction outperformed the SO(3)-equivariant architecture variant on GEOM-DRUGS precision.

Limitations

• Lower recall (diversity) than MCF-L at higher thresholds and on GEOM-XL generalization to large molecules.
• Requires post-hoc chirality correction because the O(3)-equivariant architecture cannot distinguish reflections; the SO(3) variant performed worse.
• Relies on RDKit-derived node features and bond graph inputs.

Experimental setup

Datasets: GEOM-DRUGS GEOM-QM9 GEOM-XL

Benchmarks: GEOM-DRUGS GEOM-QM9 GEOM-XL

Metrics: Coverage AMR Average Minimum RMSD Energy error dipole moment error HOMO-LUMO gap error minimum energy error

Baselines: GeoMol GeoDiff Torsional Diffusion MCF CGCF OMEGA

Builds on: flow matching TorchMD-NET Torsional Diffusion GeoDiff

Closest papers by meaning

• Equivariant Blurring Diffusion for Hierarchical Molecular Conformer Generation (NeurIPS 2024 · similarity 0.89)
• Efficient Molecular Conformer Generation with SO(3)-Averaged Flow Matching and Reflow (ICML 2025 · similarity 0.89)
• Controlled Generation with Equivariant Variational Flow Matching (ICML 2025 · similarity 0.88)
• Energy-Based Flow Matching for Generating 3D Molecular Structure (ICML 2025 · similarity 0.87)
• AlphaFold Meets Flow Matching for Generating Protein Ensembles (ICML 2024 · similarity 0.87)
• Navigating the Design Space of Equivariant Diffusion-Based Generative Models for De Novo 3D Molecule Generation (ICLR 2024 · similarity 0.86)