Research map / Probabilistic and scientific modeling / Molecules and proteins

Swallowing the Bitter Pill: Simplified Scalable Conformer Generation

ICML 2024 · Probabilistic and scientific modeling › Molecules and proteins · approach: Diffusion for molecular conformations · 8 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 Molecular Conformer Fields (MCF), a non-equivariant PerceiverIO diffusion model that directly denoises 3D atom coordinates over Laplacian-eigenvector-parameterized conformer fields, and shows that scaling model size improves conformer generation on GEOM-QM9, GEOM-DRUGS and GEOM-XL.

Problem

Molecular conformer generation: predicting diverse low-energy 3D conformers of molecules, where prior methods rely on equivariant architectures or hard-coded domain structure such as torsional angles and rule-based rotatable-bond detection.

Approach

Parameterize each conformer as a field mapping graph-Laplacian eigenvector positional encodings of atoms to 3D coordinates, and train a DDPM-style diffusion model over these fields with a PerceiverIO score network that has no rotational equivariance or torsion modeling, so performance scales with model size.

Key results

• On GEOM-DRUGS, MCF-L reaches 84.7 mean recall COV and 66.8 mean precision COV versus 72.7 and 55.2 for Torsional Diffusion.
• On GEOM-QM9, MCF-B achieves 95.0 mean recall COV and 0.103 recall AMR, against 92.8 and 0.178 for Torsional Diffusion.
• On GEOM-DRUGS, performance improves monotonically with model size (S to B to L), with MCF-L attaining the best precision AMR of 0.618.

Measured effects reported in the paper

Surprising findings

• Randomly rotating each training conformer each epoch causes only marginal performance drops, and a larger model recovers most of the gap, suggesting roto-equivariance can be traded for scale.
• MCF achieves much lower E_min error than baselines but higher error on total energy E, indicating it can generate high-energy implausible conformers.

Limitations

• Computationally expensive training and inference; default sampling uses 1000 DDPM steps.
• May not perform well in low-data regimes or on sequential problems such as MD simulations.
• Generated conformers can have high total energy E, suggesting implausible samples.

Experimental setup

Datasets: GEOM-QM9 GEOM-DRUGS GEOM-XL

Benchmarks: GEOM-QM9 GEOM-DRUGS GEOM-XL

Metrics: COV AMR precision recall xTB ensemble property error

Baselines: CGCF GeoDiff GeoMol Torsional Diffusion OMEGA

Builds on: Diffusion Probabilistic Fields DDPM PerceiverIO Torsional Diffusion DDIM

Closest papers by meaning

• DiSCO: Diffusion Schrödinger Bridge for Molecular Conformer Optimization (AAAI 2024 · similarity 0.91)
• Equivariant Blurring Diffusion for Hierarchical Molecular Conformer Generation (NeurIPS 2024 · similarity 0.91)
• Enhancing Generalizability in Molecular Conformation Generation with METRIZATION-Informed Geometric Diffusion Pretraining (AAAI 2025 · similarity 0.90)
• Unlocking hidden biomolecular conformational landscapes in diffusion models at inference time (NeurIPS 2025 · similarity 0.90)
• Efficient Molecular Conformer Generation with SO(3)-Averaged Flow Matching and Reflow (ICML 2025 · similarity 0.90)
• Sampling 3D Molecular Conformers with Diffusion Transformers (NeurIPS 2025 · similarity 0.90)