Research map / Image generation / Autoregressive and latent generation

Navigating the Design Space of Equivariant Diffusion-Based Generative Models for De Novo 3D Molecule Generation

ICLR 2024 · Image generation › Autoregressive and latent generation · approach: Equivariant diffusion for 3D molecules · 7 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.

Systematically explores design choices (x0 vs epsilon parameterization, discrete vs continuous diffusion, time-dependent loss weighting, PubChem3D pretraining, chemical features) for E(3)-equivariant graph-attention diffusion, yielding EQGAT-diff that beats EDM and MiDi on QM9/GEOM-Drugs and transfers to CrossDocked ligand design.

Problem

Suboptimal 3D de novo molecule generation with diffusion models on large molecules and limited training data; the design space of equivariant diffusion models (parameterization, loss weighting, discrete vs continuous state spaces, input features) is poorly understood.

Approach

Uses an E(3)-equivariant graph attention denoiser (EQGAT-diff) with x0-parameterization, categorical diffusion for elements and bonds, a truncated-SNR time-dependent loss weighting, PubChem3D pretraining with implicit hydrogens followed by fine-tuning, and extra chemical features (aromaticity, ring, hybridization) in the diffusion process.

Key results

• On GEOM-Drugs, EQGAT-diff with discrete x0 diffusion and time-dependent weighting reaches 91.60% molecule stability vs 87.59% with uniform weighting; with additional features and fine-tuning it reaches 95.01% molecule stability, outperforming MiDi (89.7%).
• On CrossDocked, the fine-tuned EQGAT-diff reaches a mean Vina score of -7.423 (All), better than TargetDiff (-7.318) and DiffSBDD (-6.950); TargetDiff is slightly better on Top-10% (-9.669 vs -9.571).
• Fine-tuning a PubChem3D-pretrained model on only 25% of GEOM-Drugs already reaches 90.82% molecule stability, above the 87.59% of the model trained from scratch on 100% of the data with uniform weighting.

Measured effects reported in the paper

Surprising findings

• Time-dependent loss weighting w_s(t) departs from the common uniform weighting and substantially speeds convergence.
• Pretraining with explicit hydrogens on PubChem3D performed worse after fine-tuning than implicit-hydrogen pretraining despite taking about three times as long.

Limitations

• MiDi results could not be reproduced with the official code, so baseline comparisons rely on the authors' re-training.
• Lower diversity than MolDiff in the MolDiff pipeline comparison (0.320 vs 0.422).
• Transfer results on QM9 and CrossDocked are less extensively analysed than GEOM-Drugs.

Experimental setup

Datasets: QM9 GEOM-Drugs PubChem3D CrossDocked

Benchmarks: QM9 GEOM-Drugs CrossDocked

Metrics: molecule stability validity connected components novelty uniqueness diversity AtomsTV BondsTV valency W1 bond length W1

Baselines: EDM MiDi MolDiff TargetDiff DiffSBDD

Builds on: EQGAT EDM MiDi DDPM truncated SNR weighting

Closest papers by meaning

• Equivariant Blurring Diffusion for Hierarchical Molecular Conformer Generation (NeurIPS 2024 · similarity 0.91)
• Towards Unified and Lossless Latent Space for 3D Molecular Latent Diffusion Modeling (NeurIPS 2025 · similarity 0.91)
• Equivariant Diffusion for Crystal Structure Prediction (ICML 2024 · similarity 0.91)
• All-atom Diffusion Transformers: Unified generative modelling of molecules and materials (ICML 2025 · similarity 0.90)
• Equivariant Neural Diffusion for Molecule Generation (NeurIPS 2024 · similarity 0.90)
• Learning Disentangled Equivariant Representation for Explicitly Controllable 3D Molecule Generation (AAAI 2025 · similarity 0.90)