Research map / Probabilistic and scientific modeling / Molecules and proteins
Learning Disentangled Equivariant Representation for Explicitly Controllable 3D Molecule Generation
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 E3WAE, an E(3)-equivariant Wasserstein autoencoder that factorizes the latent space of a fragment-based autoregressive 3D molecule generator into property and structural-context parts, enabling explicit control over generation.
Problem
Conditional generation of 3D drug-like molecules with explicit, independent control over molecular properties (asphericity, QED, SAS, logP) and structural context, including structure-based drug design against protein pockets.
Approach
Disentangle property and structural-context latents with an MMD-based Wasserstein regularizer plus a property prediction head, and add an alignment-based (Kabsch) coordinate loss so equivariant networks can be trained for autoregressive de novo 3D generation without reference structures.
Key results
- On property-targeting generation, E3WAE achieves the best MSE/MAE on 7 of 8 property settings across GEOM-Drugs and CrossDocked2020 (e.g., asphericity MSE 0.095 vs 0.176 for HierDiff on GEOM-Drugs).
- In structure-based generation on CrossDocked2020, E3WAE reaches Vina -5.891 for asphericity and -5.866 for QED, outperforming HierDiff and TargetDiff on both docking and property error.
- Ablation shows E3WAE (2.1M params) outperforms VAE variants E3fVAE and E3dVAE on asphericity and QED MSE/MAE.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| use E3WAE instead of HierDiff | property-targeted generation · GEOM-Drugs Asphericity | MSE: 0.176 -> 0.095 | Helped |
| use E3WAE instead of HierDiff | property-targeted generation · GEOM-Drugs QED and SAS | MSE: QED: 0.12 -> 0.072; SAS: 2.618 -> 1.563 | Helped |
| use E3WAE instead of HierDiff | property-targeted generation · GEOM-Drugs logP | MSE: 4.124 -> 4.490 | Hurt |
| use E3WAE instead of HierDiff | property-targeted generation · CrossDocked2020 | MSE: Asphericity: 0.107 -> 0.100; QED: 0.089 -> 0.062; SAS: 2.364 -> 2.356; logP: 5.752 -> 4.244 | Helped |
| replace E3WAE with FactorVAE-style E3fVAE (ablation) | E3WAE · GEOM-Drugs Asphericity | MSE: 0.095 -> 0.143 | Hurt |
| replace E3WAE with FactorVAE-style E3fVAE (ablation) | E3WAE · GEOM-Drugs QED | MSE: 0.072 -> 0.112 | Hurt |
| replace E3WAE with MI-upper-bound E3dVAE (ablation) | E3WAE · GEOM-Drugs Asphericity | MSE: 0.095 -> 0.101 | Hurt |
| replace E3WAE with MI-upper-bound E3dVAE (ablation) | E3WAE · GEOM-Drugs QED | MSE: 0.072 -> 0.098 | Hurt |
| use Kabsch alignment in coordinate reconstruction loss (ablation) | E3WAE · coordinate-loss training and validation | convergence | Helped |
Surprising findings
- E3WAE does not achieve top ranking on logP property-targeting generation, where HierDiff is better on MSE/MAE for GEOM-Drugs.
- Molecular stability on GEOM-Drugs is lower for E3WAE (25.6%) than HierDiff (31.7%).
Limitations
- Interpretability of context latent vectors remains a challenge and needs labeled supervision and domain priors.
- Generalization ability of the model is not thoroughly explored.
- Asphericity and QED cannot be fully disentangled, so multi-property control degrades slightly.
Experimental setup
Datasets: GEOM-Drugs CrossDocked2020
Benchmarks: GEOM-Drugs CrossDocked2020
Metrics: MSE MAE Vina Dock score QED RA MCF SAS delta logP delta MW atom stability
Baselines: EDM HierDiff TargetDiff G-SphereNet FactorVAE-style E3fVAE E3dVAE
Builds on: Wasserstein autoencoder HierDiff EDM G-SphereNet JT-VAE
Closest papers by meaning
- Manipulating 3D Molecules in a Fixed-Dimensional E(3)-Equivariant Latent Space (NeurIPS 2025 · similarity 0.90)
- Navigating the Design Space of Equivariant Diffusion-Based Generative Models for De Novo 3D Molecule Generation (ICLR 2024 · similarity 0.90)
- Geometric Representation Condition Improves Equivariant Molecule Generation (ICML 2025 · similarity 0.89)
- Lift Your Molecules: Molecular Graph Generation in Latent Euclidean Space (ICLR 2025 · similarity 0.89)
- Controllable 3D Molecular Generation for Structure-Based Drug Design Through Bayesian Flow Networks and Gradient Integration (NeurIPS 2025 · similarity 0.88)
- 3D Molecule Generation from Rigid Motifs via $\mathrm{SE}(3)$ Flows (NeurIPS 2026 · similarity 0.88)
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.