Research map / Probabilistic and scientific modeling / Molecules and proteins
Manipulating 3D Molecules in a Fixed-Dimensional E(3)-Equivariant Latent Space
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 MolFLAE, a VAE with a fixed-length, E(3)-equivariant latent space (virtual nodes) decoded by a Bayesian Flow Network, enabling zero-shot 3D molecule editing, interpolation and hGR drug optimization.
Problem
Latent spaces of 3D molecule generative models vary in length with atom count and are not E(3)-equivariant, which prevents interpolation and zero-shot molecule editing; existing task-specific editing models lack flexibility.
Approach
Append a fixed number (N_Z=10) of learnable virtual nodes to the molecule point cloud, encode them with an E(3)-equivariant GNN (MolCRAFT backbone) into a fixed-dimensional latent, regularize with a VAE KL term, and decode with a BFN conditioned on the latent, enabling latent-space arithmetic for atom-number editing, shape/feature swapping, and interpolation.
Key results
- On QM9, MolFLAE_100 reaches 99.4% atom stability, 92.0% molecular stability and 96.8% validity, competitive with UniGEM and GEOLDM.
- On GEOM-Drugs, MolFLAE_50 reports 86.9% atom stability, the best among compared methods in the table.
- For hGR optimization, the top-10 blended candidates outperformed BI-653048 in docking score and 8 improved CLogP over AZD2906; sample 34 gave docking -11.15 and CLogP 3.75.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use MolFLAE with 50 sampling steps instead of the strongest baseline | Unconditional 3D generation · QM9 | Atom stability (%): 99.0 -> 99.3 | Helped |
| Use MolFLAE with 50 sampling steps instead of the strongest baseline | Unconditional 3D generation · QM9 | Molecular stability (%): 89.8 -> 90.4 | Helped |
| Use MolFLAE with 50 sampling steps instead of the strongest baseline | Unconditional 3D generation · QM9 | Validity (%): 95.0 -> 95.9 | Helped |
| Use MolFLAE with 50 sampling steps instead of the strongest baseline | Unconditional 3D generation · QM9 | V×U (%): 93.2 -> 92.1 | Hurt |
| Use MolFLAE with 50 sampling steps instead of the strongest baseline | Unconditional 3D generation · QM9 | Novelty (%): 74.6 -> 77.1 | Helped |
| Use MolFLAE with 50 sampling steps instead of the strongest baseline | Unconditional 3D generation · GEOM-Drugs | Atom stability (%): 85.1 -> 86.9 | Helped |
| Use MolFLAE with 50 sampling steps instead of the strongest baseline | Unconditional 3D generation · GEOM-Drugs | Validity (%): 99.3 -> 99.2 | Hurt |
| Increase MolFLAE sampling from 50 to 100 steps | MolFLAE · QM9 | Molecular stability (%): 90.4 -> 92.0 | Helped |
| Increase MolFLAE sampling from 50 to 100 steps | MolFLAE · QM9 | V×U (%): 92.1 -> 88.9 | Hurt |
| Increase MolFLAE sampling from 50 to 100 steps | MolFLAE · GEOM-Drugs | Validity (%): 99.2 -> 99.7 | Helped |
| Generate analogs with different atom counts | MolFLAE latent decoder · ZINC-9M analog generation | MCS-IoU similarity | Mixed |
| Preserve the spatial latent component z_x rather than z_h | MolFLAE hybrid reconstruction · Molecule reconstruction | Shape similarity: 0.174 vs 0.394 | Helped |
Surprising findings
- Preserving the equivariant latent z_x gives much higher shape similarity than preserving z_h, while preserving z_h gives higher MACCS similarity, indicating partial disentanglement.
Limitations
- Latent space is only partially disentangled between shape and substructure.
- Drug optimization evaluated only computationally (docking, CLogP), without experimental validation.
- Multi-step BFN decoding with fixed atom count; molecular inpainting and superposition only sketched as future work.
Experimental setup
Datasets: QM9 GEOM-Drugs ZINC-9M
Benchmarks: QM9 GEOM-Drugs
Metrics: atom stability molecular stability validity uniqueness novelty MCS-IoU MACCS similarity shape similarity Pearson correlation of properties along interpolation docking score
Baselines: ENF G-Schnet GDM-AUG EDM EDM-Bridge GEOLDM GeoBFN UniGEM
Builds on: Bayesian Flow Network variational autoencoder MolCRAFT EDM
Closest papers by meaning
- Learning Disentangled Equivariant Representation for Explicitly Controllable 3D Molecule Generation (AAAI 2025 · similarity 0.90)
- Molecule Design by Latent Prompt Transformer (NeurIPS 2024 · similarity 0.87)
- Mol-AE: Auto-Encoder Based Molecular Representation Learning With 3D Cloze Test Objective (ICML 2024 · similarity 0.87)
- Towards Unified and Lossless Latent Space for 3D Molecular Latent Diffusion Modeling (NeurIPS 2025 · similarity 0.87)
- 3DMolFormer: A Dual-channel Framework for Structure-based Drug Discovery (ICLR 2025 · similarity 0.87)
- Score-based 3D molecule generation with neural fields (NeurIPS 2024 · similarity 0.86)
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.