Research map / Image generation / Autoregressive and latent generation
Unified Guidance for Geometry-Conditioned Molecular 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 UniGuide, a training-free self-guidance framework that steers unconditional molecular diffusion models toward geometric conditions (shapes, pockets, fragments) via condition maps.
Problem
Molecular diffusion models are typically tailored to a single downstream drug-design task and need task-specific conditional training; there is no unified way to apply unconditional models to geometric conditions such as receptor pockets, fragments, or 3D shapes.
Approach
Introduce a condition map C that transforms arbitrary geometric source conditions (point sets, surfaces, pockets, fragments) into target conditions in the diffusion configuration space, enabling self-guidance from the model's own clean-data estimate without extra networks or training; for shape conditioning it projects atoms toward a surface sampled from the reference ligand.
Key results
- On LBDD, UniGuide with EDM reaches a Ratio of 3.536 versus 2.834 for ShapeMol and 3.074 for ShapeMol+g.
- On SBDD (CrossDocked, full-atom), UniGuide achieves Vina Dock -7.921, beating DiffSBDD (-7.889) and the TargetDiff reference (-7.802).
- On linker design, UniGuide (EDM) reaches 0.49 No. Rings and 72.1 Unique, but lower Recovery (58.8) than DiffLinker (70.7).
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| guide EDM generation with the UniGuide surface condition map instead of validity guidance | EDM · MOSES test set (LBDD) | Ratio: 2.96 -> 3.53 | Helped |
| guide EDM with UniGuide rather than use ShapeMol+g position correction | LBDD generation · MOSES test set (LBDD) | Sim_S: 0.744 -> 0.749 | Helped |
| use UniGuide rather than the DiffSBDD inpainting-inspired method | DiffSBDD joint model · CrossDocked test set (SBDD) | Vina Dock: $-$7.889 -> $-$7.921 | Helped |
| use UniGuide rather than the DiffSBDD inpainting-inspired method | DiffSBDD joint model · Binding MOAD test set (SBDD) | Vina: $-$7.263 -> $-$7.661 | Helped |
| use UniGuide with EDM rather than DiffLinker | Linker design · ZINC linker design | Unique: 51.4 -> 72.1 | Helped |
| replace UniGuide's explicit surface condition map with validity guidance | EDM for LBDD · MOSES test set (LBDD) | Sim_S: 0.74 -> 0.59 | Hurt |
| replace UniGuide's explicit surface condition map with validity guidance | EDM for LBDD · MOSES test set (LBDD) | Ratio: 3.53 -> 2.96 | Hurt |
| apply resampling during unconditional joint-model generation | DiffSBDD joint model · CrossDocked and Binding MOAD | Connectivity: 0.893 -> 0.994; 0.839 -> 0.961 | Helped |
Surprising findings
- Unconditional EDM guided by UniGuide outperforms conditionally trained ShapeMol on shape-based generation, despite not using reference atom positions.
Limitations
- Requires the unconditional model to be trained on a matching configuration space.
- Does not include global graph property conditions.
- SBDD condition map lacks drift terms, so it does not match methods using explicit drift.
Experimental setup
Datasets: MOSES ZINC CrossDocked Binding MOAD
Benchmarks: MOSES reference ligands (LBDD) CrossDocked test pockets Binding MOAD test pockets ZINC linker design
Metrics: Sim_S max Sim_S Sim_G max Sim_G Ratio Diversity Vina Score Vina Min Vina Dock QED
Baselines: ShapeMol ShapeMol+g SQUID Virtual Screening 3D-SBDD Pocket2Mol TargetDiff DecompDiff DiffSBDD DiffLinker DeLinker 3DLinker
Builds on: classifier guidance EDM DiffSBDD resampling
Closest papers by meaning
- UniMoMo: Unified Generative Modeling of 3D Molecules for De Novo Binder Design (ICML 2025 · similarity 0.89)
- Enhancing Generalizability in Molecular Conformation Generation with METRIZATION-Informed Geometric Diffusion Pretraining (AAAI 2025 · similarity 0.88)
- Navigating the Design Space of Equivariant Diffusion-Based Generative Models for De Novo 3D Molecule Generation (ICLR 2024 · similarity 0.88)
- Geometric Representation Condition Improves Equivariant Molecule Generation (ICML 2025 · similarity 0.88)
- Controllable 3D Molecular Generation for Structure-Based Drug Design Through Bayesian Flow Networks and Gradient Integration (NeurIPS 2025 · similarity 0.87)
- BindGPT: A Scalable Framework for 3D Molecular Design via Language Modeling and Reinforcement Learning (AAAI 2025 · similarity 0.87)
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.