Research map / Probabilistic and scientific modeling / Molecules and proteins
ShEPhERD: Diffusing shape, electrostatics, and pharmacophores for bioisosteric drug design
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.
Introduces ShEPhERD, an SE(3)-equivariant joint diffusion model over 3D molecular graphs and their shape, electrostatic potential, and pharmacophore profiles, used via inpainting for interaction-conditioned bioisosteric design.
Problem
Generating novel 3D molecules that reproduce target shape, electrostatic, and pharmacophore interaction profiles for ligand-based bioisosteric drug design, which similarity-based virtual screening cannot explore beyond known libraries.
Approach
An SE(3)-equivariant DDPM jointly diffuses molecular graphs with point-cloud surface shape, ESP surface, and directional pharmacophore vectors (diffused as Euclidean points), and conditions generation on target profiles via inpainting; similarity is scored with custom Gaussian-overlap functions.
Key results
- Generated molecules show enriched 3D interaction similarity to target profiles, with >=94% of valid samples having graph similarity <0.2 to the target.
- For natural product ligand hopping, ShEPhERD outperforms REINVENT (10K oracle budget) and dataset virtual screening on ESP and pharmacophore similarity among SA<4.5 molecules.
- Bioactive hit diversification enriches top-1 Vina scores relative to the redocked PDB ligand for all 7 targets, and relative to a 10K docking screen for 3/7 (lowest-energy conformer) and 5/7 (bound pose) targets.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Condition on the best-scoring screened ligand instead of the experimental PDB ligand | ShEPhERD on 3ny8 · 3ny8 | mean top-10 Vina score (kcal/mol): -9.96 to -11.70 kcal/mol | Helped |
| Condition on the best-scoring screened ligand from the 10K docking screen | ShEPhERD · 7 protein targets | top-1 Vina score relative to the best-scoring screened compound: 5/7 targets | Helped |
| Condition on the experimental ligand's bound pose rather than its lowest-energy conformer | ShEPhERD · 7 protein targets | top-1 Vina score relative to the 10K docking screen: 5/7 targets | Mixed |
| Condition on the experimental ligand's lowest-energy conformer | ShEPhERD · 7 protein targets | top-1 Vina score relative to the 10K docking screen: 3/7 targets | Mixed |
Surprising findings
- Pharmacophore self-consistency is weaker than shape and ESP self-consistency, partly attributed to discrete pharmacophore types and the need to specify n4.
- Hit diversification with ShEPhERD does not necessarily beat an unbiased docking screen on top Vina scores; for 2/7 targets the screen outperforms when conditioning on the bound pose.
Limitations
- Pharmacophore generation is less self-consistent than shape/ESP and requires specifying the number of pharmacophores n4 in advance.
- Docking Vina scores are only a weak surrogate for bioactivity; no experimental validation is reported.
- Fragment merging pharmacophores and ESP were manually selected/curated.
Experimental setup
Datasets: ShEPhERD-GDB17 ShEPhERD-MOSES-aq GDB17 MOSES COCONUT ZINC
Benchmarks: ShEPhERD-GDB17 self-consistency and conditional generation natural product ligand hopping (COCONUT) Autodock Vina bioactive hit diversification (7 PDB targets) EV-D68 3C protease fragment merging
Metrics: shape similarity ESP similarity pharmacophore similarity graph similarity Vina score SA score xTB RMSD
Baselines: REINVENT virtual screening with 3D similarity AutoDock Vina docking screen
Builds on: DDPM EquiformerV2 EGNN REINVENT
Closest papers by meaning
- Controllable 3D Molecular Generation for Structure-Based Drug Design Through Bayesian Flow Networks and Gradient Integration (NeurIPS 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)
- 3DMolFormer: A Dual-channel Framework for Structure-based Drug Discovery (ICLR 2025 · similarity 0.88)
- TurboHopp: Accelerated Molecule Scaffold Hopping with Consistency Models (NeurIPS 2024 · similarity 0.88)
- Multi-domain Distribution Learning for De Novo Drug Design (ICLR 2025 · similarity 0.87)
- DecompOpt: Controllable and Decomposed Diffusion Models for Structure-based Molecular Optimization (ICLR 2024 · 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.