[Paper Review] ShEPhERD: Diffusing shape, electrostatics, and pharmacophores for bioisosteric drug design

Keir Adams, Kento Abeywardane+2 more|arXiv (Cornell University)|Oct 22, 2024

Computational Drug Discovery Methods6 citations

TL;DR

ShEPhERD is an SE(3)-equivariant diffusion model that jointly denoises/generates 3D molecular graphs and their interaction profiles (shape, ESP, pharmacophores) to enable bioisosteric design tasks. It enables interaction-aware ligand design without targeting a specific protein.

ABSTRACT

Engineering molecules to exhibit precise 3D intermolecular interactions with their environment forms the basis of chemical design. In ligand-based drug design, bioisosteric analogues of known bioactive hits are often identified by virtually screening chemical libraries with shape, electrostatic, and pharmacophore similarity scoring functions. We instead hypothesize that a generative model which learns the joint distribution over 3D molecular structures and their interaction profiles may facilitate 3D interaction-aware chemical design. We specifically design ShEPhERD, an SE(3)-equivariant diffusion model which jointly diffuses/denoises 3D molecular graphs and representations of their shapes, electrostatic potential surfaces, and (directional) pharmacophores to/from Gaussian noise. Inspired by traditional ligand discovery, we compose 3D similarity scoring functions to assess ShEPhERD's ability to conditionally generate novel molecules with desired interaction profiles. We demonstrate ShEPhERD's potential for impact via exemplary drug design tasks including natural product ligand hopping, protein-blind bioactive hit diversification, and bioisosteric fragment merging.

Motivation & Objective

• Motivate 3D interaction-aware molecular design for bioisosteric drug discovery.
• Develop a joint diffusion model that simultaneously treats 3D structure and interaction profiles.
• Enable conditional generation via inpainting to target specific interaction profiles.
• Demonstrate applications to natural product ligand hopping, hit diversification, and fragment merging.

Proposed method

• Define explicit point-cloud representations for shape, ESP, and directional pharmacophores amenable to SE(3)-equivariant diffusion.
• Formulate a joint DDPM that denoises/generates 3D molecular graphs and their interaction profiles (shape, ESP, pharmacophores).
• Employ SE(3)-equivariant neural networks (EquiformerV2/E3NN) to encode/decode heterogeneous representations.
• Use inpainting to sample interaction-conditioned molecules from targeted profiles.
• Develop 3D similarity scoring functions for self-consistency and target-profile conditioning to guide generation.
• Train on drug-like datasets and evaluate unconditional and conditional generation, including ligand hopping and bioisosteric fragment merging.

Experimental results

Research questions

• RQ1 Can a joint diffusion model generate 3D molecules together with their shape, electrostatics, and pharmacophore profiles in a symmetry-preserving way?
• RQ2 Do interaction profiles guide generation to produce molecules with high similarity to target interaction patterns after relaxation?
• RQ3 Can inpainting conditioned on interaction profiles enable bioisosteric ligand hopping, hit diversification, and fragment merging without protein target information?
• RQ4 How do shape, ESP, and pharmacophore similarities reflect self-consistency and conditioning quality in generated molecules?

Key findings

• ShEPhERD can jointly generate 3D molecules and their interaction profiles with high self-consistency between generated structures and profiles.
• Generated molecules show enriched similarity to target interaction profiles compared with random molecules, remaining stable upon geometry relaxation.
• Conditional generation via inpainting can produce diverse, interaction-consistent molecules conditioned on specific shapes, ESPs, or pharmacophores.
• The model supports ligand hopping, diversification of bioactive hits, and bioisosteric fragment merging in a context-free setting.
• Experiments trained on large drug-like datasets demonstrate practical 3D interaction-aware design capabilities for bioisosteric drug design tasks.