Loading…
Diffusion model that jointly generates a 3D molecule and its shape, electrostatic surface and pharmacophores for ligand-based drug design.
No providers recorded yet. Browse all providers
Two molecules can share almost nothing on paper — different scaffolds, different ring systems, different atom counts — and still present the same face to a protein: the same volume, the same charge across the solvent-accessible surface, the same hydrogen-bond donors pointing the same way. Medicinal chemists exploit that daily under the name bioisosterism, and usually find such analogues by screening an enumerated library with shape, electrostatic and pharmacophore similarity scoring functions. Screening can only rank molecules somebody already drew.
ShEPhERD — Shape, Electrostatics, and Pharmacophores Explicit Representation Diffusion — turns that scoring step into a generative one. Developed by Keir Adams, Kento Abeywardane, Jenna Fromer and Connor Coley at MIT and presented as an oral at ICLR 2025, it is an SE(3)-equivariant denoising diffusion model that learns the joint distribution over a 3D molecular graph and its interaction profile: a point cloud of the molecular surface, the electrostatic potential (ESP) sampled on that surface, and a set of pharmacophores carrying explicit direction vectors. Because the profile is part of what the model generates rather than a filter applied afterwards, a target profile can be held fixed while the molecule is denoised around it.
ShEPhERD is deliberately protein-blind. Where structure-based generators such as TargetDiff denoise a ligand inside a pocket, ShEPhERD conditions only on the interaction profile of a reference ligand or fragment set — the common medicinal-chemistry situation where a bioactive hit is known but no co-crystal structure is.
shepherd-score supplies differentiable 3D similarity, conformer generation, alignment and evaluation, and is usable independently to score other 3D generative models.The denoising network has three parts: embedding modules built on EquiformerV2 equivariant neural networks that encode each input state into scalar and vector node features, a joint module that couples them over a heterogeneous 3D graph plus a pooled global code, and per-state denoising modules. It is small by contemporary standards: 6,010,427 learnable parameters for the flagship MOSES-aq model and roughly 4.4M per GDB-17 variant, run over 400 denoising steps at an effective batch size of 48. Two purpose-built datasets supplied training data — ShEPhERD-GDB17, 2.8M molecules of at most 17 heavy atoms from medicinally relevant GDB-17 subsets with GFN2-xTB gas-phase conformers, and ShEPhERD-MOSES-aq, 1.6M drug-like MOSES molecules of up to 27 heavy atoms optimized in implicit water — and each model took about two weeks on two V100 GPUs. Unconditional joint generation is at least 96% valid with shape conditioning and 91–93% with electrostatics, and geometries shift under 0.1 Å heavy-atom RMSD on xTB relaxation. Conditioned on 100 held-out GDB-17 targets, at least 94% of valid samples had 2D graph similarity below 0.2 to the target while showing markedly enriched 3D similarity.
The paper works three archetypal ligand-based tasks. Natural product ligand hopping conditions on the ESP surface and pharmacophores of complex natural products and returns synthetically simpler mimics with higher 3D similarity than a REINVENT-optimized comparison set. Bioactive hit diversification generates 500 analogues each for seven PDB ligands and, docked with AutoDock Vina, enriches scores relative to a 10K-compound screen despite the model never seeing the protein. Bioisosteric fragment merging conditions on the aggregate ESP surface and 27 pharmacophores extracted from 13 fragments screened against the EV-D68 3C protease. The beneficiaries are medicinal chemists doing scaffold hopping and fragment-based design without a structure.
ShEPhERD reframes 3D similarity search as conditional generation, and its interaction-profile representation is general enough that the authors point at organocatalyst design as an extension. Code and checkpoints are MIT-licensed, with weights auto-downloaded from HuggingFace, and the separately packaged scoring library gives the field a reusable 3D evaluation toolkit. The constraints are real: sampling takes three to four minutes per ten molecules on a V100 at 400 steps, the training distributions cap out at 27 heavy atoms so larger or macrocyclic targets are out of distribution, and every reported result is in silico, with docking scores standing in as a weak surrogate for measured bioactivity.
Much of this page is generated or calculated automatically. Flag anything that looks off and we will re-run it.