ShEPhERD-2

A medicinal chemist rarely wants "a molecule that binds". They want one that donates a hydrogen bond to a particular serine, keeps a hydrophobic contact, and drops the contact driving an off-target liability. Most 3D generative models cannot be told that: they learn recognition implicitly from protein-ligand complexes, so the interactions a chemist reasons about are never design variables, only something to read off afterwards.

ShEPhERD-2, from Kento Abeywardane, Kenji Walker and Connor Coley at MIT, makes the interaction pattern the input. Its specification is an interaction profile: the solvent-accessible surface, the electrostatic potential (ESP) sampled on that surface, and pharmacophores carrying explicit direction vectors. An SE(3)-equivariant diffusion model denoises a 3D molecular graph and its interaction profile jointly, so a target profile can be held fixed by inpainting while chemically unrelated structures are generated around it. Profiles come from ligand geometry alone, so training needs no protein-ligand complexes.

It succeeds ShEPhERD, the same group's joint shape/ESP/pharmacophore diffusion model, and its additions turn a single-task analogue generator into a general design interface. One checkpoint covers hit expansion, scaffold elaboration, dual-target design, selectivity engineering and modality hopping with no task-specific retraining.

#Key Features

• Interaction profile as the specification: A chemotype-agnostic description of how a molecule engages a target, extractable from a ligand conformer, a bound pose, a fragment screen or a macrocyclic peptide.
• Absorbing dummy atoms: Superfluous atoms and pharmacophores denoise into dummy states, so one over-provisioned initialization converges to an appropriate molecular size rather than strained or overlapping geometry.
• Pharmacophore prioritization: High-priority features stay unnoised throughout denoising while the rest are inpainted stochastically, raising median 3D pharmacophore similarity on the prioritized subset to 0.808 from 0.701.
• Scaffold conditioning: Fixed atom types, charges and coordinates are carried through the trajectory rather than reconstructed, so a BRICS fragment or chosen heteroatoms survive without the strain inpainting introduces.
• AND and NOT composition: Linearly combining the denoiser's per-profile predictions at each step steers one sample toward satisfying two profiles at once, or one while avoiding another.

#Technical Details

The denoiser keeps ShEPhERD's SE(3)-equivariant layout of per-modality embedding modules, a joint module over a heterogeneous 3D graph and per-modality denoising modules, but replaces the EquiformerV2-style graph transformer blocks with faster EquiformerV3 modules using SwiGLU activations, and swaps the original diffusion process for a variance-exploding formulation under the EDM parameterization. Training used ShEPhERD-MOSES-aq: 1.6 million neutral drug-like MOSES molecules of up to 27 heavy atoms over H, C, N, O, F, Cl, Br, I, S, P and Si, each a single conformer relaxed with GFN2-xTB in implicit aqueous solvent, augmented with up to 10 dummy atoms and 5 dummy pharmacophores; 300k Adam steps at an effective batch size of 96.

Across 100 MOSES scaffold-split test molecules with 20 candidates per profile, 93.7% of samples were chemically valid after relaxation against 61.3% for ShEPhERD, at a median strain energy of 0.015 kcal/mol; an ablation attributes most of that gain to dummy atoms (84.7%) and the variance-exploding formulation (94.0%). On a single H100 the forward pass is 4.93x faster at 44.2x lower memory, and the full MOSES evaluation drops from 3.05 to 0.48 hours on an H200. On the MolGenBench hit-to-lead benchmark of 120 protein targets and 600 reference ligands, it ranks first or second on nearly every ligand-based metric, reaching interaction diversity of 0.756 at a mean reference Tanimoto of 0.199, and stays competitive on docking with pocket-conditioned models such as FLOWR.ROOT and OMTRA without seeing pocket geometry.

#Applications

Three in-silico campaigns show the range. Bioisosteric fragment merging wraps the model in an NSGA-II genetic algorithm whose individuals are interaction profiles seeded from 13 fragments bound to the EV-D68 3C protease, reaching a top-10 docking mean of -8.16 kcal/mol against -7.45 for the published 2D pharmacophore database search. AND composition of tubulin and PARP1 inhibitor profiles yields putative dual binders, and NOT composition of dabrafenib's BRAF V600E and hPXR bound poses recovers the known selective analogue's strategy of replacing a tert-butyl group with a bulky polar substituent. Medicinal chemists running fragment-based, ligand-based and polypharmacology campaigns, where a bound pose or an active is already in hand, are the beneficiaries.

#Impact

By treating interactions rather than chemical structures as the unit of specification, ShEPhERD-2 gives ligand-based design one interface across tasks usually served by separate, separately trained models. Code is MIT-licensed, weights are hosted on HuggingFace and downloaded on first use, and training data plus all evaluation outputs are archived on Zenodo, so the comparisons are reproducible. The authors state the limits plainly: every result is computational, with AutoDock Vina and Boltz-2 co-folding standing in for measured affinity; interaction profiles must be derived from ligands with known activity, so the model cannot yet propose recognition patterns from a pocket alone; synthetic accessibility sits outside the generative objective; and the training distribution caps at 27 heavy atoms.