Chemistry & ADMET · Coley Research Group, MIT
ShEPhERD-2
A 3D diffusion model that designs bioisosteric small molecules from surface interaction profiles.
What ShEPhERD-2 is
ShEPhERD-2 (Shape, Electrostatics, and Pharmacophores Explicit Representation Diffusion 2) is a 3D molecular generative model developed by the Coley Research Group at MIT. Conventional molecular similarity tools compare two-dimensional bond graphs, which often traps generative design around the initial chemical scaffold. ShEPhERD-2 sidesteps the molecular graph entirely by reading the interaction profile a molecule presents: the geometry of its solvent-accessible surface, the electrostatic potential (ESP) across that surface, and directional pharmacophore points.
Using these interaction profiles as conditioning inputs, the model diffuses new small-molecule geometries in three dimensions that display similar surface properties while allowing the underlying covalent topology to change completely. This scaffold-hopping capability supports the design of bioisosteres—molecules that elicit equivalent biological effects despite distinct chemical cores—as well as fragment merging, modality hopping, dual-target design, and selectivity engineering.
The model produces proposed 3D coordinates rather than predictions of receptor binding. Every generated atomic sample is geometry-optimised and relaxed using the semi-empirical GFN2-xTB method prior to bond perception, ensuring that returned outputs are chemically valid molecules rather than unstructured atom clouds. ShEPhERD-2 does not incorporate target protein structures or compute docking scores; its outputs represent hypotheses based purely on matched surface shape and electrostatics.
How it works
ShEPhERD-2 operates by extracting an explicit interaction profile consisting of three modalities: 3D molecular shape, electrostatic potential, and directional pharmacophores. Conditioning can be applied to all three modalities simultaneously or restricted to specific subsets, such as retaining shape while leaving electrostatics unconstrained. The model can also compose multiple interaction profiles together, prioritise critical pharmacophores, or hold specific substructures and scaffold atom indices fixed.
During generation, 3D atomic coordinates and pharmacophores are diffused based on the conditioning profile. The model allows user-specified adjustments to the number of atoms or pharmacophores relative to the reference input. After sampling, the partial charges are inferred and each structure is relaxed using GFN2-xTB to optimize molecular geometry before covalent connectivity is perceived. Generated structures can then be converted directly to RDKit Mol objects, written to SDF format, or scored against reference profiles.
Key facts
- Model type
- 3D molecular diffusion model
- Input formats
- SDF, SMILES, SMILES file, or pickled (molblock, partial charges) pairs
- Output formats
- SDF, RDKit Mol objects, and tabular SMILES with Tanimoto similarities and atom counts
- Relaxation method
- GFN2-xTB
- GPU memory
- 24 GB minimum recommended
- Supported tasks
- Molecular generation, bioisostere design, scaffold hopping, fragment merging
What you can use it for
- Scaffold hopping and bioisostere generation
- Produce novel small molecules that match the 3D surface shape, electrostatics, and pharmacophore directions of a known active ligand while featuring structurally unrelated 2D chemical scaffolds.
- Substructure-constrained elaboration
- Fix specific scaffold atom indices from a 3D reference structure and diffuse replacement functional groups or linkers that satisfy target interaction profiles.
- Pharmacophore-prioritised design
- Condition generation on a selected subset of critical pharmacophore points, allowing the model to inpaint or discard non-essential features.
- Multi-profile composition
- Combine interaction profiles from multiple molecules to perform bioisosteric fragment merging, explore modality hopping, or engineer dual-target and selective ligands.
Run it here
No installation and no GPU of your own: fill in the inputs and it runs on our servers. Results stay in your browser unless you sign in.
Runs on On-demand GPU
Runs on our servers with on-demand compute. A first run needs time to load the model; active capacity can be reused and scales down when idle.Run this model
ShEPhERD-2
Strengths and limits
Good at
- Decouples molecular surface generation from 2D bond topology, enabling genuine scaffold hopping with low 2D graph similarity.
- Allows modular conditioning on any combination of shape, electrostatic potential, and directional pharmacophores.
- Provides structural control via substructure constraints, pharmacophore prioritization masks, and multi-profile composition.
- Includes automated GFN2-xTB quantum chemical relaxation to convert raw generated atom coordinates into chemically valid molecular structures.
- Supports distributed multi-GPU sampling to handle large generation batches efficiently.
Watch out for
- Does not account for target protein structures, binding pocket geometry, or ligand-receptor docking scores.
- Requires the external semi-empirical software GFN2-xTB installed on the system PATH to perform standard molecular relaxation and charge computation.
- Pharmacophore and scaffold constraint modes require 3D structure inputs (SDF or molblock-charge files) rather than plain SMILES strings.
- Generated molecules represent physical shape and charge hypotheses, not validated affinity predictions.
Inputs and outputs
You provide
Reference molecule (SMILES)
The molecule whose interaction profile the designs should match. Drug-sized: 5 to 70 heavy atoms.
Match onoptional
Molecules to designoptional
Size change (atoms)optional
Atoms added or removed relative to the reference, hydrogens included. Far from zero, fewer samples resolve into molecules.
Seedoptional
Changes the starting point, so a different seed gives a different set of designs. It does not reproduce a previous run exactly: the model's scatter operations accumulate with GPU atomics, so two runs at the same seed diverge.
You get back
Each sample as SMILES, with its heavy-atom count and its 2D Tanimoto similarity to the reference — low means a genuinely different scaffold.
The first resolved sample in 3D, after xTB relaxation.
Licence
Commercial use permitted · MIT code / MIT weights
Verified 2026-10-29. Code: MIT, github.com/coleygroup/shepherd2/blob/main/LICENSE (Coley Research Group @ MIT), and the `shepherd-score` package it depends on is MIT too. Weights: huggingface.co/kabeywar/shepherd2 declares MIT on its model card and is not gated. GFN2-xTB, which the pipeline shells out to for partial charges and relaxation, is LGPL-3.0 and is used as an unmodified separate program.
Questions
Can I use ShEPhERD-2 commercially?
Commercial use permitted. Verified 2026-10-29. Code: MIT, github.com/coleygroup/shepherd2/blob/main/LICENSE (Coley Research Group @ MIT), and the `shepherd-score` package it depends on is MIT too. Weights: huggingface.co/kabeywar/shepherd2 declares MIT on its model card and is not gated. GFN2-xTB, which the pipeline shells out to for partial charges and relaxation, is LGPL-3.0 and is used as an unmodified separate program.
How do I run ShEPhERD-2 without installing anything?
Use the form under “Run it here” on this page: fill in the inputs and press Run model. It runs on our servers on 1× A100 40GB, typically in 2.0min, and a run costs about 56 credits. Visitors without an account get 1,000 credits a month.
Does ShEPhERD-2 require a target protein structure to generate molecules?
No. The model is ligand-based and generates molecules solely from reference interaction profiles composed of shape, electrostatics, and pharmacophores. It never sees a protein structure or docks compounds into a binding pocket.
Why is GFN2-xTB required for inference?
The diffusion model outputs unbonded 3D point clouds. GFN2-xTB geometry optimisation is used to relax these coordinates and calculate charges so that chemically valid molecular graphs and bond topologies can be perceived.
Can I condition generation on shape alone without matching electrostatic properties?
Yes. The conditioning flag accepts individual components or combinations, allowing you to match exclusively on shape, electrostatic potential, or pharmacophores.
Can ShEPhERD-2 generate molecules directly from a SMILES string?
Yes. If supplied with a SMILES string, the model automatically embeds a 3D conformer and relaxes it with xTB before extracting the interaction profile for generation. However, explicit 3D structures in SDF format are preferred and required for scaffold and pharmacophore constraint modes.
What hardware is recommended for running ShEPhERD-2?
The model specifies a minimum of 24 GB of GPU memory, with support for distributed multi-GPU generation on larger accelerators such as the NVIDIA H100 (80 GB).
How to cite it
Abeywardane, K. A., Walker, K., & Coley, C. W. (2026). Interaction Profiles as a Universal Language for Generative Molecular Design with ShEPhERD-2. bioRxiv, 2026.09.10.750648. https://doi.org/10.64898/2026.09.10.750648
Written from ShEPhERD-2's own documentation (model card, code, paper). Statements citing a figure that does not appear there are removed before publication. Licence and running details come from our registry.