Use ShEPhERD-2 Online

ShEPhERD-2: Interaction-Profile-Conditioned 3D Generative Model for Small-Molecule Drug Design

ShEPhERD-2 (Shape, Electrostatics, and Pharmacophore Explicit Representation Diffusion, version 2) is a 3D generative model that designs molecules conditioned on explicit interaction profiles. It was developed by Kento A. Abeywardane, Kenji Walker and Connor W. Coley at the Massachusetts Institute of Technology and described in a bioRxiv preprint (not yet peer reviewed). The authors state that the code is freely available under the MIT license.

An interaction profile combines molecular shape, electrostatic potential and directional pharmacophores. The authors propose that this profile is a sufficient and transferable design specification: it can be extracted from ligand conformers, bound fragments or native biomolecules and realized by many different chemical structures.

Most current 3D generative models for small molecules learn interactions implicitly from a limited corpus of protein-ligand complexes. According to the paper, they often fail to reproduce native interactions, adopt strained conformations and over-rely on nonselective hydrophobic contacts, and chemists cannot directly specify the interactions they want. ShEPhERD-2 makes interactions explicit design variables.

How ShEPhERD-2 Generates Molecules from Interaction Profiles

• Joint denoising: the model simultaneously denoises molecular structure and its interaction profile, so profiles from a reference molecule guide generation by inpainting.

• Flexible size: dummy atoms let superfluous atoms be removed during denoising, so molecule size need not be specified in advance.

• Efficiency: a variance-exploding diffusion process and a more efficient equivariant architecture reduce strain and speed up sampling.

• Pharmacophore prioritization: chemists can mark some pharmacophores as high priority while leaving others flexible.

• Substructure constraints: selected atoms or scaffolds can be fixed during denoising to support scaffold elaboration.

• Composition: AND composition seeks molecules satisfying two profiles at once, and NOT composition seeks one profile while avoiding another.

Performance of ShEPhERD-2 on Interaction-Conditioned Design

On 100 molecules from the MOSES scaffold-split test set, with 20 candidates per target profile, ShEPhERD-2 produced chemically valid molecules more often than its predecessor and reproduced target profiles more faithfully, with low strain after GFN2-xTB relaxation.

The authors also demonstrate bioisosteric fragment merging, dual-target design, selectivity engineering and modality hopping, such as small molecules that mimic peptide interactions, using a single model without task-specific retraining.

What is Tamarind Bio?

Tamarind Bio is a no-code bioinformatics platform built to give life scientists and researchers access to powerful computational tools. Many cutting-edge machine learning models are hard to deploy and use. Tamarind provides an intuitive, web-based environment that removes the complexity of high-performance computing, software dependencies and command-line interfaces.

The platform is designed for biologists, chemists and other researchers who may not have a background in programming or cloud infrastructure but want to run models on their own data. Key features include:

• A user-friendly graphical interface for setting up and launching experiments

• A robust API for integration into existing research pipelines

• An automated system for managing and scaling computational resources

Tamarind treats information and data security as a top priority, as detailed in its Trust Center and Terms of Service.

Accelerating Discovery with ShEPhERD-2 on Tamarind Bio

• Hit expansion and analoging: generate chemically distinct molecules that keep the key interactions of a known active.

• Scaffold elaboration: grow diverse substituents around a fixed substructure while preserving the interaction profile.

• Dual-target and selectivity design: compose profiles to satisfy two targets or favor one target over another.

• Peptide mimicry: design small molecules that reproduce interactions of a peptide.

How to Use ShEPhERD-2 on Tamarind Bio

1. Open the tool: log in to tamarind.bio and select ShEPhERD-2.

2. Provide a reference: supply a reference ligand or conformer from which the interaction profile is derived, in the format the interface accepts.

3. Choose conditioning: if exposed, select full-profile inpainting, pharmacophore prioritization, fixed substructures, or profile composition.

4. Set generation options: if exposed, set the number of samples and the initial atom and pharmacophore counts.

5. Run: submit the job and wait for generation to complete.

6. Download outputs: retrieve the generated 3D molecules.

7. Filter and validate: check validity, strain and similarity to the target profile, then dock or score promising candidates.

Things to Keep in Mind

• The prospective case studies rely on docking and interaction analysis as a poor proxy for binding affinity; experimental validation is still required.

• Interaction profiles currently come from ligands with known activity; inferring constraints directly from protein pockets is future work.

• Synthetic accessibility is not built into generation, and post hoc projection may relax away the designed interactions.

• The model was trained on neutral drug-like molecules with up to 27 heavy atoms from MOSES.

Source: Abeywardane KA, Walker K, Coley CW. Interaction Profiles as a Universal Language for Generative Molecular Design with ShEPhERD-2. bioRxiv preprint (2026), DOI: https://doi.org/10.64898/2026.09.10.750648. Code: https://github.com/coleygroup/shepherd2. Weights: https://huggingface.co/kabeywar/shepherd2.