ShEPhERD Scoring Functions

This repository contains the code for generating/optimizing conformers, extracting interaction profiles, aligning interaction profiles, and differentiably scoring 3D similarity. It also contains modules to evaluate conformers generated with ShEPhERD¹ and other generative models.

The formulation of the interaction profile representation, scoring, alignment, and evaluations are found in our preprint ShEPhERD: Diffusing shape, electrostatics, and pharmacophores for bioisosteric drug design. The diffusion model itself is found in a separate repository: https://github.com/coleygroup/shepherd.

_{¹ ShEPhERD: Shape, Electrostatics, and Pharmacophores Explicit Representation Diffusion}

Table of Contents

1. Documentation
2. File Structure
3. Installation
4. Requirements
5. Usage
6. Scoring and Alignment Examples
7. Evaluation Examples and Scripts
8. Data

Documentation

Full documentation is available at shepherd-score.readthedocs.io.

File Structure

.
├── shepherd_score/
│   ├── alignment/                        # Alignment package with PyTorch, JAX, and utilities
│   │   ├── utils/                        # SE(3) and PCA utilities (torch, numpy, jax)
│   │   │   ├── se3*.py                   # SE(3) transformations (torch, numpy, jax)
│   │   │   └── pca*.py                   # Principal component alignment (torch, numpy, jax)
│   │   ├── _torch.py                     # PyTorch alignment algorithms (autograd)
│   │   ├── _torch_analytical.py          # PyTorch alignment algorithms (analytical gradients)
│   │   ├── _jax.py                       # JAX alignment algorithms
│   │   └── _jax_parallel.py              # JAX parallel alignment algorithms
│   ├── alignment_jax.py                  # Backwards compatibility shim
│   ├── evaluations/                      # Evaluation suite
│   │   ├── utils/                        # Converting data types and others
│   │   ├── docking/
│   │   │   ├── pdbs/                     # PDBQT files used in *ShEPhERD* manuscript
│   │   │   ├── docking.py                # Docking classes
│   │   │   └── pipelines.py              # Docking evaluation pipelines
│   │   └── evaluate/                     # Generated conformer evaluation pipelines
│   │       ├── evals.py                  # Individual evaluation classes
│   │       └── pipelines.py              # Evaluation pipeline classes
│   ├── pharm_utils/                      # Pharmacophore definitions
│   ├── protonation/                      # Functions for protonation
│   ├── score/                            # Scoring related functions and constants
│   │   ├── analytical_gradients/         # Analytical gradient implementations
│   │   │   └── _torch.py                 # PyTorch analytical gradients for shape, ESP, pharmacophore
│   │   ├── constants.py
│   │   ├── electrostatic_scoring.py
│   │   ├── gaussian_overlap.py
│   │   └── pharmacophore_scoring.py
│   ├── conformer_generation.py           # RDKit and xtb related functions for conformers
│   ├── container/                        # Molecule, MoleculePair, MoleculePairBatch classes
│   ├── extract_profiles.py               # Functions to extract interaction profiles
│   ├── generate_point_cloud.py
│   ├── objective.py                      # Objective function used for REINVENT
│   └── visualize.py                      # Visualization tools
├── scripts/                              # Scripts for running evaluations
├── examples/                             # Jupyter notebook tutorials/examples
├── tests/
└── README.md

Installation

Via PyPI

pip install shepherd-score

Install xTB

xTB will need to be installed manually since there are no PyPi bindings. This can be done in a conda environment, but since this approach has been reported to lead to conflicts, we suggest installing from source and adding it to PATH.

With optional dependencies

# JAX support (for faster scoring and alignment)
pip install "shepherd-score[jax]"

# Include docking evaluation tools
pip install "shepherd-score[docking]"

# Everything
pip install "shepherd-score[all]"

For local development

git clone https://github.com/coleygroup/shepherd-score.git
cd shepherd-score
pip install -e ".[all]"

Requirements

This package works where PyTorch, Open3D, RDKit, and xTB can be installed for Python >=3.9. If you are coming from the ShEPhERD repository, you can use the same environment as described there. Core dependencies (installed automatically) which enables interaction profile extraction, scoring/alignment, and evaluations are listed below.

python>=3.9
numpy
torch>=1.12
open3d>=0.18
rdkit>=2023.03
pandas>=2.0
scipy>=1.10
py3Dmol
molscrub

Note: If using torch<=2.4, ensure that mkl==2024.0 with conda since there is a known issue that prevents importing torch.

Docking with Autodock Vina

Installing shepherd-score[docking] will automatically install the python bindings for Autodock Vina for the python interface. However, a manual installation of the executable of Autodock Vina v1.2.5 is required and can be found here: https://vina.scripps.edu/downloads/.

Usage

The package has convenience wrappers and base functions. Most users should reach for the convenience wrappers below; the base functions underneath are there if you need lower-level control. Scoring can be done with either NumPy or Torch, but alignment requires Torch. There are also Jax implementations for both scoring and alignment of gaussian overlap, ESP similarity, and pharmacophore similarity.

Update 8/20/25: Applicable xTB functions and evaluation pipeline evaluations are now parallelizable through the num_workers argument in the .evaluate method.

Convenience wrappers

• Molecule class
  • shepherd_score.container.Molecule accepts an RDKit Mol object (with an associated conformer) and generates its interaction profiles, exposed via two properties:
    • .surface → Surface(positions, esp, probe_radius)
    • .pharmacophore → Pharmacophore(types, positions, vectors, atom_ids, labels)
  • Pass return_atom_ids=True and/or priority_atoms=[...] to Molecule.get_pharmacophore() to populate .pharmacophore.atom_ids/.labels, e.g. for priority-weighted pharmacophore scoring
• MoleculePair class
  • shepherd_score.container.MoleculePair operates on two Molecule objects and prepares their Surface/Pharmacophore profiles for scoring and alignment
• MoleculePairBatch class
  • shepherd_score.container.MoleculePairBatch operates on a list of MoleculePair objects and enables accelerated alignment by padding all profile arrays to a common shape so a single compiled kernel is reused across every pair. Supports optional multi-CPU parallelism.

Base functions

Conformer generation

Useful conformer generation functions are found in the shepherd_score.conformer_generation module.

Interaction profile extraction

shepherd_score.pharm_utils.pharmacophore.Pharmacophore is a lightweight dataclass; it also unpacks as types, positions, vectors = get_pharmacophores(mol). The surface position/ESP arrays are assembled into a matching shepherd_score.container.profiles.Surface dataclass by Molecule (above). Most users won't call these extraction functions or construct the containers directly.

Scoring

shepherd_score.score contains the base scoring functions with seperate modules for those dependent on PyTorch (*.py), NumPy (*_np.py), and Jax (*_jax.py).

Scoring and Alignment Examples

Full jupyter notebook tutorials/examples for extraction, scoring, and alignments are found in the examples folder. Some minimal examples are below.

Extraction

Extraction of interaction profiles via the Molecule convenience wrapper.

from shepherd_score.conformer_generation import embed_conformer_from_smiles
from shepherd_score.conformer_generation import charges_from_single_point_conformer_with_xtb
from shepherd_score.container import Molecule

# Embed conformer with RDKit and partial charges from xTB
ref_mol = embed_conformer_from_smiles('Oc1ccc(CC=C)cc1', MMFF_optimize=True)
partial_charges = charges_from_single_point_conformer_with_xtb(ref_mol)

# `Molecule` extracts and owns all interaction profiles
ref_molec = Molecule(
    ref_mol,
    # optional options otherwise .surface/.pharmacophore stay empty
    num_surf_points=200,
    partial_charges=partial_charges, # If None, MMFF charges
    pharm_multi_vector=False # recommended
    # e.g., carbonyls get one HBA vector rather than two
)

# Surface point cloud + electrostatic potential
surf_pos = ref_molec.surface.positions # np.array (200,3)
esp = ref_molec.surface.esp # np.array (200,)

# Pharmacophores as a `Pharmacophore` container, unpacks as (types, positions, vectors)
# ref_molec.pharmacophore.types: np.array (P,)
# ref_molec.pharmacophore.{positions/vectors}: np.array (P,3)
pharm = ref_molec.pharmacophore

3D similarity scoring

An example of scoring the similarity of two different molecules using 3D surface, ESP, and pharmacophore similarity metrics.

from shepherd_score.score.constants import ALPHA
from shepherd_score.conformer_generation import embed_conformer_from_smiles
from shepherd_score.conformer_generation import optimize_conformer_with_xtb
from shepherd_score.container import Molecule, MoleculePair

# Embed a random conformer with RDKit
ref_mol_rdkit = embed_conformer_from_smiles('Oc1ccc(CC=C)cc1', MMFF_optimize=True)
fit_mol_rdkit = embed_conformer_from_smiles('O=CCc1ccccc1', MMFF_optimize=True)
# Local relaxation with xTB
ref_mol, _, ref_charges = optimize_conformer_with_xtb(ref_mol_rdkit)
fit_mol, _, fit_charges = optimize_conformer_with_xtb(fit_mol_rdkit)

# Extract interaction profiles
ref_molec = Molecule(ref_mol,
                     num_surf_points=200,
                     partial_charges=ref_charges,
                     pharm_multi_vector=False)
fit_molec = Molecule(fit_mol,
                     num_surf_points=200,
                     partial_charges=fit_charges,
                     pharm_multi_vector=False)

# Centers the two molecules' COM's to the origin
mp = MoleculePair(ref_molec, fit_molec, num_surf_points=200, do_center=True)

# Compute the similarity score for each interaction profile
shape_score = mp.score_with_surf(ALPHA(mp.num_surf_points))
esp_score = mp.score_with_esp(ALPHA(mp.num_surf_points), lam=0.3)
pharm_score = mp.score_with_pharm()

Alignment

Next we show alignment using the same MoleculePair class.

# Centers the two molecules' COM's to the origin
mp = MoleculePair(ref_molec, fit_molec, num_surf_points=200, do_center=True)

# Align fit_molec to ref_molec with your preferred objective function
# By default we use automatic differentiation via pytorch
surf_points_aligned = mp.align_with_surf(ALPHA(mp.num_surf_points),
                                         num_repeats=50)
surf_points_esp_aligned = mp.align_with_esp(ALPHA(mp.num_surf_points),
                                            lam=0.3,
                                            num_repeats=50)
pharm_pos_aligned, pharm_vec_aligned = mp.align_with_pharm(num_repeats=50)

# Optimal scores and SE(3) transformation matrices are stored as attributes
mp.sim_aligned_{surf/esp/pharm}
mp.transform_{surf/esp/pharm}

# Get a copy of the optimally aligned fit Molecule object
transformed_fit_molec = mp.get_transformed_molecule(
    se3_transform=mp.transform_{surf/esp/pharm}
)

Alignment of multiple MoleculePair objects can be accelerated with MoleculePairBatch with Jax installed.

from shepherd_score.container import MoleculePairBatch

batch = MoleculePairBatch(pairs)  # `pairs` is a list of MoleculePair objects

# accelerated JAX-based volumetric alignment via padding
scores, aligned = batch.align_with_vol()

# Multi-CPU parallel via shard_map (must set XLA_FLAGS *before* importing JAX)
scores, aligned = batch.align_with_vol(num_workers=4, num_buckets=4, use_shmap=True)

Evaluation Examples and Scripts

We implement three evaluations of generated 3D conformers. Evaluations can be done on an individual basis or in a pipeline. Here we show the most basic use case in the unconditional setting.

• ConfEval
  • Checks validity, pre-/post-xTB relaxation
  • Calculates 2D graph properties
• ConsistencyEval
  • Inherits from ConfEval and evaluates the consistency of the molecule's jointly generated interaction profiles with the true interaction profiles using 3D similarity scoring functions
• ConditionalEval
  • Inherits from ConfEval and evaluates the 3D similarity between generated molecules and the target molecule

Note: Evaluations can be run from any molecule's atomic numbers and positions with explicit hydrogens (i.e., straight from an xyz file).

Examples

Full jupyter notebook tutorials/examples for evaluations are found in the examples folder. Some minimal examples are below.

from shepherd_score.evaluations.evaluate import ConfEval
from shepherd_score.evaluations.evaluate import UnconditionalEvalPipeline

# ConfEval evaluates the validity of a given molecule, optimizes it with xTB,
#   and also computes various 2D graph properties
# `atom_array` np.ndarray (N,) atomic numbers of the molecule (with explicit H)
# `position_array` np.ndarray (N,3) atom coordinates for the molecule
conf_eval = ConfEval(atoms=atom_array, positions=position_array)

# Alternatively, if you have a list of molecules you want to test:
uncond_pipe = UnconditionalEvalPipeline(
    generated_mols = [(a, p) for a, p in zip(atom_arrays, position_arrays)]
)
uncond_pipe.evaluate(num_workers=4)

# Properties are stored as attributes and can be converted into pandas df's
global_series, sample_df = uncond_pipe.to_pandas()

Scripts

Scripts to evaluate ShEPhERD-generated samples can be found in the scripts directory.

Data

We provide the data used for model training, benchmarking, and all ShEPhERD-generated samples reported in the paper at this Dropbox link. There are comprehensive READMEs in the Dropbox describing the different folders.

License

This project is licensed under the MIT License -- see LICENSE file for details.

Citation

If you use or adapt shepherd_score or ShEPhERD in your work, please cite us:

@inproceedings{
adams2025shepherd,
title={Sh{EP}h{ERD}: Diffusing shape, electrostatics, and pharmacophores for bioisosteric drug design},
author={Keir Adams and Kento Abeywardane and Jenna Fromer and Connor W. Coley},
booktitle={The Thirteenth International Conference on Learning Representations},
year={2025},
url={https://openreview.net/forum?id=KSLkFYHlYg}
}