██████╗ ██╗   ██╗███████╗████████╗ █████╗ ██████╗  ██████╗
██╔══██╗╚██╗ ██╔╝██╔════╝╚══██╔══╝██╔══██╗██╔══██╗██╔════╝
██████╔╝ ╚████╔╝ ███████╗   ██║   ███████║██████╔╝██║     
██╔═══╝   ╚██╔╝  ╚════██║   ██║   ██╔══██║██╔══██╗██║     
██║        ██║   ███████║   ██║   ██║  ██║██║  ██║╚██████╗
╚═╝        ╚═╝   ╚══════╝   ╚═╝   ╚═╝  ╚═╝╚═╝  ╚═╝ ╚═════╝

Python Simulation Toolkit for Association Rate Constants

GPU-accelerated rigid body and flexible chain Brownian dynamics for bimolecular association rate constants (k_on)

---

Overview

PySTARC computes the bimolecular association rate constants (k_on) by implementing rigid body Brownian dynamics within the Northrup-Allison-McCammon formalism. k_on is the diffusion limited rate of reaching an outer surface, multiplied by the probability of subsequently reaching reactive contact before escaping. Trajectories run in parallel on the GPU with a NumPy CPU fallback. Applications include protein-ligand and protein-protein association, as well as other diffusion controlled encounters.

Method

Forces combine Adaptive Poisson-Boltzmann Solver (APBS) electrostatic grids near the receptor, a screened Coulomb (Yukawa) multipole far field outside the grid, Rotne-Prager-Yamakawa hydrodynamics, and Born desolvation. Reactions are captured by the closed form Brownian bridge crossing probability, such that a crossing between two recorded positions is detected without shrinking the timestep. Diffusion constants come from a Monte Carlo hydrodynamic radius.

Features

Performance

• Batch propagation. All trajectories advance as GPU arrays.
• Multi-GPU. Trajectories are divided across the available GPUs, while every GPU uses the same APBS grids.
• Blocked Born desolvation. The reverse-direction desolvation is evaluated in blocks so it fits in GPU memory.

Physical model

• Brownian bridge reactions. The crossing probability between two recorded positions is P = exp(-x₀·x₁ / (D_eff·Δt)), exact at a constant cost per step.
• Yukawa multipole far field. Outside the APBS grid the screened (Yukawa) potential is expanded as a multipole series through the monopole, dipole, and quadrupole terms.
• Rotne-Prager-Yamakawa hydrodynamics. The mobility tensor stays valid through the sphere overlap regime.
• Monte Carlo hydrodynamic radius. The radius comes from the solvent-excluded surface using a Kirkwood double sum, within ~1% of the analytical value.
• Bidirectional Born desolvation. Both the receptor in the ligand field and the ligand in the receptor field are evaluated, coupled by Newton's third law.
• Wilson score interval. The confidence interval holds for any P_rxn and any N ≥ 1.
• Adaptive timestep. A user configurable max_dt ceiling prevents b-surface overshoot.
• Quaternion rotation. Orientations compose directly, with no interpolation error.

Automation

• Scripted setup. setup.py builds PQR files, APBS grids, reaction criteria, and input.xml from PDB and topology files.
• Convergence diagnostics. Convergence is tracked by the relative standard error, the Wilson interval, a convergence curve, and a split-half test.
• Structured output. Each run produces trajectories, encounters, first passage times, radial densities, occupancy maps, pose clusters, fluxes, transition matrices, commitment probabilities, and energetics.
• Live progress. k_on and P_rxn are reported at a configurable interval.
• Checkpointing. Long runs can be saved and resumed.

Installation

GPU (Linux / HPC):

git clone https://github.com/anandojha/PySTARC.git
cd PySTARC
bash install_PySTARC.sh

The hsp90_inhibitors/ and ttk_inhibitors/ examples require the OpenEye Toolkits (academic license available, https://www.eyesopen.com).

Mac / CPU:

git clone https://github.com/anandojha/PySTARC.git
cd PySTARC
conda create -n PySTARC python=3.11 -y
conda activate PySTARC
conda install -c conda-forge ambertools apbs rdkit openbabel -y
conda install -c openeye openeye-toolkits -y
pip install matplotlib pdb2pqr
pip install dist/pystarc-1.1.0-py3-none-any.whl --force-reinstall

Testing

python -m pytest tests/
python -m pytest tests/ -v

Quick start

conda activate PySTARC
module load cuda                # HPC only
cd examples/two_charged_spheres
bash run.sh

Examples

See examples/ and examples/PARAMETERS.md for the parameter guide.

examples/
├── two_charged_spheres/              Analytical validation (exact Smoluchowski solution)
├── trypsin_benzamidine/              Protein-ligand (charged ligand, surface pocket)
├── beta_cyclodextrin_guests/         Host-guest (7 neutral guests, same receptor)
├── thrombin_thrombomodulin/          Protein-protein (electrostatically steered)
├── p38_mapk_sb203580/                Protein-ligand (neutral kinase inhibitor)
├── carbonic_anhydrase_inhibitors/    Protein-ligand (7 sulfonamides, 3 CA isozymes)
├── hsp90_inhibitors/                 Protein-ligand (6 HSP90 inhibitors)
├── ttk_inhibitors/                   Protein-ligand (8 TTK/MPS1 kinase inhibitors)
├── trypsin_benzamidine_multi_GPUs/   Cluster SLURM demo (single-GPU and multi-GPU)
└── barnase_barstar_chainbd/          Flexible chain BD (protein-protein, under active validation)

Requirements

• Python 3.11+
• AmberTools
• APBS
• OpenBabel
• RDKit
• OpenEye Toolkits
• NumPy
• SciPy
• Click
• Numba
• Matplotlib
• pdb2pqr
• CuPy

License

MIT

Citation

When using PySTARC, please cite:

Ojha et al. PySTARC: GPU-accelerated Brownian dynamics for bimolecular association rate constants (2026).