ermak 0.2.0

GPU-accelerated Brownian dynamics for ligand diffusion and dissociation kinetics in crowded environments

ermak (Python)

Python bindings for ermak, a Brownian-dynamics engine for ligand diffusion and dissociation kinetics in crowded environments. The core integrates the overdamped Langevin equation with the Ermak-McCammon propagator; particles are coarse-grained spheres in implicit solvent, in reduced Lennard-Jones units (kB T = 1, sigma = 1, bare D0 = 1).

pip install ermak

import ermak

# Free diffusion recovers D0.
ermak.free_diffusion_deff(d0=1.0, dt=0.01, steps=200, replicas=4000)

# Crowders slow it down.
box = 8.0
crowders = ermak.cubic_lattice(box, 4)            # 4^3 obstacles
ermak.crowded_diffusion_deff(
    d0=1.0, dt=0.01, steps=200, replicas=2000,
    box_l=box, crowders=crowders, sigma=1.0, eps=1.0,
)

# Residence time (1/k_off) climbs with the pocket barrier (Kramers).
ermak.mean_residence_time(barrier=4.0, r_b=2.0, d0=1.0, dt=1e-3,
                          max_steps=80_000, replicas=400)

# tauRAMD egress times rank the true residence times.
ermak.tauramd_egress_time(barrier=2.0, r_b=2.0, d0=1.0, dt=1e-3, accel=6.0,
                          reorient_steps=100, max_steps=200_000, replicas=400)

# Predict log k_off from descriptors with a random forest.
forest = ermak.Forest.fit(rows, log_koff, n_trees=200)
forest.predict_many(rows)

GPU

The Linux x86-64 wheel is GPU-accelerated through the CUDA driver API. It needs no CUDA toolkit to install (the driver is loaded at runtime); gpu_available() reports whether a usable device is present.

import ermak

if ermak.gpu_available():
    box = 8.0
    crowders = ermak.cubic_lattice(box, 5)
    deff = ermak.crowded_diffusion_deff_gpu(
        d0=1.0, dt=2e-4, steps=10_000, replicas=200_000,
        box_l=box, crowders=crowders, sigma=1.0, eps=1.0,
        precision="f32",     # "f64" for the reference path
    )

gpu_available() and crowded_diffusion_deff_gpu(...) exist on every wheel; on a build or machine without GPU support the latter raises a clear RuntimeError, so calling code can branch on gpu_available(). Dual-licensed under MIT or Apache-2.0.