rust-fpocket

A Rust reimplementation of fpocket — the protein binding-pocket detection algorithm — packaged as a pip-installable Python wheel.

fpocket is a widely used C program (Le Guilloux, Schmidtke & Tuffery, BMC Bioinformatics 2009) for finding druggable pockets in protein structures. It has no PyPI wheel: it must be compiled from C against qhull, or installed via conda/bioconda. fpocket-rs ports the algorithm to Rust and ships as a self-contained wheel with no system dependencies — no qhull, no C compiler, no conda.

import fpocket_rs

result = fpocket_rs.detect_pockets("1hvr.pdb")
print(result["n_pockets"])               # number of pockets found
top = result["pockets"][0]               # ranked by score
print(top["drug_score"], top["volume"])  # druggability, volume

Installation

pip install fpocket-rs

To build from source you need a Rust toolchain and maturin:

pip install maturin
maturin build --release
pip install target/wheels/fpocket_rs-*.whl

The algorithm

fpocket detects pockets geometrically, in five stages — each is a module of the Rust crate:

1. Voronoi tessellation (voronoi.rs) — a 3D Delaunay triangulation of the protein's heavy atoms; each Delaunay tetrahedron yields one Voronoi vertex (the circumcentre), a candidate alpha sphere.
2. Alpha-sphere filtering (alpha_spheres.rs) — keep spheres whose radius is in [min_radius, max_radius]; classify each apolar/polar by the electronegativity of its four contacting atoms.
3. Clustering (cluster.rs) — single-linkage clustering of the alpha spheres into pockets at clust_max_dist.
4. Descriptors (descriptors.rs) — per-pocket volume, alpha-sphere statistics, hydrophobicity/polarity/charge scores, SASA, etc.
5. Druggability scoring (druggability.rs) — a logistic-regression model over the normalized descriptors gives a druggability score in [0, 1].

The Voronoi tessellation is a dependency-free Bowyer-Watson implementation. It was verified to produce a Voronoi-vertex set identical to qhull's qvoronoi — all 10734 vertices matched within 0.05 Å on the 1AZ8 heavy-atom set.

API

def detect_pockets(
    pdb_path_or_str,        # path to a .pdb file, or PDB-format text
    *,
    min_radius=3.4,         # min alpha-sphere radius  (fpocket -m)
    max_radius=6.2,         # max alpha-sphere radius  (fpocket -M)
    clust_max_dist=2.4,     # clustering cutoff        (fpocket -D)
    min_apol_neigh=3,       # apolar-atom threshold    (fpocket -A)
    min_pock_nb_asph=15,    # min alpha spheres/pocket (fpocket -i)
    min_as_density=0.7,     # min alpha-sphere density
    mc_iter=300,            # Monte-Carlo volume iters (fpocket -v)
) -> dict

Returns {"n_pockets": int, "pockets": [...]}. Each pocket is a dict with:

The default parameters match fpocket's fparams.h (M_MIN_ASHAPE_SIZE_DEFAULT 3.4, M_MAX_ASHAPE_SIZE_DEFAULT 6.2, M_CLUST_MAX_DIST 2.4, etc.).

Parity with the fpocket C binary

Exact bit-for-bit parity through qhull is not attainable — qhull and the Bowyer-Watson triangulation resolve cospherical degeneracies differently, and some filtering tests sit on floating-point boundaries. We measured agreement honestly against the fpocket C binary (built from the Discngine/fpocket source) on three real structures:

What this means concretely:

• Every pocket fpocket-C finds is recovered, with 100 % alpha-sphere membership overlap — each C pocket's spheres are all present in the matching Rust pocket.
• Druggability scores match to a mean absolute error of 0.005–0.027. Example top pockets: 1HVR C 0.903 / Rust 0.924; 3EML C 0.984 / Rust 0.994; 1AZ8 C 0.087 / Rust 0.083.
• Hydrophobicity, polarity and charge scores are exact when the residue set matches (they depend only on residue identity).
• alpha_sphere_density, mean_local_hydrophobic_density, as_max_dist and volume agree closely.

Known differences:

• rust-fpocket reports a few extra small pockets (e.g. 10 vs 6 on 1AZ8). The extra clusters are real alpha-sphere groups; fpocket's exact set of retained alpha spheres differs by a small number due to qhull-vs-Bowyer-Watson degeneracy handling and filter-threshold boundary cases.
• Matched pockets carry 1–2 extra alpha spheres each for the same reason.
• total_sasa can differ for large/elongated pockets, since SASA scales with the (slightly larger) sphere set; this mildly affects the absolute pocket score but not pocket identity or ranking of the top pocket.

The parity tests in tests/test_parity.py reproduce these numbers; they are skipped automatically if no fpocket C binary is on PATH (or pointed at by the FPOCKET_BIN environment variable).

Development

# build + install into the current environment
maturin develop --release

# Python tests (functional + parity)
pytest tests/

# Rust unit tests
cargo test --manifest-path rust/Cargo.toml

The fpocket-ref/ directory (git-ignored) holds a clone of the fpocket C source used purely as a parity reference; clone it with git clone https://github.com/Discngine/fpocket fpocket-ref.

License

MIT — the same license as upstream fpocket. See LICENSE.

fpocket is © 2012–2018 Vincent Le Guilloux, Peter Schmidtke and Pierre Tuffery. This is an independent reimplementation, not affiliated with the fpocket authors.