Repository navigation
Releases
TopoKit v0.3.0
TopoKit v0.3.0
First public GitHub release of TopoKit, a Python toolkit for simplicial,
sequence-hyperdigraph and two-factor interaction topology.
Included
- Six composable layers: readers, builders, core, postprocessing,
visualization and workflows. - Homology, persistence, ordinary and explicitly requested persistent
Laplacians, spectral summaries and feature vectors. - Native alpha geometry using NumPy/SciPy; GUDHI remains an optional explicit
reference backend. - Fixed FS-AN protein–ligand feature extraction: 15 Å crop, 50 alpha radii,
55 element/null channels, ten summaries and 27,500 features per complex. - A manifest-based exporter with canonical schema, sample identities, hashes,
diagnostics and explicit failure records; ten small example complexes. - Three executed tutorial notebooks covering point clouds, explicit objects
and molecular/crystallographic input formats. - Agent guides and skills for topology extraction and the optional frozen
ESM-2 + CPZ sequence recipe. Sequence embeddings have 1,792 coordinates;
encoder weights are supplied separately.
Install
git clone https://github.com/ChenDdon/TopoKit.git
cd TopoKit
git checkout v0.3.0
python -m venv .venv
source .venv/bin/activate
python -m pip install .
python -m topokit infoOn Windows, activate the virtual environment with .venv\Scripts\activate.
Python 3.10+ is required. The base library depends only on NumPy and SciPy.
Install optional extras for plotting, ML, sequence encoders or supervised DL.
To run the supplied molecular examples from the repository root:
python workflows/protein_ligand_prediction/extract_features.py \
--manifest examples/protein_ligand/manifest.csv \
--output examples/output/protein_ligandThe completed matrix has shape (10, 27500); individual tensors have shape
(10, 50, 55). Output directories must be new.
Assets and validation
The wheel contains the Python library and small runtime schemas. The source
archive includes scripts, notebooks, tests, examples and documentation.
SHA256SUMS gives the checksums of both distribution files. No pretrained
weights, fitted models, generated research outputs or transfer archives are
included. This is a GitHub release, not a PyPI publication.
The release workflow publishes only after the test matrix passes on Linux,
macOS and Windows with Python 3.10 and 3.12, including installed-wheel core and
ten-complex molecular smoke checks. Local validation also covered 1,279 relevant
tests and extraction from an unpacked source archive. The optional CPZ profile,
asset hashes and CLI were checked; full pretrained-model inference was not
rerun as part of release preparation.
Scientific and data scope
Protein–ligand inputs must already be prepared in a common coordinate frame.
FS-AN computes complete ordinary L0 spectra over an alpha-radius schedule;
these descriptors are not themselves an affinity prediction. Geometry/resource
failures are explicit, and exact recipe identity matters for model reuse.
The MIT license covers first-party code. Example structures, optional weights
and dependencies retain their upstream terms. The data inventories and
NOTICE.md document provenance and unresolved redistribution metadata for
some local demonstration structures; this release does not assign them an MIT
data license. Historical H0/atom-deletion controllers preserve fixed-cohort
research protocols and are not generic alternatives to the default exporter.