Small-molecule machine learning over public datasets: toxicity, affinity, docking. Three related projects merged into one repository, each a self-contained package with its own tests and commit history (imported via subtree merge).
cd comp_tox_pipeline && pip install -e .[dev]
snakemake --cores 4 # full split -> features -> fit -> report DAG (stubbed IO)
mol-ml is the small-molecule ML repo: toxicity and affinity prediction on public datasets (ToxCast/Tox21, DAVIS) with scaffold/cold-target splits as the headline results. Sibling repos:
trust-tools (agent security
and evals), bio-qc (lab-data QC
pipelines), lab-informatics
(lab data plumbing and integrity),
llm-posttraining
(training-stage behavior work),
protein-ml (protein fitness
ML), and mol-ml (small-molecule
ML).
| Directory | What it does |
|---|---|
comp_tox_pipeline/ |
Computational toxicology on EPA ToxCast/Tox21 + PubChem. Bemis-Murcko scaffold splits are the headline metric; conformal coverage and ECE reported on held-out scaffolds. |
dti_fusion/ |
Drug-target interaction on DAVIS: fingerprint + ESM-2 fusion. Cold-target split is the headline result; random split labeled as target-identity leakage. |
dockops/ |
Reproducible docking pipeline: batch ligand prep, Vina backend, mock engine for pipeline mechanics (labeled engine="mock" on every row). |
Each package is independent. From its directory:
cd comp_tox_pipeline && PYTHONPATH=src python -m pytest tests/ -qEach subdirectory retains its own AGENTS.md with project-specific rules
(split policy, provenance requirements, mock-vs-real labeling), which still
apply.
Same domain, same conventions. Public data only, structural splits as the prospective estimate, provenance manifests on every result. One repo keeps them consistent across three projects.