Cheminformatics — QSAR & ADMET on rapamycin and its rapalogs

⚠️ Major caveat (read first)

Rapamycin and its rapalogs are the wrong molecules for de-novo ADMET ML. They are 900–1030 Da macrocycles (sirolimus, everolimus, temsirolimus, ridaforolimus, zotarolimus) that sit outside the applicability domain of essentially every publicly available trained ADMET model — which are built on small, drug-like molecules. Concretely, their nearest-neighbour Tanimoto similarity to the models' training sets is only ~0.2–0.4 (well below the ~0.3 reliability guideline), so the in-silico ADMET numbers here are extrapolation, not trustworthy predictions. Measured / label data is decisive.

This series is therefore, in large part, an honest demonstration of where ML QSAR/ADMET breaks — it computes the applicability domain and says so, rather than trusting the output. QSAR/ADMET-by-ML is most fruitful for small, in-domain drug-like molecules; these macrocycles are a deliberate stress test. (Only two things transfer: the potency model still ranks them roughly right, and BBB is correctly called "non-CNS" — because both are size/similarity-driven.)

What each notebook does

Notebook	Axis	What it does	Candid finding
rapamycin_qsar	Potency (mTOR)	RF on ~4,300 ChEMBL mTOR IC50s; random and scaffold split; predict rapalogs held-out	Qualitatively right (all potent) but under-predicts the most potent ~1 log; triage filter, not a ruler
rapamycin_ADMET_Absorption	A	BCS framing + ESOL solubility + BOILED-Egg; addendum = Tanimoto applicability-domain check of HIA/bioavailability/Caco-2 ML	Dissolution-limited (BCS II); ML models out of domain (Tanimoto ~0.2–0.4) → not usable; measured decides
rapamycin_ADMET_Distribution	D	CNS-MPO + BOILED-Egg (BBB); measured Vss/fu/B:P	BBB call works (non-CNS); Vss/B:P dominated by FKBP12-driven RBC sequestration — un-modelable from structure
rapamycin_ADMET_Metabolism	M	CYP3A4/2D6 inhibition + hepatocyte clearance ML (TDC)	Measured: CYP3A4/3A5 + P-gp substrates; models out of domain, and inhibition ≠ substrate
rapamycin_ADMET_Excretion	E	Half-life + clearance ML (TDC)	Biliary/fecal, long t½ (~62 h); half-life barely learnable (negative R²) and under-predicted
rapamycin_ADMET_Toxicity	T	hERG / Ames / DILI / LD50 ML (TDC)	Real toxicity is on-target mTOR (immunosuppression, metabolic, stomatitis) — invisible to structural QSAR

Each notebook follows the same shape (Purpose → Steps → Citations), reports its Tanimoto applicability domain, compares in-silico to measured data, and cites primary literature in ACS style (retrieved from PubMed). Public data: ChEMBL (potency) and the Therapeutics Data Commons admet_group benchmark (ADMET).

Conventions

• Notebooks are stripped of outputs before commit (nbstripout); each step's result is saved as a PNG in result_<notebook>/ so results stay viewable on GitHub.
• Data cached to data/ (git-ignored); reproducible uv environment (Python 3.11).

Quickstart

uv sync
uv run jupyter lab notebooks/     # select the "Python (cheminformatics_pubData)" kernel

If a notebook won't run, check the kernel first (wrong kernel → ModuleNotFoundError).

License

MIT — © 2026 Jin Mo Koo.