Healer ML

The first attempt was a QSAR/similarity screen over roughly 1.1 million compounds pulled from ChEMBL and ZINC. It returned nothing. The applicability domain rejected 91% of catalogue chemistry outright, and not a single surviving compound scored above p_active 0.25.

That is a real result, not a bug. Similarity-driven models can only find chemistry that resembles what they were trained on, and the known LSD1 chemotypes are narrow. If a novel scaffold binds LSD1, a ligand-based model is structurally incapable of noticing.

Docking has no such limitation: it scores against the protein, so chemistry unlike anything in ChEMBL is still in scope. The cost is that docking fails silently and confidently — which is what the rest of this project is really about.

Version 5 used a generic “remove ligands and waters” prep, which deleted FAD along with the solvent. FAD is a covalently attached cofactor, not a ligand: stripping it opens a 53-atom cavity that Vina fills happily with almost anything. Random ZINC compounds outscored real nanomolar inhibitors by 1.2 kcal/mol. Every number in that run was noise, and it looked like a working pipeline.

Version 6 tried to fix that by keeping FAD and inferring the pocket from its geometry — an offset from the FAD N5 atom. The box landed inside solid protein. Every score came back positive, meaning steric clash rather than binding, at +49 and +17 kcal/mol.

Version 8 stops guessing. It scans candidate structures for one that already contains a bound druglike heteroatom group, takes 5LHH with its reversible inhibitor 6X0, centres a 22 Å box on that ligand's own coordinates, and retains cofactors from an explicit keep-list. The box is sized against the measured ligand span, not a default.

Nothing downstream is allowed to run until the setup proves it can tell real inhibitors from random chemistry. Ten known potent reversible actives are docked against ten random library compounds: actives mean −10.49, decoys mean −8.70, a separation of +1.79 kcal/mol at p = 9.1e-5 and Cohen d = 2.58.

The active set needed one non-obvious correction. Tranylcypromine-class inhibitors — 1,950 of the 3,944 ChEMBL actives — bind covalently to FAD. Vina scores non-covalent binding only, so it cannot rank them correctly no matter how good the box is. Validating on them would have produced a false failure; they are excluded, leaving 1,994 reversible compounds.

Score separation still says nothing about whether the pose is in the right place. So the most potent reversible active is re-docked and its centroid compared against the crystallographic ligand that was stripped from the receptor: 2.2 Å apart. The pipeline is finding the actual binding site, not an incidental surface pocket.

The decoy distribution then calibrates the cutoff rather than a round number being picked by hand: −10.09 kcal/mol is the decoy mean minus 1.5 SD, which keeps 10 of 10 known actives and rejects roughly 93% of random chemistry.

300 library compounds were docked at exhaustiveness 4 — triage resolution — with a per-compound line of output, checkpointing every 10 results, and a wall-clock budget so a Colab runtime dying mid-run costs nothing. Resume reads back the checkpoint, but the receptor file is MD5-stamped and the run refuses to resume across a receptor change. A stale checkpoint had silently mixed two runs' numbers twice before that guard existed.

Vina rewards size: a floppy 45-heavy-atom molecule scores well without binding anything. Ligand efficiency — affinity per heavy atom — is the correction, thresholded at 0.30. Hits are then deduplicated to one compound per Murcko scaffold, so a single bad assumption cannot consume all of the picks. 7 compounds out of 134 successfully docked cleared both filters.

Similarity to known actives is reported but deliberately not filtered on. A hit far from known chemistry is exactly what this notebook exists to find — and also the hit most likely to be an artefact. All 7 sit below 0.4 Tanimoto to every ChEMBL active above pChEMBL 8.5.

Vina is stochastic, and exhaustiveness 4 is triage. Each survivor is re-docked at exhaustiveness 32 across three seeds; anything swinging more than 0.5 kcal/mol between seeds was a search artefact rather than a pose. All 7 held, the widest spread being ±0.17 kcal/mol. The best, CHEMBL6356, refines to −12.06 ± 0.02.

Which is where honesty matters more than the numbers. Five of the seven share a bis-aminoquinolinium motif — large, doubly cationic, conformationally floppy. That class is a well-known frequent hitter in docking, and electrostatics in a scoring function this simple flatter it. Vina also scores no desolvation penalty worth the name for permanent charges.

So the correct reading of this shortlist is: the pipeline is validated, the poses are in the right pocket, and the output is a set of hypotheses that need pose inspection against the substrate channel and FAD contacts before anyone spends money on compounds. The notebook ships a py3Dmol viewer for exactly that step, and it is the last thing the run does rather than the first.

Twelve diseases are scaffolded — lung, breast, colorectal, prostate and pancreatic cancer, glioblastoma, Alzheimer's, Parkinson's, type-2 diabetes, tuberculosis, malaria, and antimicrobial resistance. Each gets the same two-track shape, because finding a candidate and improving one have genuinely different inputs, metrics and failure modes: screening/ holds virtual screening, QSAR, hit triage and diagnostic classifiers; treatment/ holds lead optimisation, ADMET and toxicity, resistance modelling and patient stratification.

Curated inputs small enough to version live in data/ and are committed; everything a notebook produces — docked poses, checkpoints, PDBQT files, downloaded structures — goes to results/ and is gitignored. A result that matters gets promoted to data/ deliberately rather than by accident.

It is Apache 2.0 with a CLA and a contributing guide, because the point is not a portfolio piece. I am an ML engineer, not a doctor, and this needs people who know imaging, chemistry, statistics and clinical practice. Everything is public, including the two receptor preparations that produced confident nonsense.