Skip to content

the-untested.md

Latest commit

 

History

History
190 lines (159 loc) · 11.1 KB

the-untested.md

File metadata and controls

190 lines (159 loc) · 11.1 KB
Raw

the-untested: predict bioactivity of the never-tested natural world

Working name. Predict the medicinal (bioactivity) properties of natural products that were never assayed, from their molecular structure, aggregate the predictions up to the source plants/fungi, and render a plant × property certainty map. Demo foregrounds antimicrobial resistance (AMR).

The thesis in one line: train QSAR on the labelled ChEMBL compounds (176k carry a panel label), score the 227k LOTUS naturals that nobody ever tested, and show which untested plants likely carry activity, with an explicit confidence over where the model is blind.

Honesty rules

These are part of the spec. Without them, breadth produces confident wrong answers.

  1. Scaffold split, never random. Random splits leak analogs across train/test and inflate QSAR scores. Split by Bemis-Murcko scaffold so the model is scored on structurally novel molecules, the same job it does live.
  2. Beat the priors, or you earned nothing. Headline metric is lift over the taxonomic prior (same plant family as a known-active plant, from the LOTUS taxonomy columns), computed in the batch map next to every organism verdict. The folk prior from Dr. Duke's ethnobotany is a second baseline in the design, not built in v1.
  3. Applicability domain is first-class. Every prediction carries a familiarity = max Tanimoto to a capped sample of the training set in fingerprint space. A weird alkaloid the model never saw reads as a long shot, not a confident wrong answer.
  4. Coverage per plant. A plant prediction is only as trustworthy as the fraction of its chemistry we actually know. Show "40 of 60 known compounds characterized" next to every plant verdict.
  5. Binding-active is not a cure. pchembl activity in an assay ≠ medicinal in a human. Permanent, loud, on every prediction. This is a research triage tool, not medical advice.
  6. Population-split caveat. Natural products that reached ChEMBL are a biased sample. State it; read the lift over baseline, not the absolute AUC.

AI-system card

  • Prediction problem: multi-task binary classification. Molecule → activity vector across a target panel (active/inactive at a pchembl threshold per target). Multi-task so thin targets borrow strength from fat ones.
  • KPI: count of well-founded, novel bioactivity leads (untested plant × property) that survive the priors and the applicability-domain filter.
  • ML proxy metric: per-target ROC-AUC / PR-AUC / precision@k on scaffold- held-out molecules, headline = lift over the 1-NN Tanimoto baseline and the taxonomic prior; plus applicability-domain coverage.
  • Data sources: LOTUS (plant↔molecule + SMILES + taxonomy + chem-class), ChEMBL (molecule→target→pchembl label), Dr. Duke's (folk-use, the parked folk prior).
  • ML-system type: hybrid. Batch inference precomputes the full map (227k molecules × panel → activity + confidence, aggregated to 37k organisms) into feature groups; a real-time KServe deployment scores any molecule/plant on demand, including brand-new SMILES not in LOTUS.
  • Consumed: UI, the certainty map, plant/molecule search, an AMR attention rail (untested plants ranked by predicted antimicrobial activity × confidence × coverage), per-plant dossier showing the honest baselines.
  • Monitored: applicability-domain coverage of live queries, per-target calibration drift, label drift as ChEMBL releases update.

Data

source gives access scale
LOTUS frozen (Zenodo concept DOI 5794106) plant ↔ molecule links bulk .csv.gz 544k links, 227k molecules, 37k organisms
LOTUS metadata (same record) SMILES, formula, mass, xlogp, NPClassifier + ClassyFire class, full organism taxonomy (domain→species) bulk .csv.gz 227k molecules
ChEMBL 37 (EBI FTP bulk) molecule → target → pchembl activity SQLite/H5 dump (REST too flaky for bulk) 2.9M compounds; training pool = the labelled compounds on a panel target, ~42k overlap with LOTUS is the validation slice
Dr. Duke's phytochem (USDA) plant → chemical → ethnobotanical use bulk / site folk-prior baseline (parked)

Join key across all: InChIKey (exact and first-block skeleton both ~42k, so the strict exact join is enough; no fuzzy matching).

Target panel v1 (AMR-led, chosen in F2 by measured label density, panel.py is the single source of truth): 11 AMR heads, Plasmodium falciparum (malaria), Mycobacterium tuberculosis, Staphylococcus aureus, Escherichia coli, Pseudomonas aeruginosa, Enterococcus faecium, Candida albicans, the kinetoplastid parasites Trypanosoma cruzi / Leishmania donovani / Trypanosoma brucei, and Beta-lactamase, plus 3 human cytotoxicity heads (A549, HepG2, HeLa) that fatten the multi-task representation and power a selectivity read. The label-poor ESKAPE members (Klebsiella, Acinetobacter, Enterobacter) did not clear the density cut and are not in v1.

Pipelines (FTI, ordered by blocked-by)

F: Feature pipelines (model-independent transformations, MITs)

F1 · ingest-lotus: download LOTUS frozen + metadata, normalize, write two offline FGs: natural_product (InChIKey PK; SMILES, formula, exact_mass, xlogp, npclassifier_pathway/superclass/class, classyfire) and organism_compound (organism ↔ InChIKey with full taxonomy columns for the taxonomic prior). Backfill once; refresh only on LOTUS releases (rare). → skills: hops-data-sources, hops-fg. Blocked-by: none.

F2 · ingest-chembl: pull the ChEMBL 37 bulk dump (SQLite/H5, never the REST API for bulk), extract activities with a non-null pchembl_value for the target panel, binarize active/inactive per target at a fixed pchembl threshold, write offline FG compound_activity (InChIKey ↔ target ↔ label). Step one of this pipeline is the overlap measurement: count labeled molecules per target after joining ChEMBL InChIKeys to LOTUS, and fix the final target panel from the counts. → skills: hops-data-sources, hops-fg. Blocked-by: none.

F2b · pivot-labels: compound_activity is keyed by (InChIKey, target), so a feature view cannot join it on InChIKey alone. Pivot it into the wide FG compound_labels: one row per molecule, one column per panel target (1 active / 0 inactive / null unmeasured), keyed by InChIKey, so the FV joins it 1:1 to the features. → skills: hops-fg. Blocked-by: F2.

F3 · featurize-molecules: shared skew-free extractor chem_features.py (imported by this pipeline and by serving): SMILES → 2048-bit Morgan fingerprint (base64-packed) + Bemis-Murcko scaffold + 10 physchem descriptors. Featurize the LOTUS naturals and every ChEMBL compound on a panel target, write offline FG molecule_features (the model input). RDKit needs a cloned env. → skills: hops-features, hops-fg, hops-environments. Blocked-by: F1, F2.

F3b · smiles-map: the featurizer packs the fingerprint and drops SMILES, but the stage-2 GNN eats the molecule graph, built from SMILES. Persist canonical SMILES for exactly the molecules in molecule_features into the sidecar FG molecule_smiles, keyed by InChIKey, 1:1 with the features. → skills: hops-fg. Blocked-by: F3.

F4 · ingest-ethnobotany (parked): Dr. Duke's folk-use → FG folk_use, the folk-prior baseline. Not built in v1. → skills: hops-data-sources, hops-fg. Blocked-by: none.

T: Training pipeline (model-dependent transformations, MDTs, + model)

T1 · EDA: profile the LOTUS↔ChEMBL join, confirm the measured per-target label density, leakage audit. Enforce scaffold split (Bemis-Murcko), verify no InChIKey/scaffold crosses train/test, check class balance per target. Confirm the taxonomic-prior baseline computes. → skills: hops-eda, hops-eda-checklist. Blocked-by: F1, F2, F3.

T2 · train-qsar: FV qsar_fv = compound_labels (root/spine) JOIN molecule_features on InChIKey. The label FG is the root so features are fetched as-of the label event_time and a later-written label still matches. Per-target label masking (a molecule unlabeled for target k contributes no loss there).

  • Stage 1 (the bar): per-target HistGradientBoosting on the fingerprint + 10 descriptors. Registered as amr_qsar, the served champion.
  • Stage 2: multi-task Chemprop message-passing GNN on the molecule graph, served the same labels through qsar_gnn_fv (compound_labels JOIN molecule_smiles), scored on the same scaffold fold-0 test as the bar. Registered as amr_qsar_gnn.
  • Applicability-domain layer: ad_score = max Tanimoto to a capped sample of the training fingerprints, stored per prediction. Out-of-domain reads as a long shot, not a confident answer.
  • Evaluate: per-target ROC-AUC / PR-AUC / precision@k on scaffold-held-out molecules, applicability-domain coverage, and lift over the 1-NN Tanimoto baseline. Register model + eval images. Promote the GNN only if it clears the FP-GBM bar. Register every run (lineage), pick champion by advertised metric at serve time. → skills: hops-train, hops-transformations, hops-fv, hops-environments. Blocked-by: T1.

I: Inference pipelines

I1 · map-batch-inference: score all 227k LOTUS molecules × target panel with the champion → activity prob + AD confidence; aggregate to organism level (plant × property, weighted by chemistry coverage). Write FGs molecule_prediction and plant_property_map. Scheduled on model refresh. → skills: hops-batch-inference, hops-fg. Blocked-by: T2, F3.

I2 · qsar-serving: KServe deployment scoring any molecule (SMILES) or plant (name → its LOTUS compounds → aggregate) on demand, including brand-new SMILES absent from LOTUS. Uses the shared chem_features.py (no skew). Returns activity

  • confidence + the driving molecule + nearest known drug (Tanimoto) + AD flag. → skills: hops-online-inference, hops-environments. Blocked-by: T2.

I3 · app: the UI. A chemical-space certainty map (t-SNE of the fingerprints) recoloured per disease, an attention rail of the untested organisms ranked by predicted activity, a per-organism dossier that draws the molecule, links the source organism to Wikipedia, and shows the applicability-domain familiarity, a live SMILES scorer, and a broad-spectrum view that recolours the map by how many diseases a molecule hits at once and flags low-familiarity frequent-hitters. The "binding-active is not a cure" warning is permanent on every screen. Logs inputs

  • predictions for monitoring. → skills: hops-app, hops-monitoring. Blocked-by: I1, I2.

Engine note (taste sibling, later)

The pipeline is item → molecules → predict-a-property → certainty map. The food/flavor sibling (ingredient → flavor molecules → predict recipe co-occurrence) is the same engine with a different label source. Keep chem_features.py and the FG/FV shapes label-agnostic so the taste build drops in as a second target set, not a rewrite. Not built now.

Build order

F1, F2 in parallel (no deps) → F2b (needs F2) → F3 (needs F1, F2) → F3b (needs F3) → T1 → T2 → I1, I2 → I3. F4 parked.