the-untested

Which plant nobody ever tested might fight a drug-resistant infection? This predicts the antimicrobial activity of natural products that were never assayed, from molecular structure alone, and renders the answer as a chemical-space certainty map over the source plants, fungi and microbes. It trains on every compound ChEMBL has measured and scores the ~227k LOTUS naturals that no one ever put on a plate. It beats a 1-NN similarity-search baseline on every scored disease (mean ROC-AUC 0.80), and the tell that it works: ranking the untested naturals for malaria surfaces three Artemisia species, the artemisinin genus, in the top six, from structure alone.

The result

amr_qsar, a per-target gradient-boosted QSAR over the 2048-bit Morgan fingerprint and 10 descriptors, one head per pathogen. It is scored on structurally novel molecules (Bemis-Murcko scaffold split, the same job it does live) against the 1-NN Tanimoto similarity-search baseline every real QSAR must beat. The served model is amr_qsar v1; later versions register with their own metrics and are promoted by hand, not automatically.

metric	value
mean AMR ROC-AUC (scaffold-held-out)	0.80
lift over 1-NN Tanimoto	positive on every scored head, +0.05 mean
strongest head (S. aureus)	0.90
weakest kept head (M. tuberculosis)	0.69
training compounds	176k labelled

Strong heads: S. aureus 0.90, T. cruzi 0.89, L. donovani 0.85, E. coli 0.85, C. albicans 0.84, P. falciparum 0.83. Weak: P. aeruginosa 0.74, M. tuberculosis 0.69. Beta-lactamase scores 0.54 and loses to the baseline, a dud kept and flagged rather than hidden.

The validation that matters is on the untested set. Ranking the 227k naturals for antimalarial activity surfaces three Artemisia species in the top six. Artemisia is the genus of artemisinin, the frontline antimalarial, recovered from structure alone: per-molecule score 0.98 against a family taxonomic prior of 0.21.

Caveats

Read these before quoting the number anywhere.

• Binding-active is not a cure. A high score means a molecule looks like things that were active in a lab assay. It is a research triage signal, not a medicine in a human. This is loud and permanent across the app.
• Scaffold split, never random. The model is scored on molecules whose chemical scaffold was held out. Random splits leak close analogs and inflate the numbers.
• Applicability domain is first-class. Every prediction carries a familiarity, the distance to the training set in chemical space. A molecule unlike anything the model has seen reads as a long shot, not a confident answer.
• The panel is what has data. Only pathogens with enough public ChEMBL activity get a head. These are the neglected and resistant diseases people screen against, which is not the same as diseases anyone can cure.

Architecture

An FTI (feature, training, inference) system on Hopsworks. The join key across every source is the InChIKey. Training reads through a feature view so serving selects the same features the same way, with no train/serve skew: a molecule scores identically on the endpoint and in the batch map.

flowchart LR
    lotus([LOTUS]):::ext
    chembl([ChEMBL]):::ext

    subgraph FE[Feature]
        direction TB
        f1[lotus_pipeline] --> np[(natural_product)]:::hops
        f1 --> oc[(organism_compound)]:::hops
        f2[chembl_pipeline] --> ca[(compound_activity)]:::hops
        ca --> f2b[labels_pipeline] --> cl[(compound_labels)]:::hops
        np --> f3[features_pipeline]
        ca --> f3
        f3 --> mf[(molecule_features)]:::hops
        np --> f3b[smiles_pipeline] --> ms[(molecule_smiles)]:::hops
    end
    subgraph TR[Training]
        direction TB
        fv{{qsar_fv · scaffold split}}:::hops --> t1[GBM bar] --> reg[(Model Registry)]:::hops
        gfv{{qsar_gnn_fv}}:::hops --> t2[Chemprop D-MPNN] --> reg
    end
    subgraph INF[Inference]
        direction TB
        map[(plant_property_map)]:::hops
        ep[[amrscorer · KServe]]:::hops --> app[untestedmap app]
    end

    cl --> fv
    mf --> fv
    cl --> gfv
    ms --> gfv
    reg --> map --> app
    reg --> ep
    who([researcher]):::ext --> app --> ep

    classDef hops fill:#10b98118,stroke:#34d399,stroke-width:1px;
    classDef ext fill:none,stroke:#9ca3af,stroke-dasharray:4 3;

Loading

Training is staged. Stage 1 is the fingerprint GBM above, the bar. Stage 2 is a multi-task message-passing GNN (Chemprop D-MPNN) over the molecule graph, served the same labels through qsar_gnn_fv on canonical SMILES, evaluated on the exact same scaffold-held-out test set. It promotes only if it clears the stage-1 bar.

The sources:

source	gives	scale
LOTUS	organism to molecule, SMILES, taxonomy, chemical class	544k links, 227k molecules, 37k organisms
ChEMBL	molecule to target, pchembl potency label	2.9M compounds, ~42k overlap with LOTUS

The file-by-file map:

chem_features.py                 shared, skew-free: SMILES -> Morgan fp + scaffold + descriptors
panel.py                         the AMR target panel, single source of truth
pipelines/lotus_pipeline.py      F1   LOTUS -> natural_product + organism_compound   (Hopsworks job)
pipelines/chembl_pipeline.py     F2   ChEMBL bulk -> compound_activity               (Hopsworks job)
pipelines/labels_pipeline.py     F2b  pivot the panel -> wide compound_labels        (Hopsworks job)
pipelines/features_pipeline.py   F3   Morgan fingerprints -> molecule_features       (Hopsworks job)
pipelines/smiles_pipeline.py     F3b  canonical SMILES -> molecule_smiles, graph input (Hopsworks job)
pipelines/train.py               T1   qsar_fv -> amr_qsar, the GBM bar -> registry    (Hopsworks job)
pipelines/gnn_train.py           T2   qsar_gnn_fv -> Chemprop D-MPNN, gated on the bar (Hopsworks job)
pipelines/map_pipeline.py        I1   score untested naturals -> plant_property_map   (Hopsworks job)
serving/                         I2   amrscorer predictor + KServe deploy
app/                             A1   untestedmap certainty-map + discovery app
tools/                           schedule.py, build_envs.py
reqs/the-untested.md             the FTI specification

Reproduce

Clone into a Hopsworks project on the /hopsfs/... FUSE mount. Paths self-derive, nothing is hardcoded to a username. All data is free bulk, no keys.

make envs           # clone the RDKit / torch envs (featurize / train / gnn / serve / app)
make lotus-job      # F1   LOTUS map + molecule structures + taxonomy
make chembl-job     # F2   ChEMBL bioactivity labels (bulk SQLite)
make labels-job     # F2b  pivot the panel into the wide multi-task label group
make features-job   # F3   Morgan fingerprints -> molecule_features
make smiles-job     # F3b  canonical SMILES -> molecule_smiles (graph input)
make train-job      # T1   multi-task QSAR (scaffold split, vs 1-NN baseline)
make map-job        # I1   score the untested naturals -> plant_property_map
make serve          # I2   amrscorer on-demand endpoint
make app            # A1   untestedmap app

The demo

untestedmap: a chemical-space map where every dot is an untested natural product, placed by molecular shape (t-SNE of the fingerprints, so neighbours are structurally alike). Pick a disease and the map recolours by predicted activity; the attention rail ranks the plants, fungi and microbes most likely to carry it, none ever tested for it. A dossier draws the molecule, links the source organism to Wikipedia, and shows the familiarity to known chemistry so an out-of-domain guess reads as a long shot. A live box scores any pasted SMILES against every disease.

The broad-spectrum view recolours the same map by how many diseases a molecule is predicted to hit at once. It surfaces the rare multi-target naturals, and flags the ones the model has seen little of, where broad activity is as likely a frequent-hitter artifact as a real lead. Every screen repeats one rule: activity in an assay is a triage signal, not a medicine.

Full specification: reqs/the-untested.md.