EthnoHack 2026 · Track 2

Among hundreds of thousands of plant compounds, which deserve a closer look?

We took that question literally, for three bacterial pathogens, and answered it with a number next to every compound saying how far the answer can be trusted.

Built by

Team Arxigoi

Leonidas Anagnou · Eleni Nasopoulou

School of Electrical and Computer Engineering, National Technical University of Athens

Medicine has measured a great deal of chemistry against these three organisms. Almost none of it came from a plant.

Published MIC measurements are what a model like this can learn from. The plant-derived share of them is a rounding error.

Staphylococcus aureus

Gram-positive

2.4%

1,424 plant-derived of 58,191 measured compounds.

Escherichia coli

Gram-negative

2.0%

874 plant-derived of 43,099 measured compounds.

Mycobacterium tuberculosis

Acid-fast

3.0%

691 plant-derived of 22,907 measured compounds.

Against that sits 97,709 plant structures with no measurement against any of these three. The literature cannot be searched for an experiment nobody has run, so ranking them is the only move left, and it is only worth anything if it says where it stops being trustworthy.

97,709 plant compounds screened. 150 we would hand to a chemist.

Three bacterial pathogens, one model each. Each model learns from published MIC measurements in ChEMBL, then scores plant compounds from LOTUS that nobody has tested. Every score here is shown with its distance from chemistry the model already knows, because one without the other says nothing.

Plates

Staphylococcus aureus

97,709 compounds screened · 50 shortlisted · 15 beyond the read

WellA1 · rank 1 of 50AOJJSUZBOXZQNBPredicted0.909Nearest known0.986no source plant on record

Well A1 · rank 1 of 50

Predicted
0.909
against a 25% base rate in held-out plant compounds
Nearest training
0.986
a close analogue of something the model was trained on
Read as
Inside the read
The model was tested on compounds this far out.
Chemistry
Polyketides
Polycyclic aromatic polyketides

Reported in

No source plant in LOTUS. It reached the shortlist on its chemistry alone.

Skeleton

AOJJSUZBOXZQNB

Reference

No DOI on this record.

SMILES

C[C@H]1[C@H]([C@H](C[C@@H](O1)O[C@H]2C[C@@](CC3=C2C(=C4C(=C3O)C(=O)C5=C(C4=O)C(=CC=C5)OC)O)(C(=O)CO)O)N)O

The same model, read three ways. Only the last one answers the question.

Scored on validation, and on a test set with whole chemical scaffolds held out, the model looks strong. Hold out every plant-derived compound, which is what a plant screen actually asks of it, and it drops by a sixth. That last reading is the hardest of the three for all three pathogens.

Staphylococcus aureus

Gram-positive

0.719 plant holdout0.863 scaffold test0.880 validation

Plant holdout 95% interval 0.689–0.749

Escherichia coli

Gram-negative

0.682 plant holdout0.903 scaffold test0.882 validation

Plant holdout 95% interval 0.609–0.750

Mycobacterium tuberculosis

Acid-fast

0.629 plant holdout0.815 scaffold test0.846 validation

Plant holdout 95% interval 0.582–0.677

AUROC is the chance the model ranks a random active above a random inactive, where 0.5 is a coin flip; intervals are 2,000 bootstrap resamples. Scaffold overlap across the split is exactly 0.000, which is necessary and nowhere near sufficient.

Holding out ring skeletons is an easier test than it sounds, and the distances say why.

Removing shared ring skeletons removes none of the resemblance that matters. The median plant compound sits far further out than the median test compound.

Staphylococcus aureus

Gram-positive

0.328 read line0.461 median plant compound0.680 median test compound

18,634 distinct ring skeletons across 58,191 measured compounds.

Escherichia coli

Gram-negative

0.328 read line0.437 median plant compound0.677 median test compound

13,303 distinct ring skeletons across 43,099 measured compounds.

Mycobacterium tuberculosis

Acid-fast

0.328 read line0.373 median plant compound0.677 median test compound

7,401 distinct ring skeletons across 22,907 measured compounds.

Similarity to the nearest compound the model was trained on, where 1.000 is an identical fingerprint and 0.000 nothing in common. The read line is one cutoff, not three: 0.328, the 25th percentile of the S. aureus plant holdout, which is the most distant set on which this model’s performance was actually measured. It is drawn on all three scales so the three can be read against each other, and applying one organism’s cutoff to the other two is a choice we made rather than a measurement. Candidates beyond it are published, not ranked.

The chemistry medicine has measured, and how little of it ever came from a plant.

Faint points are compounds with a published MIC against the plate on view. The brighter ones a plant made. Rings are where this run's candidates fell.

1,500 measured compounds
A sample of the 58,191 with a published MIC against S. aureus.
35 of them from plants
In the whole measured set, 1,424 of 58,191. That ratio is the problem this project exists for.
50 shortlisted
Scored inside the applicability domain, and ranked.
15 beyond the read
Published, not ranked. Note where they are: not at the edges.

This figure argues against itself. The median shortlisted candidate sits 0.0029 from its nearest measured neighbour here; the median candidate beyond the read sits 0.0027. Two components carry 7.4% of the variance, so the boundary this page is built on is invisible here. That is why every well carries a Tanimoto and not a position on a map, and why we did not reach for a prettier embedding.

What this buys a lab: two to five times the hit rate of picking at random.

Precision at the top of the list is the number a chemist spends money against. Read each one against the base rate beside it.

Staphylococcus aureus

Gram-positive

Precision at top 10%
57.0%
against 25.3% picking at random95% interval 49%–65%
Enrichment, top 5%
2.45×
ceiling is 4.0× at this base rate
Precision, top 5%
62.0%
12 of every 20 syntheses
On the scaffold test
97.7%
what the easier split would have let us claim

Escherichia coli

Gram-negative

Precision at top 10%
29.9%
against 9.0% picking at random95% interval 22%–40%
Enrichment, top 5%
5.03×
ceiling is 11.1× at this base rate
Precision, top 5%
45.5%
9 of every 20 syntheses
On the scaffold test
96.8%
what the easier split would have let us claim

Mycobacterium tuberculosis

Acid-fast

Precision at top 10%
43.5%
against 23.7% picking at random95% interval 32%–55%
Enrichment, top 5%
2.29×
ceiling is 4.2× at this base rate
Precision, top 5%
54.3%
11 of every 20 syntheses
On the scaffold test
92.2%
what the easier split would have let us claim

The chemistry the model prefers tracks each pathogen’s cell envelope, and nobody told it about cell envelopes.

Three envelopes, chosen so this comparison would say something. The acid-fast list draws alkaloids and terpenoids, the Gram-negative one the polar chemistry its porins pass. The model was never given a cell wall.

S. aureusE. coliM. tuberculosis
Shikimates and Phenylpropanoids19164
Polyketides161011
Alkaloids6117
Terpenoids8313
Amino acids and Peptides496
Carbohydrates060
Fatty acids160

Counts of the 50 candidates per pathogen carrying each pathway, as classified in LOTUS. A compound with two pathways counts under both, so rows do not sum to 50.

The result we set out to find did not replicate, so it is reported as a negative.

We set out to show that holding out whole plant families is harder than holding out scaffolds. Measured four times, the magnitude never settles.

0.526 family split0.622 scaffold split0.752 random split

Mycobacterium tuberculosis MIC, 10 repeats, plant-derived compounds with exactly one reported family. Standard deviations 0.047 / 0.085 / 0.090. The family-minus-scaffold gap is -0.097, and the family split was the worse of the two in only 7 of 10 repeats.

  • Regularisation does not close the gap. Minimum leaf size swept across 1, 4 and 16 made plant-holdout AUROC fall monotonically on all three, so the drop is not a tuning artefact.
  • Tuning for AUPRC instead of AUROC changed nothing at equal budget: one inconsistent win in three, inside the noise.
  • Calibration does not transport across a prevalence shift. Fitted at 35–55% active, applied at 9–25%, and on two of three the Brier score gets worse.

What is wrong with this shortlist, in our own words.

A screen that only reports its wins is not a result. These are the defects we know about in the files on this page.

The top-ranked candidate is a near-copy of a training compound on all three plates. Reading S. aureus, E. coli, M. tuberculosis in that order here and below, their nearest training compounds sit at Tanimoto 0.986 / 0.982 / 0.976. On two of the three it is the same non-plant anthracycline-like polyketide, with no source plant and no DOI. The model is confident there because it has nearly seen it before, which is the whole phenomenon this page measures, appearing inside the deliverable. 2 / 1 / 1 candidates per plate sit above 0.90.

Plant provenance is incomplete. 12 / 8 / 3 of the 50 candidates per plate carry no source plant, family or reference. They reached the shortlist through the chemistry, and in a natural-products screen that is a gap, not a detail.

Per-compound feature attributions are not shown. The SHAP values written alongside these shortlists are numerically impossible, at magnitudes far outside the range a probability output permits, so they are excluded from this surface until the cause is found. The aggregate importances behind the written report are sound.

There is no protein-level target here. 99.6% of the MIC rows in ChEMBL for these organisms are measured against the whole organism, and only 35 single-protein targets carry any MIC at all. The model predicts whether a compound stops a bacterium growing. It says nothing about mechanism.

Nothing here is a clinical or therapeutic claim. These are model predictions over published measurements and traditional-use records, not evidence of efficacy, and no dose, preparation or use is implied or recommended.

Data and licences

Activity labels from ChEMBL (CC BY-SA). The plant-to-compound edge from LOTUS (CC0), cross-checked against GBIF for which organisms are plants. NPASS was evaluated and dropped: its MIC table is a copy of ChEMBL, so ChEMBL is cited directly.

How to reproduce this page

Every number above is read at build time from results/, never typed in by hand. The shortlists come from configuration a04b69bf83 in results/runs.jsonl, which records every rule the stage was given. The curation and split stages were re-run end to end from the raw downloads on a second machine, and all 57 ledger metrics matched to the last decimal. The model and shortlist stages have not been reproduced that way yet.

What this is for, and what we would do next.

A screen is only worth the lab time it saves, so the claim has to be the size of the evidence.

  1. What a laboratory does with it tomorrow

    A bench queue rather than a literature search: fifty compounds, each with a source plant, a reference and a distance. At 5.03× enrichment on E. coli, twenty syntheses expect several hits where random expects one. None of it has been near a laboratory yet.

  2. The protein layer we could not build

    This predicts growth inhibition and says nothing about mechanism. Protein-level activity exists in ChEMBL under IC50 and Ki, a separate and much smaller download, and it is what would turn a shortlist into a hypothesis about a target.

  3. The measurement that would settle it

    Every number here is retrospective. The honest next step is not a fourth pathogen. It is putting ten of these on a real plate and publishing the hit rate beside the predicted one, including if it is worse.

  4. Whose knowledge this is

    Traditional knowledge sits behind many of the published plant records. Anything built on this carries the Nagoya Protocol and Local Contexts as a condition of use, not as a footnote.

Who built it, and what it stands on.

Team Arxigoi: two people, one week, one repository. Every number here is read at build time from a committed result file.

  • Leonidas Anagnou
  • Eleni Nasopoulou

Both at the School of Electrical and Computer Engineering, National Technical University of Athens. Which run produced the numbers on this page, and how to re-run it, is stated under what is wrong with it.

Data
Activity labels from ChEMBL (CC BY-SA). The plant-to-compound edge from LOTUS (CC0), cross-checked against GBIF for which organisms are plants. NPASS was evaluated and dropped: its MIC table is a copy of ChEMBL, so ChEMBL is cited directly.
Built with
Python, RDKit, scikit-learn and XGBoost for the pipeline; Next.js and Tailwind for this surface; smiles-drawer for every structure, drawn in the browser from the canonical SMILES.
Generative AI
AI assistance was used while writing the pipeline, this interface and parts of the analysis. Every number, method and claim is ours to defend, and each one is traceable to a committed result file.