Model · Life sciences
Herb–Drug Interactions
A model that predicts when a medicinal plant will change how a prescribed drug behaves in the body, starting with the plants of Ayurveda. Every prediction is traced to a compound, an enzyme and a study.
Model card
- Identifier
- truth-hdi
- Area
- Life sciences
- Status
- Active
- Version
- none released
- Access
- Not yet released. Planned tier for 0.1 is open weights and code.
- Licence
- To be stated at release.
Model names start with truth- after the foundation's mission. The name is not a claim that the model's outputs are correct; each model card states what has been measured.
Intended use
- Help pharmacologists, clinicians and researchers prioritise which herb–drug pairs to review.
- Rank a patient's drugs by exposure to interference from a disclosed herb, with the mechanism named.
- Support research into enzyme-mediated interactions.
Out of scope
- Diagnosis, prescribing or treatment decisions without a qualified professional.
- Use by patients as medical advice.
- Claims about herbs or drugs outside the data described below.
How it works
The model follows the biology rather than learning only from examples. A plant contains compounds. Those compounds act on enzymes. A drug is cleared by some of the same enzymes. When the plant's compounds inhibit or induce an enzyme that clears the drug, the drug's exposure in the body can change.
The deterministic mechanism term comes first, and learned components sit on top of it. That is why every prediction can be traced back to the compound, the enzyme and the published study behind it.
When it abstains
The model is designed to rank and explain, not to give a verdict for every pair. When the evidence for a pair is too thin, it says so rather than guessing. A ranking a pharmacologist can check is the goal.
What we learned building it
Training on published interaction labels taught an early version a drug's reputation instead of its pharmacology. We measured this, changed what the model claims to do, and are building a reviewer pack so pharmacologists can create the pair-level labels the field lacks. Read the full account.
Data
- Sources
- 22 public databases covering plant chemistry, measured drug activity, liver enzymes, protein targets and pathways.
- Size
- 117 million rows after cleaning, measured in the pipeline.
- Licences
- Open sources whose licences permit research use and redistribution. Substitutions are recorded and will be published with the recipe.
Evaluation
Design
- Drugs are held out entirely, so nothing is scored on a molecule it was trained on.
- Known pharmacology, such as grapefruit and St John's wort, is tested out of sample.
- Results are reported against a permutation control, not only against chance.
- Each learned component has pass criteria stated before it is fitted.
Results
| Test | Result | Evidence |
|---|---|---|
| Mechanism model, no labelled pairs | Ranks St John's wort with cyclosporine in the top 0.7% of all pairs and names CYP3A4 as the enzyme responsible. | Internal |
| CYP3A4 inhibition potency, scaffold split | Spearman 0.567 against a pre-registered bar of 0.50. | Internal |
| Pre-registered feature gates | Four candidate features failed the check that they were not simply measuring how well studied a plant is, and were not shipped. | Internal |
Known limits
- Most published interaction records describe the drug studied, not the outcome for a specific pair. Labels at pair level are scarce.
- The model covers enzyme-mediated interactions. Other mechanisms are not yet modelled.
- Coverage is limited to plants and compounds present in the source databases.
What comes next
| When | What |
|---|---|
| Q4 2026 | Data recipe, source licences and evaluation protocol published |
| Q1 2027 | truth-hdi-0.1: mechanism model and enzyme potency model, with code |
| Q2 2027 | truth-hdi-0.2: ranking model trained on expert-reviewed labels |