Engines & discovery
Most tools answer what is known. The engines estimate what no database holds: where a molecule could bind, what an untested compound might do, whether a dose or a trial could work. The discovery tools propose new ideas and try to kill them. Each one says what its output is worth.
01The engines
You do not call these directly. The agent reaches for them when a question turns on a quantity nobody has measured, and you can ask for one by describing what you want to know: "is there anywhere on this protein a small molecule could bind?", "what dose would this need?", "would a trial this size detect the effect?".
| Engine | What it computes | What it is worth |
|---|---|---|
| structure.pockets | Candidate binding sites on a PDB or AlphaFold structure, found by geometry alone, ranked, with the residues lining each one. | On twelve ligand-bound test structures, the true site ranked first in eleven. When the structure has a bound ligand, it grades its own ranking against it. |
| dock.pose | Where a molecule can sit in a site and which residues it touches, using AutoDock Vina's scoring function over an RDKit conformer ensemble. | The geometry holds up in redocking tests. The score does not: it ranked biotin last in streptavidin in two tests. Use it for where a molecule sits, never to rank compounds. |
| qsar.build, qsar.predict | A structure–activity model fitted during the run on the potency measurements ChEMBL holds for the target, then used to score untested structures. | Validated on held-out scaffolds, not a random split. Reports its accuracy against a predict-the-mean baseline. Each prediction carries its distance from the training data, and extrapolation is labelled. |
| pk.simulate | Repeated oral dosing to steady state, coupled to target engagement through the free concentration. Answers both ways: the coverage a dose gives, and the dose a coverage needs. | Names whether potency, half-life or protein binding is the constraint. Parameters estimated from structure are roughly threefold accurate and are labelled as estimates; measured values make it a real calculation. |
| trial.simulate | Power, required sample size, and assurance — the chance of success averaged over uncertainty in the effect. Also the chance a significant result has the wrong sign, and how much it would exaggerate the effect. | Assurance is usually lower than the power figure in a protocol. It needs a stated uncertainty in the effect to compute. |
| mr.analyze | Mendelian randomisation: whether genetically higher levels of a target change disease risk, by IVW, MR-Egger and weighted median together. | Reports instrument strength. Disagreement between the three methods is itself reported as a finding. |
| surrogate.validate | Whether moving a biomarker has predicted moving the clinical outcome across trials: trial-level R² with a bootstrap interval, and the smallest marker effect that predicts a real outcome effect. | Returns no R² below five trials. When you supply no trials, a model lists them from memory; the trial labels come back with the result so you can check them. |
| program.base_rates, program.simulate | What happened to programmes like this one, counted from the two bundled datasets, and a simulation of cost, time and the phase where a programme most likely fails. | Counting, not modelling. Handles programmes still running, and separates scientific failure from trials that stopped for operational reasons. Cost figures are published literature values and labelled as such. |
| efficacy.estimate | Composes five links — the target causes the disease, there is a pocket, the molecule engages it, a tolerable dose reaches it, the endpoint predicts benefit — against the reference class's approval rate. | Returns a probability with an interval and the link worth settling next. Only the genetic link moves the estimate, because it is the only one with a measured effect on outcomes; a missing link widens the interval. |
| design.audit | Checks an observational study's design for defects that manufacture an association, such as immortal time or selection on events after baseline. | The checks are deterministic. The negative controls it proposes come from a model, and the audit is re-run with them in place. |
02The gauntlet
A proposal that looks promising and one that is promising differ in one thing: whether anything was tried that could have shown it was wrong. The gauntlet takes a target and a disease — optionally a compound, a route and dose, an expected effect and a biomarker readout — and runs the checks that end drug programmes, cheapest first.
| Gate | Fails when |
|---|---|
| Prior failure | There is a recorded falsification of this exact claim. |
| Novelty | An existing publication or patent makes the same claim. |
| Tractability | There is no cavity, or its druggability score is below 0.4. |
| Chemical matter | There are too few measured compounds to fit a model, or the model is no better than predicting the mean. |
| Potency | A named compound is predicted weaker than 1 µM, or is outside the model's domain. |
| Dose | Free drug stays above the IC50 for under half the dosing interval. |
| Detectability | No feasible sample size exists, or assurance is below 50%. |
| Prior plausibility | The reference class succeeds under 5% of the time and nothing was said about why this proposal is different. |
| Surrogate validity | The biomarker readout has a trial-level R² below 0.7, or has never been validated. Runs only when a readout is named. |
Every gate reports the number that would have failed it. The result leads with what failed. A proposal survives only if nothing failed and at least four gates passed. If nothing failed but too few gates could run, it is reported as untested, not as a pass. Gates that did not run, or passed on an estimate, come back as experiments still outstanding.
03Proposing new ideas
Ask for a direction rather than an answer, and the agent can run a discovery process instead of a literature summary.
| Tool | What it does |
|---|---|
| discover.propose | Generates falsifiable hypotheses from evidence it retrieves during the run, screens each against prior art, and returns the survivors together with what was rejected as already known. |
| discover.campaign | The full loop: generate across six strategies, gate for novelty, review with four adversarial personas, rank by a pairwise tournament, evolve the survivors, and repeat. Each round ends with a critique of what the whole set assumes and misses. It is the most expensive tool available. |
| discover.novelty | Checks one claim against Europe PMC, ClinicalTrials.gov, the Negative Results Index and, for a molecule, exact-structure and nearest-neighbour search in PubChem. |
| kg.build, kg.link_predict | Builds a graph of a literature from resolved entities, and predicts missing links: pairs that share neighbours but never appear together. It needs a topic broad enough to span two literatures. |
Some details that keep the process honest:
- The six strategies are literature gaps, Swanson-style link prediction, failed-trial mining, mechanism transfer, constraint inversion and orthogonal transfer. Each is seeded from data retrieved in the run, not from the model's memory. A hypothesis without a falsification criterion is dropped.
- The novelty gate resolves doubt toward "already known". The number rejected is reported next to the survivors.
- The four reviewers — mechanistic, translational, skeptic and methodologist — score on axes that trade against each other.
- Every tournament comparison is judged in both orders. If the verdict flips, the comparison is discarded rather than counted.
- Campaigns read and write the falsification ledger, so an idea already vetoed is not offered back as new.
04Designing and repurposing compounds
drug.design proposes new compounds for a target. It pulls measured actives from ChEMBL, picks efficient seeds across distinct chemotypes, applies published medicinal-chemistry transformations two generations deep, filters for drug-likeness, structural alerts and synthesisability, checks every survivor against PubChem for novelty, and ranks on a transparent composite whose parts are all reported.
It does not predict binding affinity, and says so in its output. A candidate is a structure worth making and testing, justified by where it came from — not a prediction that it works.
drug.repurpose looks for approved drugs for a new indication by three independent routes: reversal of the disease's transcriptional signature (LINCS L1000), signature enrichment, and genes with human genetic evidence that already have approved drugs. Candidates found by more than one route are flagged.
Around these sit the chemistry tools — structure lookup, a medicinal-chemistry profile, alert screening, similarity as a novelty test, analogue design, and ADMET estimates from published models reported separately rather than as one score.
05Reading engine output
The engines run on one rule: the model proposes, the arithmetic decides. A model may suggest inputs — which trials exist, which negative controls to use — but no number a model states is used as a conclusion. Where inputs are estimates, the output says so.