Create active-learning campaign
Select a Discovery programme from the shared context bar. Campaign creation will be blocked when approved assay evidence is absent.
Governed learning · Beta
Learn from approved experiments, not ratings.
Plan assay-grounded acquisition campaigns with uncertainty, diversity, cost, synthesis feasibility, temporal holdouts, leakage controls, champion–challenger review, and explicit human deployment approval.
Mandatory controls
Approved assay data onlyDraft, unreviewed, or quality-flagged measurements cannot silently train a model.
Temporal and chemical holdoutsProspective performance must be measured on data unavailable during model fitting.
Leakage and duplicate controlsCompound, scaffold, target-family, pocket-similarity, and replicate leakage must be reviewed.
Champion–challenger comparisonA new model replaces the current model only after bounded, versioned comparison.
Human deployment approvalThe platform can propose a model update; it cannot autonomously promote it into a regulated or programme-critical context.
Assay Data Gateway
Use Discovery Programme OS to register protocol version, candidate batch, plate, well, endpoint, units, qualifier, replicates, controls, detection limits, curve-fit diagnostics, operator, experiment date, QC flags, reviewer, and approval status.
Data that cannot train directly
Star ratings, comments, manually typed “actual affinity,” unversioned spreadsheets, unreviewed curve fits, mixed units, unknown compound batches, and non-independent replicates are feedback signals—not training-grade assay evidence.
Acquisition policy library
Uncertainty sampling
Prioritize compounds where calibrated prediction uncertainty is highest.
Expected improvement
Balance predicted improvement with uncertainty against a defined objective.
Diversity aware
Avoid spending the entire batch on one narrow chemical neighbourhood.
Pareto acquisition
Select across potency, selectivity, ADME, safety, and synthesis trade-offs.
Cost sensitive
Discount candidates by assay, synthesis, material, and opportunity cost.
Synthesis aware
Prefer informative compounds that can realistically be made and purified.
Prediction Feedback Prototype
This legacy endpoint adjusts a heuristic using manually supplied feedback. It is retained for research comparison only and is not the governed active-learning system.
No result.
Prototype statistics
Not loaded.