What does active learning do in drug discovery?

Active learning uses the current model, uncertainty, and candidate pool to propose informative compounds for the next measurement cycle. The proposals still require scientific and operational review.

Does active learning replace medicinal chemistry selection?

No. It supplies a documented selection signal that should be balanced with chemistry, assay, safety, novelty, cost, and program strategy.

Design-make-test-analyze

Keep every learning cycle connected to the evidence available at the time.

Q: Why preserve assay and model versions?

A prediction and selection decision can only be interpreted correctly when the training data, assay definition, model version, and candidate pool at that time are known.

Link candidate design, synthesis or procurement, assay results, model updates, uncertainty, and next-compound selection so teams can understand not only what was chosen, but why.

DMTAActive learningAssay feedbackUncertainty samplingModel updatesDecision lineage

Project memory across cycles

Turn each assay result into usable, reviewable learning.

MolexIO connects molecular identity, assay definition, measured result, model version, candidate pool, and selection rationale. The platform distinguishes a model recommendation from the final scientific decision and preserves compounds that were not selected, preventing retrospective certainty.

Design context

Retain the parent, transformation, hypothesis, target, and objective behind each proposed molecule.

Make status

Track procurement or synthesis context and distinguish planned, available, attempted, and completed work.

Assay-ready data

Attach measured values to explicit assay definitions, units, qualifiers, dates, and molecular identities.

Model lineage

Preserve the training snapshot, features, validation, applicability, and version that produced each prediction.

Selection strategies

Balance predicted performance, uncertainty, diversity, coverage, and constraints in a documented candidate proposal.

Cycle comparison

Review what changed between rounds in chemistry, measurements, model behavior, uncertainty, and program priorities.

Measurements stay measuredAssay results are never relabeled as model outputs or vice versa.

Non-selections remain visibleThe candidate pool and decision context prevent survivor-only interpretation.

Human decision retainedActive learning proposes; authorized project teams decide.

A closed learning loop

Advance the program and improve the next question.

01
Design
Generate or nominate candidates against explicit hypotheses, objectives, and project constraints.
02
Make
Evaluate route or supply evidence and record the actual molecules prepared or acquired.
03
Test
Capture measured outcomes with assay identity, units, qualifiers, controls, and provenance.
04
Analyze and select
Update evidence and models, inspect uncertainty, and document the next candidate set.

Scientific boundary

Active learning optimizes a selection rule, not the whole discovery program.

Recommendations reflect the available training data, candidate pool, uncertainty method, objective, and constraints. Biased assays, narrow chemical space, data leakage, inconsistent endpoints, or a poorly chosen objective can produce misleading priorities. Expert review and prospective measurement remain central to every cycle.

Questions before use

DMTA and active learning FAQ

What does active learning do?

It proposes informative next measurements using the current model, uncertainty, and candidate pool.

Does it replace medicinal chemistry selection?

No. Chemistry, assay, safety, novelty, cost, and program strategy still require human judgment.

Why preserve assay and model versions?

They are necessary to reconstruct what evidence existed when a prediction and selection were made.

Make each DMTA decision explainable to the next cycle.

Connect the molecule, measurement, model, uncertainty, and selection rationale in one project record.

Request a tailored demo →