Active learning in autonomous discovery — what to make next

Abstract

Active learning is the part of autonomous discovery that asks: given everything we know, what compound do we make next? We describe how the active-learner in Aether ^[1] picks the next batch: a cost-aware acquisition function divides expected information gain by expected experimental cost; batch acquisition imposes a diversity penalty across structural space; a separate critic agent audits the choices for signs of assay-exploitation; and the planner respects real cycle-time constraints — plate-reader queues, reagent lead times, chemistry-team bandwidth — that dominate in practice. We document the methodology, simulated-benchmark performance, the failure modes the design avoids, and a real nine-week customer campaign ^[4] that closed on a measured primary lead from a cold start.

1. The question and the stakes

Active learning is the part of autonomous discovery that asks: given everything we know, what compound do we make next? Get it right and a campaign closes in weeks; get it wrong and the agent will spend a quarter chasing structural patterns that the assay calls hits without being real hits. The stakes are real — discovery-stage biotechs run on cycles measured in months, and a wrong-headed acquisition strategy compounds across rounds.

We frame the problem as constrained Bayesian optimisation over a structured chemical space, with the design constraint that the bench operates on different timelines for different experiments. The model maintains a posterior over compound properties (potency, selectivity, ADMET liabilities ^[2]) and chooses the next batch to maximise expected information gain per unit of resource consumed — not per unit of compound.

2. Cost-aware acquisition

Standard expected-improvement acquisition maximises information gain per experiment. That is the wrong objective when experiments cost different amounts. A counter-screen costs twice as much as a primary plate; an in-vivo pharmacokinetics study costs orders of magnitude more. Our acquisition function divides expected information gain by expected cost, with the cost drawn from the runtime's planner ^[3] rather than from a hard-coded table.

Cost is itself a function of state. A reagent that is in the freezer right now is cheap; the same reagent ordered fresh is expensive and has a two-week lead time. An assay queue with three pending plates this week is more expensive (in scheduling terms) than the same queue with zero pending plates. The planner updates costs continuously; the acquisition function uses current costs.

On simulated benchmarks where ground-truth optimal acquisition is known, the cost-aware variant outperforms naive expected-improvement by roughly 2× in measured information gain per dollar across realistic campaign budgets. The improvement comes from picking cheaper experiments more aggressively early in the campaign — when uncertainty is high enough that any cheap measurement is valuable — and saving expensive experiments for the few compounds that matter most by mid-campaign.

3. Batch acquisition with diversity

Real campaigns pick batches, not single compounds. The bench runs N plates in parallel; the chemistry team synthesises M compounds at once; pure sequential one-compound-at-a-time acquisition is operationally impossible.

Standard batch acquisition picks the top-K compounds under the acquisition score. This is the wrong objective for an active-learner because the top-K under a noisy acquisition score are correlated — they share structural features, they probe the same region of chemical space, and they collectively contribute less information than their individual scores suggest. We use determinantal-point-process acquisition with an explicit diversity penalty across structural features (Morgan fingerprints at radius 2, plus a learned similarity kernel over the chemotype space).

Calibration of the diversity penalty matters. Too low and the batch collapses; too high and the batch ignores signal. We tuned the penalty on retrospective campaigns and re-tune it per programme during the first three rounds, when the posterior is still wide enough that the optimum is identifiable.

4. The exploit-the-assay failure mode

An active-learning agent with no diversity penalty will, given a noisy assay, eventually find a structural pattern that the assay calls a hit even though it isn't a real hit. The mechanism is straightforward: any biased measurement provides a signal the agent can learn; learning the bias is cheaper than discovering real biology; the agent will rationally pursue the cheaper path.

The consequence is the failure mode that ends bad campaigns. The agent enthusiastically synthesises variations on a pattern; the assay enthusiastically calls them hits; the chemistry team enthusiastically advances the series; the in-vivo confirmation experiment kills them all, two months and several million dollars later.

Diversity acquisition reduces this risk but does not eliminate it. We add a second mechanism: the critic agent.

5. The critic agent

A separate critic agent audits the active-learner's choices for signs of pattern-exploitation before the synthesis order commits. The critic is a different model than the active-learner — different training distribution, different architecture, different prior — and it looks specifically for the failure-mode patterns we have catalogued from historical campaigns.

Concrete failure-mode signatures the critic flags.

• Batches converging on a narrow structural neighbourhood without orthogonal-assay confirmation.
• Apparent hits clustered around a particular sub-structure that is known to interfere with the assay readout (e.g., compounds that auto-fluoresce in a fluorescence-based assay).
• Counter-screen evidence inconsistent with primary-assay evidence — typically a sign that the primary signal is artefactual.
• Posterior uncertainty shrinking faster than experimental evidence supports — a sign that the model is over-confident.

The critic flags suspicious series for orthogonal confirmation before the agent commits the synthesis order. Operators can override the flag; overrides are recorded. In our experience, the critic catches roughly one campaign in five before the operator would have caught it on their own; in the remaining four campaigns the critic flag confirms what the operator already suspected, which is still useful as a check.

6. Cycle time, not just uncertainty

In a real campaign the binding constraint is rarely model uncertainty. It is plate-reader queue depth, reagent lead time, or the chemistry team's bandwidth. The runtime plans around these — preferring compounds the assay can read tomorrow over compounds with marginally higher expected-information but a two-week reagent lead time ^[3].

Operationally, this means the acquisition function does not see chemistry as an unconstrained discrete optimisation over compound space. It sees chemistry as a constrained scheduling problem. The planner maintains a current view of bench state (queues, reagents, instrument availability) and chemistry-team state (synthesis-team load, lead times); the acquisition function picks compounds that fit the current state, not compounds that maximise unconstrained acquisition.

The trade-off is real — sometimes the highest-acquisition compound is one that takes two weeks to source, and waiting may be the right call. We expose the trade-off in the operator console: each batch is presented with an estimated cycle-time and a counterfactual of the highest-acquisition batch that would have been picked under unconstrained acquisition. Operators choose; the agent does not override the operator.

7. Campaign result — fibrosis programme

A discovery-stage customer ran a fibrosis programme on the platform from cold start. They had a validated target and an in-house chemistry team but no prior compound series. Twelve compounds went to bench; three measured as primary hits, four as secondary leads. The active-learner made twelve choices in roughly thirteen rounds; the platform handled scheduling, chain-of-custody, and eBR capture ^[4].

From campaign start to first measured primary lead was nine weeks. This is meaningfully shorter than the team's previous campaigns on the same target class — typically 5–7 months from start to first measured hit. The shortening is dominated by the cost-aware acquisition and the cycle-time-aware planner, not by raw model accuracy.

The campaign also surfaced one critic flag that the operator initially overrode and subsequently confirmed: a fluorescence-quenching artefact in the primary assay had been calling a non-binder a hit. The flag was raised after round seven; the operator overrode on operational pressure; round nine's counter-screen results vindicated the flag. The series was dropped; no downstream waste resulted.

8. Limitations

Three meaningful limitations. First, cost-aware acquisition depends on accurate cost estimates from the planner; in cold-start situations where cost calibration is poor, the acquisition reverts to a near-uniform cost weighting that performs no better than naive expected-improvement. Second, the critic agent is itself a model and can miss novel failure modes that are not in its training distribution. Third, the methodology is tuned for small-molecule discovery; biologics campaigns have different cost structures and we are extending the framework to handle them separately ^[2].

9. Reproducibility

The simulated benchmark cases used to characterise cost-aware acquisition are open under the Research tier. Customer-campaign details are anonymised in this paper; full traces are available under NDA to academic groups studying active-learning methodology. The acquisition function and critic agent are documented in the platform's reference ^[3].

10. Conclusion

Active learning works in autonomous discovery when the acquisition function respects cost, batch diversity is enforced, the critic checks for assay-exploitation, and the planner respects real cycle-time constraints. Without any one of these, the failure modes dominate. With all of them, a discovery-stage campaign can close on a measured primary lead in nine weeks from cold start. The methodology is one of the load-bearing pieces of the autonomous-lab story; it is not the only one ^{[3, 4]}.