Imagine a discovery team with ten thousand virtual molecules and budget to synthesize twenty. The conventional answer is simple: rank the molecules by predicted potency and make the top twenty.
That is often the wrong strategy.
The top twenty may be nearly identical analogues occupying the same narrow region of chemical space. Their potency predictions may be confidently wrong because the model is extrapolating beyond its training data. They may all fail for the same unmodeled reason -- poor permeability, reactive metabolism, insolubility, or synthetic instability. Even if several are active, the team may learn almost nothing about how to design the next series.
Early discovery is not merely an optimization problem. It is a learning problem under severe experimental constraints.
Potency is an outcome, not a strategy
Potency matters, but optimizing it too early can create a local optimum: a chemically fragile series with beautiful biochemical activity and no credible path to a medicine. The best next experiment may be a modestly predicted compound that tests whether a new scaffold works, whether a polar substitution improves exposure, or whether an apparent structure-activity relationship is real.
The difference is between exploitation and exploration. Exploitation chooses what currently looks best. Exploration chooses what could change the model of the problem. A good campaign needs both.
Uncertainty is not a defect -- it is a resource
Most prediction dashboards emphasize the mean: predicted IC50, clearance, solubility, or probability of toxicity. Active learning cares equally about uncertainty. Two compounds with the same predicted potency can have very different experimental value:
- One sits in a well-studied chemical series where the model is confident.
- The other sits near an unexplored boundary where a single measurement could confirm a new opportunity or eliminate an entire region.
The second compound may be more valuable even if its expected potency is lower. Its experiment has a higher chance of changing what the team believes.
This is the core of Bayesian optimization: combine the expected outcome with uncertainty through an acquisition function. Different acquisition strategies express different risk preferences -- improve the current best, sample where uncertainty is highest, or maximize expected information gain.
The best experiment is not always the one most likely to succeed. It is the one most likely to improve the next decision.
Drug discovery is multi-objective by definition
A molecule is rarely selected because it wins one metric. It must balance potency, selectivity, solubility, permeability, metabolic stability, safety margin, synthetic accessibility, solid-state behavior, and exposure at the intended tissue. These objectives frequently conflict.
For that reason, a serious optimization system should not hide trade-offs inside one arbitrary weighted score. It should expose a Pareto frontier: candidates for which improving one objective would require sacrificing another. The team can then see distinct strategic options -- the potent but metabolically soft compound, the less potent but highly soluble compound, the novel scaffold with wider uncertainty, and the synthetically simple backup.
Weights can still be used, but they should reflect the program’s current stage and target product profile. The importance of oral exposure, brain penetration, duration, or selectivity can change dramatically across indications and modalities.
The next experiment may be an assay, not a molecule
Active learning is often described as choosing the next compound. In practice, the larger decision is a pair: which candidate should be tested in which assay?
A cheap, noisy assay may be appropriate for broad exploration. An expensive mechanistic assay may be justified when it can resolve a critical contradiction. A cassette PK study may reduce more uncertainty than another round of biochemical potency. A solid-state screen may prevent an apparently excellent molecule from failing months later because the selected form cannot be manufactured reproducibly.
The expected value of an experiment therefore depends on:
- The uncertainty it can reduce
- The importance of that uncertainty to the next decision
- The probability the experiment is technically informative
- Cost, time, material consumption, and ethical burden
- Whether the result discriminates between competing hypotheses
Closed-loop discovery is a discipline, not a robot
The popular image of closed-loop discovery is a robotic laboratory designing, synthesizing, and testing molecules autonomously. Automation can shorten the cycle, but the deeper innovation is the loop itself:
- Represent the current evidence and uncertainty.
- Generate candidate experiments.
- Estimate their expected decision value.
- Run a deliberately small batch.
- Update the model with successes, failures, and technical caveats.
- Repeat until the program reaches a decision threshold.
A human-run laboratory can follow this discipline. A fully automated laboratory can violate it by optimizing the wrong objective at enormous speed. The quality of the loop matters more than the number of robots.
Measure learning velocity, not screening volume
Discovery organizations often celebrate compounds synthesized, assays completed, or virtual molecules screened. Those are throughput measures. They do not reveal whether the program is converging.
More meaningful metrics include:
- Reduction in uncertainty per experiment
- Time required to reject a weak hypothesis
- Number of design cycles to reach a target profile
- Information gained per unit of cost or biological material
- How often recommended experiments change the program decision
The long-term advantage of AI in discovery may not be that it predicts the winning molecule in one shot. It may be that it helps teams learn the shape of a difficult problem with far fewer physical experiments -- and knows when the remaining uncertainty is small enough to act.
