A surrogate model iteratively picks the most informative candidates for synthesis and assay, then retrains on the results — the closed DMTA loop that replaces brute-force HTS.
Modern make-on-demand and virtual chemical libraries (Enamine REAL, ZINC20, GDB-17) contain anywhere from hundreds of millions to billions of synthesizable compounds. Experimentally testing even a small fraction is physically and financially impossible — so drug discovery starts with almost no ground truth and must decide, intelligently, what to measure first.
The count of drug-like small molecules obeying Lipinski-style rules has been estimated at 10⁶⁰ or more — vastly larger than the number of atoms in the solar system. Even restricting to make-on-demand catalogs like Enamine REAL (currently >36 billion enumerable structures) or curated virtual libraries like ZINC20 (~1.4 billion purchasable compounds), no organization can synthesize and test more than a vanishingly small slice.
Each compound is typically represented not by its raw structure but by a learned or hand-crafted embedding — Morgan/ECFP fingerprints, graph neural network latents, or 3D pharmacophore features — that places chemically similar molecules near each other in a continuous "chemical space." The scatter plot used throughout this simulation is exactly this: an abstract 2D projection (think UMAP or t-SNE) of that high-dimensional space, where proximity implies structural and often functional similarity.
A single Enamine REAL Space search now spans over 36 billion synthesizable compounds — roughly five times the world's population — yet a typical primary screen tests only a few hundred thousand of them.
Traditional High-Throughput Screening (HTS) tests every compound in a fixed library against a target, at a cost of roughly $0.05–$5 per data point depending on assay complexity. Screening a modest 500,000-compound deck therefore costs anywhere from $25,000 to several million dollars — and only covers an infinitesimal corner of the accessible chemical universe.
Brute force also wastes resources: most tested compounds are inactive, and a fixed library encodes no ability to explore beyond what was pre-plated. If the true hit lies in a region the library never sampled, no amount of additional screening within that library will find it.
Active learning does not require a large starting dataset. A seed set of 50–500 molecules — sometimes a legacy screening deck, sometimes a diversity-selected random sample — is enough to bootstrap a first-pass surrogate model. What matters is that the seed set spans the chemical space broadly enough that the model can estimate, even crudely, "how much do I not know" everywhere else.
In the visualization, the initial handful of colored points scattered among thousands of gray, unlabeled candidates represents exactly this seed set: sparse, imperfect, but sufficient to start the iterative Design-Make-Test-Analyze (DMTA) cycle that pharma teams run in every modern discovery program.
The surrogate model is the engine of active learning: a statistical model — Gaussian Process, Random Forest, or deep ensemble — trained on the (small) labeled set that predicts both an activity value and a calibrated uncertainty for every unlabeled candidate. Uncertainty, not just prediction, is the key output: it tells the loop where it does not yet know enough.
A Gaussian Process (GP) defines a distribution over functions and, conditioned on observed data, yields a posterior mean (the prediction) and posterior variance (the uncertainty) at every point in input space. Using a Tanimoto or RBF kernel over molecular fingerprints, the GP naturally produces low uncertainty near molecules structurally similar to ones already tested, and high uncertainty in unexplored regions — precisely the "fog" visualized in this simulation, thick far from labeled points and thin nearby.
GPs scale poorly with dataset size (roughly O(n³) for exact inference), so as the labeled set grows past a few thousand compounds, sparse GP approximations or alternative surrogates take over.
A deep ensemble of 5–20 independently trained neural networks, disagreeing more on unfamiliar molecules than familiar ones, is now the dominant uncertainty-estimation approach at AI-native drug discovery companies — it scales to millions of compounds where exact Gaussian Processes cannot.
Random Forests estimate uncertainty from the variance across individual decision trees' predictions — a cheap, robust, and surprisingly effective proxy for epistemic uncertainty on tabular molecular descriptors. Deep ensembles extend the same idea to neural networks: train several models with different random initializations and/or bootstrapped data, and use the spread of their predictions as the uncertainty signal.
Both approaches sidestep the cubic scaling of exact GPs, making them the practical default once a program's labeled set grows into the thousands — which happens quickly once several rounds of the active learning loop have run.
A surrogate model that is highly accurate on average but badly miscalibrated — confidently wrong in specific regions — will misdirect the acquisition function and waste experimental budget. Calibration is typically checked by comparing predicted variance against observed residual variance on held-out folds; well-calibrated models keep this error under roughly 10%.
In practice, teams re-calibrate or ensemble multiple surrogate architectures specifically to guard against the failure mode where the model is certain about a region it has never actually seen — the single most damaging error an active learning loop can make.
Prediction and uncertainty alone do not choose what to test next — an acquisition function combines them into a single score per candidate, and the top-ranked batch becomes next round's experiments. The choice of acquisition function encodes an explicit exploration-exploitation tradeoff: explore to reduce uncertainty broadly, or exploit to refine around molecules that already look promising.
Pure exploitation — always picking the molecule with the best predicted activity — risks getting stuck exploiting a local optimum the model overestimates, having never verified it near the boundary. Pure exploration — always picking the most uncertain molecule — wastes budget mapping regions of chemical space that are structurally interesting but biologically irrelevant.
Every serious acquisition strategy is a negotiation between the two. Bayesian optimization theory formalizes this negotiation: the acquisition function is itself an optimization objective, computed cheaply over the surrogate model, whose argmax tells you the single most valuable next experiment (or top-k for batch selection).
Real DMTA cycles rarely test one molecule at a time — synthesis and assay throughput make batches of 10 to 200 compounds far more efficient. Naively selecting the top-k by acquisition score risks redundancy: the highest-scoring candidates often cluster in the same region of chemical space. Batch-aware acquisition (e.g. determinantal point processes, kriging believer, or simple diversity-penalized re-ranking) explicitly spreads the batch across distinct clusters so each experiment contributes independent information.
In this simulation, adjusting the batch size slider changes how many points are pulled into each selection round, and the acquisition strategy slider changes whether that batch clusters near the current best guess (greedy) or spreads across high-uncertainty regions (uncertainty-only).
Published retrospective analyses of active learning applied to DMTA cycles report reaching equivalent potency to brute-force HTS using 10 to 100 times fewer synthesized and tested compounds — turning multi-year, multi-million-dollar campaigns into weeks-scale, six-figure ones.
AI-native drug discovery companies — Recursion, Insitro, Exscientia (now Recursion after their 2024 merger), and others — have built entire discovery platforms around this loop: a phenotypic or biochemical surrogate model proposes the next experimental batch, robotic synthesis and automated assay pipelines execute it within days, and the resulting data retrains the model before the next round begins. Exscientia reported reaching clinical candidates with roughly 70% fewer synthesized compounds than typical medicinal chemistry campaigns; Recursion's phenomics platform runs millions of automated cellular assays specifically to keep the labeled set growing fast enough to feed continuous retraining.
| Product | Indication | Trial Design | Key Result |
|---|---|---|---|
| Pure Exploitation (Greedy) | Highest predicted activity | argmax(predicted); ignores uncertainty entirely | Fastest convergence once the model is already accurate; ideal for late-stage lead optimization |
| Pure Exploration (Uncertainty Sampling) | Highest model uncertainty | argmax(uncertainty); ignores predicted value entirely | Maximizes information gain; best in early rounds when data is sparse |
| Upper Confidence Bound (UCB) | predicted + κ·uncertainty | Linear combination with tunable exploration weight κ | Simple, tunable dial between explore and exploit; strong empirical baseline |
| Expected Improvement (EI) | Expected gain over current best | Integral over the predictive distribution above the incumbent best value | Naturally decays exploration as the optimum is approached |
| Thompson Sampling | Random draw from posterior | Sample a plausible model from the posterior; pick its argmax | Balances explore/exploit automatically; parallelizes cleanly for batch selection |
The acquisition function's output is only a hypothesis until it meets the wet lab. Selected candidates are synthesized (or sourced from make-on-demand catalogs), assayed for the property of interest, and the resulting measurements become the new ground-truth labels that the next model iteration will train on — the "Make" and "Test" steps of the Design-Make-Test-Analyze cycle.
Turning an acquisition function's output into real data requires an actual synthesis route. Many active learning campaigns now restrict candidate selection to make-on-demand chemical space (like Enamine REAL) specifically so that every top-ranked molecule is guaranteed synthesizable within days via validated reaction protocols, rather than requiring bespoke medicinal chemistry. Where custom synthesis is unavoidable, retrosynthesis-aware filters down-weight candidates with no feasible route before they ever reach the acquisition ranking.
Once synthesized, compounds enter the same assay pipeline used to build the original seed set — biochemical binding, cellular potency, ADMET, or phenotypic readouts — so that new labels are directly comparable to old ones.
Modern discovery platforms pair active learning software with robotic liquid handling, automated flow chemistry, and high-throughput assay readers so that a full DMTA round — from acquisition function output to new labeled data back in the training set — completes in as little as two to six weeks, compared to many months for manually planned campaigns.
Even with automation, synthesis is not perfect: roughly 60–90% of proposed compounds are successfully made on the first attempt, and the loop must gracefully handle failures by either retrying with an alternate route or simply treating the compound as missing data for that round.
Automated DMTA platforms integrating robotic synthesis with active learning have compressed lead optimization cycle times from the traditional 6–8 weeks per round down to under 2 weeks in some reported industrial deployments — nearly a 4x acceleration in iteration speed.
At $500–$5,000 per synthesized-and-tested compound (custom synthesis plus assay), every experimental decision has real financial weight. This is precisely why the acquisition function's job — choosing the batch that yields the most model-improving information per dollar — is the economic core of the entire active learning strategy, not just a modeling nicety. A poorly chosen batch wastes tens of thousands of dollars confirming what the model already predicted with confidence.
New labels are folded back into the training set, the surrogate model retrains, and the cycle repeats: uncertainty contracts, predicted-activity peaks sharpen, and the batch of molecules selected each round drifts from broad exploration toward tight refinement around the discovered optimum. After enough rounds, a potent hit emerges having synthesized and tested a tiny fraction of what brute-force screening would have required.
Each retraining round adds informative labels precisely where the model was least certain or most likely to be wrong, so the surrogate's error contracts fastest in exactly the regions active learning chose to sample. Over successive rounds, the acquisition function naturally shifts its own behavior: early rounds propose broadly scattered, high-uncertainty batches; later rounds propose tightly clustered, high-predicted-value batches near the emerging optimum — because the uncertainty fog itself has retreated everywhere except right at the frontier of the best-known region.
This self-correcting property is what separates active learning from a fixed random or diversity-based screening plan: the sampling strategy adapts to what has already been learned, round over round.
Retrospective benchmarking studies that replay historical HTS campaigns through simulated active learning loops consistently find that 10 to 100 times fewer compounds need to be tested to reach a given potency threshold compared to exhaustively screening the full library. The magnitude of the saving depends on how peaked the true activity landscape is: a landscape with one sharp, isolated optimum (much like the single dominant peak used in this simulation) benefits the most, since brute-force screening wastes the overwhelming majority of its budget on the flat, inactive majority of chemical space.
Exscientia's DSP-1181, developed with an AI-driven active learning design loop, reached a clinical candidate after synthesizing roughly 350 compounds — compared to the 2,500+ compounds typically required by traditional medicinal chemistry programs, a reduction of about 85%.
What began as a research technique is now standard infrastructure at AI-native drug discovery companies. Recursion's phenomics platform, Insitro's disease-model-driven screening, and Exscientia's (now part of Recursion) design engine all run some variant of this same predict-select-test-retrain loop, differing mainly in what surrogate model architecture and what experimental readout they use. As assay automation and generative chemistry mature further, the loop is expected to close faster and require progressively less human intervention between rounds — pushing discovery timelines from years toward months for well-automated targets.