Active Learning for Scientists: What the Literature Actually Says

An NMR run can cost an order of magnitude more and take 5 to 10 times longer than a plate assay, so wet-lab query costs differ materially. Of roughly 60 active learning papers we reviewed while building Discovery Loop's acquisition layer, 47 treat every query as equally costly. Their main difference is the information returned. Neither premise fits a wet lab, and the literature has only partly adapted.

This is not a rejection of active learning. The field is useful and expanding. But deploying it in wet-lab research means separating transferable results from methods tied to simpler assumptions. That separation is wider than most product descriptions suggest. We reached these conclusions by studying the methods and testing several with real experimental data.

What We Read and Why

We examined three groups of papers: foundational machine learning methods such as pool-based sampling, query-by-committee, and expected error reduction; Bayesian optimization for chemistry and materials; and a smaller group on experimental design for biological assays and drug discovery. Bayesian optimization was the closest fit to our use case. The general ML literature was the least directly useful.

This was not an exhaustive academic survey. We were not mapping the field for publication. We wanted an operational answer: which methods can support an experiment loop with expensive, noisy queries, batch-arriving labels, and a partly structured search space? Some regions cannot be visited because of synthesis constraints or instrument limits.

The Assumption Gap

Most active learning papers rely on three assumptions that fail in wet labs.

First is cheap querying. In the usual pool-based setup, a learner sends selected unlabeled points to an oracle, often a human annotator or fast function, and receives an answer quickly. In a wet lab, the oracle is a physical experiment that may run for 3 hours or 3 days, fail midway, and consume meaningful funds. Costs vary by query as a matter of course, while most methods make them uniform.

Second is low-noise labeling. Many methods presume the oracle is correct, or introduces only small additive noise. Biological experiments often produce much noisier measurements. An assay coefficient of variation of 20 to 30% is common, and some protocols require replicates before the result is interpretable. In our domains, one experiment often cannot provide a dependable label for a condition.

Third is online selection. Conventional active learning chooses one query, receives its label, updates the model, then chooses again. Wet labs usually operate in batches. You cannot queue 8 HPLC injections one at a time while waiting for each result. The instrument runs a sequence, and the next sequence must be prepared before those results arrive. Batch active learning addresses this, but the literature contains fewer batch methods than single-query methods.

What Does Not Transfer Well

Query-by-committee maintains several models and selects points on which they disagree most. It performs well in simulation, but struggles with small initial datasets and substantial experimental noise. Early discovery loops often begin with fewer than 20 prior measurements. An ensemble trained on 20 noisy observations may show little meaningful diversity. Agreement then reflects shared blind spots, not a well-mapped search space.

Greedy expected error reduction is also a poor fit for batches. Choosing the point that most lowers expected model error is practical for one query, but batch construction requires recomputing that quantity after every choice. Large candidate pools can make this computation the bottleneck. Iterative batch selection is also vulnerable to ordering artifacts when many candidate conditions are nearly duplicates.

Pure exploitation, selecting whatever the current model ranks highest, rapidly settles on local optima and says little about the remaining space. This deserves emphasis because naive implementations often default to it. Scientific response surfaces may contain multiple optima, while model uncertainty is high at the start of the loop, making exploitation alone a poor discovery strategy.

What Does Transfer

With suitable configuration, Bayesian optimization using Gaussian process surrogates carries over well to wet-lab work. Gaussian processes are better suited to small datasets than data-hungry deep neural networks and offer calibrated uncertainty estimates for acquisition functions. The critical choice is a prior kernel that captures domain expectations about response-surface smoothness and length scale. A poor kernel will not necessarily break the loop, but it can produce overconfident extrapolation in sparse regions.

Expected improvement, or EI, and upper confidence bound, or UCB, both transfer, but their tradeoffs differ. EI is conservative, focusing budget near the best known region and exploring when the expected return justifies it. UCB explores more because its uncertainty bonus can be adjusted with a hyperparameter. When the optimum may lie far from current conditions, our benchmarks favor UCB with a moderate exploration bonus over EI.

Noisy expected improvement explicitly includes measurement noise in its calculation and is underused despite being highly useful for noisy experiments. Standard EI treats the observed best result as a close proxy for the true best. Noisy EI instead considers the chance that the observed value is above or below the underlying value. In practice, it is less eager to exploit apparent optima inflated by noise, reducing convergence on a false local maximum.

The Acquisition Function Problem in Practice

Selecting an acquisition function in theory is simpler than selecting one for a particular experimental system. The decision depends on experiment cost relative to the total budget, measurement noise, prior structural knowledge of the response surface, and batch requirements. Relevant structure can include smoothness, known inactive regions, and symmetry. Batch requirements determine how many experiments the lab can run in parallel each cycle.

We chose a configurable acquisition layer instead of one permanent function. Its default is UCB with a lab-controlled exploration parameter, initialized to a moderate value that has worked across the domains we tested. For projects with fewer than 30 experiments total, we suggest more exploration to avoid committing early to a poor region. For larger budgets, the standard setting is more suitable. With noisy assays, we apply a noise-discount factor that lowers the effective improvement assigned to results with high variance coefficients.

There is no universal setting, and we do not claim our defaults fit every lab. They are practical starting points based on published work and our testing. Labs with strong prior knowledge about response-surface geometry should tune them accordingly.

How We Adapted These Ideas for the Loop

Our largest change to standard methods addresses batch selection. Instead of slow greedy sequencing across a large candidate set, we combine UCB acquisition scores with a variant of a determinantal point process, or DPP. The DPP penalizes candidates too similar to those already selected, encouraging diversity within the batch. The loop therefore avoids queuing five near-identical experiments when a broader batch could reduce uncertainty across more of the space.

We also introduced explicit cost weighting. Every candidate carries an estimated cost based on instrument time and reagent use. We divide its acquisition value by that cost weight. When expected information gain is comparable, this lowers the priority of expensive experiments, applying a basic cost-sensitive search principle that receives limited attention in standard active learning.

The third change covers missing results. Experiments fail, instruments malfunction, and samples become contaminated. Standard active learning expects every query to return a label. In automated wet-lab work, 5 to 15% of experiments may yield no data or invalid data. The model must update from successful members of a partial batch without interpreting instrument malfunction or sample loss as evidence about the chemical or parameter space.

The review showed that active learning in wet labs is harder than the standard formulation suggests. That is not a reason to discard it. We need to account for experimental failure modes rather than accept benchmark assumptions unchanged. Our loop excludes failed runs from the scientific update instead of treating them as observations.