Active Learning

This page describes the Active Learning row of the Papers.md matrix: iterative design of experiments in which the model selects which experiments to run next, closing a predict, experiment, retrain loop around a surrogate. The row’s authoritative scope is its Taxonomy.md definition; this page synthesizes what currently sits in it.

Scope boundary

The taxonomy places this row and Bayesian Optimization in a deliberate containment relation: Bayesian optimization is a specific acquisition-driven instance, and this row is the broader iterative-selection framing. In practice the distinction that matters when reading a paper is the surrogate. A Gaussian process with an explicit acquisition function is the BO row; a gradient-boosting or ensemble surrogate whose top-ranked predictions are simply sent to the bench is this one. A paper that does both, such as #58, sits in both.

What the row is really about is experiment budget. Every reference below exists because the search space is astronomically larger than what a lab can test, and the paper’s contribution is a defensible way of choosing the next plate. That makes it the most directly transferable row on the method axis for cultivated-meat media work, and the one where the wet-lab detail in the methods sections is worth as much as the algorithm.

Media Optimization

#24 (Hashizume et al. 2022, bioRxiv, Ying lab): the pilot study for this thread, fine-tuning 29 components of a HeLa-S3 medium with gradient-boosting decision trees. Its most useful contribution is a warning rather than a result. Optimization raised cellular NAD(P)H abundance without raising cell concentration, because the conventional assay used to score cultures measures the former as a proxy for the latter. Any media campaign scored on a metabolic-activity readout can optimize the readout instead of the cells. The paper also details experimental-error control that recurs across the whole thread: dozens of identical cell stocks frozen at once, stock solutions used exactly once, media mixed immediately before use. Code at hashizume711/medium_optimization.
#23 (Zhang et al. 2023, Applied Microbiology, Ying lab): applies the same GBDT-plus-active-learning loop to a different objective, selective growth, tuning eleven MRS components to favour Lactobacillus plantarum over E. coli or the reverse. Growth is scored on two parameters read off the curves, exponential rate and maximal yield, from 1,660 growth curves at 30-minute intervals over 48 hours. Both parameters were maximally differentiated between the two strains, and analysis of the accumulated data identified which components are decision-making for specificity. Relevant to cell-ag wherever a co-culture or a contamination-resistant formulation is the goal rather than raw growth.
#25 (Ozawa et al. 2025, Biochemical Engineering Journal, Ying lab): the mature mammalian version, reformulating a 31-component EMEM for HeLa-S3, with cell concentration measured on a particle size analyzer rather than a metabolic proxy, closing the loop on #24’s caveat. It compares readout modes explicitly, using 96-hour yields in one and 168-hour in the other, and evaluates on both raw and EMEM-normalized fold-change targets, which is what surfaces the growth-rate versus final-yield tradeoff. RNA sequencing was then used to confirm that the optimized media shift cells toward proliferation, an independent check that the optimizer found biology rather than an artefact. Code at yuki020527/medium_optimization.
#169 (Hashizume & Ying 2025, New Biotechnology): the most elaborate of the thread, a “biology-aware” platform combining error-aware preprocessing, an ensemble surrogate and genetic-algorithm selection, applied to a 57-component serum-free medium for CHO-K1 across 364 experimentally tested media. Reported roughly 60% higher cell concentration than commercial serum-free alternatives, and the improvement was specific to CHO-K1 rather than general, which the authors present as evidence of targeting rather than of a universally better medium. It also measures ammonia and lactate in the supernatant to check for metabolic stress, so the optimized medium is assessed on more than the objective it was optimized for. Also in Ensemble Learning and Genetic Algorithms.
#250 (Hashizume et al. 2026, Journal of Bioscience and Bioengineering, Ying lab): sequential active learning for monoclonal-antibody production in CHO cells, adjusting 44 components of a serum-free medium against IgG titer. The methodological point is about what the loop is allowed to learn from: biological insights that emerged during experimentation, osmolality control and amino-acid composition, were not in the initial variable set and were progressively folded into the optimization. That is a different discipline from a fixed design space, and the paper argues it is what makes the approach practical under limited experimental resources.
#58 (Narayanan et al. 2025, Nature Communications, MIT Love lab): a Bayesian-optimization framework applied to two problems, cytokine supplementation to maintain viability and subpopulation distribution in human PBMCs, and recombinant-protein production in Komagataella phaffii. Reported 3 to 30 times fewer experiments than standard design of experiments would need, with transfer learning used to add design factors mid-campaign rather than restarting. Described more fully in Bayesian Optimization; it is here because the iterative exploration-exploitation loop is the contribution. Code at NHarini-1995/CellCultureBayesianOptimization; data on figshare.

The Ying-lab sequence (#24, #23, #25, #169, #250) is unusual in the matrix for being a single group iterating on one method across five papers, and reading it in order is the fastest way to see which parts of an active-learning media campaign are the algorithm and which are the assay discipline. The answer, on their evidence, is mostly the latter.

Cellular Engineering

#63 METIS (Pandi et al. 2022, Nature Communications, Erb lab): a general active-learning workflow, and the reference implementation of the loop’s core algorithm. An ensemble of 20 XGBoost models with hyperparameters found by randomized search is trained on all prior rounds’ data, 100,000 candidate combinations are scored, and the top candidates are ranked by an upper-confidence-bound score that trades expected yield against ensemble disagreement. Before committing to that design the authors ran a simulation study, training a “gold regressor” on 1,017 real data points and using it to stand in for the wet lab so that MLP, DNN, linear and XGBoost surrogates and different batch sizes could be compared without spending experiments. Applied to cell-free transcription-translation, genetic circuits, and a 27-variable synthetic CO₂-fixation cycle, where it explored a 10²⁵ space in 1,000 experiments to produce the most efficient CETCH cascade reported at the time. It also quantifies factor importance, surfacing bottlenecks rather than only optima, and runs in Colab so no installation is required. Also in this row under AI Tooling.
#97 PerTurboAgent (Hao et al. 2025, bioRxiv, Regev lab): active learning where the surrogate is an LLM agent. The task is choosing which genes to perturb in the next round of a sequential Perturb-seq campaign, formalized precisely: a phenotype is defined by a set of associated descriptive genes, hit genes are those whose perturbation shifts that set beyond a z-score threshold, and the agent must find them over N rounds selecting m genes each time, with no gene selectable twice. Two design choices distinguish it from prior work. The agent sees expression profiles of both control and perturbed cells and analyses the raw data itself rather than receiving summaries; and it carries an action memory, storing each action-result pair within a round so a multi-step plan can adapt mid-round. A phenotype score quantifies how strongly each selected perturbation moved the target. Also in Domain-Specific Biomedical Agents.

AI Tooling / Methodology

#63 METIS (Pandi et al. 2022): also placed here because the workflow is general-purpose rather than tied to one biological system. See the description under Cellular Engineering.
#182 The Robot Scientist (King et al. 2004, Nature): the earliest reference in the row and still the sharpest framing of why active learning matters. Its experiment-selection strategy was competitive with human performance while cutting cost 3-fold against cheapest-first and 100-fold against random selection, over a metabolite set whose per-experiment costs spanned three orders of magnitude. The paper is also careful about what “performance” means here, using average performance across competing hypotheses precisely because it rewards a learner that discriminates between them rather than one that is confidently right about a single one. Also in Robot Scientists & Lab Automation.

Adjacent methods

Bayesian Optimization: the acquisition-driven special case, and the other half of most media-optimization campaigns.
Ensemble Learning: the gradient-boosting and ensemble surrogates most of these loops are built on.
Genetic Algorithms: the selection layer in #169, and an alternative search strategy over the same design spaces.
Robot Scientists & Lab Automation: what the loop becomes when the experiment is executed without a human.
Domain-Specific Biomedical Agents: where an LLM replaces the trained acquisition function.