Bayesian Optimization

This page describes the Bayesian Optimization row of the Papers.md matrix: sequential, surrogate-model-based optimization in which a probabilistic model plus an acquisition function chooses the next experiment to run. The row’s authoritative scope is its Taxonomy.md definition; this page synthesizes what currently sits in it.

Scope boundary

Two exclusions do most of the work. Bayesian optimization is not Bayesian inference: posterior or flux estimation that quantifies uncertainty without using it to select the next experiment is not this row. And the acquisition function is what separates it from Active Learning, which the taxonomy frames as the broader iterative-selection family containing this one as its acquisition-driven instance. In practice a paper here has a probabilistic surrogate (usually a Gaussian process) and an explicit rule for turning predicted mean and variance into a decision about what to run next.

This is the workhorse row for cellular agriculture, and the only one where the field’s own literature leads rather than borrows. Most of its references sit in Media Optimization, several were written specifically about cultivated meat, and the recurring finding across them is not that a particular acquisition function wins but that the objective function is the hard part.

Media Optimization

The Cosenza / Block / Baar programme

#211 (Cosenza, Block & Baar 2021, Biotechnology Journal): the starting point, and the most useful negative result in the row. A hybrid nonlinear design of experiments, coordinate search for local refinement plus a truncated genetic algorithm over a radial-basis-function surrogate, reduced the cost of a 30-component C2C12 proliferation medium in 70 experiments against a traditional DOE’s 103. But because the objective was a 48-hour AlamarBlue reading weighted by medium cost, the result held only at a single passage and did not generalize to growth over multiple passages. The paper says so plainly and draws the lesson: objective functions have to align with process goals. Everything downstream in this thread is a response to that. Also in Genetic Algorithms.
#3 (Cosenza et al. 2022, Biotechnology and Bioengineering): the fix, and the row’s methodological centrepiece. Rapid growth assays are convenient but inaccurate; robust cell counts are accurate but so slow they limit how many experiments you can run. Multi-information-source BO models both, fusing a cheap low-fidelity source (2,000 cells in a 96-well plate, AlamarBlue) with an expensive high-fidelity one (60,000 cells in a six-well plate, trypan blue counting) into an uncertainty-weighted desirability function, so the algorithm decides which assay to spend on as well as which medium. Over 14 components it produced media with 181% more cells than a commercial variant at similar cost, in 38% fewer experiments than an efficient DOE, and the optimum generalized to four passages, which is precisely what #211 could not do. Code at ZacharyCosenza/GradStuff_Cosenza.
#2 (Cosenza et al. 2023, Engineering in Life Sciences): extends to two objectives at once, minimizing cost while maximizing growth, using a hypervolume-improvement acquisition to trace the Pareto front rather than collapsing the tradeoff into one number. The wet-lab preparation deserves attention on its own: the C2C12 line was adapted to serum-free conditions gradually over passages into E8 medium, then Matrigel was needed once E8 exceeded 90%, then dilution cloning at one cell per well isolated a subclone that survived without Matrigel at all. The design space groups the 30-plus basal components by function (amino acids, vitamins, salts, trace metals, DNA precursors, fatty acids) while varying suspected drivers individually. Result: several media with more than 100% more growth than control, and one with 23% more growth at 62.5% of control cost, sustained well past the study period.
#18 (Cosenza 2022, UC Davis dissertation): the connecting document, which states the arc explicitly: long-term cell growth is not fully correlated with the short-term chemical assays cell culture normally uses, and the multi-information-source Bayesian model was built to correlate the two so long-term experiments could be reserved for the most promising regions while cheap short-term ones explore freely. It also documents the earlier RBF-plus-truncated-GA machinery in detail, including the convergence criterion (normalized Euclidean distance between successive best queries below 0.2) and k-means clustering of candidates into the final batch. Also in Deep Learning.

Beyond muscle cells

#58 (Narayanan et al. 2025, Nature Communications, MIT Love lab): generalizes the machinery in two directions that matter for real campaigns. A Gaussian-process kernel handles categorical variables directly, so choice-of-basal-medium sits in the same optimization as concentrations; and transfer learning lets new design factors be added mid-campaign without discarding prior data. Applied to cytokine supplementation for human PBMC viability and subpopulation distribution, and to three recombinant proteins in K. phaffii, reporting 3 to 30 times fewer experiments than standard DOE. The experimental setup uses C.BIRD microplate bioreactors with 6 to 8 biological replicates per condition. Also in Active Learning. Code at NHarini-1995/CellCultureBayesianOptimization.
#249 (Ndahiro et al. 2025, iScience): adds a physical constraint the other papers leave implicit. Media formulations proposed by an optimizer can be chemically infeasible, so this method folds thermodynamics-based solubility constraints into the BO so every suggestion can actually be mixed, and supports parallel experiments to match the throughput of automated bioreactors. Validated in silico and in an AMBR automated cell-culture system on CHO-K1 VRC01 cells, with a discipline worth copying: cells were passaged in the new formulation for at least three passages before the measured run, so reported titers could not come from a transient stress response to the media change. Improved titers against classical DOE.
#15 (Yoshida et al. 2023, Journal of Bioscience and Bioengineering): a deep neural network as the surrogate rather than a Gaussian process, optimizing 31 components for GFP expression in E. coli, trained on 81 media designed by an orthogonal-array scheme at three levels each. Its honest reporting is the value: the first round of BO-designed media showed large errors between predicted and measured expression, so the network was retrained on the validation data and re-predicted, at which point predictions fit and the best medium reached 1.4-fold the best initial composition. A worked example of the first iteration of a surrogate loop being wrong and the loop being what fixes it. Also in Deep Learning.
#212 (Tu et al. 2022, Scientific Reports): blank-slate optimization of bull semen cryopreservation extender media, comparing artificial neural networks against Gaussian process regression over eleven components, with GPR coupled to differential evolution. Median post-thaw motility rose from 52.6 ± 6.9% to 68.3 ± 6.0% between generations 7 and 17, across ejaculates from 68 Holstein bulls with quality thresholds applied before inclusion. Relevant to cell-ag beyond dairy genetics because cryopreservation of a working cell bank is an unavoidable step in any cultivated-meat process, and the paper offers itself as a template for other cell types. Also in Deep Learning.
#251 (Emerson et al. 2026, arXiv): multi-objective BO for cryoprotectant cocktails, where the tradeoff is intrinsic: a formulation must be concentrated enough to suppress ice formation and dilute enough not to kill the cells. Seven component CPAs (glycerol, DMSO, ethylene glycol, 1,2- and 1,3-propanediol, 3-methyl-1,2-propanediol, urea) over a high-throughput T24 assay with a reported Z-factor of 0.67, starting from 535 measured data points against a full combinatorial space of 48,198 candidates. Selection maximizes expected Pareto improvement under uncertainty. Reported hypervolume gains of 9.5% over a naive strategy and 4.5% over a strong baseline, and in synthetic studies it matched the prior state of the art using 30% of the evaluations, which the authors translate to roughly ten weeks of experimental time.
#16 (Kanda et al. 2022, eLife, RIKEN): batch Bayesian optimization driving a LabDroid robot through 143 culture conditions in 111 days from a space of 200 million combinations, ending 88% better than the pre-optimized protocol on pigmentation score. Described more fully in Robot Scientists & Lab Automation; it is the only reference in this row where the loop runs without a human executing it.

Cellular Engineering

#16 (Kanda et al. 2022, eLife): the same iPSC-RPE differentiation campaign, placed here as well because what is being optimized is a differentiation protocol rather than a growth medium. See above.

Bioprocess & Scale-Up

#59 (Antonakoudis & Richelle 2026, npj Systems Biology and Applications): reduces a genome-scale metabolic model to something compact enough for dynamic bioprocess simulation and digital twins, with Bayesian flux estimation propagating uncertainty from noisy, sparse exo-metabolomics into the reduction itself. The argument for why that matters is precise: existing metabolomics-guided reductions treat estimated uptake and secretion rates as exact within a sampling interval, which causes both over-pruning (discarding pathways that should stay feasible) and under-pruning (retaining reactions that should be inactive). Applied to time-course data from 12 fed-batch CHO cultures, it produced one reduced model feasible across all conditions that reproduced observed extracellular fluxes and preserved broad metabolic functionality from extracellular data alone.
#256 (Choi et al. 2025, Biotechnology and Bioengineering): BO over CFD simulations rather than over experiments, optimizing fermentation reactor design and operating conditions across five variables with a Gaussian-process surrogate using an RBF kernel and hyperparameters fitted by marginal log likelihood. Each objective evaluation is a CFD run, which is the reason BO is worth using: the evaluations are expensive whether or not they involve cells. It couples fermentation kinetics to hydrodynamics and reports that wide baffles, a narrow impeller gap, slow gas flow and high agitation together improve gas distribution and reduce stagnant zones. Directly applicable to scaling a cultivated-meat bioreactor, where concentration gradients are exactly the problem that appears on scale-up.

Adjacent methods

Active Learning: the containing family, and where the ensemble-surrogate media campaigns sit.
Genetic Algorithms: the search layer in the earlier Cosenza work and in several hybrid designers.
Deep Learning: where a neural surrogate rather than a Gaussian process is used.
Ensemble Learning: the tree-ensemble surrogate alternative.
Robot Scientists & Lab Automation: the same loop with the human removed.