Media Optimization

Reducing the cost of cell-culture media is one of the binding constraints on commercial-scale cellular agriculture. Media supplies the nutrients cells need and accounts for a large share of production cost, and designing it means navigating cost-versus-growth trade-offs across an enormous combinatorial space of components and concentrations. A recurring sub-goal is removing fetal bovine serum (FBS) — historically central but expensive, ethically fraught, and a source of variability and contamination risk (Xu et al., 2014, ref #21; Nikkhah et al., 2023, ref #1). The papers below show a clear technical progression in using AI to search this space efficiently, clustered by the method family each relies on.

Bayesian optimization of media formulations

Bayesian optimization (BO) — a Gaussian-process surrogate plus an acquisition function that proposes the most informative next experiments — is the most developed thread, largely from the Cosenza/Block/Baar program on muscle-cell media. Cosenza et al. (2022, ref #3) used multi-information-source BO, fusing cheap assays with high-fidelity cell counts, to optimize a 14-component medium for C2C12 cells; it reached a desirability of 0.94 in 81 experiments where a classical design-of-experiments approach reached only 0.40 in 132, delivering 181% more cell growth at just 1.6% higher cost in 38% fewer experiments, and flagged FBS as by far the most critical component. Cosenza et al. (2023, ref #2) extended this to a multi-objective, serum-free formulation, using a hypervolume-improvement acquisition to map the cost-versus-growth Pareto front and surfacing a medium with 23% more growth at 62.5% of control cost that held up over five passages. Narayanan et al. (2025, ref #58) generalized the machinery with a Gaussian-process kernel that handles categorical variables directly and transfer learning to add factors mid-campaign, raising human-PBMC viability to 75–80% (from ~60%) and lifting recombinant-protein productivity 2–3× in Komagataella phaffii, reporting roughly 3× — and up to 10–30× with categorical factors — fewer experiments than DOE. Yoshida et al. (2023, ref #15) combined the two ideas at the surrogate level: a deep neural network trained on an orthogonal-array screen of 31 components (reaching R² ≈ 0.99) was then searched by BO to maximize protein expression in E. coli, yielding media 1.4–1.65× better than the best training formulation across a search space of 3³¹ combinations. A related demonstration, Kanda et al. (2022, ref #16), pairs batch BO with an autonomous robot (the Maholo LabDroid) — though it optimizes a seven-parameter iPSC-to-retinal-pigment-epithelium differentiation protocol rather than a media formulation, improving the pigmentation outcome by 88% over the manual protocol, and is best read as the closed-loop “robot + BO” template that media campaigns can adopt.

Hybrid surrogate and evolutionary search

A second thread builds general-purpose search engines for expensive, high-dimensional, non-linear design problems. Cosenza & Block (2021, ref #17) is the methodological foundation: a hybrid that evolves a radial-basis-function surrogate while drawing query points from both a neural-network-guided genetic algorithm and a dynamic coordinate search, performing at least as well as its constituent algorithms on 19 of 20 synthetic high-dimensional test functions, including under simulated experimental noise. Cosenza, Block & Baar (2021, ref #211) applies essentially this designer to the cell-ag problem directly: a hybrid nonlinear design-of-experiments method (coordinate search plus a truncated genetic algorithm over a radial-basis-function surrogate) optimized a 30-component muscle-cell proliferation medium for cost-weighted growth, matching traditional-DOE quality in 70 experiments instead of 103 — while candidly showing that a single-passage 48-hour objective did not generalize to multi-passage growth, an early and honest statement of the validation problem that recurs across this area. Nikkhah et al. (2023, ref #1) applied a comparable composite — response-surface design, radial-basis-function neural networks, and the NSGA-II genetic algorithm — to reduced-serum media for zebrafish embryonic cells, distinctively folding environmental impact into the objective alongside cost and growth: its optimized formulations are projected to cut global-warming potential by 60–65% and cost by 20–24% while raising growth 13–51%, with PDGF, selenium, and serum identified as the most influential factors. Two further studies extend evolutionary media search to non-standard animal-cell systems: Munroe et al. (2019, ref #210) used a genetic algorithm to evolve the amino-acid composition of a basal medium for primary marine-sponge cells across four generations of ~30 formulations, roughly doubling metabolic activity over the basal control; and Tu et al. (2022, ref #212) coupled differential evolution with a Gaussian-process surrogate to design bull-sperm cryopreservation extenders, raising median post-thaw motility from ~53% to ~68%. Both cryopreservation and sponge culture are analogous rather than native to cultivated meat, but the surrogate-plus-evolutionary-optimizer template is exactly the one muscle-cell media campaigns use. Cosenza (2022, ref #18) is the UC Davis dissertation that consolidates this cultivated-meat media program into a single citable source.

Active learning with tree and ensemble models

A third thread, largely from the Ying lab, treats media design as iterative active learning: a model predicts promising formulations, the most informative are tested, and the data refine the next round. Hashizume et al. (2022, ref #24) introduced the approach with gradient-boosting decision trees on a 29-component medium for HeLa-S3 cells, producing reduced-FBS formulations and noting that an NAD(P)H-based assay can inflate apparent gains. Zhang et al. (2023, ref #23) adapted the same GBDT-plus-active-learning loop to selective media, tuning eleven components to favor one bacterial strain over another. Ozawa et al. (2025, ref #25) published the mature mammalian version, reformulating a 31-component EMEM for HeLa-S3 while cautioning about a growth-rate-versus-yield trade-off in early-readout modes and using transcriptomics to confirm the optimized media shift cells toward proliferation. Hashizume & Ying (2025, ref #169) is the most elaborate: a “biology-aware” platform combining an ensemble surrogate (gradient boosting, support-vector regression, k-nearest neighbors, and a neural network), genetic-algorithm selection, and error-aware preprocessing to reformulate a 57-component serum-free medium for CHO-K1, achieving roughly 60% higher cell concentration than a commercial serum-free medium and using SHAP to explain which component groups drove the gains. A deliberately one-shot counterpoint is Grzesik & Warth (2021, ref #213): rather than iterating, it runs a single definitive-screening design over twelve components, fits per-donor regression models (R² > 0.92), searches 10⁵ candidate formulations in silico, and uses k-means clustering to pick donor-robust media — optimizing primary human T-cell expansion in one experimental round, a useful template when donor-to-donor variability rather than search cost is the binding constraint.

Classical ML and explainable feature selection

A fourth thread uses classical ML to identify the few decisive components rather than to search exhaustively. Xu et al. (2014, ref #21) paired a Plackett-Burman screen with genetic-algorithm-tuned support-vector regression (reaching R² ≈ 0.98) and a mean-impact-value analysis to rank seven supplements for serum-free CHO-K1 media, with both methods agreeing on zinc sulfate, BSA, and ferric citrate. Gangwar et al. (2024, ref #170) built an explainable-AI pipeline — random-forest and SHAP feature attribution feeding a suite of fifteen regressors — to predict charge-variant critical quality attributes of an antibody from CHO media metal-ion content, identifying iron and zinc as the dominant levers. Both are CHO biopharma rather than cultivated meat, but the feature-attribution logic transfers directly to pinpointing rate-limiting media components.

Tools and data

Media-optimization tooling: the amii-cell-ag-tools suite (the Media Optimization & Cell Line Engineering section in Software.md) provides active-learning utilities aimed at cell-media problems; several papers above also ship their own code, linked from their entries in Papers.md.
Training data: serum-free-media and satellite-cell differentiation datasets for cultivated beef — the kind of data these methods train on — are catalogued in Datasets/Cow.md, and the CHO media-component data underlying ref #170 (and the CHO serum-free work in refs #21 and #169) is inventoried in Datasets/CHOReference.md.

Open challenges for cell-ag

The methods are converging, but the cell-ag-native evidence base is thin. The directly relevant mammalian work concentrates on a few muscle and stem-cell lines (C2C12, satellite cells, iPSCs); a large fraction of the most methodologically advanced studies are microbial or CHO-biopharma systems that are analogous rather than native to cultivated meat. Serum replacement is the throughline across refs #1, #2, #169, yet there is still no shared, standardized benchmark dataset or formulation that the field can optimize against and compare on. Recurring methodological cautions — assay artifacts that inflate apparent growth, single-passage results that under-predict long-term performance, and growth-rate-versus-yield trade-offs — mean validation discipline matters as much as the optimizer. The most distinctive opportunity is the multi-objective framing in refs #1 and #2: optimizing simultaneously for growth, cost, and environmental impact, on cultivated-meat-relevant cell types, against an open and reusable media-performance dataset.