Genetic Algorithms

This page describes the Genetic Algorithms row of the Papers.md matrix: population-based evolutionary search, covering genetic algorithms, evolutionary strategies, NSGA, and differential evolution. The row’s authoritative scope is its Taxonomy.md definition; this page synthesizes what currently sits in it.

Scope boundary

The row is defined by the search mechanism, a population of candidate solutions improved by selection, crossover and mutation. That makes it unusually promiscuous: a genetic algorithm is nearly always bolted onto something else, and almost every reference here also sits in the row for its surrogate model. The GA is the optimizer; the neural network, support-vector machine, radial-basis-function model or biophysical simulator is what it optimizes against.

The mechanism alone does not earn the row. Taxonomy.md sets the test as where the search’s output appears: the row is earned when what the search returned is part of what the paper reports, and not when that output only configures the machinery behind the reported result, which covers a search over hyperparameters, over model architecture, and over which technique to apply. Architecture search is the case that reads both ways, since optimising an architecture also produces a fitted model, and Taxonomy.md writes it out as a third shape rather than leaving it to judgment: the row is earned only where the method producing that model is the paper’s stated contribution and its results characterise the model the search returned. The question it gives you is what the paper would still be reporting if the search had been replaced by a default configuration. Both kinds are real optimisation and both are often described at length, so the length of the description settles nothing.

The distinction from Bayesian Optimization is what happens to uncertainty. A GA searches a surrogate’s predicted surface and needs many cheap evaluations, so it fits when the surrogate is fast and the design space is large and combinatorial. Bayesian optimization models its own uncertainty and uses it to decide, so it fits when every evaluation is a wet-lab experiment. Papers that need both, notably the Cosenza work, hybridize them.

Media Optimization

#17 (Cosenza & Block 2021, Engineering Optimization): the methodological foundation for the cell-ag media thread, and deliberately not a biology paper. It builds a hybrid framework in which a cubic radial-basis-function surrogate with a linear tail stores prior knowledge, a truncated genetic algorithm searches it globally, and a dynamic coordinate search refines locally. The argument for the GA half is stated plainly: its randomness lets it explore global and local regions at once, which matters when the surrogate is untrustworthy because the space is sparsely sampled. Evaluated on 20 high-dimensional synthetic test functions chosen to mimic real physical and biological processes in dimensionality and nonlinearity, where the hybrid performed at least as well as either constituent on 19 of them, including under simulated experimental noise.
#211 (Cosenza, Block & Baar 2021, Biotechnology Journal): that framework applied to a 30-component C2C12 muscle-cell proliferation medium, reaching traditional-DOE quality in 70 experiments against 103, and the most useful negative result in the row. 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. The follow-up work that responds to it sits in Bayesian Optimization.
#18 (Cosenza 2022, UC Davis dissertation): the connecting document for this thread, and the fullest account of the search layer itself. It documents the radial-basis-function surrogate driven by a truncated genetic algorithm in detail, including the convergence criterion (normalized Euclidean distance between successive best queries below 0.2) and the k-means clustering of candidates into the final batch. Described in Bayesian Optimization, where its multi-information-source half is set out.
#212 (Tu et al. 2022, Scientific Reports): blank-slate optimization of bull-semen cryopreservation extender media over eleven components, where the search layer is differential evolution coupled to a Gaussian-process regressor and compared against artificial neural networks. 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 cryopreserving 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.
#1 (Nikkhah et al. 2023, Science of The Total Environment): optimizes a reduced-serum medium for a zebrafish ZEM2S cell line against three objectives at once, and is the only reference in the row where one of them is environmental. Response-surface methodology designs the experiments over seven variables (IGF, FGF, TGF, PDGF, selenium, ascorbic acid, serum); three separate radial-basis-function neural networks then predict growth rate, cost and global warming potential; and a multi-objective genetic algorithm searches those three surrogates jointly, with the resulting formulations validated experimentally. Folding GWP into the objective is a direct response to culture media being identified as a dominant contributor to the footprint and cost of cultivated meat.
#169 (Hashizume & Ying 2025, New Biotechnology): uses genetic-algorithm selection as the search layer over an ensemble surrogate in a 57-component serum-free CHO-K1 campaign, described in Active Learning. Also in Ensemble Learning.
#210 (Munroe et al. 2019, In Vitro Cellular & Developmental Biology - Animal): a GA run in vitro, with four generations executed as wet-lab experiments rather than against a surrogate, optimizing the amino-acid composition of a basal medium for cells of the marine sponge Dysidea etheria. It reports a twofold increase in metabolic activity over the basal control. Two things make it worth reading in a cell-ag context. There are no sponge cell lines and no defined sponge-specific medium, so this is media design with essentially no prior, which is the position a team faces on a novel species. And running the algorithm’s generations as physical experiments, rather than searching a model, is what most of the rest of this row avoids because it is slow, so the paper is a useful reference point for how few generations that actually takes.
#252 (Pi et al. 2020, Journal of Biomechanical Engineering): a control-parameter study rather than an application, asking how population size, mutation factor, crossover rate and algorithm variant (random, best, local-to-best, and local-to-best with self-adaptive modification) affect differential evolution when optimizing DMSO-free cryoprotectants. The training model is post-thaw recovery of Jurkat cells across glycerol, creatine and sucrose concentrations, a four-dimensional surface with multiple saddle nodes, chosen precisely because it is hard. Self-adaptive modification performed best and was least sensitive to mutation-factor and crossover choices. Useful to anyone about to run a GA on a cryopreservation or media problem and wondering which knobs matter.

Bioprocess & Scale-Up

#30 (Peng et al. 2013, Bioresource Technology): the earliest reference in the row and still a clean template. A “5-10-1” artificial neural network maps five online-recorded fermentation parameters, and crucially fermentation time as an explicit input, to bacteriocin antibacterial activity in a 5-L stirred tank, reaching R² above 0.95. Because time is an input rather than an index, a genetic algorithm can then optimize the parameters stage by stage and emit a whole control trajectory rather than a single setpoint, improving production by 26%. The reactor is treated explicitly as a black box, ignoring the physiology, which is the modelling assumption the whole ANN-plus-GA pattern rests on. Also in Deep Learning.
#31 (Zhang et al. 2020, Bioresource Technology): microbial lipid production from cellulosic ethanol wastewater by Rhodotorula glutinis, where the GA optimizes over a support-vector-machine model rather than a neural one. The comparison is the point: on a small experimental dataset the SVM fitted and generalized better than a back-propagation ANN, and the GA then found conditions reaching 11.87 g/L biomass and 2.18 g/L lipid. Initial COD and glucose content dominated. Also in SVM.
#209 (Takahashi et al. 2016, Process Biochemistry): inverts the usual arrangement, using the GA to optimize the neural network’s own parameters rather than the process. Each individual in the population is a set of neural computing parameters, evolved by selection, crossover and mutation, for a network describing rabies-virus production in BHK-21 cells over temperature, multiplicity of infection, and infection and harvest times. The interesting design question it answers is how few experimental runs a uniform design can get away with: architectures were simulated against experimental fractions of 18, 24, 30, 36 and 42 runs, and the GA-optimized network calibrated well (R above 0.85) across all of them. This is the row’s architecture-search case, and it earns the row on the third shape rather than despite the exclusion: the combined ANN-GA method is what the paper sets out to develop, and its results characterise the network the search returned.
#258 (Brinc & Belič 2019, Journal of Biotechnology): a GA optimizing 14 process parameters at once for mammalian fed-batch culture in a robotic 15-mL ambr micro-bioreactor system, across CHO-K1-derived and CHO-DG44 lines expressing a monoclonal antibody and an Fc-fusion protein. The parameter set is exactly what a process-development team faces: eight medium components, four feed components, plus pH and dissolved oxygen. The paper’s argument against classical design of experiments is dimensionality, and its experimental discipline is worth copying, with two replicate control runs in every experiment to measure the uncontrolled variability (medium lot-to-lot and similar) that would otherwise be read as signal.

Scaffolding

#34 (Andrews et al. 2025, arXiv): closes the loop opened by its sibling #35, which trained a fast surrogate of a biophysical tissue-growth simulation. Here an evolutionary strategy searches mould tethering designs for high cellular alignment and uniform density, initialized with 1,000 random candidates, each encoded, rendered to a two-channel image, and scored by the RAPTOR surrogate rather than by a full simulation. Parents are chosen by tournament selection, then combined by crossover and perturbed by mutation. That surrogate is what makes the search affordable at all: evaluating each candidate by direct simulation would be prohibitive. The paper names cultivated meat first among its applications, alongside pharmaceutical assays and regenerative medicine.

Sensory Prediction

#342 (Androutsos et al. 2024, npj Science of Food): a multi-objective machine-learning method predicting sweet, bitter, umami and other from molecular structure, assembled from the VirtuousSweetBitter dataset, the UMP442 umami set and ChemTastesDB, with all SMILES checked against PubChem and duplicates and malformed entries removed. The curation decisions are stated rather than hidden: multitaste compounds were dropped for lack of numbers, and sour and salty were not treated as classes at all because only 38 and 12 compounds were available, which is a useful marker of where the field’s data actually runs out. Also in Ensemble Learning.
#269 iUmami-SCM (Charoenkwan et al. 2020, Journal of Chemical Information and Modeling): predicts umami peptides from sequence using a scoring card method, where the model is a table of propensity scores over amino acids and dipeptides. That is what puts it in this row rather than making the search an implementation detail: the evolutionary search produces the scores themselves, so its output is the paper’s result. The fitness function linearly combines AUC with the Pearson correlation between the initial and optimised score sets, and the loop is fully specified: 40 sets of scores initialised to include the starting table, a 20-point crossover between the best set and the others, real-valued mutation at Pm = 0.01 sparing the best set, and 20 generations as the stop condition. A useful contrast with the refs this row excludes, where an evolutionary search only tunes a model someone else’s method defines. Also in Linear & Regularized Models.

Adjacent methods

Bayesian Optimization: the uncertainty-aware alternative when every evaluation is a physical experiment, and the other half of the hybrid designers here.
Active Learning: where the GA appears as the selection layer over an ensemble surrogate.
Deep Learning and SVM: the surrogate models most of these GAs search against.
GAN / VAE: the surrogate behind the scaffolding search in #34, trained in #35.
Reinforcement Learning: the other route to a bioreactor control trajectory, learned as a policy rather than evolved.