4.3 Experimental design and surrogate models for growth media
A major problem in cultivated meat is that industrially relevant cells
are grown with media with upwards of 30 – 50 different metabolites
[13]. Additionally, cellular responses to metabolites not only
depend on reactor conditions like temperature, pH and mixing, but on
metabolite-metabolite interactions, causing response nonlinearity.
Lacking verified system-level and mathematical approaches to
optimization, tools for experimental methods that can be used in the lab
quickly are highly valuable. A traditional way that modeling has
addressed the issue of high dimensionality (many inputs) is through
combining statistical design of experiments (DOE) with linear or
polynomial regression in a family of methods called response surface
methods (RSM) [14]. However, both DOE and sequential RSM tend to
become infeasible at input dimensions > 5 [15] and
linear and low-order polynomial models often are unable to capture
nonlinear effects of responses in design problems.
Nonlinearity has been overcome in the past by using heuristic optimizers
such as genetic algorithms [16,17] (dim ~ 10).
Higher dimensional problems in media optimization have been solved using
neural networks [18] (dim ~ 20) and are especially
useful in their ability to capture nonlinear and high dimensional
responses. Combining many of these separate ideas is the field of
meta-modeling. Usually (i) a DOE or other space-filling design of
experiments is conducted then (ii) a polynomial, neural network, kriging
model [19], SVR and/or regression/classification method [20]
maps input-output responses and (iii) a nonlinear solver, trust-region
algorithm, or search rule chooses a new set of experiments in sequence
based on an objective function. Finally, information theoretic criteria
for function approximation has been applied to media design [18] and
process optimization [21,22]. As the purpose of these methods is to
leverage modeling techniques to discover better media designs while
minimizing cost/experimentation, model agreement is important so
researchers should consider liberal use of cross-validation,
hyperparameter optimization, and model ensemble methods.