Augmentation Models of Stem Cell Culture Data for the Application of Machine Learning

Bioprocesses for stem cell-based therapeutics are resource- and time-intensive, hindering the generation of sufficient data for machine learning-informed development. We describe a framework combining data augmentation, multivariate regression, and feature importance analysis to investigate the relationship between metabolites and critical quality attributes (CQAs) of cultured stem cells. A first-principles model (FP), a hybrid model using neural ordinary differential equations (ODEs) with stoichiometric constraints (HD), and a purely statistical neural ODE model (NODEAM) were considered. Probing these models through data augmentation, we generated synthetic data and amplified the biological information encoded in each modality, directly linking architectural choices to their ability to capture relevant dynamics. For all models, the validation error remained relatively constant, and the convergence of the feature importance tensor followed a power law with the number of augmented runs. Temporal feature importance and functional data analysis of variance revealed key time windows during which glucose and lactate showed strong correlation with CQAs. The HD model provided the best fidelity and accuracy, underscoring the value of combining mechanistic and statistical modeling for improved interpretability at lower complexity. Overall, this framework can yield insights into the physiology of cultivated cells and can be adapted for various culture-based biomanufacturing systems.