Biokinetic Priors Boost Data-Scarce Bioprocess Modeling.

Deep learning has transformed drug discovery, but its application in biomanufacturing is hampered by severe data scarcity. Bioreactor experiments are costly, time-consuming, and rarely publicly shared, leaving researchers with limited datasets. However, the bioprocess domain is rich in established biokinetic ordinary differential equation (ODE) models that describe microbial growth. This paper explores how to effectively integrate this valuable prior knowledge into neural networks. The study systematically compares two methods: a data-level prior, which involves pre-training a generic decoder on simulated ODE curves, and an architecture-level prior, where the ODE is directly embedded within the decoder's structure. Across 11 datasets and 7 microbial species, both approaches consistently outperform neural networks trained without such prior knowledge. A key finding is their substitutability: a generic decoder pre-trained on simulations performs comparably to a fully bio-structured decoder trained on real data. This suggests that simulation pre-training offers a simple, data-efficient strategy for applying deep learning effectively in data-scarce bioprocess environments.