(53e) Biochemical Kinetics-Informed Ann Modeling for Cyanobacterial Biological Processes

C-phycocyanin (C-PC), an algal bio-pigment, is in high demand in the food and pharmaceutical industries due to its anti-inflammatory, neuroprotective, and antioxidant properties. Currently, C-PC is primarily obtained through extraction and purification from cyanobacterial cultures. However, the limited biomass concentration and phycocyanin productivity in photobioreactors (PBRs) have hindered the commercialization of C-PC production. To enhance C-PC yield, fed-batch optimization techniques are employed to control cyanobacterial growth. Typically, these optimization studies rely on first-principles models that accurately predict the behavior of the biosystem. Yet, due to the intricate metabolic mechanisms in microorganisms, creating precise first-principles models based on metabolic kinetics can be exceptionally challenging. To address this challenge, artificial neural networks (ANNs) have recently been applied to simulate and control dynamic bioprocesses. ANNs offer an advantage by treating the bioprocess as a "black box," constructing models solely from data. However, this approach has its strengths and limitations. While ANNs can establish models for biological processes, their accuracy depends on the range of available modeling data. Since biochemical kinetics information is not incorporated into ANN modeling, the relationships among variables in ANN-based models may not always be accurate. There has been limited emphasis on improving interpretability and generalization in previous studies.

In this study, we propose an ANN modeling approach for cyanobacterial biological processes that considers biochemical kinetics. To infuse the neural network with biological process information, we convert biochemical kinetics into constraints and embed them into the loss function. By minimizing this new loss function, we optimize the neural network parameters, allowing the model to converge towards the true biological process state during training. We validate our proposed method using experimental data from existing literature.