Explainable Deep Learning Model for Predicting Cell Viability and Antibody Productivity in Culture Processes

This paper proposes a machine learning framework for predicting key in antibody production in cell culture processes, focusing on antibody product concentration and cell viability. Three types of input data-process variables, output variables and their integration-were applied to Random Forest and Deep Neural Network models. Experimental results demonstrated that integrated datasets improved antibody product concentration prediction accuracy, while output variables were more effective for cell viability prediction. Shapley additive explanations analysis was employed to interpret variable contributions, revealing distinct drivers for each quality attribute. For antibody product concentration, effect of metabolic indicators such as glutamate and ammonia were dominant, while viability was mainly influenced by culture conditions and glutamine levels. These findings highlight that input composition strategies should be tailored to the target attribute and that combining prediction accuracy with interpretability provides actionable insights for process optimization. The proposed approach contributes to smart biomanufacturing and supports digital twin-based control strategies.