Phase-Incremental Decision Trees for Multi-Phase Feature Selection and Interaction in Biologics Manufacturing

Abstract

Data from cell culture processes contain myriad parameters arriving sequentially in phases which may hold vital information for optimizing process runs and ameliorating manufacturing yield. This study analyzed temporal process data from 249 cell culture production batches of an active pharmaceutical ingredient at Roche's Location A manufacturing facility. The titer manufactured is utilized for Roche's Product X, a prescription drug that can treat adults with cancer. We aim to optimize the upstream production phase titer in Chinese hamster ovary cell manufacturing by identifying the most influential features. A phase-incremental (PI) decision tree method is proposed for feature selection and interaction exploration, being model and loss function agnostic while promoting early feature importance for prediction and process control. In this case study, the method is applied to Ensemble of Gradient Boosting Machines, using adjusted R-squared as the penalized loss function. The result leads to better process understanding and enables earlier control in the manufacturing.