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

Area of Science:

• Biopharmaceutical Manufacturing
• Process Analytical Technology (PAT)
• Data Science in Biotechnology

Background:

• Cell culture processes generate complex temporal data crucial for optimizing manufacturing yield.
• Understanding these parameters is vital for producing active pharmaceutical ingredients like Roche's Product X, a cancer treatment.
• Current methods may not fully leverage sequential data for early process control.

Purpose of the Study:

• To optimize the upstream production phase titer in Chinese hamster ovary (CHO) cell manufacturing.
• To identify the most influential features within temporal process data for enhanced prediction and control.
• To develop a novel method for feature selection and interaction exploration in bioprocesses.

Main Methods:

• Analysis of temporal process data from 249 cell culture production batches.
• Proposal and application of a phase-incremental (PI) decision tree method for feature selection.
• Utilizing Ensemble of Gradient Boosting Machines with adjusted R-squared as the penalized loss function.

Main Results:

• The phase-incremental decision tree method effectively identified influential features for titer prediction.
• The approach demonstrated model and loss function agnosticism, promoting early feature importance.
• Application to Gradient Boosting Machines led to improved process understanding and earlier control capabilities.

Conclusions:

• The proposed phase-incremental decision tree method enhances feature selection and interaction exploration in cell culture processes.
• This data-driven approach facilitates earlier process control, leading to optimized manufacturing yield.
• The findings contribute to better process understanding and control in biopharmaceutical production for cancer therapies.