(601b) Source Analysis of Process Variability in Multi-Step Bio-Process Manufacturing

Commercial bio-processes are comprised of multiple-step batch operations. However, its productivity is often affected by abnormal variations that lead to undesirable results. Though multivariate statistical methods have been in use in recent years to monitor and detect faults for bio-processes, it is still difficulty to identify the causes of multiple-step process variability.

In this paper, we analyze cell culture manufacturing variability in multiple-step processes using machine learning techniques, with the objective to identify the causes of high lactate production in recombinant cells. The proposed method is composed of three parts: 1) a unsupervised learning step that finds the most separable subspace for high / low lactate batches by principal component analysis (PCA); 2) a clustering step that classifies different causes in the space found in Step 1; and 3) a fault diagnosis step that applies robust linear discriminant analysis (LDA) to normal data and each class of fault data and characterizes faults based on robust LDA contribution analysis. The proposed method is demonstrated on a multi-stage batch process from Genentech, showing its ability to categorize and diagnose the sources of multi-step process viability.