Advanced adaptive modelling approaches in the evolution of vector/cell manufacturing processes

Abstract

The field of cell gene therapy has seen significant progress in recent years. The last decade has seen the licensing of the first Cell Gene Therapy (CGT) treatments in Europe and clinical trials have demonstrated safety and efficacy in the treatment of numerous severe inherited diseases of the blood, immune and nervous systems. Specifically, autologous viral vector-based CGT treatments have been the most successful to date. However, the manufacturing processes for these CGT treatments are at an early stage of development, and high levels of complexity, process variability and a lack of advanced process and product understanding in vector/cell manufacturing are hindering the development of new processes and treatments. Here, Multivariate Data Analysis (MVDA) and Machine Learning (ML) techniques, which have not yet been widely exploited for the development of CGT processes, were leveraged to address some of the main hurdles in the development and optimisation of CGT processes. Principal component analysis (PCA) was primarily used for feature extraction to understand the main correlations and sources of variability within the process data, and to evaluate the similarities and differences between batches. Additionally, a sparse PCA algorithm was developed to ease the interpretation of the principal components with a large number of variables present in the dataset. Predictive modelling techniques were utilized to model the relationships between process variables and critical quality attributes (CQAs) of the viral vector and cell drug products. The infectious titres of lentiviral vector (LV) products from both adherent cell cultures and suspension cell cultures were modelled and predicted successfully and critical process variables were identified with statistically significant correlations to this CQA. In cell drug product manufacturing, the LV copy number in the patient’s transduced cells was also modelled and process parameters in LV manufacturing and cell drug product manufacturing were linked to this CQA. Overall, the modelling process recovered valuable information from historical process data from the early stages of process development. This data frequently remains unexploited, due to its commonly truncated and unstructured nature; however, this work showed that MVDA/ML techniques can yield beneficial insights despite less than ideal data structure and features.