Extensive evaluation of machine learning models and data preprocessings for Raman modeling in bioprocessing

Abstract

Raman spectroscopy is a very promising tool for monitoring key analytes in mammalian cell culture fermentations in real time. However, major challenges are associated with this promising technology in aqueous bioprocessing matrixes such as a strong background fluorescence, which is typically addressed by computational preprocessing of the raw Raman spectra. In this study, we present an extensive combinatorial assessment of various machine learning algorithms with numerous popular preprocessing methods on the performance and robustness of Raman-based real-time predictions of key analytes. We show that preprocessing methods have a large influence on algorithmic performance. Furthermore, there is a large variance across the various combinations of preprocessing steps and tested machine learning regression algorithms. We demonstrate that neural networks and random forest regression show very good performance and robustness across different, bioprocess relevant analytes. They significantly outperform partial least squares regression, the most widely used regression algorithm in the field. Overall, this extensive study provides a sound basis for building robust, next-generation models for monitoring analytes in real time based on Raman spectroscopy.