Data-driven Linear Predictor based on Maximum Likelihood Nonnegative Matrix Decomposition for Batch Cultures of Hybridoma Cells

Abstract

This paper presents an original design of low-rank linear predictors of nonlinear process state variables based on nonnegative matrix decomposition (NMD). Therefore, this predictor is data-driven and does not require an accurate model description of the process. In addition, measurement errors are considered, conferring maximum likelihood (ML) properties to the estimator and resulting in a maximum likelihood nonnegative matrix decomposition (MLNMD) formulation. The latter is validated in simulation with a model developed by the authors, describing monoclonal antibody (MAb) production from sequential batch hybridoma cell cultures that are further validated with real-life experimental data. To this end, two available experimental data sets are used for direct and cross-validation, highlighting the good predictive properties of the method.