Development of machine learning methods for predicting cell viability and mortality in bioreactors.

Cécile SHAYKHIAN

Thierry BASTOGNE

Pr. Bensoussan Danièle (d.bensoussan@chru-nancy.fr) , Pr Olmos Eric (eric.olmos@univ-lorraine.fr)

The PhD subject will focus on the development of methods and tools for identifying dynamic models associated with two output variables: cell quantity and cell mortality. It is structured around two complementary research axes, both utilizing the same datasets. The two application axes will rely on the same hierarchical model, consisting of (1) an external first-principles biological model, and (2) an internal learning model based on Bayesian neural networks. This research is part of the Bioproduction France 2030 initiative, aiming to develop a new measurement technology, as well as a proposed CNRS MITI Avancées Thérapeutiques project (currently under submission) that seeks funding for complementary measurement technologies to enhance the datasets used for training our learning models. The PhD work will be specifically applied to the production of CD34+ eMDSC cells, contributing to the development of a new cell therapy. Axis 1: Intensified Experimental Designs. The optimization of produced cell density is currently performed using statistical design of experiments (DoE) methods. These methods aim to select a minimal yet efficient number of trials to identify algebraic models that can predict optimal operating conditions. A typical design of experiments includes 10 to 20 different trials. In the case of a bioreactor, a single culture session can last between 2 and 4 weeks, depending on the production mode. This means that a full experimental campaign based on such a design can take at least 20 weeks, which is a major constraint limiting its implementation in bioproduction. Our goal is to develop an alternative inspired by Bayer et al. (Bayer et al., 2020; Nold et al., 2022), which proposes a hybrid approach combining an artificial neural network and a first-principles biological model. This approach enables multiple experimental conditions to be tested within a single production session, reducing the overall duration of a testing campaign by at least a factor of three. Building on this work, we also aim to explore the use of Bayesian neural networks, which we are already applying in the field of medical radiotherapy. Axis 2: Software Sensor for Cell Mortality. This second objective complements Axis 1 by focusing on a different variable to model⬔cell mortality⬔while also reusing the data collected during the first phase. Currently, there is no reliable technology to measure and characterize cell mortality in real-time during a bioproduction cycle. However, understanding cell mortality is crucial, as it significantly impacts the quality of production batches and enables optimal adjustment of critical process parameters, ensuring control over both cell quantity (Axis 1) and viability (Axis 2). The development of software sensors for bioreactors has gained renewed interest in recent years, thanks to the integration of new analytical technologies such as Raman and near-infrared spectrometry (Gustavsson, 2018; Pérez et al., 2020; Reyes et al., 2022). For this study, we aim to utilize more widely available and cost-effective sensors, such as dielectric and turbidity probes, to broaden the applicability of our results. Additionally, we will supplement the learning dataset with daily cell mortality measurements and characterizations from manual sampling. Our work will be based on the research by Kroll et al. (Kroll et al., 2017) as a starting point, and we will seek to enhance the accuracy and robustness of the results by developing a hybrid dynamic model of cell mortality. The hierarchical model structure will be inspired by the approach explored in Axis 1, combining Bayesian neural networks with a first-principles biological model.

Biology, Signals and Systems in Cancer and Neuroscience