A recurrent neural network for soft sensor development using <scp>CHO</scp> stable pools in fed‐batch process for <scp>SARS</scp>‐<scp>CoV</scp>‐2 spike protein production as a vaccine antigen

Fed-batch recombinant therapeutic protein (RTP) production processes utilizing Chinese Hamster Ovary (CHO) cells can take a long period of time (>10 days). Within this period, not all critical features may be measured routinely, and in fact, some are only measured once the process is terminated, complicating decision making. As a consequence, utilizing routine current day bioreactor online data to aid in next day predictions is a promising strategy for model predictive control-based feeding strategies. The article details the development of a proposed soft sensor that merges current day bioreactor online data and offline historical sampling data to generate predictions about the next day of the production process. This approach demonstrated the ability to track product titer, cell growth, key metabolites, and cumulative glucose consumption across the 17-day process with low normalized root mean squared error (nRMSE = 0.24) and low normalized mean absolute error (nMAE = 0.18) as well as high linearity with respect to ground data (average R² = 0.97). It was also demonstrated that the same model architecture could effectively soft sense product titer and metabolic profiles (glucose, lactate, ammonia) without having sampling day's offline data as inputs to the model. This suggests that the proposed model could act as a true soft sensor of hard-to-determine variables such as the trimeric SARS-CoV-2 spike protein that relies on end-of-process measurements to acquire the data (labor-intensive semi-quantitative SDS-PAGE gels or ELISA assay). Instantaneous specific glucose consumption rates were also predicted and showed good agreement with experimental measurements, further offering opportunities for online glucose control.