Generic and specific recurrent neural network models: Applications for large and small scale biopharmaceutical...

Area of Science:

• Biotechnology and Bioprocessing
• Computational Biology
• Chemical Engineering

Background:

• Accurate estimation of time-varying upstream process outcomes is crucial but challenging.
• Existing parametric, semi-parametric, and non-parametric methods have limitations in reliability and applicability.
• Recurrent neural networks (RNNs) offer a potential solution for complex bioprocess modeling.

Purpose of the Study:

• To develop and evaluate generic and product-specific RNN models for calculating upstream process parameters.
• To assess the temporal evolution of growth and metabolite-related parameters.
• To demonstrate the utility of RNNs for both large-scale manufacturing and early-stage development.

Main Methods:

• Implementation of generic and product-specific recurrent neural network (RNN) models.
• Application of models for computing growth and metabolite-related upstream process parameters and their temporal dynamics.
• Validation of model accuracy against experimental data, including root-mean squared error of prediction analysis.

Main Results:

• RNN models achieved high accuracy in predicting product titer and other upstream outcomes compared to experimental data.
• Calculated root-mean squared errors of prediction were significantly lower than experimental standard deviations.
• A generic RNN model successfully simulated process outcomes under varying temperatures for platform processes.

Conclusions:

• RNN models provide a highly accurate and broadly applicable approach for upstream bioprocess monitoring and control.
• The straightforward implementation and high predictive capability make RNNs a promising alternative to traditional methods.
• RNNs offer significant benefits for process development, scale-up, and manufacturing, particularly for platform processes.