Hybrid deep modeling of a CHO-K1 fed-batch process: combining first-principles with deep neural networks

Area of Science:

• Biotechnology
• Process Engineering
• Computational Biology

Background:

• Hybrid modeling, integrating first-principles and machine learning, is crucial for Biopharma 4.0.
• Chinese Hamster Ovary (CHO) cells are vital for industrial glycoprotein production.
• Previous hybrid models often used shallow Feedforward Neural Networks (FFNNs).

Purpose of the Study:

• To compare deep versus shallow hybrid modeling for CHO cell process development.
• To evaluate the impact of FFNN depth on model performance.
• To assess predictive capabilities for glycoprotein production processes.

Main Methods:

• Utilized data from 24 fed-batch cultivations of a CHO-K1 cell line.
• Compared hybrid models with FFNNs of varying depths (3-5 layers).
• Employed classical training (Levenberg-Marquardt) and deep learning training (ADAM).

Main Results:

• Deep hybrid models showed systematic generalization improvements over shallow models.
• Training and testing errors decreased by 14.0% and 23.6%, respectively.
• The deep model accurately predicted 30 state variables and key metabolic shifts.

Conclusions:

• Deep hybrid modeling offers superior performance for CHO cell bioprocess development.
• Increased computational time for deep models is offset by enhanced accuracy.
• Deep hybrid models are expected to accelerate the development of digital twins in biopharma.