Toward Machine Learning-Guided CHO Bioprocess and Media Optimization for Improved Titer and Glycosylation

Area of Science:

• Biotechnology
• Bioprocessing
• Computational Biology

Background:

• Controlling critical quality attributes (CQAs) like titer and glycosylation in bioprocessing is crucial but challenging due to complex media and process parameters.
• Conventional design of experiments (DOE) methods struggle with the high-dimensional and nonlinear nature of biomanufacturing design spaces.

Purpose of the Study:

• To develop a computational workflow using machine learning (ML) to predict key CQAs in Chinese hamster ovary (CHO) fed-batch cultures.
• To identify impactful features and generate combinatorial media designs for optimizing bioprocesses.
• To accelerate CHO process development and efficiently explore complex biomanufacturing parameter spaces.

Main Methods:

• Assembled a comprehensive glycan-focused CHO fed-batch dataset.
• Trained ML models to predict titer and major glycan metrics (mannosylation, fucosylation, galactosylation).
• Employed a hybrid ML + knowledge-based strategy for feature selection and utilized a ML surrogate model with simulated annealing for active learning.

Main Results:

• ML models achieved high prediction accuracy for titer (R² ≈ 0.93) and glycosylation (R² ≈ 0.79-0.95) directly from initial media and process parameters.
• Identified 15 actionable features influencing titer and glycosylation, independent of nucleotide sugar supplementation.
• Successfully proposed a media composition and process parameter combination that reduced mannosylation by 10% while increasing titer.

Conclusions:

• ML-enhanced DOE significantly accelerates CHO process development and optimizes critical quality attributes.
• The developed computational workflow efficiently explores complex biomanufacturing design spaces.
• This approach enables precise control over bioprocess outcomes, improving efficiency and product quality.