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

Precise control of critical quality attributes, including titer and glycosylation, is essential in bioprocessing, yet conventional design‑of‑experiments methods are challenged by the high-dimensional, nonlinear design space for media and process parameters. We assemble a comprehensive glycan‑focused Chinese hamster ovary (CHO) fed‑batch dataset and develop a computational workflow (i) to train machine learning (ML) models to predict key CQAs, (ii) apply a hybrid ML + knowledge-based strategy to select potentially impactful features, and (iii) generate combinatorial media designs. The resulting models predict final titer (R² ≈ 0.93) and major glycan metrics-mannosylation, fucosylation, galactosylation (R² ≈ 0.79-0.95)-directly from initial media composition and process parameters without requiring spent media analysis. Feature selection shortlisted 20 features out of 76 for a second-tier validation, from which 15 were confirmed as actionable levers impacting titer and glycosylation, uncovering glycan effects independent of nucleotide sugar supplementation. Finally, we incorporated our workflow, utilizing a ML surrogate model coupled with simulated annealing, in a proof‑of‑concept active learning step, successfully proposing a media composition and process parameter combination that reduced mannosylation by 10% while increasing titer. Together, these results underscore how ML‑enhanced DOE can accelerate CHO process development and explore complex biomanufacturing spaces with greater efficiency.