òòò½Íø

Science

Medicines like cancer-fighting antibodies are made using living cells, which require just the right mix of nutrients and conditions to grow effectively and produce high-quality drugs. Identifying the best “recipe” is normally slow and expensive, as it requires many experiments. In our study, we trained machine learning models to learn how different ingredients and conditions affect cell growth and product quality. We then used these models to suggest better combinations for subsequent testing - speeding up the search for optimal conditions using fewer experiments.

Societal Impact

This work demonstrates how machine learning can accelerate the development of biologic medicines, such as life-saving antibodies and therapeutic proteins - by making cell culture optimization faster, smarter, and less resource-intensive. By reducing the number of experiments needed to achieve high product quality and yield, this approach can help lower manufacturing costs, shorten development timelines, and improve access to affordable biologics worldwide. In our single-round proof-of-concept, we achieved a ~6% increase in drug yield and a ~10% reduction in undesirable product variants, illustrating the potential for meaningful gains in production efficiency and quality. Ultimately, these improvements support a more responsive and sustainable biomanufacturing ecosystem, benefitting patients, healthcare systems, and global health equity.

Technical Summary

We present a machine learning-guided workflow for optimizing CHO fed-batch bioprocesses using predictive models trained on initial media composition and process parameters (pH, DO, feed %). These models accurately predict Day 14 Critical Quality Attributes (CQAs) – titer, mannosylation, fucosylation, and galactosylation - with R² values ranging from 0.80 to 0.95. We applied a hybrid feature selection approach integrating Machine Learning-based (ML) importance scores and domain knowledge to shortlist 20 influential factors, 15 of which were subsequently validated in follow-up experiments. The results revealed both expected and novel influences on glycosylation, even in the absence of nucleotide sugars.

In the second stage, we used the trained models within an active learning framework to guide in silico optimization via simulated annealing. This allowed us to propose combinatorial media and process conditions that resulted in a 10.5% reduction in mannosylation while increasing titer by 6.3%. Our approach demonstrates how predictive ML models can be iteratively refined and deployed to streamline Design of Experiments (DOE), guide multi-objective optimization, and accelerate CQA-focused process development in mammalian cell culture systems.

References

Walsh, I., Shozui, F., Sato, A., Park, U., Cho, H., Chia, C., "Toward Machine Learning-Guided CHO Bioprocess and Media Optimization for Improved Titer and Glycosylation," in Biotechnology Journal, vol. 20, no. 11, pp. e70149, Oct. 2025.