Monoclonal antibody (mAb) manufacturing must continually improve to keep up with increasing demands. To do this, biomanufacturers can deploy machine learning tools to augment traditional process optimization methods that struggle with the complexities caused by multiple parameters and nonlinear co-dependencies.
Researchers recently presented a machine-learning approach to identify how varying specific process parameters affect upstream yields. Findings were based on industrial-scale data from 65 batches provided by a global contract development and manufacturing organization.
The study—conducted by Breno Renato Strüssmann, a PhD student at Technical University of Denmark (DTU) and the paper’s first author, along with Lars Hvam, PhD, professor, DTU, and Anderson Rodrigo de Queiroz, PhD, associate professor, North Carolina State University—compared the accuracy of random forest regression, gradient boosting machine, and support vector regression in predicting yield. They focused on three yield indicators: bioreactor final weight, harvest titer, and packed cell volume.
Key parameters affecting yield
“The findings are of immediate practical use for biomanufacturers,” Strüssmann tells GEN. “Existing production data can be leveraged to identify the process parameters that most strongly affect yield.” Here, they were thaw media warming time and tyrosine or nutrient additions during key culture phases. Knowing this “allows manufacturers to focus their control efforts on the most influential variables, supporting data-driven decision-making even before deploying new experiments or automation tools.”
“It was interesting to see how small timing variations early in the process could significantly affect final yield,” Strüssmann says. Varying key process parameters by 10% during the thaw step warming time affected harvest by 1.09%. Making those variations during the weight of tyrosine addition at early and/or mid-phases affected harvest by 0.6%. A 10% variation during the growth medium addition volumes during the late phase A affected harvest 0.45%, and doing so at mid-phase B affected harvest 0.4%. Controlling these parameters precisely, therefore, is vital to improving production yield.
“It was also notable that support vector regression outperformed the other models in predicting bioreactor final weight, even when using a relatively small industrial dataset,” he adds.
While emphasizing the need for additional, specific research, the researchers suggested this approach could ultimately be integrated into automated or supervisory control frameworks for upstream bioprocessing and data-driven decision-making in biomanufacturing.
“By focusing on empirical patterns rather than predefined kinetic models, this work demonstrates how machine learning-based analysis of historical data can provide actionable insights for upstream optimization,” they note.
Therefore, they caution, when modeling mAb production, it’s important “to incorporate process constraints and control logic along with interpretable or hybrid modeling frameworks to enable practical deployment in regulated biomanufacturing environments.”