òòò½Íø

From Left: Dr Andy Tan, A/Prof Andre Choo, Dr Ian Walsh, Dr Sean Chia and Dr Zach Pang

 

Science

Antibody (Ab) therapeutics are engineered proteins that can treat various conditions, including cancers, autoimmune diseases, and infectious diseases. There are special molecules found on antibodies called glycans, which can change the effectiveness of the Ab treatment in patients. It is therefore a requirement by drug regulators that glycans are identified and their amount measured, often using very expensive instruments like mass spectrometers (MS). In bioprocessing, Abs are usually produced using Chinese Hamster Ovary cells in a nutrient rich liquid called media, in which they consume the nutrients, grow healthily, and secrete the Ab product. It is well known that the nutrients consumed by the cells can change the glycan type and abundance and consequently the Ab quality. Therefore, in this work the team attempted to use machine learning (ML), a type of artificial intelligence, and use the nutrient consumption information to predict the glycan types and levels.

 

Societal Impact

The ML predictions are shown to be very accurate suggesting that we can tell if the Abs are high or low quality even without the expensive MS glycan analyses. We could use these models to monitor the Ab quality throughout the biomanufacturing process or prioritize batches into high and low quality before the Abs are purified. In the future, the models could form part of a control strategy to ensure high quality Abs and part of media optimization strategy.

 

Technical Summary

N-glycosylation can significantly impact the quality of monoclonal antibody (mAb) therapeutics. In bioprocessing, one way to influence N-glycosylation patterns is by modifying the media used to grow mAb-producing cells. This study explores the use of machine learning (ML) to predict the abundances of N-glycan types based on the growth media components. The ML models utilize a dataset from daily glycomic analyses of Anti-HER fed-batch bioreactor cell cultures grown under twelve different conditions, including variations in dissolved oxygen, pH, temperature, and two different commercial media. By quantifying spent media and using cell consumption rates as inputs to the ML model, we identified a small subset of media components (18 out of 167 mass spectrometry peaks) that are crucial for predicting N-glycan relative abundances in Chinese Hamster Ovary cell cultures.

The results showed excellent prediction performance using the ML models (Regression - correlations between 0.80–0.92; Classification – area under the Receiver Operating Characteristic curve between 75.0–97.2), indicating that the ML can accurately infer N-glycan critical quality attributes from extracellular media. This approach has potential applications in biomanufacturing, particularly in process development, and both downstream and upstream bioprocessing.

Figure 1. There are new variables of the bioreactor media that can be used to predict N-glycan quality of mAbs. Analysing the bioreactor media can determine the quality an antibody therapeutic.

 

References

Lakshmanan, M., Chia, S., Pang, K. T., Sim, L. C., Teo, G., Mak, S. Y., ... & Walsh, I. (2024). Antibody Glycan Quality Predicted from CHO Cell Culture Media Markers and Machine Learning. Computational and Structural Biotechnology Journal.