Machine-learning models could help drug firms remove viral contaminants from products more effectively as long as they have access to high-quality training data, say biomanufacturing researchers.
Anion exchange chromatography (AEX) is an effective tool for virus removal. The approach separates molecules based on charge using a positively charged resin to attract negatively charged molecules present in the process stream. The difficulty is that developing an effective AEX protocol for each new process requires a large number of characterization experiments, which are time-consuming and expensive.
But machine learning can streamline AEX protocol development, according to a study by scientists at Catalent Biologics and industrial artificial intelligence firm, Quartic.ai.
“Developing anion exchange chromatography processes with robust viral clearance tends to be challenging due to a limited understanding of the impact of various parameters on the viral clearance performance,” write the scientists. “Moreover, only a handful of viral clearance studies can be performed during process development due to high costs, long timeframe, and executional difficulty associated with these studies. Therefore, the impact of various process parameters and feed characteristics on the viral clearance performance cannot be fully characterized.
“In such scenarios, in silico models, such as machine-learning models, trained on historical datasets can be used to quickly, cheaply, and fully explore the process design space and suggest appropriate process improvements, resulting in the accelerated development of effective viral clearance processes and enabling predictive monitoring during manufacturing.”
Model development
In the study, the researchers used data from viral clearance experiments performed by Catalent over the past 30 years to train a model to predict how various AEX protocols would perform. The 104 training data points included information from 30 recombinant protein products and two model viruses: the mouse minute virus (MMV) and the xenotropic murine leukemia virus (XMuLV).
To test the resulting ML model, the team compared generated virus log reduction values with data points from the training data set and found that for approximately 70% of the data points the predicted values closely aligned with the experimental values.
Having detailed data was vital, according to the authors, who suggest firms interested in using machine learning in this way will need the same sort of high-quality information to achieve comparable results.
“The primary challenge in implementing ML approaches in bioprocessing is the availability of actionable data. Biopharmaceutical manufacturers require contextualized and comprehensive viral clearance datasets to be able to train and validate ML-based models,” according to the scientists.
A robust, automated framework for storage, pre-processing, and training is also necessary to keep models updated along with interfaces that let users to design, interpret, and deploy them effectively. But, even with these requirements, machine learning has real potential in helping process developers to optimize viral clearance and other downstream steps.
“The adoption of ML-based approaches in designing, developing, characterizing, and controlling bioprocesses is rising, and many companies now offer end-to-end solutions for data collection, storage, analytics, and model deployment” the authors conclude.