To improve the outcome of bioprocessing, many manufactures rely in part on computation, including artificial intelligence (AI) and machine learning (ML). In a recent article, Anurag S. Rathore, PhD, professor of chemical engineering at the Indian Institute of Technology in New Delhi, and his colleagues discussed how AI-ML can improve the bioprocessing of monoclonal antibodies.
“Manufacturing in biopharma is uniquely challenging due to the inherent complexity of the products being made,” says Rathore. “The production and purification processes of biopharmaceuticals are also complex, consisting of a number of upstream and downstream unit operations, including cell culture, clarification, chromatography, membrane separations, enzyme reactions, refolding, etc.”
Bioprocessors can apply techniques based on AI-ML to a process’s development, optimization, monitoring, control, automation, and correction.
“Unlike mechanistic approaches to process control that rely on deep fundamental data that is generally not available from common sensors, AI-ML approaches are data-driven and can be developed using the huge process variable datasets that are collected by process sensors that are typically present in the manufacturing facilities,” Rathore explains. “In the current digital era, data-driven AI-ML approaches—such as artificial neural networks, fuzzy logic, multivariate process evolution models, and expert self-learning control systems—have promising applications for biopharmaceutical manufacturing.”
As an example, he says, “AI-ML can be a key enabler of the shift from operation based on quality-by-testing to quality-by-design.”
Generating bioprocessing models
In the work by Rathore and his colleagues on monoclonal antibodies, the scientists used a variety of AI-ML methods to generate bioprocessing models. As Rathore says, “Techniques are tuned for optimal performance and compared based on,” a collection of statistical parameters.
For instance, the scientists compared decision-tree (DT) and random-forest (RF) methods for real-time process monitoring. “DT is prone to overfitting with high variance and bias while RF—being an ensemble of decision trees—has the ability to mitigate this setback,” Rathore says. “For real-time predictions, RF error percentage of predicted outcomes was less than 5% in all situations.”
Such approaches could be used by commercial bioprocessors in various ways.
“The AI-ML models compress the chromatography profiles, which form a signature of each cycle, into immediately usable information about critical quality attributes,” Rathore says. “This approach offers a way of tracking quality on a cycle-by-cycle basis without the need for introducing a surge tank post-chromatography in which multiple elution cycles are mixed and samples are periodically drawn for offline analysis.” For Bioprocessing 4.0, real-time analysis lies at the heart of improving the quality of products.
