Use of artificial intelligence (AI) in bioprocessing is set to increase. According to a Deloitte survey, biopharma is undergoing a “digital transformation” with 58% of executives quizzed identifying the adoption of innovative technologies, AI, and automation as a top priority.
Some of the advantages of AI for biopharmaceutical manufacturers are obvious, says Fabian Mohr, an advanced manufacturing systems researcher at the Massachusetts Institute of Technology (MIT). He cites AI’s ability to predict the impact processing changes will have on product critical quality attributes (CQAs) as an example.
“CQAs are crucial indicators for the quality of the final product, especially for the highly regulated sector of biopharmaceutical manufacturing which needs to guarantee high quality of their products,” he tells GEN. “Sometimes those CQAs cannot be measured directly or it is expensive to measure in-line or at-line. In this case, AI can be used to predict these CQAs instead.”
Process control is another application. According to Mohr, operating variables from early batch production can be used to train AI-based control systems.
“A model can be developed based on a training data set to construct a binary classifier predicting whether the final product will be in-spec or out-of-spec. When new data on operations are collected, the supervised classifier constructed from past data can be used to assess whether the final product is likely to meet specifications,” he explains. “In case it is not likely that the batch will be in-spec, time and resources can be saved by not proceeding with downstream processing. AI also has the potential to more quickly find the causes of process deviations.”
But an AI system is more than just a technology for modeling and crunching numbers. A well-trained AI system can help biopharmaceutical companies make better choices about how they analyze data from manufacturing processes.
The key is selecting the right analytical method for the right data, according to Mohr.
“In general, there is a plethora of different data analytics methods for a given objective and none of those methods is superior to the others in all circumstances,” he points out. “Each of the different methods has certain inherent assumptions that make the methods most suitable for certain types of problems.”
The challenge is that data analytics practitioners tend to either favor the methods with which they are most familiar or try a limited range of approaches. But this approach can result in less-than-optimal predictions that do not exploit the most powerful methods for the particular problem.
To address this, Mohr and colleagues developed a machine learning approach that selects the analysis method after looking for specific characteristics in early data.
“Based on those characteristics, the approach selects a subgroup of algorithms suitable for the dataset. In the next step, a rigorous cross-validation procedure is applied to determine the best algorithm of the subset,” he says. “Best practices in data analytics are incorporated into the approach to generate the most accurate model that can be constructed from the given dataset.”