Although artificial intelligence (AI) and machine learning (ML) get applied to many things—from bacteria to toothbrush design—there are usually challenges along the way. To find out about the bottlenecks in bioprocessing, I asked four experts: What is the biggest challenge of using AI in bioprocessing, what makes it so challenging, and how could it be addressed? Here’s what they said.
Richard D. Braatz, PhD, Edwin R. Gilliland Professor; professor of chemical engineering, MIT: Few bioprocesses generate data of sufficient quality and quantity for generic AI software to be applicable. The high dimensionality, complexity, dynamics, and variability of most bioprocessing limit the ability of generic AI software to make reliable decisions. This situation can be addressed through the development of sensor technologies that improve the quality and quantity of data, and the development of AI methods that take bioprocess domain knowledge into account.
Anurag S. Rathore, PhD, professor and head, department of chemical engineering, Indian Institute of Technology, Delhi: The biggest challenge of implementing AI in bioprocessing, perhaps, is the fact that AI models are purely data based and as such lack any mechanistic insights into the process that they attempt to model. As a result, one would need to recalibrate the model every time there are changes in the system—for example, if we change the chromatography equipment from one vendor to another or scale up or technology transfer.
Another challenge is that you need a significant amount of data which might be a challenge during the early stages of product development. AI modeling is best suited for commercial manufacturing when systems are relatively constant and a lot of data is being generated. Also, in my opinion, hybrid models—combinations of mechanistic and data based—would be preferred over purely data-based models.
Wei Xie, PhD, assistant professor, mechanical and industrial engineering, Northeastern University: The primary goal of using AI in bioprocessing is to enhance scientific understanding and process predictions with limited physical experimental data, while providing interpretable insights for optimizing processes and supporting automation.
Black-box modeling and AI/ML approaches do not incorporate bioprocessing mechanisms and cannot provide input-output causal information. This hinders interpretability, data fusion, manufacturing systems integration, and sample-efficient learning—crucial for guiding the design of informative experiments.
Emerging sensing technologies, such as optical sensors and multi-omics assays, can facilitate real-time monitoring of biomanufacturing processes at molecular and cellular scales. Multi-scale hybrid, which is mechanistic plus ML, modeling, and optimal learning strategies, accounting for all sources of uncertainties due to limited knowledge and sensing capability, can decode the fundamental mechanisms of bioprocessing, support data integration, and guide the most informative design of experiments for process optimization.
Barkha Singhal, PhD, assistant professor, school of biotechnology, Gautam Buddha University: The biggest challenge of using AI in bioprocessing is the dynamicity and complexity of living systems. Heterogeneous data arise from multiple sources, such as fermentation parameters, growth and product-formation kinetics, metabolic pathways, genomics, transcriptomics, proteomics, and metabolomics. Integrating these data is complex in AI-driven models and algorithms based on training data sets and the black box approach. The nonlinear behavior of biological systems poses significant challenges for real-time predictions for AI tools in bioprocessing.
Therefore, there is a pressing need to develop comprehensive databases for bioprocessing, with consistent data labeling and formatting needed. Developing advanced AI algorithms for assessing temporal data and nonlinear dynamics in a time-dependent manner is crucial for bioprocesses. Interdisciplinary collaborations between bioprocess engineers, biologists, and AI experts are required to generate the real high-quality data that can be used in training sets and successfully implemented in real-time prediction of the complexity of bioprocesses.
So, despite the potential of AI in bioprocessing, many bottlenecks remain. Like most transitions, it will take time before bioprocessors can really get the most out of AI-based tools.
