Although artificial intelligence (AI) and machine learning (ML) are hot buzz words in science today, much more work could be done in applying these techniques to bioprocessing. Harini Narayanan, PhD, a postdoc and AI-expert at the Massachusetts Institute of Technology (MIT), sees three major benefits of applying AI to the bioprocessing of drugs.
“First, it provides the ability to strategically navigate the highly complex and combinatorial design space and subsequently makes the bioprocess development more efficient,” she says. “Second, using AI/ML-based surrogates to digitally mimic the real system leads to the reduction of resources and accelerates the process development timeline.”
Last, she points out that AI/ML creates “the opportunity to generalize the protocol for process development—mediated by algorithms and not influenced by subjective choices—that is translatable from one drug product to another and from current to future drug products.”
In particular, Harini works on hybrid modeling, which combines two different modeling paradigms: data-driven modeling, such as ML models that solely rely on data; and knowledge-driven modeling, such as a system of ordinary differential equations that solely rely on the mathematical formulation of domain knowledge.
Hybrid modeling has been receiving much attention
“The field of hybrid modeling has gained a lot of attention in bioprocessing, both for upstream and downstream processes,” she continues. “Such models require much less training data, are much better in extrapolation, and subsequently also capture physically relevant behavior.”
She confirmed these benefits in upstream and downstream bioprocessing operations. Her work also demonstrated the value of AI/ML in several aspects of bioprocessing, including optimization, monitoring, and control.
More recently, Harini proposed the concept of degree of hybridization, which she describes as “the optimal ratio between the amount of knowledge to be added and flexibility to be allowed for data-driven ML modeling.” Her work showed that the optimal ratio depends on the objective, such as optimizing a bioprocess or enhancing the understanding of it.
However, she adds: “Though supporting data with process or domain knowledge is helpful, there is a difference between adding generalized trends versus specific biases to the model with the latter being more detrimental than useful.”
To drive more use of AI/ML, Harini says that bioprocessors must be willing “to move away from the conventional methods of bioprocess development so that the complete potential of AI can be harnessed.”
As an example, she says that “a space-filling design probing the design space uniformly would generate an information-rich dataset to start the model building activities, and that this approach surpasses the classic design-of-experiments method. Subsequently, the models can be used to iteratively suggest suitable next experiments. This accelerates the search for optimal operating conditions while reducing the resources required,” as she has shown in, for example, optimizing a formulation.
Although transitioning bioprocessing to more AI/ML-based approaches sounds like a massive task, Harini says: “I strongly believe that this effort is more in the direction of taking a step back to make a higher jump, especially if the AI-based pipelines are developed with the intention to serve as general-purpose protocols.”