Artificial intelligence has been making a difference in the discovery lab for a decade. Recently, consulting group BCG suggested that AI played a pivotal role in the identification of more than 150 current drug candidates, 15 of which are in trials.
But in the biomanufacturing facility, it’s a different story. Instances of AI use in drug production are few and far between. And this is a missed opportunity, according to Kiefer Eaton, from industrial artificial intelligence developer Basetwo AI, particularly because drug makers already have many of the elements they need to benefit from AI.
“AI has been widely adopted for drug development and has recently seen a lot of growth for commercial pharmaceutical activities. However, biomanufacturing remains underserved despite the plethora of data to leverage,” he says. “AI has a unique opportunity in manufacturing through the creation of ‘hybrid process models’ through physics-informed AI. These hybrid models combine the power of engineering knowledge with machine learning to create digital twins— virtual representations of physical processes or unit operations.”
Eaton cites the ability to optimize processes in a virtual environment without consuming resources as one potential application.
“Applying AI in process development requires two things: process knowledge and data. Leveraging process through engineering equations allows us to contextualize the training of the model that learns from the data,” he continues. “This means we can use less data than would otherwise be required to train a machine learning model since the process knowledge fills in a lot of the gaps. This even allows us to use data from related processes at different sites or scales.”
For example, data from a fed-batch process can be used to train a perfusion bioreactor model by changing the equations, notes Eaton, who adds, “This significantly improves cross-learning of processes for manufacturers.”
So the drug industry has the data it needs to use AI in manufacturing. But setting up an industrial AI architecture is about more than just data, points out Eaton, who suggests drug makers will need to invest in internal training and outside expertise.
“A big challenge getting the right stakeholders and subject matter experts to collaborate on a common platform to tap into the strengths of both groups. For example, a data scientist may reach for Python to build a predictive process model using AI while an engineer would stick to MatLab or other traditional pieces of modeling software,” he tells GEN.
“Beyond this, sometimes manufacturers don’t have the capability to deploy AI at scale which prevents adoption altogether, especially if there’s a lack of data science expertise to manage these AI models throughout their entire lifecycle.”
Having a technology that minimizes coding requirements is also a good idea, according to Eaton, citing Basetwo’s own AI system, which has already been deployed by a Johnson & Johnson Innovation team, as an example.
“We built Basetwo from the ground-up with engineers in mind. An engineer can build a data pipeline using our drag-and-drop interface, visualize data in a spreadsheet view, and create a hybrid model using our equation editor with Excel-like syntax, all without writing a single line of code,” he says. “This makes unlocking the power of hybrid modeling and AI accessible to engineers and improves collaboration between data scientists and engineers.”