A group from Imperial College London has developed an approach that combines artificial intelligence (AI) with more traditional mechanistic models, to predict the yields and critical quality attributes of antibody products.
“These sorts of tools might help [drug developers] develop processes faster and explore different process conditions without exhaustive experiments,” explains Cleo Kontoravdi, PhD, an associate professor in biosystems engineering at Imperial College. “And, for manufacturers, our work can help them with regulatory approval where they need to show they have a good handle on [product] quality and reproducibility.”
Traditionally, some companies use mechanistic models to predict glycosylation and other critical quality parameters, Kontoravdi points out. However, these models are time-consuming to develop and require good knowledge of processes taking part in cells.
In contrast, other companies have a highly-AI-driven approach, but they still often rely on cumbersome experimental measurements, she says.
Tapping into a neural network
Kontoravdi and her team developed a mechanistic model using in-house and published data, and then fed the results into a neural network—a type of AI—which modeled how cells use components of culture media to produce and decorate antibodies with sugars. The team discovered that it reduced errors by 30% compared to a solely mechanistic model.
Unlike a mechanistic model, the neural network can also deal with more complex products, such as fusion proteins that have sugars attached to more than one site, because changes in sugars at any one site affect the 3D structure of the protein. “Sugars are linked to protein folding, so—if you change the 3D structure of a protein—all the model parameters go out of the window,” according to Kontoravdi.
The neural network could reliably predict the effect of up to four specific enzymes being knocked out in the cells, she continues: “It has very good predictability if you start modeling what happens when you modify cells in a site-specific way with CRISPR, which you’re not able to do with a mechanistic approach.”
The team is collaborating with several companies that have taken their approach in-house and applied it to their own data. Kontoravdi also predicts that such modeling will become more common in the future due to regulators in the U.S. and Europe wanting analytical data.