A biopharma company specializing in difficult-to-produce biologics has developed a deep learning model to rapidly predict protein expression using their moss-based production process. The model can reduce the time needed for early cell line development from three to four months to one day, according to Benjamin Fode, PhD, head of genetic engineering, Eleva, who added that it’s a relatively novel use of neural networks and will be applied in the company’s R&D programs.
“My personal view is that these models are mainly used for antibody discovery and figuring out structural features. They’re not currently used that often for predicting expression, especially not for the whole production process,” he says. “We’re not the only ones doing this, but it is quite new and at the cutting edge of current industry practice.”
The model does the equivalent of screening 8,000 to 10,000 constructs to see which perform best for protein production at larger scales, notes Fode. The best performing constructs can then be taken into cell line development in a wet lab.
Normal workflow
“The normal workflow often uses shallow learning models like design-of-experiments to test the expression of 40 to 100 constructs,” he continues. “But even when you have a multi-well based assay and can screen 96-well plates in a semi-automated way, your ability to perform large numbers of experiments remains limited.”
In contrast, he points out, the model can estimate protein expression from thousands of constructs in a matter of minutes. Having tested the model, it identified the best and least good variants and had a 74% overlap with real-life data on the average performers.
“The key point to take away is that we would have chosen the same constructs as the deep learning prediction model, and the outcome was comparable to what we get from four months’ wet lab work,” he says.
Having developed a model for protein expression, Fode says Eleva is now moving into investigating other factors–apart from protein expression–that affect product yield.
“We’ve managed to really speed up the process by focusing on some key factors in play in protein production,” he explains, stating that future work might include examining subsequent steps in the production process including downstream processing as part of the model.
