Machine learning is usually applied to huge datasets, but researchers from the ETH Zurich in Switzerland are using it to reduce the number of experiments needed for biopharmaceutical formulation. “When people think about machine learning, they think about big data, but here we’re doing it on smaller data to screen millions of different possibilities for experiments,” explains Paolo Arosio, PhD, assistant professor in biochemical engineering at ETH Zurich.
Unlike traditional design-of-experiment (DoE) approaches, he says, the algorithm can learn from experiments and transfer knowledge from one experiment to the next.
“You don’t need to start from scratch again,” he continues. “And when there are multiple objectives or targets, you can make trade-offs, and find a compromise.”
The algorithm, he says, provides formulation designers with the equivalent of a car accelerator—helping them to reach the best formulation faster, and find solutions that would not be possible by intuition. The proof-of-concept of the algorithm was tested on single-chain antibody products but can be used for any formulation. It could help companies explore a wider range of excipients–from a handful to any of the thousands that he says are authorized by the FDA.
Machine learning tools, such as his algorithm, help to optimize multiple properties of the ideal formulation, he points out, which can include stability, viscosity, self-association, and non-specific interactions. This is especially important when there are multiple properties to optimize, such as in new classes of products, e.g., cellular and vesicle-based pharmaceuticals.
“The more complex the product, the better this approach, so we’re starting to use it for exosomes, and it can be used for gene and cell therapies, as well as vaccines,” he explains.
Arosio gave a presentation about his work at Bioprocessing Summit Europe in March. It was based on a paper published in Molecular Pharmaceutics and lead authored by Harini Narayanan, Arosio’s doctoral student.