A Japanese-government initiative to support artificial intelligence research is paying dividends in the field of biopharmaceuticals. The biotech start-up RevolKa accelerates protein engineering using machine learning, and has recently attracted 350 million yen (roughly US$2.5 million) in funding and partnerships with several drug manufacturers.
Taking the direct approach
Using evolutionary processes to create new proteins is a specialty of Mitsuo Umetsu, a professor in biomolecular engineering at Tohoku University in Japan. By introducing random mutations into genes and iteratively raising new generations of proteins with the help of bacterial biofactories, Umetsu and his team have created antibodies that stick to surfaces such as plastics or ceramics, for example, as well as enzymes that transform cellulose into sugars for producing bioethanol.
However, one problem with this ‘directed evolution’ approach is that it can generate large libraries of DNA sequences that only differ slightly from each other. If the DNA library is too large, it becomes challenging to experimentally screen for proteins of interest. Conversely, a library without enough sequences may not hold any relevant variants.
“In directed evolution, the probability of success depends on the scale of investment,” says Umetsu. “That’s why I felt machine learning could guide us to optimal proteins without being limited by the scale of the DNA library.”
In 2016, Umetsu joined the Center for Advanced Intelligence Project of RIKEN, Japan’s largest national research institute, to team up with machine learning experts from across the country. This partnership produced a different slant on typical protein engineering experiments. After creating a DNA library, the team characterized the variants and used the results to train a machine learning model. The new algorithms were then used to steer subsequent rounds of gene mutations towards a target protein.
Through experiments that included changing the emission colours of fluorescent proteins and enhancing the activity of peptidyl enzymes, the researchers discovered that machine learning could significantly reduce the typical DNA library size used in directed evolution.
“When we carefully acquired high-quality data, we only needed about 100 samples to train the machine learning to propose the optimal protein and enzyme sequences, compared to 10,000 or so needed with conventional techniques,” explains Umetsu.
Change for the better
While Umetsu was conducting his machine learning research, he met Shiro Kataoka, an executive with more than 30 years of experience in biopharmaceuticals. Together, they decided to commercialize the machine learning technology and apply it to optimizing preclinical drug candidates.
“In lead optimization, properties such as biological activity, structural stability and expression should be changed at the same time, and it’s difficult to address this with large-scale libraries,” says Kataoka, who is RevolKa’s president and CEO. “Our technology can simultaneously solve multiple properties with minimal training data, which drastically cuts the time and effort needed to find solutions.”
Umetsu’s and Kataoka’s plan for RevolKa, which is named after the Ainu word for raise (‘reska’) and the Latin word evolve, is for it to become a global leader in the development of optimized biopharmaceuticals.
“Protein pharmaceuticals have been diversified and there are various characteristics that should be improved,” says Umetsu. “We’re using our recent round of funding to evolve our technology to meet the needs of our partner companies and to help them develop innovative bioproducts.”