(Nanowerk News) Researchers in Japan have developed a machine learning process that simultaneously designs new molecules and suggests chemical reactions to create them. The team from the Institute of Statistical Mathematics (ISM) in Tokyo, published their results in the journal Advanced Materials Science and Technology: Methods (“Bayesian methods for simultaneously designing molecules and synthetic reaction networks”).
Many research groups are making significant progress using artificial intelligence (AI) and machine learning methods to design feasible molecular structures with desired properties, but progress in putting design concepts into practice has been slow. The biggest hurdle is the technical difficulty in finding chemical reactions that can create molecules designed with practicable efficiency and cost for real-world use.
“Our new machine learning algorithm and associated software systems can design molecules with any desired properties and suggest synthetic routes to manufacture them from an extensive list of commercially available compounds,” said statistical mathematician Ryo Yoshida, leader of the research group.
The process uses a statistical approach called Bayesian inference that works with vast data sets about different options for starting materials and reaction pathways. The possible starting materials are any combination of the millions of compounds that can be purchased easily. Computer algorithms evaluate a large number of possible reactions and reaction networks to find a synthetic route to a compound with the properties it has been instructed to go for. Expert chemists can then review the results to test and refine what the AI proposes. AI makes suggestions while humans decide which is best.
“In case studies for designing drug-like molecules, this method shows great performance,” said Yoshida. It also designs routes to industrially useful lubricant molecules.
“We hope our work will accelerate the process of data-driven discovery of new materials,” concluded Yoshida. To support this goal, the ISM team has made the software that implements their machine learning system available to all researchers on the GitHub website.
Current successes have focused solely on small molecule designs. The team now plans to investigate adapting the procedure to design polymers. Many of the most important industrial and biological compounds are polymers, but it is proving difficult to fabricate new versions proposed by machine learning due to challenges in finding reactions to construct designs. The design discovery and simultaneous reaction options offered by this new technology can break through that barrier.