Researchers in Japan have developed a machine learning process. It simultaneously designs new molecules and suggests the chemical reactions to make them. The team consists of scientists from the Institute of Statistical Mathematics (ISM) in Tokyo. They published the results in the journal Science and Technology of Advanced Materials: Methods
Many research groups are making significant progress in using artificial intelligence (AI) and machine learning. With these, they are designing feasible molecular structures with desired properties. But progress in putting the design concepts into practice is slow. The biggest obstacle is the technical difficulties in finding chemical reactions that can make the designed molecules. Efficiency, cost and practicality for real-world use must be considered.
New algorithm
"Our new machine learning algorithm and associated software system can design molecules with all desired properties. On top of that, they propose synthetic routes to make them from an extensive list of commercially available compounds." So says statistical mathematician Ryo Yoshida, leader of the research group.
Machine learning and scientists
The process uses an approach called Bayesian inference. It works with a huge set of data on different options for starting materials and reaction routes. The possible bases: all combinations of the millions of compounds that are readily available. The computer algorithm evaluates the huge range of feasible reactions and reaction networks. It does this to discover a synthetic route to a compound with the properties sought. Expert chemists can then review the results to test and refine what the AI proposes. The AI makes the proposals while humans decide which solution is the best.
Medicines and lubricants
“A case study examined the design of molecules similar to those used in medicines. The method demonstrated outstanding performance,” says Yoshida. Routes were also designed to produce industrially viable lubricant molecules.
Accelerating
Yoshida: "We hope that our work will accelerate the process of data-driven discovery of a wide range of new materials." In support of this goal, the team made the software implementing their machine learning available to all researchers on the website GitHub.
Polymers
The current success focused only on the design of small molecules. The team now plans to investigate whether they can adapt the procedure to design polymers. Many of the most important industrial and biological compounds are polymers. But making new versions suggested by machine learning has proved difficult. This is because of challenges in finding reactions to build the designs on. This new technology can change that.
Image by Gerd Altmann via Pixabay
Also read: AI and federated learning can improve detection of brain tumours by 33%
