AI and Chemistry: Accelerating Drug Discovery and New Material Development
Discovering new materials that show improved, up to now unknown or unexpected properties for industrial or medical uses is something resembling the holy grail of modern chemistry. And new materials today mean: new compounds, whose atoms are ingeniously combined to form ever more complex molecule structures that might yield products with desired physical properties or enable some valuable chemical reactions.
Now there is a way of doing this atomic-scale puzzle in novel ways, which haven’t been explored yet by human chemists: by deploying artificial intelligence in the guise of deep learning networks. Judging the flood of scientific publications on promising discoveries and methodologies from universities and industrial firms around the world, AI is turning into a most potent tool in the process of predicting, discovering, and testing new organic and inorganic materials. Especially when it comes to the very capital-intensive business of drug discovery.
Three young researchers at Florida State University have succeeded to devise a powerful predictive AI model for combining three elements into molecules with similar – or better – chemical properties by training a deep learning neural network on analyzing 50,000 known crystalline structures. The user may simply enter a set of promising elements and after just a few seconds the AI-based predictive model returns a list of possible molecular structures, which then can be experimentally tested for their effective results.
On a similar track is the organic chemist and AI researcher Marwin Segler at the University of Muenster in Germany. Using a method called “retrosynthesis“, going backwards from viable known compounds and analyzing what starting materials and reactions could possibly produce these, he and his colleagues use a deep self-learning network to vastly accelerate this process from hours and days down to minutes. This way, they have investigated about 12.4 million of organic-chemistry reactions to find some appropriate starting reagents, or elements.
Will this possibly spell the end of traditional lab chemists who use their intuition and experience in discovering new drugs? “No”, assures Segler: “It will be an assistant for the chemist who wants to make molecules as quickly as possible.”
In this vein, at Glasgow University (UK), a team of researchers is using a robot together with their AI system to analyze every possible reaction of all the promising compounds thought up by the machine-learning system. To prove this, a set of 72 significant chemical reactions of known products was given to the AI system to train it and let it choose, which reactions it would investigate to predict the untested behavior of newly structured products – with an accuracy of more than 80 percent – and speedily discarding any non-promising materials.
So, time to market and test accuracy are the main advantages when using AI in the lengthy process of drug discovery and in the development of industrial materials. “By realizing only 10 percent of the total number of reactions”, one researcher states, “we can predict the remaining 90 percent without needing to carry out the experiments.”
A replacement of the brain power of human researchers is not on the agenda.