Chemists want to build a map of every possible molecule. This could accelerate the discovery of new compounds for everything from drugs to materials, but determining the viability and properties of so many molecules is a huge challenge 鈥� one that artificial intelligence may just be able to help with.
The idea for the map would be that nearby molecules have similar properties. For example, compounds that respond to heat or to other chemicals in similar ways would be close together.
Different properties would correspond to different dimensions in the map. This means that if researchers knew of, say, a catalyst that contains a toxic ingredient, they could move along the toxicity dimension to find a safer chemical that work just as well.
Advertisement
At the recent meeting in San Francisco, California, researchers suggested that AI may now be good enough to help overcome the vast task of putting such a map together.
鈥淚f we know the right questions to ask of AI, it鈥檚 possible to think about coming up with a discovery of a new molecule in days or weeks that would have normally taken decades,鈥� said at the Massachusetts Institute of Technology in her keynote address to the conference.
Some maps already exist for this purpose, such as a smaller than 17 atoms or , which has nearly 50,000 drugs and environmentally problematic compounds organised into a three-dimensional grid of chemical properties.
However, these are very small in comparison to a map of every molecule. For just small, drug-like compounds, for instance, there could be as many as 1060 聽molecules, each with multiple properties relevant for creating them.
Ongoing efforts to build such a map focus on taking a typical molecule, like a well-understood catalyst, then asking AI to explore the stability of each compound made by swapping a few of its atoms for another element. This fills out the surrounding of the original molecule in one plane of the map, populating its neighbourhood.
Kulik studies materials that can be used for solar energy storage or to accelerate the conversion of greenhouse gases into easier-to-deal-with liquids, and she has successfully used AI to speed up the discovery of the most useful compounds in each category among millions of possibilities.
At ACS she pointed to the success of AlphaFold, an AI that predicted the structure of almost every protein on the planet, as evidence that something similar may soon be possible for chemistry.
However, building a chemical map will be more difficult because some data that is needed isn鈥檛 easily accessible, said at the University of Southampton during his ACS keynote address. 鈥淧rocedures that did not work or compounds that were not published, it鈥檚 all sitting in somebody鈥檚 thesis or lab notebook,鈥� he says.
