Researchers are beginning to harness the power of artificial intelligence (AI) to help them discover and develop new materials and solutions in chemistry and other scientific fields.
|Researchers have designed a machine learning algorithm that predicts the outcome of chemical reactions with much higher accuracy than trained chemists and suggests ways to make complex molecules, removing a significant hurdle in drug discovery. (Image source: University of Cambridge)|
As part of this trend, a team at the University of Cambridge has developed an algorithm that can predict the outcome of complex chemical reactions with more than 90 percent accuracy to aid in the discovery and behavioral prediction of new medications, they said. The algorithm is a key part of the technology platform for drug discovery that researchers at the university have developed.There have been three key challenges to developing new medications that scientists historically have faced—to design them, to make them, and to test them, Alpha Lee, a researcher in Cambridge’s Cavendish Laboratory, told Design News.
We need to predict which molecules are likely to be active against the desired target, find ways to make those molecules in the lab, and judiciously design experiments to find the best molecules quickly and cheaply,” he told us.
While there has been much progress to meet the first and third aspects of medication creation, finding ways to make the molecules is still “a time-consuming challenge,” Lee acknowledged to Design News.
“AI might be able to come up with interesting molecules, but realizing those designs in the lab is difficult,” he told us.
How to Make ‘Making’ Easier
This is why the most recent work of the team focused on the “make stage” of developing new medications, Lee told Design News. Specifically, the algorithm his team developed can accurately predict both the outcomes of chemical reactions and how to make complex molecules from simpler building blocks.
“The accuracy of our model is beyond what one would expect from an experienced chemist,” Lee told Design News.
Calling the platform developed by his team a “GPS for chemistry,” Lee said the algorithm uses tools in pattern recognition to recognize how chemical groups in molecules react by training the model on millions of reactions published in patents.
“We consider chemical-reaction prediction as a machine translation problem. Molecules are represented as texts,” Lee told Design News. “The reacting molecules are considered as one language, while the product is considered as a different language. The model then uses the patterns in the text to learn how to translate between the two languages.”
To Know or Not to Know
The result is a model that provides chemists with a chemical “map”–in the form of a sequence of chemical reactions–that leads to the desired product starting from easily purchasable building blocks, Lee told us.
“Importantly, our model can accurately estimate the confidence of its own prediction–it knows what it knows, but also knows what it doesn’t know,” he told Design News. “As experiments are time-consuming, accurate prediction of confidence is crucial to avoid pursuing expensive pathways that eventually end in failure.”
Researchers published two papers on their recent work—one in the journal ACS Central Science and another in the journal Chemical Communications.
Lee and his team believe the technology platform they developed “will significantly reduce the time and cost of drug discovery,” he told Design News. The technology also is applicable to other areas of chemistry, such as the whole organic molecules industry, including agrochemicals, organic semiconductors, polymers, fragrances, and other areas of research, Lee told us.
Ultimately, Lee hopes to achieve what he calls a “molecular industrial revolution,” accelerating the cycle of designing new chemicals and materials from concept to completion by combining AI with physics and chemistry, he told Design News
“I believe the synergy between machine learning, physics, and chemistry will change the way we think about designing functional molecules, from an artisanal craft to a systematic approach,” Lee told us.