AI for Molecular Design
Machine-learning algorithms are speeding up the search for novel drugs and materials
By Jeff Carbeck
Want to design a new material for solar energy, a drug to fight cancer or a compound that stops a virus from attacking a crop? First, you must tackle two challenges: finding the right chemical structure for the substance and determining which chemical reactions will link up the right atoms into the desired molecules or combinations of molecules.
Traditionally answers have come from sophisticated guesswork aided by serendipity. The process is extremely time-consuming and involves many failed attempts. A synthesis plan, for instance, can have hundreds of individual steps, many of which will produce undesired side reactions or by-products or simply not work at all. Now, though, artificial intelligence is starting to increase the efficiency of both design and synthesis, making the enterprise faster, easier and cheaper while reducing chemical waste.
In AI, machine-learning algorithms analyze all known past experiments that have attempted to discover and synthesize the substances of interest—those that worked and, importantly, those that failed. Based on the patterns they discern, the algorithms predict the structures of potentially useful new molecules and possible ways of manufacturing them. No single machine-learning tool can do all this at the push of a button, but AI technologies are moving rapidly into the real-world design of drug molecules and materials.
An AI tool developed by researchers at the University of Münster in Germany, for example, repeatedly simulates the 12.4 million known single-step chemical reactions to come up with a multistep synthetic route—planning it 30 times faster than humans do.
In the pharmaceutical arena, an AI-based technology called generative machine learning is also exciting. Most pharmaceutical companies store millions of compounds and screen them for the potential to serve as new drugs. But even with robotics and lab-automation tools, this screening process is slow and yields relatively few hits. Further, the “libraries” collectively include only a tiny fraction of the more than 1030theoretically possible molecules. Using a data set describing the chemical structures of known drugs (and drug candidates), as well as their properties, machine-learning tools can construct virtual libraries of new compounds that have similar, and potentially more useful, characteristics. This capability is starting to dramatically accelerate the identification of drug leads.
Close to 100 start-ups are already exploring AI for drug discovery. Among them are Insilico Medicine, Kebotix and BenevolentAI; the last recently raised $115 million to extend its AI technology to the discovery of drugs for motor neuron disease, Parkinson’s and other hard-to-treat disorders. BenevolentAI is applying artificial intelligence to the entire drug development process—from the discovery of new molecules to the design and analysis of clinical trials meant to demonstrate safety and effectiveness in humans.
In the field of materials, ventures such as Citrine Informatics are using approaches similar to those of pharmaceutical makers and are partnering with large companies, including BASF and Panasonic, to speed innovation. The U.S. government is also supporting research into AI-enabled design. Since 2011 it has invested more than $250 million in the Materials Genome Initiative, which is establishing an infrastructure that includes AI and other computing approaches to accelerate the development of advanced materials.
Past experience teaches that new materials and chemicals can pose unforeseen risks to health and safety. Fortunately, AI approaches should be able to anticipate and reduce these undesirable outcomes. The technologies seem poised to markedly increase the speed and efficacy with which novel molecules and materials are discovered and brought to the market—where they may provide such benefits as improved health care and agriculture, greater conservation of resources, and enhanced production and storage of renewable energy.
Source: Scientific American