New Deep-Learning Software Knows How To Make Desired Organic Molecules (nature.com)

dryriver shares a report from Nature about a neural network-based, deep-learning software that is as good as trained chemists in figuring out what reagents and reactions may lead to the successful creation of a desired organic molecule: Chemists have a new lab assistant: artificial intelligence. Researchers have developed a "deep learning" computer program that produces blueprints for the sequences of reactions needed to create small organic molecules, such as drug compounds. The pathways that the tool suggests look just as good on paper as those devised by human chemists. The tool, described in Nature on March 28, is not the first software to wield artificial intelligence (AI) instead of human skill and intuition. Yet chemists hail the development as a milestone, saying that it could speed up the process of drug discovery and make organic chemistry more efficient. Chemists have conventionally scoured lists of reactions recorded by others, and drawn on their own intuition to work out a step-by-step pathway to make a particular compound. They usually work backwards, starting with the molecule they want to create and then analyzing which readily available reagents and sequences of reactions could be used to synthesize it -- a process known as retrosynthesis, which can take hours or even days of planning. The new AI tool, developed by Marwin Segler, an organic chemist and artificial-intelligence researcher at the University of Munster in Germany, and his colleagues, uses deep-learning neural networks to imbibe essentially all known single-step organic-chemistry reactions -- about 12.4 million of them. This enables it to predict the chemical reactions that can be used in any single step. The tool repeatedly applies these neural networks in planning a multi-step synthesis, deconstructing the desired molecule until it ends up with the available starting reagents.

Comments

Re:Significant and usefule, but ...

[physicsphairy]

But, one of the things I wonder about is the potential for reduced understanding and insight among the people using it, and where it might lead.

To give you an idea of the present state of chemistry, we only recently measured the energy of a transition state, imaged atoms and molecules, or directly observed hydrogen bonds. New insights into the behavior of water is common reading. As for syntheses, the reaction mechanisms drawn are at best guesses and many times syntheses reasonable in theory are found not to work in practice. Basically, we chemists do not have much fundamental understanding so much as a practical intuition for how chemical systems behave. But improved understanding and classification often go hand-in-hand in science, so I think it likely that the output of these algorithms will actually improve our human-level interpretations.