Planning chemical syntheses with deep neural networks and symbolic AI

• Published in: Nature (London) Vol. 555; no. 7698; pp. 604 - 610
• Main Authors: Segler, Marwin H. S., Preuss, Mike, Waller, Mark P.
• Format: Journal Article
• Language: English
• Published: London Nature Publishing Group UK 29.03.2018; Nature Publishing Group
• Subjects: 639/638/403/933; 639/638/549/973; 639/638/630; 639/705/117; Artificial intelligence; Artificial neural networks; Automation; Chemical reactions; Chemists; Computer networks; Computer simulation; Design; Heuristic; Humanities and Social Sciences; Methods; multidisciplinary; Neural networks; Neural stem cells; Organic chemistry; Organic compound synthesis; Planning; Problem solving; Science; Searching
• ISSN: 0028-0836; 1476-4687
• DOI: 10.1038/nature25978

To plan the syntheses of small organic molecules, chemists use retrosynthesis, a problem-solving technique in which target molecules are recursively transformed into increasingly simpler precursors. Computer-aided retrosynthesis would be a valuable tool but at present it is slow and provides results of unsatisfactory quality. Here we use Monte Carlo tree search and symbolic artificial intelligence (AI) to discover retrosynthetic routes. We combined Monte Carlo tree search with an expansion policy network that guides the search, and a filter network to pre-select the most promising retrosynthetic steps. These deep neural networks were trained on essentially all reactions ever published in organic chemistry. Our system solves for almost twice as many molecules, thirty times faster than the traditional computer-aided search method, which is based on extracted rules and hand-designed heuristics. In a double-blind AB test, chemists on average considered our computer-generated routes to be equivalent to reported literature routes. Deep neural networks and Monte Carlo tree search can plan chemical syntheses by training models on a huge database of published reactions; their predicted synthetic routes cannot be distinguished from those a human chemist would design. Computers teach themselves to make molecules Chemical reaction databases that are automatically filled from the literature have made the planning of chemical syntheses, whereby target molecules are broken down into smaller and smaller building blocks, vastly easier over the past few decades. However, humans must still search these databases manually to find the best way to make a molecule. This involves many steps and choices. Some degree of automation has been achieved by encoding 'rules' of synthesis into computer programs, but this is time consuming owing to the numerous rules and subtleties involved. Here, Mark Waller and colleagues apply deep neural networks to plan chemical syntheses. They trained an algorithm on essentially every reaction published before 2015 so that it could learn the 'rules' itself and then predict synthetic routes to various small molecules not included in the training set. In blind testing, trained chemists could not distinguish between the solutions found by the algorithm and those taken from the literature.