In the recent years, Deep Learning with neural networks has transformed computer vision, speech processing (built in in all modern smartphones), and many other pattern recognition tasks. Earlier last year, a system based on neural networks has even beaten the world champion in Go, a strategic game far more complex than chess. Remarkably however, smart assistants are basically absent from the chemical lab.<br/><br/>In our work, we combine Deep Learning with the core concept of named reactions, or more generally, reaction types. Our model can learn overnight which are the most likely reactions to make a molecule for retrosynthesis, or under which reaction type the starting materials will likely react (Forward Reaction Prediction). The key point is that we train the machine with essentially the complete published knowledge of organic chemistry (&gt;3.5 million reactions acquired from the Reaxys database). This allows our model to learn reactivity conflicts and the selectivity of transformations from data with an unprecedented precision. Before, this information had to be laboriously encoded by hand.<br/><br/>In the largest scale comparison of reaction prediction systems ever conducted, we show that our approach has a higher than 95% accuracy when allowed to suggest up to 10 different routes for a target molecule on a test set of around 1 million reactions. Furthermore, preliminary studies indicate that coupling the neural networks with modern tree search techniques, it outperforms traditional computational synthesis planning with handcoded heuristics.

Retrosynthesis and reaction prediction with deep neural networks

254th National Meeting (2017)

The official archive and repository for all ACS Meetings & Expositions abstracts and select posters or presentations.