Automated synthesis planning has become an important tool in drug discovery. One exciting application is de novo drug discovery, in which automated synthesis planning can be used to prioritize synthetically accessible molecules. However, it can take minutes to plan synthesis for a drug-like molecule on a single CPU. While relatively fast, this can be too slow for applications such as de novo drug discovery. To address this challenge, recent work has focused on speeding up automated synthesis planning software. Bringing down processing time would also more broadly improve user experience of retrosynthesis software. Two approaches have recently emerged. The first idea is to precompute outputs of synthesis planning software for a large number of molecules. We can then train a neural network on such a dataset to rapidly predict outcomes of retrosynthesis software for novel compounds. The second approach is to use reinforcement learning techniques to learn from previously executed plans to prioritize successful choices in future searches, as done by Kim et al. and Schreck et al. In this talk, we will discuss these and other techniques for rapid estimation of synthetic accessibility. To compare them on even grounds, we trained two recently introduced techniques on the same dataset, and evaluated them against a comprehensive set of benchmarks.

Rapid estimation of synthetic accessibility based on retrosynthesis software

Fall 2022 abstracts, posters, and presentations will launch in mid-September.

ACS Fall 2022

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