AiZynthFinder: Developments and Learnings from Three Years of Industrial Application
Samuel Genheden
AstraZeneca R&D, Gothenburg, Vastra Gotaland County, Sweden
AiZynthFinder is a software package for AI-assisted retrosynthesis analysis that saw its initial release three years ago. Initially, it supported a Monte Carlo tree search guided by a reaction template-recommendation policy to predict synthetic routes. However, during the three years since its initial release the codebase has undergone substantial development, chiefly to support an industrial synthesis platform. In this talk, I will give an outline of these developments, presenting for instance work on filter policies, generic expansion policies, and comparisons of tree search algorithms. I will focus on recent work on retraining models and selecting the best tree search hyperparameters. Finally, I will reflect on the application of this software in an industrial setting and what we have learned from this. I will discuss large-scale predictive power, reaction space utilization, and features of predicted routes. To conclude, I will outline some grand challenges for AI-assisted retrosynthesis analysis.
