Artificial applicability labels for improving policies in retrosynthesis prediction

Automated retrosynthetic planning algorithms are a research area of increasing importance. Automated reaction-template extraction from large datasets, in conjunction with neural-network-enhanced tree-search algorithms, can find plausible routes to target compounds in seconds. However, the current method for training neural networks to predict suitable templates for a given target product leads to many predictions that are not applicable in silico. Most templates in the top 50 suggested templates cannot be applied to the target molecule to perform the virtual reaction. Here, we describe how to generate data and train a neural network policy that predicts whether templates are applicable or not. First, we generate a massive training dataset by applying each retrosynthetic template to each product from our reaction database. Second, we train a neural network to perform near-perfect prediction of the applicability labels on a held-out test set. The trained network is then joined with a policy model trained to predict and prioritize templates using the labels from the original dataset. The combined model was found to outperform the policy model in a route-finding task using 1700 compounds from our internal drug-discovery projects.