A Quick Policy to Filter Reactions Based on Feasibility in AI-Guided Retrosynthetic Planning

A neural network-guided Monte Carlo tree search (MCTS) has been shown to be a promising algorithm

for computer-aided synthesis prediction. Here we train and investigate a filter policy that removes

unfeasible reactions from the search. We investigate three different methods to generate negative

data that the filter policy model can be trained on, and we show that these methods are

complementary to each other. Therefore the most robust model is one that combines all generated

negative data. The filter in itself is quick (< 0.1 s per prediction on average) but using the filter policy

in the MCTS results in a doubling of the search time. However it only leads to a small reduction in the

success rate of finding synthetic routes (< 1%) and we are able show that using filter policy the MCTS

algorithm produce more promising routes, although the predicted routes are more complex. The filter

policy has been integrated in the AiZynthFinder software.

Abstract