Self-supervised molecular pretraining strategy for reaction prediction in low-resource scenarios

In the face of low-resource reaction training samples, we construct a chemical platform for addressing small-scale reaction prediction problem. By using a self-supervised molecular pretraining strategy, the chemical information from 1 billion molecules can be delivered to small-scale reaction prediction. To demonstrate the broad applicability of our approach, we apply our model to three different name reaction prediction tasks. In the Baeyer-Villiger, Heck and Sharpless asymmetric epoxidation reactions, the accuracies increase by 5.7%, 10.8%, 4.8% on average, respectively.