Self-supervised molecular pretraining strategy for low-resource reaction prediction 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 pretraining strategy called MASS, the transformer model can absorb the chemical information about 1 billion molecules and then finetunes on small-scale reaction prediction, which is different from previous works that only rely on reaction samples. To demonstrate the broad applicability of our approach, we adopt three dif-ferent name reactions in our work. In the Baeyer-Villiger, Heck and Sharpless asymmetric epoxidation reaction prediction tasks, the average accuracies increase by 5.7%, 10.8%, 4.8% respectively, marking an important step to low-resource reaction prediction.