Fig. 3. Framework of molecular interaction graph neural network (MIGNN) and the design of the interaction module.

The MIGNN (molecular interaction graph neural network) first processes the SEMGs (steric- and electronics-embedded molecular graph) of the reaction components via the attention, linear, convolution, max pooling, and flatten layers, which leads to a one-dimensional reaction vector. Subsequently, the reaction vector is duplicated into two copies. One copy undergoes an interaction module that allows the information of each reaction component to interact with those of the other components, resulting in an interaction vector. This interaction vector and the other copy of the reaction vector are concatenated as the final one-dimensional vector, which passes the attention and linear layers to provide the prediction value.