Fig. 2. Overview of the proposed MGraphDTA. The MGNN and MCNN were used to extract multiscale features of the input drug graph and protein sequence, respectively. The output multiscale features of the two encoders were fused respectively and then concatenated to obtain a combined representation of the drug–target pair. Finally, the combined representation was fed into a MLP to predict binding affinity. The Grad-AAM uses the gradient information flowing into the last graph convolutional layer of MGNN to understand the importance of each neuron for a decision of affinity.
