Fig. 1.

Overall framework. The main steps are as follows. First, a feature extractor network to map inputs $x:=(x^1,x^2)\in \mathcal{X}$ to a latent space. Then, each pair of drugs is represented as a feature vector by concatenating the corresponding latent features of the drugs. Last, the feature vectors representing the drug pairs are fed into a deep neural network to train the predictor to uncover potential DDIs