Fig. 3From: Learning self-supervised molecular representations for drug–drug interaction predictionOverview of SMR-DDI. Step 1: The molecules are sampled from ChEMBL22, and SMILES enumeration is applied to generate a randomized view of the molecule. A base encoder network \({\text{e}}(.)\) and a projection head \({\text{g}}(.)\) are trained to maximize the similarity between the canonical and randomized SMILES using the InfoNCE contrastive loss. Step 2: After the training process is completed, the representation \(h\) is transferred for DDI prediction. The latent features of each drug pair are combined to create a vector that is fed into a feed-forward neural network to predict DDIsBack to article page