SAM-DTI: A Spatial Attention Model for Drug-Target Interaction Prediction

Drug-target protein interaction (DTI) prediction is an important area of research in drug repurposing and discovery. Laboratory experiments for exploring DTI space need a lot of time and money due to the high volume of potential compounds. The past decades advancements in deep learning have made progress in efficient prediction of DTIs. This paper proposes a spatial attention model for predicting drug-target interactions (SAM-DTI) from labeled SMILES strings and protein sequences. SAM-DTI (i) tries to encapsulate the input features with more relevance to the biological binding of drug-target interactions, rather than models that solely extract features from individual inputs, and (ii) emphasizes giving dynamic importance to each pair of drug-target interactions. The model follows a knowledgeinspired sub-structure vocabulary to preprocess the inputs. It predicts accurate DTI values from the interactions between each SMILES sub-structure and protein substructure pair. SAM-DTI achieves competitive performance compared to state-of-the-art methods.