Multilevel Attention Models for Drug Target Binding Affinity Prediction

Drug-Target Binding Affinity (DTBA) prediction is one class of Drug-Target Interaction problem (DTI), where the focus is to predict the binding strength of a drug-target pair. Several machine learning approaches have been developed for this purpose. However, almost all rely on the use of increasingly sophisticated inputs to improve the obtained results besides that they don’t allow any analysis or interpretation due to their black-box characteristic. This work is an attempt to address these limitations by leveraging the use of attention mechanisms with convolution-deconvolution architecture. In this paper, we define two multilevel attention-based models for DTBA prediction. Our two approaches attempt to get advantage of the attention concept, by probing different abstraction levels of drug-target feature maps. We evaluate the performance of our methods on two benchmark datasets, KIBA and Davis. The results show that both approaches are very effective. Compared to other well-known methods, they achieved excellent results regarding the considered performance metrics, while using merely sequences as inputs and providing a potential way of results interpretation.