Motivation: Drug-target binding affinity (DTA) reflects the strength of the drug-target interaction; therefore, predicting the DTA can considerably benefit drug discovery by narrowing the search space and pruning drug-target (DT) pairs with low binding affinity scores. Representation learning using deep neural networks has achieved promising performance compared with traditional machine learning methods; hence, extensive research efforts have been made in learning the feature representation of proteins and compounds. However, such feature representation learning relies on a large-scale labelled dataset, which is not always available.
Results: We present an end-to-end deep learning framework, ELECTRA-DTA, to predict the binding affinity of drug-target pairs. This framework incorporates an unsupervised learning mechanism to train two ELECTRA-based contextual embedding models, one for protein amino acids and the other for compound SMILES string encoding. In addition, ELECTRA-DTA leverages a squeeze-and-excitation (SE) convolutional neural network block stacked over three fully connected layers to further capture the sequential and spatial features of the protein sequence and SMILES for the DTA regression task. Experimental evaluations show that ELECTRA-DTA outperforms various state-of-the-art DTA prediction models, especially with the challenging, interaction-sparse BindingDB dataset. In target selection and drug repurposing for COVID-19, ELECTRA-DTA also offers competitive performance, suggesting its potential in speeding drug discovery and generalizability for other compound- or protein-related computational tasks.