Efficient drug–target affinity prediction via interaction features and parallel CNN–BiLSTM with attention

Drug-Target Affinity (DTA) prediction is critical for reducing failure rates in drug discovery, but existing deep learning methods often trade efficiency for accuracy. Existing CNN-LSTM methods for DTA have convolutional neural networks (CNNs) and long short-term memory (LSTM) in series. CNNs capture local patterns, while LSTMs model long-range dependencies. In series CNN-LSTM architectures, sequential dependencies are only modelled after convolutional compression, leading to loss of raw order information and limiting long-range interaction capture. Existing graph neural networks (GNNs) for DTA, on the other hand, capture structural interactions more explicitly but require large numbers of parameters, long training times, and high computational resources. To address the challenges, we propose EDTA (Efficient Deep Learning for Drug-Target Affinity prediction), a lighter architecture that combines CNNs and bidirectional LSTM (BiLSTM) in parallel and also uses attention mechanism to simultaneously capture both local structural patterns and global sequential dependencies. This design ensures that important interactions are exploited without the computational overhead. On benchmark datasets, EDTA achieves r_m² values of 0.783 (Davis) and 0.787 (KIBA), outperforming state-of-the-art DTA methods while using fewer parameters, less memory, and up to five-fold faster inference. A virtual screening experiment on the Database of Useful Decoys: Enhanced (DUD-E) dataset further confirms its effectiveness in distinguishing binders from decoys. By emphasizing both efficiency and strong r_m² performance, EDTA demonstrates that accurate DTA prediction can be achieved without sacrificing scalability or sustainability, offering a more practical solution for modern drug discovery.