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

Area of Science:

• Computational chemistry
• Bioinformatics
• Drug discovery

Background:

• Drug-Target Affinity (DTA) prediction is crucial for efficient drug discovery.
• Current deep learning methods (CNN-LSTM, GNNs) face challenges with accuracy, efficiency, and computational resources.
• Existing serial CNN-LSTM models lose raw order information, limiting long-range dependency capture.

Purpose of the Study:

• To develop a more efficient and accurate deep learning model for DTA prediction.
• To address the limitations of existing methods in capturing both local patterns and global dependencies.
• To provide a scalable and sustainable solution for modern drug discovery.

Main Methods:

• Proposed EDTA (Efficient Deep Learning for Drug-Target Affinity prediction) architecture.
• Utilized a parallel combination of CNNs and bidirectional LSTM (BiLSTM).
• Incorporated an attention mechanism to simultaneously capture local and global dependencies.

Main Results:

• Achieved high performance on benchmark datasets: r_m² of 0.783 (Davis) and 0.787 (KIBA).
• Outperformed state-of-the-art DTA methods in accuracy.
• Demonstrated significant improvements in efficiency: fewer parameters, less memory, and up to five-fold faster inference.
• Successfully distinguished binders from decoys in virtual screening on the DUD-E dataset.

Conclusions:

• EDTA offers a practical and effective solution for DTA prediction, balancing accuracy and efficiency.
• The parallel CNN-BiLSTM architecture with attention captures essential interactions without computational burden.
• EDTA represents a significant advancement for scalable and sustainable drug discovery.