DeepMCL-DTI: predicting drug-target interactions using multi-channel deep learning with attention mechanism

Area of Science:

• Computational Biology
• Pharmacology
• Artificial Intelligence

Background:

• Drug-target interaction (DTI) prediction is vital for accelerating drug discovery and development.
• Current deep learning methods often use two-channel architectures, limiting feature extraction from original input data.
• There is a need for advanced models that can more effectively learn drug and target features for improved DTI prediction.

Purpose of the Study:

• To introduce DeepMCL-DTI, an attention-based multi-channel deep learning model for enhanced DTI prediction.
• To address the limitations of existing two-channel architectures in fully leveraging input data.
• To improve the accuracy and efficiency of identifying potential drug candidates.

Main Methods:

• Developed DeepMCL-DTI, featuring four distinct feature extraction channels: Graph Sample and Aggregate (GSA) and convolutional neural network (CNN) for drugs, and ProtBert and bidirectional convolutional long short-term memory (BiLSTM) for proteins.
• Incorporated an interact-attention module to model DTI across spatial and channel dimensions.
• Validated the model on the DrugBank and Davis datasets.

Main Results:

• DeepMCL-DTI demonstrated superior performance compared to state-of-the-art methods on benchmark datasets.
• The multi-channel approach effectively captured complex features from both drug and target data.
• A case study on angiotensin-converting enzyme 2 receptor validated its utility in drug discovery pre-screening.

Conclusions:

• DeepMCL-DTI represents a significant advancement in deep learning for DTI prediction.
• The proposed model architecture effectively overcomes limitations of previous methods.
• DeepMCL-DTI shows promise as a valuable tool for accelerating the early stages of drug discovery.