Predicting Drug-Target Interactions with Deep-Embedding Learning of Graphs and Sequences.

Area of Science:

• Computational chemistry
• Bioinformatics
• Drug discovery

Background:

• Drug-target interactions (DTIs) are crucial for drug discovery.
• Experimental screening for DTIs is costly and time-consuming.
• Computational methods offer an efficient alternative for DTI prediction.

Purpose of the Study:

• To develop an end-to-end deep learning system for predicting DTIs.
• To enhance the accuracy and efficiency of DTI prediction models.
• To provide biological insights into drug-target binding through attention mechanisms.

Main Methods:

• Utilized a graph neural network with an attention mechanism.
• Employed an attentive bidirectional long short-term memory (BiLSTM) network.
• Integrated bidirectional encoder representations from transformers (BERT) for protein feature extraction.
• Applied local breadth-first search (BFS) for molecular subgraph information learning.

Main Results:

• Achieved significant performance improvements in AUC (2.4%) and recall (9.4%) on unbalanced datasets.
• Demonstrated the model's capability to effectively screen potential drug candidates for specific proteins.
• Attention weight visualization provided interpretable biological insights.

Conclusions:

• The proposed end-to-end representation learning model offers a powerful tool for DTI prediction.
• This computational approach accelerates the drug discovery pipeline.
• The model's interpretability aids in understanding drug-target relationships.