BiGraph-DTA: Predicting drug-target interactions of hepatoprotective agents with graph convolutional networks

Area of Science:

• Pharmacology and Cheminformatics
• Computational Biology and Drug Discovery

Background:

• Accurate prediction of drug-target affinity (DTA) is essential for identifying effective hepatoprotective agents to combat liver diseases.
• Existing computational methods often struggle to capture the complex interactions between drug molecules and protein targets.

Purpose of the Study:

• To develop and validate BiGraph-DTA, a novel predictive model for DTA scoring of hepatoprotective compounds.
• To enhance the identification of potential drug candidates for liver disease therapy.

Main Methods:

• Utilized a hybrid deep learning architecture combining graph convolutional networks (GCNs) and bidirectional long short-term memory (BiLSTM) networks.
• Processed molecular structures as graphs and protein sequences as sequential data.
• Trained and evaluated the model on a curated dataset of 21,421 hepatoprotective interactions from ChEMBL.

Main Results:

• BiGraph-DTA significantly outperformed traditional machine learning (Random Forest, XGBoost) and existing deep learning models (DeepDTA, GraphDTA).
• Achieved a mean squared error of 0.7885, an R-squared value of 0.7208, and a concordance index of 0.8508.
• Demonstrated superior ability in capturing complex dependencies and interactions for DTA prediction.

Conclusions:

• The BiGraph-DTA model offers a robust, data-driven framework for accelerating the discovery of novel hepatoprotective compounds.
• This approach holds significant potential for expediting the development of new therapeutics for liver diseases.
• Highlights the power of integrating graph and sequential deep learning for complex drug discovery challenges.