Predicting drug-target interactions using machine learning with improved data balancing and feature engineering.

Area of Science:

• Computational drug discovery
• Bioinformatics
• Machine Learning

Background:

• Drug-Target Interaction (DTI) prediction is crucial but challenged by data imbalance and complex biochemical representations.
• Accurate DTI prediction accelerates therapeutic development and pharmaceutical research.

Purpose of the Study:

• To develop a novel hybrid framework for enhanced Drug-Target Interaction (DTI) prediction.
• To address data imbalance and improve the accuracy and sensitivity of DTI prediction models.

Main Methods:

• A hybrid framework combining machine learning (ML) and deep learning (DL) was developed.
• Feature engineering utilized MACCS keys for drug structures and amino acid/dipeptide compositions for target properties.
• Generative Adversarial Networks (GANs) were employed to balance imbalanced datasets, and Random Forest Classifier (RFC) was used for prediction.

Main Results:

• The GAN+RFC model demonstrated high performance across BindingDB-Kd, BindingDB-Ki, and BindingDB-IC50 datasets.
• Achieved up to 97.46% accuracy, 97.49% precision, and 99.42% ROC-AUC on BindingDB-Kd.
• Showcased significant improvements in sensitivity and reduced false negatives due to GANs.

Conclusions:

• The proposed GAN-based hybrid framework significantly enhances DTI prediction accuracy and robustness.
• This approach sets a new benchmark in computational drug discovery by effectively handling data imbalance and complex features.
• The framework's scalability and generalizability offer substantial contributions to therapeutic development.