UnbiasedDTI: Mitigating Real-World Bias of Drug-Target Interaction Prediction by Using Deep Ensemble-Balanced...

Area of Science:

• Computational chemistry and cheminformatics
• Bioinformatics and computational biology
• Machine learning in drug discovery

Background:

• In vitro drug-target interaction (DTI) studies are costly and slow.
• Computational DTI prediction offers efficiency but faces challenges.
• Class imbalance, with far more negative than positive interactions, biases predictive models.

Purpose of the Study:

• To develop a computational framework for DTI prediction that addresses class imbalance.
• To evaluate the impact of data balancing on DTI prediction performance.
• To validate computational predictions through experimental methods.

Main Methods:

• An ensemble of deep learning models was employed for DTI prediction.
• Data balancing techniques were specifically applied to mitigate class imbalance.
• Model performance was assessed computationally and experimentally using the BindingDB dataset.

Main Results:

• The proposed balanced deep learning model significantly outperformed unbalanced models.
• Balancing the dataset reduced bias towards the majority negative class, enhancing predictive accuracy.
• Experimental validation confirmed the superior performance of the balanced computational approach.

Conclusions:

• Data balancing is critical for improving the accuracy of computational DTI prediction.
• Deep learning models, when properly balanced, provide a powerful tool for drug discovery.
• Experimental validation is essential for credible DTI prediction, especially with unbalanced datasets.