Comparative Studies on Resampling Techniques in Machine Learning and Deep Learning Models for Drug-Target Interaction...

Area of Science:

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

Background:

• Predicting drug-target interactions (DTIs) is crucial for efficient drug discovery.
• Machine learning (ML) and deep learning (DL) methods are increasingly vital for DTI prediction.
• High-dimensional and imbalanced datasets pose significant challenges for ML/DL model training.

Purpose of the Study:

• To compare various data resampling techniques for addressing class imbalance in DTI prediction.
• To evaluate the effectiveness of deep learning methods in overcoming class imbalance for DTI prediction.
• To analyze DTI prediction performance across ten cancer-related activity classes from the BindingDB dataset.

Main Methods:

• Comparative analysis of data resampling techniques, including Random Undersampling (RUS) and SVM-SMOTE.
• Application of machine learning classifiers (Random Forest, Gaussian Naïve Bayes) and a deep learning model (Multilayer Perceptron).
• Binary classification evaluation of model performance on imbalanced DTI datasets.

Main Results:

• Random Undersampling (RUS) significantly degrades model performance on imbalanced DTI datasets, proving unreliable.
• SVM-SMOTE, when combined with Random Forest and Gaussian Naïve Bayes, achieves high F1 scores for imbalanced activity classes.
• The Multilayer Perceptron deep learning model demonstrates high F1 scores across all activity classes, irrespective of resampling.

Conclusions:

• SVM-SMOTE is a recommended resampling method for ML-based DTI prediction on imbalanced data.
• Deep learning models, particularly Multilayer Perceptron, offer robust performance in DTI prediction without requiring data resampling.
• The findings guide the selection of appropriate methods for building accurate DTI prediction models in drug discovery.