Algorithm-Level Data-Guided Correction for Class Imbalance in Biological Machine Learning Predictions: Protein Interactions as a Case

Ebrahim  Barzegari; Parviz  Abdolmaleki

doi:10.18502/jbe.v11i3.21374

Ebrahim Barzegari Medical Biology Research Center, Health Technology Institute, Kermanshah University of Medical Sciences, Kermanshah, Iran.
Parviz Abdolmaleki Department of Biophysics, Faculty of Biological Sciences, Tarbiat Modares University, Tehran, Iran.

DOI: https://doi.org/10.18502/jbe.v11i3.21374

Keywords: Machine learning; Bioinformatics; Statistical bias; Random forests; Protein-protein Interaction domains

Abstract

Introduction: In real-world biomedical applications of data mining, machine learning and artificial intelligence, there are situations where the widespread problem of class imbalance cannot be addressed by data-level methods such as over- or under-sampling. Correct and efficient use of algorithm-level methods, on the other hand, needs paying heed to data structure and content. This study aims to devise and examine simple methods for addressing the imbalanced class distribution issue in predicting the protein-protein interaction (PPI) sites in membrane proteins as a biomedical case experiment.

Methods: Using an adopted dataset of membrane protein complexes and a retrieved validation set, a class-weighted random forests (CWRF) classifier model was built for predicting interfacial residues from positional frequencies and an evolutionary index.

Results: Among several class weighting methods, a data imbalance-emulating weighting method for the CWRF model achieved an area under the receiver operating characteristics curve (AUC) of 0.815 (95% CI: 0.805-0.823) in the independent test prediction and 0.802 (95% CI: 0.794-0.809) in the prediction for the external validation set, which outperformed previous similar studies. A case prediction confirmed the practical utility of this method.

Conclusion: The proposed approach implies potential applications in other fields of biomedicine and beyond. It also highlights the role of algorithm-data interplay in addressing the class imbalance.