Identification and analysis of driver missense mutations using rotation forest with feature selection

Area of Science:

• Genomics
• Bioinformatics
• Computational Biology

Background:

• Identifying cancer-associated mutations, or driver mutations, is crucial for understanding cancer genome function, including oncogene activation and tumor suppressor gene inactivation.
• Existing supervised machine learning approaches for driver mutation prediction often lack clarity on the importance of specific features.
• This highlights a need for effective feature selection methods in cancer genomics research.

Purpose of the Study:

• To propose and evaluate a novel feature selection method, termed DX, for identifying important features in cancer driver mutation prediction.
• To assess the performance of the DX method using the rotation forest algorithm.
• To demonstrate the generalizability of the DX method across different cancer datasets.

Main Methods:

• A novel feature selection method (DX) was developed using a set of 126 candidate features.
• The rotation forest algorithm was employed for experimental validation.
• The method was trained and tested on datasets from the COSMIC and Swiss-Prot databases.

Main Results:

• The DX method, utilizing the 11 top-ranked features, achieved high prediction performance: 88.03% accuracy, 93.9% precision, and 81.35% recall.
• The selected features proved effective in identifying cancer-associated mutations.
• Comparative analysis on TP53, EGFR, and Cosmic2plus datasets confirmed the method's generality.

Conclusions:

• The DX feature selection method is effective for identifying critical features in cancer driver mutation prediction.
• The rotation forest algorithm combined with DX yields high predictive performance.
• The proposed method demonstrates broad applicability across various cancer-related datasets.