A Comprehensive Evaluation Framework for Benchmarking Multi-Objective Feature Selection in Omics-Based Biomarker...

Area of Science:

• Bioinformatics
• Computational Biology
• Genomics

Background:

• Machine learning (ML) is used for cancer subtype classification via gene expression biomarkers.
• Current biomarker models lack reproducibility and optimal feature set sizes, hindering clinical translation.

Purpose of the Study:

• To address the multi-objective problem of optimizing biomarker classification performance and feature set size.
• To evaluate ML-driven feature subset selection algorithms for reproducible and cost-effective cancer biomarkers.

Main Methods:

• Applied seven ML feature subset selection algorithms to eight large-scale cancer transcriptome datasets.
• Utilized a benchmark including training and external validation sets with metrics for accuracy, diversity, and gene stability.
• Proposed a new evaluation metric generalizing hypervolume for cross-validation studies.

Main Results:

• Achieved biomarkers with 0.8 balanced accuracy in external datasets for breast, kidney, and ovarian cancers using 4, 2, and 7 features, respectively.
• Genetic algorithms generally outperformed other methods.
• NSGA2-CH and NSGA2-CHS demonstrated superior performance in most cases.

Conclusions:

• Optimized gene expression biomarkers can achieve high accuracy with minimal features, enhancing clinical utility.
• Genetic algorithms, particularly NSGA2-CH and NSGA2-CHS, are effective for developing reproducible cancer biomarkers.
• The proposed evaluation metric aids in assessing multi-objective optimization performance in cross-validation.