Ensemble Feature Selection Compares to Meta-analysis for Breast Cancer Biomarker Identification from Microarray Data

Abstract

Identifying stable and precise biomarkers is a key challenge in precision medicine. A promising approach in this direction is exploring omics data, such as transcriptome generated by microarrays, to discover candidate biomarkers. This, however, involves the fundamental issue of finding the most discriminative features in high-dimensional datasets. We proposed a homogeneous ensemble feature selection (EFS) method to extract candidate biomarkers of breast cancer from microarray datasets. Ensemble diversity is introduced by bootstraps and by the integration of seven microarray studies. As a baseline method, we used the random effect model meta-analysis, a state-of-the-art approach in the integrative analysis of microarrays for biomarkers discovery. We compared five feature selection (FS) methods as base selectors and four algorithms as base classifiers. Our results showed that the variance FS method is the most stable among the tested methods regardless of the classifier and that stability is higher within datasets than across datasets, indicating high sample heterogeneity among studies. The predictive performance of the top 20 genes selected with both approaches was evaluated with six independent microarray studies, and in four of these, we observed a superior performance of our EFS approach as compared to meta-analysis. EFS recall was as high as 85%, and the median F1-scores surpassed 80% for most of our experiments. We conclude that homogeneous EFS is a promising methodology for candidate biomarkers identification, demonstrating stability and predictive performance as satisfactory as the statistical reference method.