Semisupervised approach for dominant gene selection and classification

BACKGROUND: Semisupervised learning has attracted significant interest in gene expression analysis due to its ability to improve classification performance under limited labeled data, a common challenge arising from high dimensionality and small sample sizes. Existing semisupervised methods often rely on manifold learning to propagate label information; however, weak label connections in early learning stages can reduce their effectiveness. RESULTS: To address these limitations, we propose a semisupervised approach for dominant gene selection and classification (SADGSC) that employs feature decomposition to identify essential features while suppressing noise and redundancy. The proposed method adopts hinge loss instead of square loss for label prediction and constructs a projection matrix based on confidence-weighted features, leading to improved classification accuracy and interpretability. Biological analysis of the top-ranked genes (e.g., PGR, KRT14, TOX3, and FGF10) reveals enrichment in hormone signaling epithelial-mesenchymal transition, apoptosis, and oxidative stress pathways, demonstrating the biological relevance of the selected genes. Experimental evaluations conducted on eleven datasets, primarily gene expression datasets, show that SADGSC achieves competitive performance compared with state-of-the-art methods. CONCLUSIONS: The proposed SADGSC framework effectively addresses the challenges of limited labeled data and high dimensionality in gene expression analysis, providing strong classification performance and biologically meaningful gene selection. These findings highlight the potential of semisupervised learning for dominant gene discovery and robust classification in high-dimensional biological datasets.