Recent advances in high-throughput DNA sequencing have enabled the generation of extensive microarray datasets from tissue samples, facilitating the identification of disease-specific biomarkers. The high dimensionality of these datasets, characterised by a large number of genomic features and limited sample sizes, poses significant challenges for conventional feature selection methods in identifying robust, biologically relevant biomarkers within practical computational timeframes. To address these limitations, this study proposes a hybrid computational framework that integrates Kernel Principal Component Analysis (KPCA) with an enhanced Gravitational Search Algorithm (GSA), termed OBKGSA. KPCA is initially applied to extract biologically meaningful, nonlinearly separable gene subsets from high-dimensional expression data. Subsequently, opposition-based learning (OBL) is integrated into the GSA, yielding the OBKGSA algorithm that improves population diversity and convergence efficiency during the search for optimal biomarker combinations. The proposed OBKGSA method was validated on six publicly available microarray cancer datasets and benchmarked against four established nature-inspired classification approaches. Experimental results demonstrate that OBKGSA consistently achieves superior classification accuracy using minimal feature subsets, outperforming existing methods in cancer identification and classification. Specifically, the model achieved good accuracies of 98.80% using only 10 optimal genes on the SRBCT dataset and 97.89% using 9 genes on the Lung Cancer dataset. • A hybrid dimensionality reduction and metaheuristic framework that balances biological relevance and computational efficiency. • The OBKGSA algorithm, which leverages opposition-based learning to enhance population diversity and avoid premature convergence in feature selection. • Comprehensive experimental validation across six benchmark microarray cancer datasets using four nature-inspired classifiers, demonstrating that OBKGSA consistently outperforms existing state-of-the-art metaheuristic approaches in terms of classification accuracy and robustness. • Outperforms existing feature selection methods in both efficiency and predictive power.
Shukla et al. (Sun,) studied this question.