Multigene genetic programming for cancer biomarker discovery in DNA microarrays

Abstract

The discovery of cancer biomarkers based on deoxyribonucleic acid (DNA) microarray data is hindered by dimensionality, with the number of genes significantly exceeding the available tissue samples, leading to overfitted classification models. This study proposes a multigene genetic programming (MGGP) method for gene selection that automatically evolves tree‐based models to identify minimal discriminative gene subsets from microarray data with high dimensionality. By integrating an inter‐gene correlation penalty into a multi‐objective fitness function, the framework optimized predictive power while improving classifier generalization to unseen data. We validated the MGGP gene selection approach on four benchmark cancer datasets: colon cancer, prostate tumor, leukemia, and lymphoma, demonstrating a considerable reduction in dimensionality, low root‐mean‐square error (RMSE), high values, and statistically significant gene selection frequencies (permutation test ). These results confirm that MGGP provides an effective evolutionary framework for robust biomarker discovery using high‐dimensional genomic data.