A hybrid filter-wrapper gene selection approach based on vertical federated learning using deep neural networks and discrete manta ray foraging optimization algorithm
摘要
DNA microarray technology enables the comprehensive profiling of gene expression, an essential capability for diagnosing multifaceted diseases such as cancer. However, inherent challenges—including high dimensionality, limited sample sizes, and complex inter-gene interactions—significantly impede the attainment of accurate analytical results. To overcome these limitations, this paper proposes an interactive, hybrid gene selection framework that integrates vertical federated learning, a dual filter-wrapper strategy, and deep learning methodologies. At the core of this framework is the Discrete Manta Ray Foraging Optimization (DMRFO) algorithm. Within the federated architecture, the first node executes filter-based feature selection—driven by entropy reduction and mutual information criteria—while the second node performs wrapper-based (cover-based) selection, optimized for minimizing classification error and feature cardinality. The distinct feature subsets generated by each node are subsequently aggregated and rigorously refined through a deep learning kernel embedded within the DMRFO algorithm, ensuring the isolation of an optimal gene subset. Extensive validation across five benchmark gene expression datasets (Leukemia, Lymphoma, MLL, Ovary, and SRBCT) yielded outstanding accuracy scores of