<p>Biomedical datasets with a High-Dimensional Low-Sample-Size (HDLSS) structure, such as gene expression and microarray data, contain extensive redundancy and irrelevant information that severely degrade classification performance. Many existing feature selection methods face challenges in reducing dimensionality without sacrificing predictive accuracy. In this paper, a two-stage hybrid filter–wrapper framework is proposed for genomic HDLSS data. In the first stage, a graph-based filter models feature subsets as paths in a fully connected weighted graph, where edges represent relevance and redundancy relationships. By exploiting the properties of matrix power series, the values of paths (i.e., feature subsets) with arbitrary lengths are evaluated through the design of a dynamic threshold for feature filtering. Features are ranked based on the values of the paths in which a given feature participates. In the second stage, a trigonometry-based optimization algorithm called the Tangent–Cotangent Optimization Algorithm (TCOA) and its binary variant (BTCOA), using the absolute value of the cosine function, perform stable multi-objective optimization in a wrapper-based manner. The computational structure of this framework, which involves iterative matrix operations and the simultaneous evaluation of multiple feature subsets, suggests that the proposed approach can be well suited for implementation in High Performance Computing (HPC) and supercomputing environments. Experimental results on 26 high-dimensional biomedical datasets demonstrate that BTCOA consistently outperforms 32 well-established methods in terms of feature selection effectiveness and classification accuracy.</p>