Predictive Performance is Often Insensitive to Feature Selection in High-Dimensional Biological Classification

Feature selection (FS) is commonly assumed to improve predictive performance and identify meaningful features in high-dimensional data. We systematically evaluated this assumption across $30$ classification benchmarks, primarily drawn from computational genomics, including microarray, bulk RNA-Seq, mass spectrometry, and imaging datasets. Across these datasets, we observe that small random subsets of features (0.02--1.0\% of available features) frequently achieve predictive performance that is comparable to, and in some cases statistically indistinguishable from, models trained on full feature set. Notably, performance variability across random subsets of a given size is often low, suggesting substantial redundancy in the predictive signal. Together, these results suggest that, for many widely used high-dimensional biological benchmarks, predictive accuracy alone is not sufficient to justify claims about the importance of specific selected features. They also underscore the need for rigorous validation before interpreting selected features as biologically meaningful or actionable, particularly in computational genomics.