Nested Evaluation of Sparse Multi-Objective Evolutionary Gene-Panel Discovery for Diabetic Nephropathy

Background: Diabetic nephropathy gene panels come from small kidney transcriptome cohorts, where screening and adaptive search inflate performance. Evolutionary multi-objective search balances panel size, accuracy, and biological coherence, but dense initialization slows it, and its benefit over ranking or regularized regression is rarely tested under nested evaluation. Methods: Six evolutionary algorithms optimizing leanness, inner balanced accuracy, and Gene Ontology (GO) coherence were evaluated with four repeats of nested five-fold cross-validation on GSE96804. Harmonic initialization was compared with Bernoulli, bounded uniform, and SparseEA initializers; GO was removed or randomized. Comparators were random, top-ranked differentially expressed gene (top-DEG), regularized, and wrapper panels and a permuted-label null. Panels were transferred to GSE30528 and GSE142025. Results: Panels contained 2.10–2.70 genes. Harmonic initialization improved hypervolume over Bernoulli and matched existing sparse initializers. Outer area under the curve (AUC) was 0.973–0.991, comparable to top-DEG and regularized selectors; permuted labels gave 0.510. Real GO raised maximum archive coherence from 0.044 to 0.960, and a GO-aware rule raised representative coherence from 0.098 to 0.620 at equal AUC. External AUC was 0.612–0.726 and 0.849–0.964. Conclusions: Harmonic initialization yields compact Pareto fronts without a size cap, GO guidance supplies coherent panels at equal discrimination, and regularized selection gives the most reproducible panel.