A Leakage-Controlled Benchmark of Metaheuristic Multi-Omics Feature Selection for Reconstructing TCGA Molecular Subtypes in Endometrial Cancer

Background: The four molecular subtypes of endometrial carcinoma defined by The Cancer Genome Atlas (TCGA) can be reconstructed from multi-omics data, but these data contain about 460,000 features for 369 patients, the subtypes are imbalanced, and feature selection easily leaks test-set information. Methods: We developed POISE-X, a leakage-controlled, imbalance-aware benchmarking pipeline with an interchangeable metaheuristic search engine. Using RNA-seq, miRNA, DNA methylation, copy-number, mutation, and clinical data from 369 TCGA-UCEC patients, we compared 19 optimizers under nested five-fold cross-validation, with in-fold prefiltering to 2505 candidate features and a macro-averaged F1 fitness function. Results: Performance differed across optimizers in a Friedman test on fold means (p=1.989×10−6). No selected panel exceeded logistic regression on the full candidate pool, which reached a macro-averaged F1 of 0.902; the compact configuration chosen by our selection rule, POISE-GWO, retained 443 features with a macro-averaged F1 of 0.880, and the selection rule itself scored 0.881 under leave-one-fold-out evaluation, and an ANOVA-ranked panel of 450 features reached 0.886. These results held without clinical variables and, in the two searches tested, for parsimony weights of 0.8–0.99. A surrogate of the clinical molecular classification built from POLE exonuclease-domain mutation, mismatch-repair deficiency, and TP53 mutation reached a macro-averaged F1 of 0.924, higher on average than every multi-omics model. Conclusions: In this cohort, wrapper search therefore compressed the feature set without adding discriminative power, and the selected panels reconstructed an existing classification, not a new treatment-relevant signature. POISE-X is a framework for benchmarking feature-selection strategies; any selected panel requires external validation before clinical use.