Quadratic Unconstrained Binary Optimization With Simulated Annealing for Combinatorial Genomic Feature Selection in Gynecologic Cancer

BackgroundCombinations of somatic alterations can carry information that is not evident from any one variant, and a discovery procedure is more useful if it recovers relevant features without flooding later validation with large candidate sets.We evaluated a pairwise-aware Quadratic Unconstrained Binary Optimization (QUBO) formulation, solved by classical simulated annealing, for genomic feature selection in gynecologic cancer. MethodologyThe retrospective cohort comprised 183 patients with somatic tumor profiling, represented by 34 gene-byvariant classification indicators.High tumor grade was available for 174 patients, including 101 with grade 3 tumors.QUBO was compared with LASSO, Elastic Net, and XGBoost.On the clinical outcome, each method's selected feature count was compared with the distribution obtained from 100 label permutations, with p-values reported using the add-one Monte Carlo convention.Because the clinical cohort has no independently established true feature set, we planted a two-feature joint-risk signal on a modified copy of the clinical matrix and scored recovery of the two designated columns by recall, precision, and F1 over 50 replications per signal setting.We also recorded how often those designated columns were still selected when the planted combination carried no extra risk (target-wise null selection rate).Larger datasets were generated by resampling the original matrix. ResultsQUBO achieved higher feature-level F1 than LASSO in each of the four tabulated synthetic settings: 0.40 versus 0.24, 0.64 versus 0.41, 0.76 versus 0.65, and 0.79 versus 0.62.Under the null construction, the targetwise selection rate was 0.10 for QUBO compared with 0.57 for LASSO, 0.68 for Elastic Net, and 0.75 for XGBoost, with mean selected feature counts of 2.0 versus 15.4, 14.5, and 14.1.In the clinical application, QUBO selected nine features compared with a permutation mean of 1.4 (p = 0.010), including features in ARID1A, PTEN, PIK3CA, TP53, and KRAS.A co-occurrence support rule expressed as a fraction of cohort size, rather than a fixed count of 10, was required to preserve this compact null behavior at resampled sizes up to 1,000. ConclusionsUnder the tested configurations, QUBO recovered the two constituents of a planted joint-risk pattern with higher F1 than LASSO, selected the designated columns less often under a null outcome, and produced a nine-feature clinical set concentrated in established gynecologic cancer genes.The study is a clinically motivated methods evaluation.It does not report a validated predictive biomarker, conventional false discovery rate control, or a quantum hardware advantage.