Speaker
Description
High-order combinations of genetic and environmental factors may reveal risk-associated subgroups in oral malignant disorders, but their exhaustive evaluation grows combinatorially, while heuristic methods provide no guarantee of global optimality. We formulate corrected-odds-ratio barcode discovery as a feature-level fractional quadratic unconstrained binary optimization (QUBO) problem. Binary variables encode selected factor states, profile-match auxiliary variables identify participants satisfying the complete barcode, and exactly-k and one-state-per-factor constraints enforce biologically valid solutions. The Haldane–Anscombe corrected odds ratio is optimized through Dinkelbach iterations.
The framework was evaluated using a public Taiwanese cohort of 576 participants, comprising 242 patients with oral or pharyngeal cancer, 70 patients with oral potentially malignant disorders (OPMD), and 264 controls. The search space included seven CYP26 single-nucleotide polymorphisms and seven demographic or environmental factors. For barcode sizes k=2,…,7, global-best and sequential nested-best solutions were obtained for cancer versus control, OPMD versus control, and cancer versus OPMD. Exhaustive enumeration, greedy search, binary particle swarm optimization, and simulated annealing were used for validation.
Across 36 global and nested scenarios, the minimum-energy QUBO solutions exactly matched the exhaustive corrected-OR optima. The generated models contained up to 2,119 binary variables and 55,613 nonzero quadratic coefficients. Global and nested paths coincided for cancer-versus-control and OPMD-versus-control, but diverged from k=3 in the cross-sectional cancer-versus-OPMD analysis. In this contrast, the global corrected OR reached 15.81 for k=4–6, with 24 cancer and no OPMD matches, whereas the nested path remained at 6.37 through k=5. This demonstrates that medically interpretable nested extensions do not necessarily preserve global optimality. Across the global problems, mean exact-optimum hit rates were 100%, 66.7%, 84.4%, and 95.6% for exhaustive enumeration, greedy search, binary particle swarm optimization, and simulated annealing, respectively.
At the present 14-factor scale, exhaustive enumeration remained the fastest exact reference. The results therefore establish an exactly validated real-data proof of concept for fractional-QUBO SNP–environment barcode optimization, rather than evidence of quantum advantage, and provide a foundation for subsequent large-scale digital- and quantum-annealing studies.
Keywords: fractional QUBO; SNP–environment barcode; corrected odds ratio; oral cancer; OPMD; combinatorial optimization