BIMS-GA: Adaptive Multi-Objective Optimization for Network-Aware Pancreatic Ductal Adenocarcinoma Biomarker Discovery

Abstract

Pancreatic ductal adenocarcinoma (PDAC) has a five-year relative survival of 8%. Feature selection optimizes accuracy, survival concordance, and network structure separately; whether uniting them matters has not been determined. The BIMS-GA optimizes the cross-validated AUC, Cox concordance, and STRING v11.0 protein–protein interaction coherence in one fitness function and was compared against fourteen alternative methods on GSE62452 (n = 130) and eight external evaluation sets. Using a nested protocol, the fifteen methods had pooled out-of-fold AUCs spanning 0.877–0.900; the BIMS-GA ranked thirteenth with 0.8855. No DeLong comparisons survived correction. Stability did not persist either: the ten-seed Jaccard index of 0.804 exceeded every comparator but fell to 0.521 in a second seed block. Network coherence separated them by two orders of magnitude, 0.480 against 0.005–0.192, in an ordering unbroken across twenty runs. Removing the network term raised the discovery AUC by 0.0117 and lowered external prognostic concordance by 0.0149 in all ten runs, acting as a regularizer. The twenty-gene panel comprises one laminin-332 and an integrin adhesion program with no isolated member, stratifying survival in four cohorts (hazard ratios of 2.92–4.13 fitted within each cohort; 1.02–2.17 transferred). Network coherence, not discrimination, is the axis that remained separate in every run examined, while short-run stability did not.