Data-driven subtyping and differential glufosfamide benefit in pancreatic adenocarcinoma.

Abstract

755 Background: Glufosfamide couples a glucose moiety to an ifosfamide mustard to exploit tumor glucose transport and deliver an alkylator, with prior pancreatic cancer activity reported in early trials. Variation in baseline characteristics may identify subgroups with differential benefit. We applied bfLEAP, an ensemble clustering and stability-selection framework, to a glufosfamide clinical dataset (TH-CR-302) to discover reproducible subpopulations and test treatment-by-subgroup interactions for overall survival (OS). Methods: A post hoc analysis of TH-CR-302, a randomized phase 3 clinical trial evaluating glufosfamide vs. best supportive care (BSC), was conducted using an ensemble clustering and stability-selection framework (bfLEAP), to identify patient subgroups. A patient-to-patient similarity network was constructed from baseline demographics, screening labs, and pre-randomization clinical features. Clusters were then derived from this network by identifying recurring patient groupings. Patients without consistent groupings were not assigned a cluster. The primary endpoint was OS, with death as the event of interest. We identified clusters on pooled arms, then estimated arm-specific outcomes within clusters. Age and sex adjusted Cox models were used to evaluate the association between treatment assignment and OS. Hazard ratios (HRs) with 95% confidence intervals (CIs) were reported. Statistical analyses were performed in Python using the lifelines package. Results: Four distinct clusters were identified from 134/281 (47%) patients: Cluster A (n=12), Cluster B (n=26), Cluster C (n=24), and Cluster D (n=72). Cluster A, characterized by lower baseline glucose values and higher baseline neutrophil and monocyte counts, was associated with improved OS vs BSC [HR 7.26, 95% CI 1.21,43.69]. Global interaction P-value = 0.03. These findings suggest clinically relevant treatment effect heterogeneity. Direction consistent with overall trial signals for glufosfamide. Conclusions: Utilization of bfLEAP successfully identified patient subgroups within existing glufosfamide clinical trial data. Treatment effect heterogeneity was identified amongst clusters, identifying possibly early predictors of outcomes. These findings highlight the potential of data-driven clustering approaches to refine patient stratification and guide the development of personalized treatment strategies. Cox proportional hazards analysis: Cluster A vs. best supportive care. Coefficient Hazard Ratio 95% CI z p Treatment 7.26 (1.21, 43.69) 2.16 0.03 Age 1.01 (0.95, 1.07) 0.17 0.86 Sex 2.06 (0.42, 10.15) 0.89 0.37