Skip to main content
SSO 2026 Annual Meeting Main banner

P94. TP53 Mutant Variant Allele Frequency Can Improve Machine Learning Prediction of Post-operative Survival Outcomes in Pancreatic Cancer

Saturday, March 7, 2026
9:07 AM - 9:14 AM MST
Location: 105A-C
Introduction

Even after neoadjuvant therapy (NAT) and resection for localized pancreatic ductal adenocarcinoma (PDAC), overall survival (OS) varies greatly; while clinical factors such as positive lymph nodes (LNs) can stratify risk of poor outcome, there is need for greater precision. Comprehensive genomic profiling (CGP) of resected tumor has potential to bridge this gap. Here, we demonstrate a novel method of combining machine learning (ML) models with variant allelic frequency (VAF; % detected mutant alleles) to identify drivers of OS.

Methods

We identified all localized PDAC patients who completed NAT, resection, and had CGP data (from TEMPUS xT 648 gene panel). Using OS from surgery as a binary variable (above vs. below median), eXtreme Gradient Boosting (XGBoost) ML framework was utilized to extract features of importance using clinicopathologic variables, as well as VAFs for any present pathogenic mutations (with VAF=0% for wildtype [wt]). Model performance was evaluated using area under the receiver operating characteristic curve (AUC) and Shapley additive explanation plots (SHAP), with Kaplan-Meier curves for clinical validation.

Results

110 patients who completed NAT and resection had CGP data; 89 (81%) were KRAS mutated (G12D 33%, G12V 23%, and G12R 18%), and 67 (61%) had pathogenic TP53 mutation (mut). Initial XGBoost models using only known pathogenic mut VAFs identified TP53 as the highest impact feature. Addition of TP53 VAF, when combined with time-of-surgery clinical variables (e.g. age, comorbidity index, pathologic LN and T stage, lymphovascular or perineural invasion), improved prediction of OS (AUC = 0.81), vs. clinical variables alone (AUC = 0.73; Fig.a,b). SHAP analysis showed high TP53 VAF and LN+ status as the two highest contributing features for poor OS. When categorized by TP53 and LN status, TP53 mut/LN+ patients had significantly worse OS than all other groups (median 11.0 mo [95%CI 7.4-15.3] vs. 23.0 mo [20.4-32.1], p< 0.0001); there was no difference in OS between TP53 wt/LN+, TP53 mut/LN-, and TP53 wt/LN- patients (Fig.c). When stratified into high vs. low VAF by median (6.5%), only high VAF patients had worse OS (10.7 mo [6.7-22.0]) compared to wildtype (25.9 mo [16.5-38.7], p=0.05; Fig.d).

Conclusions

VAF analysis from CGP can uncover novel predictive targets for post-surgical outcomes. TP53 mut, particularly when present with LN+, confers worst OS and may be a clonally dependent process (based on VAF). Post-surgical TP53 mut/LN+ cohorts should be stratified as high risk and be considered for adjuvant treatment and clinical trials.

Learning Objectives:

  • Understand how comprehensive genomic profiling data can be utilized to generate predictive models of outcome.
  • Understand how XGBoost models can be utilized to identify genes of interest in prognostication for outcomes in pancreatic adenocarcinoma after surgical resection.
  • Understand how TP53 mutational status and variant allelic frequency may be a predictor for poor overall survival after surgery, particularly in the patients who have positive pathologic nodes.