Abstract A073: TP53 mutant variant allele frequency improves machine-learning prediction of post-operative survival in pancreatic cancer and marks a proliferative transcriptional program
Kshitij S. Gaur, Imaad Said, Eugene Chen, Megan Zeller, Mohammad Aldakkak, Matthew A. Sochor, Bhabishya Neupane, Mandana Kamgar, Alexandra Phan, Janice Zhao, Samih Thalji, William Hall, Beth Erickson, Christina Small-Tom, Callisia Clarke, Kathleen K . Christians, Nikki Lytle, Thomas McFall, Douglas B. Evans, Anai Kothari, Yongwoo D. SeoAbstract
Background:
Prognostication after curative-intent resection of pancreatic ductal adenocarcinoma (PDAC) relies on clinicopathologic staging, yet post-operative outcomes remain heterogeneous. Comprehensive genomic profiling (CGP) is now routine, but its high dimensionality is poorly suited to conventional analysis. We tested whether integrating CGP with clinicopathologic data using machine learning improves survival prediction, and whether TP53 mutant variant allele frequency (VAF), a continuous measure of clonal burden, adds prognostic value beyond binary mutation status.
Methods:
We studied 110 patients from the MCW Surgical Oncology Tissue Bank who completed neoadjuvant therapy and underwent resection, with primary-tumor CGP (Tempus xT, 648-gene panel); 78 had matched bulk RNA sequencing. Gradient-boosted decision trees (XGBoost) with 10-fold cross-validation predicted above- versus below-median overall survival (OS; median 18.2 months), with SHAP (Shapley additive explanations) feature-importance values. Kaplan-Meier and Cox models evaluated TP53 status and VAF. Whole-transcriptome differential expression (DESeq2) and VAF-gradient gene set enrichment analysis (GSEA) characterized TP53-associated transcription. Transcriptional findings were externally validated in TCGA-PAAD (n=177) and CPTAC-PDAC (n=140).
Results:
Among all CGP features, TP53 and KRAS carried the greatest feature importance, and TP53 mutation status rivaled nodal status as a predictor of OS. Adding TP53 to clinicopathologic features improved the cross-validated area under the curve for our prediction model by 0.09 (0.729 to 0.813). TP53 mutation and node positivity stratified OS independently (median OS: TP53-mutant/node-positive 11.0, mutant/node-negative 16.5, wild-type/node-positive 21.9, wild-type/node-negative 28.2 months). Modeled continuously, higher TP53 VAF was associated with worse OS dose-dependently (Cox hazard ratio 1.35 per 10% increase in VAF). On GSEA, higher TP53 VAF tracked a proliferation-dominant transcriptional program (E2F targets, G2M checkpoint, MYC; normalized enrichment 3.5-4.0) with reciprocal depletion of epithelial-mesenchymal and immune signatures; this proliferative phenotype replicated in TCGA. DESeq2 identified LYPD2 as the only gene overexpressed in TP53-mutant tumors, and LYPD2 mRNA upregulation replicated independently in both TCGA) and CPTAC (p=0.0098).
Conclusions:
Integrating CGP with clinicopathologic data via machine learning improves post-operative risk stratification in PDAC, and TP53 VAF appears to carry dose-dependent prognostic information beyond binary mutation status. TP53 mutation is coupled to a proliferation-dominant transcriptional program and to LYPD2 overexpression, both externally validated across independent cohorts, nominating LYPD2 for further mechanistic study. These findings show how machine learning can render CGP clinically actionable to guide surgical risk stratification and precision-medicine trial enrollment.
Citation Format:
Kshitij S. Gaur, Imaad Said, Eugene Chen, Megan Zeller, Mohammad Aldakkak, Matthew A. Sochor, Bhabishya Neupane, Mandana Kamgar, Alexandra Phan, Janice Zhao, Samih Thalji, William Hall, Beth Erickson, Christina Small-Tom, Callisia Clarke, Kathleen K . Christians, Nikki Lytle, Thomas McFall, Douglas B. Evans, Anai Kothari, Yongwoo D. Seo. TP53 mutant variant allele frequency improves machine-learning prediction of post-operative survival in pancreatic cancer and marks a proliferative transcriptional program [abstract]. In: Proceedings of the AACR Conference on Pancreatic Cancer: New Frontiers in Biology and Therapeutic Development; 2026 Sep 25-28; San Diego, CA. Philadelphia (PA): AACR; Cancer Res 2026;86(18_Suppl_2):Abstract nr A073.