More isn't always better: Expanding and refining prognostic methylation signatures for pancreatic ductal adenocarcinoma.

769 Background: In our previous study, we developed protein-informed methylation signatures from a chemoresistance gene panel (RGp, n=122) curated from the literature, identifying 28 signatures with strong prognostic value in The Cancer Genome Atlas (TCGA) pancreatic ductal adenocarcinoma (PDA) database. To enhance the clinical utility of this panel, we incorporated methylation changes from genes with known prognostic (PGp) and diagnostic (DGp) significance in PDA to form a composite gene panel (CGp-PDA) and repeated our analysis. We hypothesized that epigenetic changes in these genes reflect protein expression and that deriving signatures from them could help identify patients at risk of treatment failure. Methods: We assembled a composite gene panel (CGp-PDA) of 206 genes (122 RGp + 26 PGp + 82 DGp) through a literature review spanning from January 1970 to April 2024. These PGp and DGp genes are unique, as cell-free DNA methylation changes in them have established prognostic and diagnostic value, respectively in PDA. The same cohort of pancreatic ductal adenocarcinoma (PDA) patients (n=184) from our previous study was analyzed. We applied elastic net multivariate analysis to calculate survival outcomes and generated Kaplan-Meier plots using our gene panel, optimizing values for 𝛼 and λ over 100 iterations. Results: Our analysis identified 20 signatures with cytosines followed by a guanine residue (CpG) site in genes from our panel. The number of CpG sites in these signatures ranged from 3 to 4248. The most effective signature featured 17 CpG sites (22 months [m] vs. 67m, p=0.03), followed by signatures with 3 (16m vs. 49m, p=0.009) and 12 (17m vs. 36m, p=0.02) CpG sites. The 4248 CpG site signature demonstrated the poorest prognostic value among the signatures (17m vs. 21m, p=0.01). Some signatures included multiple CpG sites within a single gene. Notably, KCNH2 reappeared in several signatures, as in the prior study. Expanding our RGp to CGp-PDA did not produce more (20 vs. 28) signatures or improved prognostic value. Conclusions: Targeted methylation signatures, supported by robust preclinical or translational evidence of their impact on treatment resistance, can be clinically valuable for informing treatment selection and prognosis. We may need to shift our focus from machine learning-based broad methylation or transcriptomic analyses to more targeted investigations.