HDMAX2-surv: high-dimensional mediation analysis of survival data with application to pancreatic cancer

Abstract

MotivationHigh-dimensional mediation analysis is critical for dissecting causal pathways in complex diseases. However, existing methods struggle to handle censored survival outcomes and unobserved confounders in epigenetic data. There is a pressing need for scalable, statistically robust frameworks to identify epigenetic mediators of exposure-outcome relationships, especially in cancer research where survival data are central and molecular pathways are highly complex and interconnected. ResultsWe introduce HDMAX2-surv, a two-step framework extending HDMAX2 to survival analysis. HDMAX2-surv integrates (i) latent factor modeling to adjust for unobserved confounders, (ii) flexible survival models (Aalen additive hazards and accelerated failure-time models). Simulations demonstrated superior performance over state-of-the-art methods (e.g., HIMA) in mediator selection and effect estimation. We integrated this approach with causal discovery frameworks and immune deconvolution algorithms to dissect multi-pathway mediation mechanisms. Applied to TCGA pancreatic adenocarcinoma data (n=112), HDMAX2-surv identified 36 aggregated methylated regions (AMRs) mediating tobacco exposure effects on survival, including immune-mediated pathways undetectable via gene expression alone. Availability and implementationHDMAX2-surv is implemented in R and available on GitHub (https://github.com/bcm-uga/tims-pdac), with documentation and example scripts for reproducibility.