Hierarchical Multi-Omics AFT Mediation Network (HOMAN): Modeling the Mitochondrial‑Immune Axis as a Causal Mediator of Epigenetic Aging on Disability‑Free Survival

Hypothesis: Hierarchical Multi-Omics AFT Mediation Network (HOMAN) improves healthspan prediction by explicitly modeling the mitochondrial‑immune axis as a causal mediator between epigenetic aging signatures and disability‑free survival, using late‑integrated autoencoders, penalized AFT regression, and causal mediation analysis.

Recent work shows that accelerated failure time (AFT) models outperform Cox‑based neural survival models when proportional hazards are violated in aging cohorts [1]. Deep autoencoders robustly compress multi-omics data, and late integration of omics‑specific features consistently yields superior predictive performance [2][3][5]. However, current pipelines stop at prediction; they do not translate multi-omics risk scores into mechanistic pathways that can be intervened upon. Moreover, the shift from lifespan to healthspan endpoints (e.g., disability‑free survival) demands models that not only forecast risk but also identify modifiable mediators [4].

We hypothesize that epigenetic age acceleration (e.g., GrimAge residuals) influences disability‑free survival largely through a mitochondrial‑immune cascade: epigenetic dysregulation alters mitochondrial DNA copy number and reactive oxygen species production, which in turn activates NLRP3 inflammasome signaling and elevates circulating IL‑6 and TNF‑α, ultimately accelerating functional decline. This cascade represents a testable, biologically grounded mediator that links molecular aging to clinical healthspan.

HOMAN operationalizes this hypothesis in three stages: (1) Layer‑specific variational autoencoders learn low‑dimensional representations of epigenomics, transcriptomics, proteomics, metabolomics, and microbiomics data, preserving late‑integration advantages [2][3][5]; (2) A penalized AFT regression with minimax concave penalty predicts disability‑free survival from the concatenated latent features, providing unbiased effect estimates when hazards are non‑proportional [1]; (3) Causal mediation analysis decomposes the total effect of epigenetic age acceleration on survival into direct and indirect effects mediated by the latent mitochondrial‑immune module (constructed from correlated mitochondrial proteomics, metabolomics, and cytokine levels), using bootstrap confidence intervals for inference [6].

The hypothesis is falsifiable: if HOMAN does not achieve a statistically significant improvement in concordance index (C‑index) ≥0.05 over a baseline late‑integrated AFT model without mediation, or if the indirect effect through the mitochondrial‑immune axis is not significant (p>0.05) after correcting for multiple testing, the hypothesis is rejected. Conversely, a significant indirect effect representing ≥30% of the total effect would support the mechanistic claim.

Implementation requires a longitudinal cohort with baseline multi-omics measures and biennial healthspan assessments (e.g., NHANES or the Framingham Offspring Study). Code will be made publicly available, and sensitivity analyses will test alternative mediators (e.g., senescent‑cell secretome) to ensure specificity. By uniting predictive survival modeling with causal mediation, HOMAN bridges the gap between high‑dimensional omics and actionable healthspan interventions.