Integrating Dynamic Epigenetic Clocks with Physics-Inspired Degradation Priors in Joint Transformer-MOFA-Cox Models

Background Recent transformer‑based survival models achieve high group‑level discrimination (C‑index 84.17 vs 76.60) but fail to explain most individual‑level longevity heterogeneity, especially among men, non‑Hispanic Blacks, and low‑education groups [4]. This gap suggests that static omics snapshots miss time‑varying biological processes that drive differential aging trajectories.

Hypothesis We propose that incorporating dynamic epigenetic clock signals (e.g., longitudinal changes in DNA methylation at CpG sites linked to mitochondrial function and cellular senescence) as latent trajectories, regularized by physics‑inspired degradation priors derived from materials science fatigue models, within a joint MOFA‑Cox/transformer framework will substantially improve individual‑level mortality prediction and reduce sociodemographic bias.

Mechanistic Rationale

1. Epigenetic drift reflects cumulative exposure to metabolic stress, inflammation, and toxin load, which vary across life course and sociodemographic contexts [5]
2. Materials‑science fatigue laws (e.g., Paris’ law for crack growth) provide a principled way to model hazard acceleration as a function of accumulated damage, offering a biologically interpretable prior for non‑linear hazard functions [6]
3. Joint modeling of longitudinal omics (MOFA) with survival (Cox) already links biomarker trajectories to risk [7]; adding transformer attention enables capture of non‑linear interactions across time‑varying covariates [1]
4. Physics‑inspired priors constrain the hazard surface, reducing over‑fitting in high‑dimensional settings and improving generalization to under‑represented groups.

Testable Predictions

• Models that add longitudinal epigenetic clock features will increase the C‑index for individual prediction by ≥5 % points in the HRS cohort, with the largest gains in subgroups previously poorly predicted (men, non‑Hispanic Blacks, low‑education).
• The physics‑inspired prior will yield calibrated hazard estimates (Brier score reduction ≥10 %) compared to unregularized transformer‑MOFA‑Cox baselines.
• Feature importance analysis will show that time‑varying methylation at mitochondria‑related CpGs contributes more to risk variation than static baseline omics.

Experimental Design

1. Extract quarterly DNA methylation data (EPIC array) from a subsample of HRS participants with ≥5 years follow‑up.
2. Compute epigenetic age acceleration residuals (e.g., GrimAge) at each wave to form longitudinal trajectories.
3. Build three models: (a) baseline Transformer‑MOFA‑Cox (omics + clinical), (b) + static epigenetic baseline, (c) + longitudinal epigenetic trajectories with physics‑inspired degradation prior (implemented via a hazard‑regularization term derived from Paris’ law).
4. Evaluate using time‑dependent C‑index, Integrated Brier Score, and calibration plots across subgroups.
5. Perform sensitivity analysis varying the prior strength to assess robustness.

Implications If validated, this approach would bridge the gap between population‑level discrimination and individual‑level precision, offering a mechanistic, interpretable tool for healthspan‑focused interventions and highlighting epigenomic dynamics as a key driver of longevity inequality.