Hypothesis

Transformer‑based neural Cox models that ingest longitudinal multi‑omics trajectories and treat epigenetic age acceleration as a time‑varying covariate will outperform static‑baseline DeepSurv and CoxPH models in predicting mortality and multimorbidity onset in community‑dwelling aging cohorts.

Rationale

Recent work shows transformer survival nets excel at capturing non‑linear hazard functions and time‑varying covariates in oncology [1][2], yet these architectures have not been transferred to aging biology where biomarkers such as DNA‑methylation clocks, proteomic signatures, and transcriptomic modules evolve non‑linearly over years [3]. Pathway‑preserving transformer designs (e.g., cGAS‑STING‑guided attention) already demonstrate how biological structure can be embedded without sacrificing predictive power [4]. By modeling methylation age acceleration (ΔAge) as a dynamic predictor that updates at each omics wave, we can directly test whether the rate of biological aging, rather than its baseline offset, drives future hazard.

Novel Mechanistic Insight

We propose that inflammasome‑driven nucleic‑acid sensing (cGAS‑STING) acts as a upstream modulator that couples cytosolic DNA damage signals to epigenetic drift. In the transformer, attention heads weighted by cGAS‑STING pathway activity will learn to amplify omics features that precede inflammasome activation, thereby capturing a causal cascade: DNA damage → cGAS‑STING activation → NAD⁺ decline → altered DNA methylation → increased mortality risk.

Testable Predictions

1. Prediction 1: In the Framingham Heart Study offspring cohort (n≈3,000 with tri‑annual omics), a pathway‑preserving transformer that ingests baseline and follow‑up methylome, proteome, and transcriptome will achieve a C‑index ≥0.78 for 5‑year all‑cause mortality, a statistically significant improvement (≥0.04) over DeepSurv trained on baseline ΔAge only (p<0.01, bootstrap).
2. Prediction 2: SHAP‑derived feature importance will show that the change in ΔAge between visits (ΔΔAge) contributes more to risk prediction than the baseline ΔAge value, confirming the dynamic nature of epigenetic aging.
3. Prediction 3: Attention weights linked to the cGAS‑STING module will correlate positively with circulating IFN‑stimulated gene signatures and predict imminent inflammasome activation (elevated IL‑1β) within the subsequent 6‑month window.

Experimental Design (Falsifiable)

• Data: Use publicly available longitudinal multi‑omics from the Framingham Offspring Study and the NHS/HPFS cohorts, with at least three omics waves per participant over 8 years.
• Models: (a) Baseline CoxPH, (b) DeepSurv (static baseline omics), (c) Transformer‑Cox with time‑varying omics (no pathway constraints), (d) Pathway‑preserving transformer‑cGAS‑STING attention (our method).
• Evaluation: Time‑dependent C‑index, integrated Brier score, calibration plots. Use 5‑fold cross‑validation stratified by sex and baseline age.
• Statistical Test: Compare C‑indices via paired bootstrap (10 000 reps). Null hypothesis: no difference between (c) and (d). Reject if p<0.01.
• Falsification: If (d) fails to outperform (c) or shows no meaningful SHAP signal for ΔΔAge or cGAS‑STING attention, the hypothesis is refuted.

Impact

Confirming this hypothesis would provide a mechanistically interpretable deep survival framework that bridges cancer‑grade transformer oncology tools to aging research, enabling dynamic risk stratification and early intervention targeting inflammasome‑epigenetic axes.