Hypothesis

Embedding longitudinal multi‑omics trajectories into a hybrid model that combines a pathway‑aware transformer encoder with a permanental process hazard layer will produce significantly more accurate and biologically interpretable mortality risk predictions across species than existing DeepSurv, DeepHit, or DySurv baselines.

Mechanistic Rationale

Aging hazards are non‑stationary and driven by cascading molecular damage that converges on conserved pathways (e.g., inflammation, proteostasis, mitochondrial dysfunction). Transformer‑based encoders excel at capturing long‑range dependencies and preserving pathway structure when supplied with guided attention masks derived from curated gene‑set databases [7]. Permanental processes model the square‑root of a Gaussian process on the hazard scale, allowing flexible, non‑proportional time‑varying effects while scaling linearly with the number of events via the representer theorem [4]. By feeding transformer‑generated pathway embeddings as the covariate input to a permanental hazard layer, the model can learn how temporal shifts in pathway activity warp the baseline hazard in a mathematically tractable way. This architecture also inherits the causal handling strengths of TV‑SurvCaus (RNNs with representation balancing) when extended to treatment‑response longitudinal studies [6], enabling interrogation of intervention effects on pathway‑specific hazard trajectories.

Experimental Design

1. Data – Assemble two longitudinal cohorts: (a) human aging multi‑omics (e.g., blood proteomics, transcriptomics, metabolomics from the NIH LINCS Aging Project or similar, with quarterly sampling over 5 years) and (b) mouse lifespan multi‑omics from the Interventions Testing Program (ITP) with matched time points. Both datasets include all‑cause mortality outcomes.
2. Model Variants – (i) Baseline DeepSurv Cox PH, (ii) DeepHit, (iii) DySurv conditional VAE, (iv) Permanental process alone with raw omics summaries, (v) Pathway‑aware transformer alone feeding a Cox PH layer, (vi) Full hybrid (Transformer → Permanental hazard). All models receive identical time‑varying covariates.
3. Training – Use stratified 5‑fold cross‑validation within each species, optimizing hyper‑parameters via Bayesian search. For cross‑species transfer, train on human data, fine‑tune on mouse data (and vice‑versa) using a small learning rate.
4. Evaluation – Primary metric: time‑dependent C‑index at 1‑, 3‑, and 5‑year horizons. Secondary: Integrated Brier Score (IBS) and calibration plots. Statistical significance assessed via paired bootstrap (p<0.05).
5. Interpretability – Extract pathway‑specific hazard contributions from permanental layer; validate against known aging hallmarks using enrichment analysis.

Predictions and Falsifiability

• Prediction 1: The full hybrid model will achieve a time‑dependent C‑index improvement of ≥0.02 over the best baseline (DySurv) in both human and mouse held‑out tests.
• Prediction 2: Cross‑species fine‑tuning will reduce mouse‑sample hazard prediction error (MAE) by at least 15 % compared to training on mouse data alone, indicating transferable pathway hazard signatures.
• Prediction 3: Pathway‑specific hazard weights will show significant enrichment for inflammation and mitochondrial dysfunction pathways (FDR<0.05), providing mechanistic plausibility.

Falsification Criteria: If the hybrid model fails to outperform baselines by the stipulated margins in both species, or if cross‑species transfer yields no significant MAE reduction, the hypothesis is falsified. Likewise, if pathway hazard weights do not align with established aging hallmarks, the mechanistic link between transformer‑encoded pathway dynamics and permanental hazard modulation would be refuted.