Hypothesis

We propose that a manifold‑aligned variational autoencoder (VAE) combined with a transformer encoder can learn a shared, noise‑robust latent space across transcriptome, metabolome, lipidome, and microbiome data that captures causal drivers of aging rather than mere correlates. By explicitly modeling technical variation as a separate manifold and enforcing alignment of biological manifolds across cohorts, the model will isolate factors whose perturbation predicts changes in physiological age acceleration independent of chronological age.

Mechanistic Rationale

Recent multi‑omics clocks achieve high correlation with chronological age but their latent representations conflate technical batch effects with true biological signals [1,2]. We argue that aging manifests as a low‑dimensional trajectory in a conserved biological manifold, while platform‑specific noise occupies orthogonal dimensions. A VAE that learns two complementary latent subspaces—one for shared biology, one for technical covariates—can disentangle these sources. The transformer encoder then models temporal dependencies within the biological subspace, enabling the prediction of future states from baseline omics. This structure provides a mechanistic link: perturbations that shift the biological subspace (e.g., inhibition of mTOR or activation of NAD+ salvage) should produce measurable changes in the predicted age acceleration score, whereas changes confined to the technical subspace will not.

Testable Predictions

1. Cross‑cohort generalization: When trained on UK Biobank and CALERIE data with manifold alignment, the model’s age acceleration scores will retain >0.85 correlation with chronological age in an independent cohort (e.g., Human Phenotype Project) without retraining, whereas a standard LightGBM clock will drop below 0.70 due to batch effects.
2. Intervention sensitivity: In a randomized caloric restriction trial, the model’s predicted age acceleration will decrease significantly after 6 months (effect size > 0.4 SD) and this change will correlate with improvements in insulin sensitivity and grip strength, while the technical subspace remains unchanged.
3. Causal validation: CRISPR‑based knock‑down of a key mitochondrial gene (e.g., COX7A2) in cultured human fibroblasts will shift the biological latent vector toward higher predicted age acceleration, whereas knock‑down of a ribosomal protein linked to technical variance will not affect the score.

Experimental Design

• Data collection: Obtain multi‑omics profiles (RNA‑seq, metabolomics, lipidomics, 16S rRNA) from three independent cohorts totaling >5 000 participants, with longitudinal follow‑up at baseline, 12 months, and 24 months.
• Model architecture: Parallel VAEs for each omics modality produce modality‑specific latent vectors; a shared biological VAE is trained via a manifold alignment loss (Maximum Mean Discrepancy) across modalities and cohorts; a separate technical VAE captures batch effects. A transformer processes the shared biological sequence over time to output age acceleration.
• Analysis: Compare predictive performance, compute effect size of intervention on age acceleration, and perform perturbation experiments in vitro to validate causal mapping.

If the manifold‑aligned VAE‑Transformer framework fulfills these predictions, it will demonstrate that aging clocks can move beyond correlation to provide actionable, mechanistically interpretable biomarkers of biological age and intervention response.