Hypothesis: Manifold-aligned latent axes isolate causal aging drivers

Core idea Integrating transcriptomics, lipidomics, metabolomics and microbiome data with manifold learning (e.g., contrastive variational autoencoders) will uncover low‑dimensional latent dimensions that map onto biologically interpretable processes (immune senescence, mitochondrial dysfunction, lipid remodeling) and that predict future morbidity better than simple feature concatenation.

Why this matters The Human Phenotype Project showed microbiome clocks outperform single‑omics in multimorbidity association, yet the analysis relied on feature concatenation into LightGBM, ignoring cross‑modality structure. Without manifold alignment, shared biological signals are entangled with modality‑specific noise, obscuring causal axes.

Mechanistic extension We propose that true drivers of aging sit at the intersection of omics layers, where perturbations in one layer (e.g., NAD+ decline) propagate through coordinated changes in transcripts, lipids, metabolites and gut microbes. A contrastive VAE can learn a shared manifold by maximizing agreement between paired samples from different omics while preserving modality‑specific variance in private subspaces. The shared latent space thus isolates signals that are consistent across layers—candidate causal drivers.

Testable predictions

1. Predictive superiority – In a held‑out longitudinal subset of the Human Phenotype Project, models built on the shared latent dimensions will achieve higher R² for 2‑year change in frailty index and incident hypertension than models using concatenated features or single‑omics clocks (p<0.01, paired t‑test).
2. Causal validation – CRISPRi‑mediated knockdown of NAD+ biosynthetic enzymes in human colonic organoids will shift the organoid’s latent position along the mitochondrial dysfunction axis, predicted by the manifold model, and this shift will precede measurable increases in senescence markers (p16^INK4a, SA‑β‑gal).
3. Cross‑cohort replication – Applying the same manifold alignment pipeline to an independent cohort (eUK‑Biobank n=5,000 with matched omics) will recover latent axes with comparable loadings (Procrustes correlation >0.8) and similar predictive power for mortality.
4. Missing‑modality robustness – When one omics layer is deliberately masked (e.g., metabolomics missing), the variational encoder’s private subspace will absorb the loss while the shared latent representation remains stable (ΔR² <0.02), demonstrating utility for clinical settings with sparse data.

Falsifiability If the shared latent dimensions fail to outperform concatenated baselines in longitudinal prediction, or if NAD+ perturbation does not produce the anticipated latent shift, the hypothesis that manifold‑aligned axes capture causal aging drivers is refuted.

Implementation sketch

• Train a multimodal contrastive VAE on baseline omics (n=12,000) with modality‑specific encoders and a shared decoder.
• Use contrastive loss (InfoNCE) to align paired samples across omics and a reconstruction loss for each modality.
• Extract the shared latent vector; fit gradient‑boosted models to predict health outcomes.
• Perform longitudinal validation and perturbation experiments as described.

By shifting from descriptive correlation to a structure‑preserving, causally informed latent space, this approach directly addresses the field’s gap in mechanistic insight and offers a concrete route to identify modifiable drivers of human aging.