Preserving Cross-Omics Covariance in Batch Correction Reveals a Microbiome‑Epithelial Axis that Drives Tissue‑Specific Epigenetic Aging

Mia Diallo· Mar 24, 2026

Mechanism: A coupled batch correction method (cVAE) preserves the vital links between gut microbiome metabolites and host epigenetic aging, unlike naive per-omics approaches that discard these signals. Readout: Readout: This leads to significantly improved multi-omics age clock accuracy and better prediction of disease risk, which can be modulated by interventions like butyrate supplementation.

Hypothesis

Multi‑omics aging clocks lose predictive power when batch correction removes true biological covariance between host epigenetics and microbiome‑derived metabolites. We hypothesize that a coupled batch‑correction strategy that explicitly preserves cross‑omics manifold alignment will uncover a causal microbiome‑epithelial axis: gut‑derived short‑chain fatty acids (SCFAs) and endotoxin levels modulate DNA‑methylation dynamics in intestinal epithelial cells, accelerating epigenetic aging locally and systemically. This axis explains why multi‑omics clocks show tissue‑specific patterns and why naive per‑omics correction attenuates those signals.

Mechanistic Rationale

Microbial metabolites as epigenetic modulators – Butyrate and propionate inhibit histone deacetylases and influence TET enzyme activity, altering CpG methylation rates in colonocytes【4】.
Age‑related gut barrier decline – Increased permeability lets lipopolysaccharide (LPS) enter circulation, triggering low‑grade inflammation that feeds back on DNA‑methylation clocks in blood and liver【1】.
Covariance as signal, not noise – In young individuals, metabolite levels and methylation states co‑vary to maintain homeostasis; with age, this coupling shifts, creating a divergent manifold that captures biological age beyond what any single omic layer shows【2】.
Current correction methods erase this coupling – Per‑omics batch correction (e.g., ComBat, per‑layer VAEs) treats each data type independently, discarding the very cross‑omic correlations that constitute the microbiome‑epithelial signal【3】.

Testable Predictions

Prediction 1: Applying a coupled variational autoencoder (cVAE) with a shared latent space and cycle‑consistency loss to paired methylome‑metabolome data from the same biopsy will retain higher mutual information between the two layers than standard per‑omic correction, measurable via retained canonical correlation analysis (CCA) scores.
Prediction 2: Clocks built on the cVAE‑corrected multi‑omics latent space will show significantly improved association with incident colorectal cancer and inflammatory bowel disease (HR > 1.3 per SD increase) compared with clocks built on per‑omic corrected data, in a prospective cohort of ≥2,000 adults followed for 5 years.
Prediction 3: Pharmacological modulation of the axis (e.g., butyrate supplementation or LPS‑neutralizing antibodies) in a randomized trial will shift the multi‑omics age acceleration residual in the direction predicted by the baseline metabolite‑methylation covariance, while leaving single‑omic clocks unchanged.

Experimental Design

Cohort: Recruit 1,200 participants aged 40‑80 with baseline colon biopsies, blood metabolomics, and stool 16S/metagenomics; follow for 5 years for clinical outcomes.

Processing: Generate parallel DNA‑methylation (EPIC) and targeted metabolomics (SCFAs, bile acids, LPS‑binding protein) from the same mucosal sample.
Batch correction: Compare three pipelines – (a) per‑omic ComBat, (b) per‑omic VAE (MAPBATCH), (c) coupled VAE with shared latent dimension and cross‑modal reconstruction loss.
Clock building: Train Gradient Boosting models (LightGBM) on each corrected dataset to predict chronological age; compute AgeAccel residuals.
Outcome testing: Use Cox proportional hazards models to test AgeAccel for prediction of incident colorectal adenocarcinoma, adjusting for age, sex, BMI, smoking.
Intervention sub‑study: In a nested RCT (n=200), assign participants to butyrate (4 g/day) or placebo for 12 weeks; re‑measure multi‑omics layers and evaluate changes in AgeAccel residuals.

Falsifiability

If the coupled VAE does not retain higher cross‑omics mutual information than per‑omic methods, or if clocks derived from it fail to outperform per‑omic clocks in predicting tissue‑specific morbidity, the hypothesis is falsified. Likewise, if butyrate supplementation does not alter the multi‑omics AgeAccel residual while leaving single‑omic clocks unchanged, the proposed mechanistic link between microbiome metabolites and epigenetic aging is unsupported.

Broader Implications

Confirming this hypothesis would shift batch correction from a nuisance‑removal step to a signal‑preserving operation that captures inter‑system communication. It would justify designing aging clocks as multi‑modal dynamical systems rather than static omic summaries, opening avenues for microbiome‑targeted interventions to modulate epigenetic aging trajectories.

[1] https://arxiv.org/html/2510.12384v1 [2] https://pmc.ncbi.nlm.nih.gov/articles/PMC12510163/ [3] https://www.youtube.com/watch?v=i-a4BjAn90E [4] https://pmc.ncbi.nlm.nih.gov/articles/PMC10547252/

Hypothesis

Mechanistic Rationale

Microbial metabolites as epigenetic modulators – Butyrate and propionate inhibit histone deacetylases and influence TET enzyme activity, altering CpG methylation rates in colonocytes【4】.

Age‑related gut barrier decline – Increased permeability lets lipopolysaccharide (LPS) enter circulation, triggering low‑grade inflammation that feeds back on DNA‑methylation clocks in blood and liver【1】.

Covariance as signal, not noise – In young individuals, metabolite levels and methylation states co‑vary to maintain homeostasis; with age, this coupling shifts, creating a divergent manifold that captures biological age beyond what any single omic layer shows【2】.

Current correction methods erase this coupling – Per‑omics batch correction (e.g., ComBat, per‑layer VAEs) treats each data type independently, discarding the very cross‑omic correlations that constitute the microbiome‑epithelial signal【3】.

Testable Predictions

Prediction 1: Applying a coupled variational autoencoder (cVAE) with a shared latent space and cycle‑consistency loss to paired methylome‑metabolome data from the same biopsy will retain higher mutual information between the two layers than standard per‑omic correction, measurable via retained canonical correlation analysis (CCA) scores.

Prediction 2: Clocks built on the cVAE‑corrected multi‑omics latent space will show significantly improved association with incident colorectal cancer and inflammatory bowel disease (HR > 1.3 per SD increase) compared with clocks built on per‑omic corrected data, in a prospective cohort of ≥2,000 adults followed for 5 years.

Prediction 3: Pharmacological modulation of the axis (e.g., butyrate supplementation or LPS‑neutralizing antibodies) in a randomized trial will shift the multi‑omics age acceleration residual in the direction predicted by the baseline metabolite‑methylation covariance, while leaving single‑omic clocks unchanged.

Experimental Design

Cohort: Recruit 1,200 participants aged 40‑80 with baseline colon biopsies, blood metabolomics, and stool 16S/metagenomics; follow for 5 years for clinical outcomes.

Processing: Generate parallel DNA‑methylation (EPIC) and targeted metabolomics (SCFAs, bile acids, LPS‑binding protein) from the same mucosal sample.

Batch correction: Compare three pipelines – (a) per‑omic ComBat, (b) per‑omic VAE (MAPBATCH), (c) coupled VAE with shared latent dimension and cross‑modal reconstruction loss.

Clock building: Train Gradient Boosting models (LightGBM) on each corrected dataset to predict chronological age; compute AgeAccel residuals.

Outcome testing: Use Cox proportional hazards models to test AgeAccel for prediction of incident colorectal adenocarcinoma, adjusting for age, sex, BMI, smoking.

Intervention sub‑study: In a nested RCT (n=200), assign participants to butyrate (4 g/day) or placebo for 12 weeks; re‑measure multi‑omics layers and evaluate changes in AgeAccel residuals.

Falsifiability

Broader Implications