Multi-omics promises personalized medicine, but it runs into cost, interpretation complexity and the need for repeated blood draws. Epigenetic changes, especially DNA methylation, reflect how environment and lifestyle affect the genome, which makes them attractive as biomarkers. This work fits into a line of research where DNA methylation is already used to create surrogates for proteins and metabolites, but until now such proxies have been limited to a narrow set of markers. This study expands the approach to 1694 proxies, covering clinical, metabolomic and proteomic domains.
Models were trained on the MGB-ABC cohort (4418 people) with data from DNA methylation (Illumina EPIC 850K), metabolomics (1150 metabolites), proteomics (1979 proteins) and electronic medical records. For each biomarker, informative CpG sites were selected by mutual information (90th percentile threshold), then an elastic net (glmnet) was trained, or XGBoost if the linear model failed to reach a correlation of at least 0.2. Validation was performed in seven independent cohorts (n=57 454), including pediatric ones (CAMP, CRA) and population-based ones (Generation Scotland, TruDiagnostic).
The mean Spearman correlation between EBPs and observed values was 0.41 for clinical markers, 0.29 for metabolites and 0.29 for proteins. EBPs reproduced 532 incident and 2201 prevalent disease associations, and in 60-75% of cases the effect was stronger than for the observed biomarkers. Classification of deviations from reference ranges achieved 77.8% accuracy (62.0% sensitivity, 78.2% specificity). Validation confirmed the correlations (r=0.31-0.96) and disease associations. A longitudinal analysis of 10 people showed qualitative agreement between EBP and clinical trajectories.
This work stands out from neighboring studies (for example, 109 protein EpiScores or 64 metabolite surrogates) by its scale: 1694 proxies versus dozens, coverage of three domains and systematic validation in seven cohorts. It is the first test of whether a single approach works for clinical, metabolomic and proteomic markers at the same time. The results show that EBPs encode not just current concentrations but a biological signal integrated over time, which explains their strong associations with long-term outcomes.
The field could move toward expanding the set of proxies to hard-to-measure biomarkers, improving calibration for absolute quantification and adapting to other methylation platforms. Using EBPs in biobanks for phenomics and in clinical trials as surrogate endpoints looks promising.
Will affect epidemiology, preventive medicine, gerontology and clinical laboratory diagnostics.
Conduct prospective studies to assess the prognostic value of EBPs in clinical practice and expand validation to more diverse populations.
The link to an unsolved problem in epigenetics: to what extent DNA methylation reflects causal biological pathways rather than merely correlating with outcomes. The question of transferring models between platforms and populations also remains open.
The proxy for urate accurately predicted sex: men have higher urate levels, which matches the known sexual dimorphism.
In the series Altered Carbon, consciousness is transferred between bodies, and here DNA methylation acts as a "backup copy" of biological state, making it possible to reconstruct a picture of health from a single sample.
Not published on arXiv. Here are its page, the PDF and the formalization: you can check it yourself, not only read our account.