Misalignment of age clocks

Biological aging is a complex non-linear process, with markedly distinct starting and end points, yet the biomarkers of its progression remain elusive. A key assumption of most machine learning (ML) approaches for age clocks is that predictive biomedical features can be identified via mathematical transformations of data to favor a linear transition from start to end, even if they erase any natural biological pattern. It is given that expected correlations, e.g., time lived (age) and time left to live (mortality), would persist in such mathematically optimized models, biologically meaningful or not. Here, we further clarify the workings of the clocks, explain the trade-off between mathematical optimization and biological interpretability, and discuss a hallmark of aging, inflammaging, that age clocks struggle to detect. We expand on the negative consequences of incoherence in linear models where some DNA methylation (DNAm) features increase with aging and disease, while others correspondingly decrease, yet positive weights are assigned to both. We quantify the misalignment between major DNAm clocks and actual changes in DNAm, providing an interactive visualization of these errors for each model. We demonstrate that major conventional age clocks are both incoherent and skewed toward leukocyte fractions and that rectifying incoherence makes the model balanced and not skewed toward neutrophils and better detects inflammaging. We briefly outline non-linear ML age clocks and the advantages of identifying a natural trajectory of aging directly from the primary data.