💻 bioinformatics

Long-Read epigenetic clocks identify improved brain aging predictions

This study introduces improved brain aging prediction models by leveraging Oxford Nanopore long-read sequencing of over 28 million CpG sites across diverse African and European ancestry cohorts, demonstrating that these long-read epigenetic clocks outperform existing array-based methods and highlight the critical importance of inclusive, ancestry-aware training datasets.

Original authors: Grant, S. M., Eger, S. J., Makarious, M. B., Meredith, M., Moller, A., Grant-Peters, M., Hicks, A., Mandal, A., Auluck, P., Villegas-Lanau, A., Mejia-Cupajita, B., Acosta-Uribe, J., Aguillon, D., Leon

Published 2026-08-17

📖 4 min read☕ Coffee break read

CC0 1.0

Original authors: Grant, S. M., Eger, S. J., Makarious, M. B., Meredith, M., Moller, A., Grant-Peters, M., Hicks, A., Mandal, A., Auluck, P., Villegas-Lanau, A., Mejia-Cupajita, B., Acosta-Uribe, J., Aguillon, D., Leonard, H., Kuznetsov, N., Weller, C., Reed, X., Catching, A., Jain, M., Ferrucci, L., Kosik, K. S., Cookson, M. R., Ryten, M., Nalls, M. A., Billingsley, K. J.

Original paper dedicated to the public domain under CC0 1.0 (https://creativecommons.org/publicdomain/zero/1.0/). ⚕️ This is an AI-generated explanation of a preprint that has not been peer-reviewed. It is not medical advice. Do not make health decisions based on this content. Read full disclaimer

Imagine your body is a bustling city, and every cell is a building with its own unique blueprint. Over time, weather, traffic, and daily wear-and-tear leave marks on these buildings. In the world of biology, one of the most reliable ways to see how much "wear and tear" a cell has experienced is by looking at its DNA methylation. Think of methylation as tiny sticky notes or highlighters that the cell attaches to its DNA instructions. As we get older, the pattern of these sticky notes changes in a very predictable way. Scientists have been using these patterns to build "epigenetic clocks"—essentially, biological stopwatches that tell us how old a person's cells feel, which can be different from their actual birthday.

For a long time, these clocks were built using a map that only showed a few major streets (about 5% of the DNA), and they were mostly drawn using data from people of European ancestry. This meant the clocks were like a GPS that worked great in London but got lost in Lagos or Mumbai, and it missed all the tiny alleyways where the real action happens. Furthermore, most of these clocks were tested on blood samples, but when it comes to diseases like Alzheimer's, we really need to know what's happening in the brain. The big question scientists have been asking is: Can we build a better, more accurate clock for the brain that works for everyone, no matter their background, by looking at the entire map of the DNA instead of just the main streets?

This paper says, "Yes, we can, and here's how." The researchers decided to stop using the old, limited maps and instead used a super-powerful new tool called Oxford Nanopore long-read sequencing. Imagine that instead of reading a book by looking at one letter at a time through a tiny keyhole, this new tool lets you read entire paragraphs and chapters in one go. This allowed them to see over 28 million sticky notes (CpG sites) across the DNA of brain tissue from two very different groups of people: one group of European ancestry and another of African or mixed ancestry.

Instead of just counting individual sticky notes, which can be a bit noisy and confusing, the team grouped them together into neighborhoods. They looked at "promoters" (the on/off switches for genes) and "whole-genome windows" (broad slices of the DNA landscape). They fed this massive, detailed data into a smart computer system called GenoML, which acted like a super-talent scout, trying out dozens of different math tricks to find the one that could best predict a person's age.

The results were impressive. The new "long-read clocks" were much better at guessing the age of brain cells than the old, standard clocks. In fact, when they tested these new clocks on a group of people from Colombia (who weren't part of the original training group), the models built on the "promoter neighborhoods" still worked very well, showing they could travel across different populations. However, the models that looked at the broader "whole-genome windows" got a bit confused when they tried to predict the age of the Colombian group, often guessing they were much older than they actually were. This suggests that while the broad landscape of DNA changes in ways that might be unique to specific populations, the "neighborhood" switches are more universal.

The study also dug into why these clocks work. They found that the clocks weren't just reacting to how long the body had been sitting in a freezer after death or the mix of different cell types in the brain. Instead, they were picking up on real biological aging signals. The clocks highlighted specific genes and pathways that seem to get "tired" or "stressed" as we age, including genes related to how neurons talk to each other and how the brain develops. Interestingly, they found that some aging signals were shared across all groups, while others were specific to the ancestry of the group, proving that we need diverse data to build a complete picture of human aging.

In short, this paper shows that by using a high-resolution, whole-DNA view and including people from different backgrounds, we can build much more accurate and fair biological clocks for the brain. It suggests that the future of aging research isn't just about counting more dots, but about understanding the neighborhoods those dots live in, and making sure we have a map that works for everyone.

Technical Summary: Long-Read Epigenetic Clocks for Improved Brain Aging Predictions

Problem Statement
Current epigenetic clocks, which estimate biological aging via DNA methylation, face significant limitations in generalizability and resolution. Most existing models rely on array-based data (capturing <5% of CpG sites) derived predominantly from peripheral tissues of individuals of European ancestry. This restricts their applicability to brain tissue and non-European populations. Furthermore, linear regression models often fail to capture the non-linear complexities of aging, and reliance on individual CpG sites introduces stochastic variability. There is a critical need for high-resolution, ancestry-aware, and tissue-specific aging clocks that leverage the full genomic context provided by long-read sequencing.

Methodology
The study utilized Oxford Nanopore Technologies (ONT) long-read sequencing (LRS) to generate genome-wide methylation profiles from prefrontal cortex (PFC) samples. The dataset comprised two discovery cohorts: the North American Brain Expression Consortium (NABEC; 189 individuals of European ancestry) and the Human Brain Collection Core (HBCC; 132 individuals of African or admixed ancestry). An external validation cohort was drawn from the Grupo de Neurociencias de Antioquia Brain Bank in Colombia (14 individuals).

The technical workflow involved:

Feature Aggregation: Instead of using individual CpG sites, the authors aggregated methylation signals across two types of genomic regions to reduce noise and preserve biological context:
- Promoter-based features: Aggregating signal over EPDnew promoter regions (expanded by 1kb).
- Whole-genome window-based features: Aggregating signal over non-overlapping 1kb windows containing at least 50 CpGs.
Machine Learning Pipeline: Models were developed using GenoML, an automated machine learning platform that competitively selects the best-performing architecture from a pool of linear and non-linear algorithms. The pipeline included rigorous feature selection (filtering for age association and removing high-correlation redundancy) and nested cross-validation.
Benchmarking: The new "Long-Read clocks" (LR-clocks) were benchmarked against established clocks (Horvath, PhenoAge, DunedinPACE, Lu et al., and Shireby) using both the original CpG sites and the aggregated promoter/window features derived from the long-read data.
Downstream Analysis: Feature importance was assessed using SHAP (Shapley Additive exPlanations). Functional enrichment analyses included Gene Ontology (GO), Expression-Weighted Cell Type Enrichment (EWCE), and transcription factor motif enrichment (HOMER) to identify biological mechanisms.

Key Results

Performance Superiority: LR-clocks significantly outperformed benchmark models. Aggregating signal into promoter-based and window-based features yielded higher accuracy ( $R^2$ $R^{2}$ ) than models using individual CpG sites.
- Promoter-based LR-clocks: Achieved $R^2$ of 0.862 (HBCC), 0.905 (NABEC), and 0.905 (Combined).
- Window-based LR-clocks: Achieved $R^2$ of 0.901 (HBCC), 0.912 (NABEC), and 0.917 (Combined).
- The window-based models generally performed best, followed by promoter-based, with CpG-based models performing the worst.
Ancestry and Tissue Specificity: Models trained on brain-specific data (Shireby clock) outperformed pan-tissue clocks. While cohort-specific models performed well, combined-cohort models demonstrated strong generalizability, though with some variation in feature selection between African and European ancestry groups.
Robustness to Confounders: Age acceleration estimates derived from the LR-clocks were not significantly driven by post-mortem interval (PMI) or estimated cell type composition, suggesting the clocks capture genuine biological aging signals rather than sampling artifacts. A consistent negative association was observed between inhibitory neuron fraction and age acceleration.
Biological Insights:
- Shared and Specific Pathways: Enrichment analyses identified shared mechanisms (e.g., transcriptional regulation, REST motifs in hypermethylated regions) and cohort-specific differences. For instance, PCDH (protocadherin) cluster promoters were strong predictors in NABEC/Combined models but largely absent in HBCC models.
- Cell Type Associations: Genes mapped to hypomethylated window features were consistently enriched in astrocytes, while hypermethylated window features were enriched in oligodendrocyte progenitor cells (OPCs) across all cohorts.
External Validation: In the Colombian cohort, promoter-based LR-clocks maintained strong correlations ( $R^2$ > 0.72), whereas window-based models showed weaker correlations and over-prediction of age, suggesting promoter-level patterns may be more conserved across populations than non-promoter genomic regions.

Significance and Claims
The paper claims that long-read methylation profiling enables the construction of accurate, ancestry-aware aging clocks for the brain that overcome the limitations of array-based data. Key contributions include:

Resolution: Demonstrating that aggregating methylation signals across promoters and whole-genome windows, rather than relying on sparse CpG sites, substantially improves prediction accuracy and cross-cohort generalizability.
Inclusivity: Highlighting the necessity of diverse training datasets (including African and European ancestry) to build equitable models, as feature importance and pathway enrichment varied significantly between cohorts.
Biological Interpretability: Providing a framework that preserves biological interpretability (unlike PCA-based models) while identifying specific cell types (e.g., inhibitory neurons, astrocytes) and transcriptional regulators (e.g., REST, HOX genes) implicated in brain aging.
Tissue Specificity: Reinforcing that brain-specific models are superior to multi-tissue models for brain aging predictions.

The authors conclude that while their workflows offer a powerful tool for studying brain aging, future iterations will benefit from larger sample sizes, harmonized sequencing chemistries, and integration with comprehensive phenotypic data to further resolve population-specific epigenetic drift.

More like this