🤖 machine learning

Learning Multi-Relational Graph Representations for DNA Methylation-Based Biological Age Estimation

The paper introduces RelAge-GNN, a multi-relational graph neural network framework that integrates co-methylation patterns, genomic co-localization, and gene-level associations to improve the accuracy and biological interpretability of DNA methylation-based age estimation compared to existing methods.

Original authors: Qing Qing, Xikun Zhang, Zhongyuan Zhang, Jiarui Liu, Xingtong Yu, Xiaotao Shen, Ziqi Xu, Qixin Zhang, Zhe Wang, Renqiang Luo

Published 2026-05-11

📖 4 min read☕ Coffee break read

CC BY 4.0

Original authors: Qing Qing, Xikun Zhang, Zhongyuan Zhang, Jiarui Liu, Xingtong Yu, Xiaotao Shen, Ziqi Xu, Qixin Zhang, Zhe Wang, Renqiang Luo

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine your body has an internal "biological clock" that ticks differently than the calendar on your wall. While your chronological age is just the number of years since you were born, your biological age is a measure of how worn out or healthy your cells actually are. Some people look their age; others look older or younger than they are.

To figure out this biological age, scientists look at DNA methylation. Think of DNA as a massive instruction manual for your body. Methylation is like sticky notes placed on specific pages of that manual. These notes tell the cell whether to read a page (turn a gene on) or ignore it (turn a gene off). As we age, the pattern of these sticky notes changes in very specific ways.

The Problem with Old Methods

For a long time, scientists tried to guess your biological age by looking at these sticky notes one by one, like checking a grocery list item by item. They treated every note as an independent fact.

The paper argues this is like trying to understand a symphony by listening to each instrument in isolation. You might hear a violin, but you miss the harmony, the rhythm, and how the instruments talk to each other. In reality, sticky notes (CpG sites) don't just sit there; they influence each other. Some are on the same page (same gene), some are in the same chapter (same chromosome), and some move in perfect sync (co-methylation).

The New Solution: RelAge-GNN

The authors propose a new tool called RelAge-GNN. Instead of a grocery list, they built a social network map for these sticky notes.

Here is how they built it, using three different types of "friendships" between the notes:

The "Sync" Group (Co-methylation): Notes that always change together, like two friends who always dress the same way.
The "Roommate" Group (Same Chromosome): Notes that live on the same chromosome, like neighbors in the same apartment building.
The "Family" Group (Same Gene): Notes that belong to the exact same gene, like siblings in the same family.

How the AI Works

The model uses a special type of AI called a Graph Neural Network (GNN). Imagine three different detectives, each specializing in one type of friendship:

Detective A only looks at the "Sync" group.
Detective B only looks at the "Roommates."
Detective C only looks at the "Families."

Each detective gathers clues and writes a report. But here is the clever part: instead of just averaging their reports, the model has a smart manager (a "gating mechanism"). This manager looks at the specific person being analyzed and decides, "For this person, Detective A's clues are the most important, but for that person, Detective B's clues matter more."

This allows the model to adapt. It doesn't force a single rule for everyone; it dynamically weighs which biological relationships matter most for each individual.

What They Found

The team tested this new "detective team" on thousands of blood samples from healthy people and people with diseases like ovarian cancer, schizophrenia, and osteoporosis.

Better Accuracy: The new model was better at predicting biological age than previous methods, including the famous "Horvath clock" and other AI models. It made fewer mistakes.
Spotting "Fast Aging": The model was particularly good at detecting age acceleration. This is when a person's biological age is much higher than their actual age. The model successfully identified that people with certain diseases (like ovarian cancer) often had "older" biological clocks than healthy people of the same age.
Understanding the "Why": Because the model separates the three types of relationships, the researchers could see which clues were most important. They found that the "Sync" group (notes that change together) was the most critical factor, confirming that the relationships between the notes are just as important as the notes themselves.

In Summary

This paper introduces a smarter way to read the body's biological clock. Instead of reading the DNA notes in isolation, it maps out how they interact with each other. By using a flexible system that weighs different types of biological relationships, it creates a more accurate picture of how fast or slow a person is aging, offering a clearer view of their health status.

Technical Summary: RelAge-GNN for DNA Methylation-Based Biological Age Estimation

Problem Statement
Aging clocks estimate biological age from biomarkers to assess physiological state and disease risk, distinct from chronological age. DNA methylation is a primary biomarker due to its stability and strong correlation with aging. However, existing AI-based aging clock methods face significant limitations:

Independence Assumption: Most approaches treat CpG sites as independent features, ignoring complex, heterogeneous biological relationships such as co-methylation, genomic proximity, and gene-level associations.
Rigid Modeling: While some graph-based methods exist, they often rely on a single graph or fixed strategies to combine different relationship types. This fails to distinguish the varying contributions of different biological relations across samples and limits model flexibility.
Interpretability: Deep learning models often lack clear interpretability regarding how specific biological structures contribute to predictions.

Methodology: RelAge-GNN
The authors propose RelAge-GNN, a multi-relational graph neural network framework designed to explicitly model the structural dependencies among CpG sites. The methodology consists of three core components:

Multi-Relational Graph Construction:
Instead of a single graph, the framework constructs three complementary graphs, each retaining a specific semantic relationship:
- Co-methylation Graph ( $G_1$ ): Edges represent continuous correlation values between CpG sites (thresholded to ensure sparsity).
- Genomic Co-localization Graph ( $G_2$ ): Edges represent Boolean connections between CpG sites located on the same chromosome.
- Gene-Level Association Graph ( $G_3$ ): Edges represent Boolean connections between CpG sites located within the same gene.
  Node features include methylation values, CpG island attributes (location, length), flanking sequence information, and distance to transcription start sites.
Three-Branch Encoder with PNAConv:
Each graph is processed by an independent GNN branch using PNAConv (Principal Neighbourhood Aggregation). PNAConv is chosen for its ability to handle varying node degree distributions by employing multiple aggregators (mean, max, min, std) and scalers. This allows each branch to learn relation-specific structural representations ( $H_1, H_2, H_3$ ) without conflating heterogeneous edge types.
Learnable Gated Fusion Mechanism:
To address the issue of fixed combination strategies, the model employs a learnable gating network.
- The outputs of the three branches are concatenated ( $H_{cat}$ ).
- A gating network (MLP + Softmax) generates adaptive weights ( $\alpha_1, \alpha_2, \alpha_3$ ) for each node.
- The final representation ( $H_f$ ) is a weighted sum of the branch outputs: $H_f = \sum \alpha_k \odot H_k$ .
  This enables the model to dynamically adjust the contribution of each biological relationship at the node level based on the specific sample context.
Prediction and Interpretability:
The fused representation passes through a compression layer and an MLP regression head to predict age. For interpretability, the authors extend the GNNExplainer framework to quantify the importance of:
- Individual CpG nodes.
- Node attributes (features).
- The three distinct graph branches (relational structures).

Key Contributions

Identification of Limitations: The paper highlights that current methods fail to capture complex biological dependencies and often use rigid fusion strategies that cannot adapt to sample-specific variations in relationship importance.
Novel Framework: RelAge-GNN introduces a multi-relational architecture that explicitly models co-methylation, same-chromosome, and same-gene associations via independent branches.
Adaptive Fusion: The introduction of a learnable gating mechanism allows for dynamic, node-level adjustment of relationship weights, improving flexibility over fixed-weight baselines.
Enhanced Interpretability: The framework provides a mechanism to quantify the contribution of specific biological relations and CpG sites, offering insights into the drivers of age prediction.

Experimental Results
The model was evaluated on large-scale datasets comprising 3,707 healthy samples and 624 disease samples (ovarian cancer, schizophrenia, osteoporosis) from blood tissue.

Predictive Performance: RelAge-GNN achieved competitive accuracy compared to state-of-the-art baselines (Horvath, AltumAge, DeepMAge, GraphAge). It demonstrated a higher Fitting Regression Coefficient (FRC = 0.962) and lower Mean Absolute Error (MAE = 3.43) and Mean Squared Error (MSE = 29.46) than the baselines.
Age Acceleration (AA) Sensitivity: The model showed superior sensitivity in detecting age acceleration across diverse disease cohorts. For instance, in postmenopausal ovarian cancer, it identified approximately 40% of samples with positive AA, significantly outperforming GraphAge (<20%) and AltumAge (<10%).
Ablation Studies: Removing any single graph branch degraded performance, with the co-methylation graph ( $G_1$ ) and gene-level graph ( $G_3$ ) being particularly critical. The full multi-branch model with gating outperformed single-graph configurations, confirming the value of modeling heterogeneous relations.
Interpretability Findings:
- Methylation values were the most important node attributes.
- CpG island information (start/end positions, length) and flanking sequences were also key factors, suggesting a hierarchical pattern of signal plus structural context.
- Graph Importance: The co-methylation graph contributed most significantly to predictions, followed by the gene-level and same-chromosome graphs.

Significance and Claims
The paper claims that explicitly modeling multi-relational structures is an effective direction for improving aging clock models. By moving beyond the assumption of independent features and fixed graph combinations, RelAge-GNN captures heterogeneous structural patterns that better reflect biological aging mechanisms. The model's ability to adaptively fuse different biological relations and provide granular interpretability suggests potential utility for disease characterization and offers biologically meaningful directions for future aging research. The authors note that while the current framework is robust on blood tissue, extending it to multi-tissue settings and incorporating more fine-grained biological interactions remains a future direction.