TL;DR
This paper introduces iTARGET, a novel two-phase approach for predicting biological age from DNA methylation patterns. By first clustering individuals into age groups and then applying tailored prediction models to each group, iTARGET significantly improves accuracy and interpretability, outperforming existing methods. This work offers a more nuanced understanding of how epigenetic markers change with age across different life stages.
What Problem Does This Paper Address?
Predicting a person's biological age from their DNA methylation patterns is a key goal in aging research, offering insights into healthspan and disease risk. However, existing methods face challenges due to two main issues:
- Epigenetic Correlation Drift (ECD): The relationship between DNA methylation levels at specific sites and chronological age isn't constant. It changes dynamically throughout life, accelerating in early development and late life, and stabilizing in adulthood. Traditional models often assume a uniform rate of change, missing these nuances.
- Heterogeneity Among CpGs (HAC): Different DNA methylation sites (CpGs) exhibit unique patterns of change over time. Some might be highly correlated with age in one life stage but not another, making a single set of "age-predicting" CpGs insufficient for the entire lifespan.
These issues make it difficult to achieve both high accuracy and interpretability in age prediction models.
Methods
The authors propose a two-phase algorithm called iTARGET:
- Age Group Clustering (Phase 1):
- They use Facebook AI Similarity Search (FAISS), a powerful library for efficient similarity searching in high-dimensional data, to cluster DNA methylation profiles into age groups. This is done without dimensionality reduction, leveraging FAISS's ability to handle large datasets.
- Two age grouping strategies are explored:
- Decade-sized intervals: [0-10), [10-20), etc., for interpretability.
- Biologically significant inflection points: Based on research identifying key aging transitions (e.g., [0-34), [34-60), [60-78), 78+).
- Age Group-Specific Prediction (Phase 2):
- Within each identified age group, they perform age-group-specific CpG correlation calculation and feature selection. This involves calculating Pearson correlation coefficients between CpG methylation levels and age for samples within that specific group.
- The top 30 CpG sites with the strongest correlations within each age group are selected. This selection aims to balance interpretability and predictive performance, as studies have shown that a limited set of CpGs can achieve high accuracy.
- Explainable Boosting Machines (EBMs) are then trained for each age group using these selected CpGs. EBMs are generalized additive models that incorporate boosting for accuracy and are designed for interpretability, allowing for the detection of individual feature effects and pairwise interactions.
Datasets
The study utilizes a publicly available dataset of DNA methylation profiles from 11,910 blood-derived methylomes from donors aged 0 to 103 years, as detailed in the study by Varshavsky et al. [10]. This dataset was assembled from the Illumina Infinium HumanMethylation450 BeadChip and MethylationEPIC v2.0 array. The data was split into 80% for training and 20% for testing, with careful random selection within age groups to ensure representation. Potential age-related CpG sites were filtered down to 2,374.
Key Insights and Ideas
- Age-specific biomarkers are crucial: The correlation between CpG sites and age is not static. Identifying "decade-specific age-related biomarkers" is key to capturing these dynamic relationships.
- Interactions matter: EBMs can reveal synergistic interactions between CpG sites, offering deeper insights into the biological mechanisms of aging beyond individual marker effects.
- Clustering simplifies complexity: By segmenting the lifespan into age groups, the complex, non-linear relationships of DNA methylation with age are simplified within each segment, leading to more accurate and interpretable models.
- A limited set of CpGs can be powerful: Focusing on the top 30 CpGs per age group, rather than a vast number, strikes a good balance between predictive power and the ability to understand why the model makes a prediction.
Key Results
- Superior Accuracy: The iTARGET-Decade approach achieved the lowest Mean Absolute Error (MAE) of 3.7752 and an RMSE of 5.8164 among all evaluated models, outperforming traditional epigenetic clocks and classic linear regression models.
- Outperforming Baselines:
- Compared to traditional clocks like Horvathv1 (MAE 13.2526) and Hannum (MAE 6.6072), iTARGET-Decade showed significantly lower errors.
- Even compared to machine learning models like Linear Lasso (MAE 4.4514) and Gaussian Process Regression (MAE 4.8943), iTARGET-Decade demonstrated better precision.
- Interpretability: The method identifies specific CpG sites and their interactions within each age group, providing clear biological insights. For example, Figures 3, 4, and 5 showcase the global and local feature importance and interaction effects for the 20-30 age group.
- Ideal Scenario Performance: In an "iTARGET-ideal" scenario where age group classification is perfect, the model achieves exceptionally low error rates (MAE = 1.5744, RMSE = 2.5867), demonstrating its theoretical potential.
Limitations and Future Work
- Ideal vs. Real-World Classification: The "iTARGET-ideal" scenario highlights the model's potential but relies on perfect age group classification in the first stage. While the FAISS clustering is effective, achieving perfect classification in real-world forensic applications without prior age estimation might be challenging.
- Future Directions: The paper suggests that interactions between CpGs can be automatically detected, providing an index for future exploration into the biology of synergistic interactions between CpGs and genes. Further refinement of the age group identification step could bring practical applications closer to the ideal performance.
Summary updated