GT-Mamba: A Topology-Aware Graph-State Space Model for Robust and Interpretable Epigenetic Age Prediction.

Area of Science:

• Computational Biology
• Genomics
• Machine Learning

Background:

• Epigenetic clocks traditionally face accuracy-interpretability trade-offs and require dataset-specific corrections for cohort generalization.
• Existing models often struggle with heterogeneity and missing data in real-world applications.

Purpose of the Study:

• To introduce GT-Mamba, a novel architecture designed to overcome the limitations of current epigenetic clocks.
• To enhance predictive accuracy and biological interpretability while improving generalization across diverse cohorts.

Main Methods:

• Developed GT-Mamba, integrating a Structure-Aware Graph Transformer with the Mamba state space model.
• The architecture captures CpG topological correlations and genome-wide long-range dependencies.
• Validated performance across heterogeneous independent cohorts and addressed feature missingness.

Main Results:

• GT-Mamba demonstrated strong out-of-the-box robustness, achieving a weighted average Mean Absolute Error (MAE) of 4.43 years.
• The model generalized effectively to EPIC 850k arrays, even with partial feature missingness.
• Ablation studies confirmed graph topology's contribution to noise robustness, with mechanistic analysis revealing links to developmental and functional processes.

Conclusions:

• GT-Mamba offers a robust and interpretable solution for epigenetic age prediction, outperforming existing methods in generalization.
• The model's ability to capture complex genomic dependencies enhances its applicability across various biological contexts and datasets.