MethConvTransformer: A Deep Learning Framework for Cross-Tissue Alzheimer's Disease Detection.

Area of Science:

• Epigenetics
• Neuroscience
• Computational Biology

Background:

• Alzheimer's disease (AD) involves complex epigenetic changes, particularly DNA methylation, offering potential for noninvasive biomarkers.
• Current DNA methylation signatures for AD lack reproducibility due to tissue and study variations, hindering clinical translation.

Purpose of the Study:

• To develop a deep learning framework, MethConvTransformer, for robust Alzheimer's disease biomarker discovery by integrating DNA methylation data from multiple tissues.
• To improve the accuracy and generalizability of epigenetic biomarkers for early AD detection.

Main Methods:

• Developed MethConvTransformer, a transformer-based deep learning model integrating DNA methylation profiles from brain and peripheral tissues.
• Employed CpG-wise linear projection, convolutional layers, and self-attention to capture dependencies, incorporating covariates and tissue embeddings.
• Validated the model on six GEO datasets and the ADNI cohort, comparing performance against conventional machine learning baselines.

Main Results:

• MethConvTransformer consistently outperformed baseline models in discrimination and generalization across datasets.
• Interpretability analyses identified biologically relevant AD-associated methylation patterns linked to immune signaling, metabolism, and cellular organization.
• The model demonstrated robust, cross-tissue biomarker discovery capabilities for Alzheimer's disease.

Conclusions:

• MethConvTransformer provides a powerful, interpretable framework for discovering reliable, cross-tissue epigenetic biomarkers for Alzheimer's disease.
• The findings advance reproducible methylation-based diagnostics and offer new hypotheses for AD mechanisms.
• This approach holds significant promise for improving early AD detection and understanding disease pathology.