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

TL;DR

MethConvTransformer introduces a cross-tissue transformer framework that jointly models brain and peripheral DNA methylation to detect Alzheimer's disease. It uses a CpG-wise linear projection, convolutional downsampling, and self-attention, augmented by covariate and tissue embeddings, to capture local and long-range methylation patterns while remaining interpretable. Across six GEO cohorts and an ADNI validation set, the model outperforms baselines and reveals multi-resolution biomarkers linked to immune signaling, glycosylation, lipid metabolism, and ER/Golgi organization. The work demonstrates robust cross-tissue biomarker discovery with multi-level interpretability, offering testable hypotheses for disease mechanisms and pathways toward translational methylation-based diagnostics.

Abstract

Alzheimer's disease (AD) is a multifactorial neurodegenerative disorder characterized by progressive cognitive decline and widespread epigenetic dysregulation in the brain. DNA methylation, as a stable yet dynamic epigenetic modification, holds promise as a noninvasive biomarker for early AD detection. However, methylation signatures vary substantially across tissues and studies, limiting reproducibility and translational utility. To address these challenges, we develop MethConvTransformer, a transformer-based deep learning framework that integrates DNA methylation profiles from both brain and peripheral tissues to enable biomarker discovery. The model couples a CpG-wise linear projection with convolutional and self-attention layers to capture local and long-range dependencies among CpG sites, while incorporating subject-level covariates and tissue embeddings to disentangle shared and region-specific methylation effects. In experiments across six GEO datasets and an independent ADNI validation cohort, our model consistently outperforms conventional machine-learning baselines, achieving superior discrimination and generalization. Moreover, interpretability analyses using linear projection, SHAP, and Grad-CAM++ reveal biologically meaningful methylation patterns aligned with AD-associated pathways, including immune receptor signaling, glycosylation, lipid metabolism, and endomembrane (ER/Golgi) organization. Together, these results indicate that MethConvTransformer delivers robust, cross-tissue epigenetic biomarkers for AD while providing multi-resolution interpretability, thereby advancing reproducible methylation-based diagnostics and offering testable hypotheses on disease mechanisms.

Figures (6)

• Figure 1: Overview of the MethConvTransformer framework for DNA methylation analysis: (A) Preprocessing pipeline converting raw IDAT files from multiple brain regions and blood into normalized, batch-corrected $\beta$-matrices with selected CpG features; (B) MethConvTransformer architecture combining CpG-wise linear projection, local convolutional encoding, and Transformer-based long-range attention, jointly integrated with covariate and tissue embeddings for AD prediction; (C) Multi-resolution interpretability module capturing CpG-level importance, regional co-methylation patterns (Grad-CAM++), long-range dependencies (attention maps), and subject-specific feature attribution (SHAP).
• Figure 2: Differential methylation profiles across individual datasets. (a-d): Volcano plots display sitewise differential methylation (raw $p$ values, $|\Delta\beta|>0.2$) and show that genome-wide significant DMPs are sparse. (e-h): DMR-based gene set enrichment analyses ($-\log_{10}$ FDR) reveal heterogeneous pathway enrichment across datasets. Together, these results indicate that isolated site-level testing yields limited and inconsistent signal, motivating integrative modeling.
• Figure 3: Comparison between CpG linear projection weights and SHAP-based feature attributions. The blue curve shows the absolute coefficients from the linear projection layer, representing intrinsic feature relevance within the network. The orange dashed curve denotes the mean SHAP values across samples, summarizing context-dependent marginal effects. Both profiles reveal that most CpGs exert weak individual effects, and predictive information arises primarily from distributed, cooperative methylation patterns rather than from isolated loci.
• Figure 4: Attention pattern in the last Transformer layer. Heatmap of self-attention weights from one representative attention head, illustrating pairwise dependencies among CpG embeddings. Most attention values are near zero, indicating sparse and localized interactions. Periodic cross-shaped hotspots highlight CpG clusters that mutually reinforce each other’s representations, suggesting region-specific co-methylation or shared functional regulation captured by the model.
• Figure 5: Grad-CAM++ interpretability of methylation-based AD classification.(A) Tissue-specific Grad-CAM++ activation maps showing averaged attribution magnitudes across 10 tissues. Rows correspond to tissues and columns to CpG indices at region level. (B) Overall Grad-CAM++ importance profile aggregated across tissues, highlighting CpG clusters contributing to AD classification.