Key Takeaway

• The fusion model achieved ~93% accuracy, surpassing SNP‑only (87%) and MRI‑only (92%).

• Multimodal integration improves diagnosis and uncovers biologically relevant SNPs/genes, paving the way for biomarker discovery in Alzheimer’s disease.

Introduction

• Alzheimer’s disease involves both genetic variation and structural brain changes.

• Unimodal approaches miss cross‑modal interactions.

• Aim: build a multimodal deep learning pipeline integrating MRI and SNP data for classification and biomarker discovery.

Methods

• Cohort: 757 ADNI‑1 participants → 308 after QC (137 AD, 171 CN).

• Genetics: 339,255 SNPs filtered → 4,478 via GWAS → one‑hot encoded.

• MRI: bias correction, skull stripping, normalisation, resized to 152×184×152 voxels.

• Pipeline: SNP features via Transformer encoder, MRI features via PeleeNet, fusion via Cross‑Fusion Transformer Module.

Results

• Fusion model: ~93% accuracy (vs SNP‑only 87%, MRI‑only 92%).

• eQTL analysis: ANXA5, MTR, SDK1, MTMR14, and 13 SNPs on Chromosome 19 near APOE.

• Integrated Gradients: 43 high‑priority SNPs linked to apoptosis, synaptic adhesion, vascular signalling, and chromatin remodelling.

Conclusions

• Multimodal integration boosts diagnostic precision beyond unimodal models.

• Provides a biologically grounded framework for identifying region‑specific genetic and structural signatures.

• Highlights genes and pathways central to Alzheimer’s, reinforcing the value of combining imaging and genomics for both diagnosis and biomarker discovery.

• Serves as a stepping stone toward clinically relevant, interpretable AI in neurodegenerative research.