Research rationale
Alzheimer’s disease (AD) involves complex gene-brain interactions, but traditional GWAS and neuroimaging rely on large labeled datasets, limiting insights into rare atypical/young-onset forms. Self-supervised learning (SSL) offers a data-efficient alternative by learning representations from unlabeled data. However, existing SSL lacks neuroscience-specific biases. Our NSSL framework addresses this gap by integrating structural MRI (sMRI) and SNP data with brain-informed algorithms, enabling discovery of novel mechanisms in less represented cohorts.
Research Objectives
Develop NSSL Framework: Design and implement a multimodal SSL model embedding inductive biases (e.g., anatomical priors, genetic heritability) for joint sMRI-SNP representation learning.
Pretrain on Large Datasets: Train NSSL on disease-agnostic public cohorts (e.g., UK Biobank, ADNI) to generate transferable, interpretable priors capturing intrinsic gene-brain couplings.
Transfer to Rare Cohorts: Fine-tune pretrained representations on small atypical/young-onset AD datasets to uncover subtype-specific mechanisms (e.g., amyloid-tau-vascular interactions).
Validate and Interpret: Assess biological plausibility via downstream tasks (e.g., phenotype prediction, pathway enrichment) and interpretability tools (e.g., attention maps).
Research Questions
Can neuroscience-informed SSL yield representations that robustly encode gene-brain interactions across AD heterogeneity?
Do pretrained NSSL priors enhance discovery in small cohorts, revealing novel gene-brain mechanisms related to atypical AD?