Knowledge-Augmented Genomics Transformers for Mechanistic Links to AD Dementia

Описание: Alzheimer’s disease (AD) is a progressive neurodegenerative disorder marked by cognitive decline and heterogeneous, cell-type-specific molecular changes. Extracting mechanistic insights across cohorts from large scale AD genomic data remains difficult. We propose to fine-tune a transformer foundation model for single-cell analysis (scGPT) on more than 4 million single-nucleus RNA-seq profiles from two deeply characterized AD resources—the Religious Orders Study and Memory and Aging Project (ROSMAP) and the Seattle Alzheimer’s Disease Brain Cell Atlas (SEA-AD), while harmonizing cognitive and neuropathology measures to enable integrative modeling. Our approach embeds prior biological knowledge into the model via knowledge-guided attention masks and multimodal fusion (e.g., pathway, GWAS, and perturbation priors). We will pair model explanations (Shapley values, pathway enrichment) with conformal prediction to quantify uncertainty and flag low-confidence calls. To ensure generalizability, we will conduct cross-cohort benchmarking and external validation against classical bioinformatics baselines. The pipeline will emphasize reproducibility (open code, standardized QC, and data harmonization recipes) and deliver interpretable gene- and pathway-level hypotheses for amyloid/tau burden and cognitive impairment.