2026 Theses Doctoral

Bayesian frameworks for functionally informed rare variant association tests in Alzheimer's disease

Das, Anjali

Rare genetic variants represent a largely underexplored source of disease risk, offering unique insight into the mechanisms underlying complex traits that common variants alone cannot capture. Their study is complicated by statistical challenges inherent to low variant frequency, the vast size of the noncoding genome, and difficulties determining which variants are functionally relevant. This thesis presents a series of methodological contributions unified by the goal of improving functionally informed rare variant (RV) association testing in Alzheimer's disease (AD), through probabilistic modeling, functional annotation integration, and scalable inference. The first contribution, gruyere, introduces an empirical Bayesian framework that learns global, trait-specific annotation weights to improve variant prioritization in genome-wide RV association tests. Applied to whole-genome sequencing (WGS) data from the Alzheimer's Disease Sequencing Project (ADSP), gruyere identifies 13 significant associations not detected by existing methods, and establishes that deep learning-based predictions of splicing, transcription factor binding, and chromatin state are informative priors for noncoding RV effects in AD. Building on this foundation, parmigiano generalizes the annotation-informed framework into a flexible integration layer for existing RV association tests, jointly learning annotation weights and a global variant selection threshold within a Bayesian hierarchical model. Applied to a larger ADSP release comprising 12,900 cases and 23,846 controls, parmigiano increases association yield across five existing RV association tests, uncovers 23 candidate AD genes including SIGLEC10 and HUNK, and produces associations that replicate more reliably in held-out data than those from unannotated methods. Finally, emmental extends these ideas into the transcriptomic setting, introducing a hierarchical Bayesian framework for incorporating RVs into transcriptome-wide association studies (TWAS). A key limitation of existing TWAS methods is their restriction to variants observed in both the reference and target datasets, which systematically excludes rare variation. emmental addresses this by modeling variant effect sizes as a nonlinear function of functional annotations, enabling expression imputation for variants absent from the reference panel. Trained on brain-specific matched WGS and RNA-seq data from BigBrain and applied to ADSP, emmental demonstrates improved power for gene-trait discovery in AD. Together, these contributions advance a unified perspective on rare variant association testing through Bayesian hierarchical modeling and functional annotation integration as core principles for improving gene discovery in Alzheimer's disease and complex traits more broadly.