AAIC Conference (2025)
Soorin Yim, Kyungwook Lee, Dongyun Kim, Sungjoon Park, Doyeong Hwang, Kiyoung Kim, Danniel Gatti(Jackson Laboratory), Elissa Chesler(Jackson Laboratory), Amy Dunn(Jackson Laboratory), Kristen O'Conell(Jackson Laboratory)
Abstract
Background
Alzheimer’s disease (AD) is characterized by complex, cell-type-specific molecular changes. Single-nucleus RNA sequencing (snRNA-seq) enables detailed analysis of these alterations, offering insights into AD pathogenesis. Predictive models that classify disease status from snRNA-seq data can aid interpretation by linking expression patterns to patient phenotypes. However, most existing models ignore cell types, limiting both accuracy and biological interpretability.
Methods
We applied a recently developed hierarchical multiple instance learning (MIL) framework to improve patient-level phenotype prediction from snRNA-seq data. This model introduces a two-step pooling mechanism: first aggregating representations within each cell type, and then across cell types, thereby incorporating biologically meaningful hierarchy into the learning process. Aggregation can be done by either taking average (mean), or attention-based, resulting in four models. We evaluated these models on the ROSMAP cohort, which includes postmortem brain snRNA-seq profiles from individuals with AD and cognitively normal controls.
Results
Among four models, Cell Type Attention (CTA) achieved superior predictive performance on the ROSMAP dataset, improving the area under the ROC curve (AUC) for AD vs. control classification. Importantly, CTA offers improved interpretability by enabling attribution of prediction importance at cell-type levels. Using this framework, we identified key cell types contributing to the classification outcome, highlighting disease-relevant populations such as astrocytes. These findings demonstrate the value of incorporating cell-type structure in phenotype prediction and suggest new avenues for exploring cellular mechanisms of AD.
Conclusion
By applying a hierarchical MIL to Alzheimer’s snRNA-seq data, we demonstrate enhanced predictive performance and interpretability over existing models. This approach not only improves classification of AD status but also facilitates the identification of cell types most associated with disease, offering insights that may support biomarker discovery and therapeutic development.