ABS 115: Discovering vulnerable cell types in the earliest stages of Alzheimer’s disease

Caleb P. Schultz ¹ , Kyle J. Travaglini ¹ , Yi Ding ¹ , JT Mahoney ¹ , Mariano I. Gabitto ¹ , SEA-AD Consortium ¹ ² ³ , Jennie L. Close ¹ , Rebecca D. Hodge ¹ , Jeremy A. Miller ¹ , Michael Hawrylycz ¹ , C. Dirk Keene2, Ed S. Lein ¹

¹ Allen Institute for Brain Science
² University of Washington
³ Kaiser Permanente. Seattle, WA, USA

Van Wickle (2025) Volume 1, ABS013

Introduction: Alzheimer’s disease (AD) affects over six million Americans and is projected to double by 2050. Neuropathological changes, particularly tau protein aggregation in the hippocampus (HIP) and entorhinal cortex (EC), precede clinical symptoms by decades and are strongly correlated with cognitive decline. Identifying cell type-specific vulnerabilities to tau pathology in these early-affected regions is critical for understanding AD progression and developing targeted interventions. Methods: As part of the SEA-AD consortium, single-nucleus RNA sequencing (snRNA-seq) data from ~10 million cells across 12 brain regions and 84 donors were analyzed to classify and characterize neurons and non-neuronal cells in HIP and EC. Dimensionality reduction and clustering were performed using scVI (single-cell variational inference), a deep generative model utilizing gene expression variance. Subsequent cell-type annotation employed scANVI, a semi-supervised extension of scVI, transferring knowledge from annotated neocortical datasets to uncharacterized hippocampal and entorhinal cells. An iterative labeling process incorporated differential gene expression analysis, reference atlases, and literature review to resolve ambiguous clusters. The cell-type taxonomy was refined using robust clustering algorithm HICAT (Hierarchical Iterative Clustering for the Analysis of Transcriptomics), enabling the identification of biologically plausible supertypes. Results: The study expanded the SEA-AD Brain Cell Atlas with detailed subclass and supertype annotations for excitatory neurons in HIP and EC. These classifications were integrated with a continuous pseudo-progression score (CPS), derived from tau and beta-amyloid neuropathology, to quantify disease severity. Bayesian compositional analysis of the fine-grained taxonomy and CPS using scCODA revealed four excitatory supertypes in the lateral entorhinal cortex with significant decreases in abundance between dementia and non-dementia patients. Conclusions: This work identifies specific excitatory neuron populations vulnerable to early AD pathology, offering a framework for targeted mechanistic studies and therapeutic development. The integration of deep generative modeling with neuropathological data represents a scalable approach for mapping disease-related cellular trajectories in neurodegenerative disorders.

Methods: As part of the SEA-AD consortium, single-nucleus RNA sequencing (snRNA-seq) data from ~10 million cells across 12 brain regions and 84 donors were analyzed to classify and characterize neurons and non-neuronal cells in HIP and EC. Dimensionality reduction and clustering were performed using scVI (single-cell variational inference), a deep generative model utilizing gene expression variance. Subsequent cell-type annotation employed scANVI, a semi-supervised extension of scVI, transferring knowledge from annotated neocortical datasets to uncharacterized hippocampal and entorhinal cells. An iterative labeling process incorporated differential gene expression analysis, reference atlases, and literature review to resolve ambiguous clusters. The cell-type taxonomy was refined using robust clustering algorithm HICAT (Hierarchical Iterative Clustering for the Analysis of Transcriptomics), enabling the identification of biologically plausible supertypes.

Results: The study expanded the SEA-AD Brain Cell Atlas with detailed subclass and supertype annotations for excitatory neurons in HIP and EC. These classifications were integrated with a continuous pseudo-progression score (CPS), derived from tau and beta-amyloid neuropathology, to quantify disease severity. Bayesian compositional analysis of the fine-grained taxonomy and CPS using scCODA revealed four excitatory supertypes in the lateral entorhinal cortex with significant decreases in abundance between dementia and non-dementia patients.

Discussion: This work identifies specific excitatory neuron populations vulnerable to early AD pathology, offering a framework for targeted mechanistic studies and therapeutic development. The integration of deep generative modeling with neuropathological data represents a scalable approach for mapping disease-related cellular trajectories in neurodegenerative disorders.