Searching for the cellular underpinnings of the selective vulnerability to tauopathic insults in Alzheimer's disease

Neurodegenerative diseases such as Alzheimer's disease exhibit pathological changes in the brain that proceed in a stereotyped and regionally specific fashion. However, the cellular underpinnings of regional vulnerability are poorly understood, in part because whole-brain maps of a comprehensive collection of cell types have been inaccessible. Here, we deployed a recent cell-type mapping pipeline, Matrix Inversion and Subset Selection (MISS), to determine the brain-wide distributions of pan-hippocampal and neocortical cells in the mouse, and then used these maps to identify general principles of cell-type-based selective vulnerability in PS19 mouse models. We found that hippocampal glutamatergic neurons as a whole were significantly positively associated with regional tau deposition, suggesting vulnerability, while cortical glutamatergic and GABAergic neurons were negatively associated. We also identified oligodendrocytes as the single most strongly negatively associated cell type. Further, cell-type distributions were more predictive of end-time-point tau pathology than AD-risk-gene expression. Using gene ontology analysis, we found that the genes that are directly correlated to tau pathology are functionally distinct from those that constitutively embody the vulnerable cells. In short, we have elucidated cell-type correlates of tau deposition across mouse models of tauopathy, advancing our understanding of selective cellular vulnerability at a whole-brain level.

Methods
Gene expression

The scRNAseq data used to generate the cell-type maps come from Yao, et al. for the Allen Institute for Brain Science (AIBS), which sequenced approximately 1.3 million individual cells sampled comprehensively throughout the neocortex and hippocampal formation at 10x sequencing depth (Yao et al, 2021, Cell). Using a standard Jaccard-Louvain clustering algorithm, the authors jointly and hierarchically clustered these samples at three taxonomic levels: class (n = 4), subclass (n = 42), and cluster (n = 387). The full annotation and gene expression profile of each sample, as well as trimmed mean expression across cell-type clusters, are publicly available (https://portal.brain-map.org/atlases-and-data/rnaseq/mouse-whole-cortex-and-hippocampus-10x). Here we used this trimmed means by cluster dataset, as the Matrix Inversion and Subset Selection (MISS) algorithm only requires the consensus profiles of cell types per cluster. Utilizing the hierarchical taxonomy provided by the authors as described above, we grouped the 387 individual clusters into subclasses as we have done previously\cite{Mezias2022}, resulting in 42 unique neuronal and non-neuronal cell types spanning four major classes: cortical glutamatergic, hippocampal glutamatergic, GABAergic, and non-neuronal (herein referred to as the Yao cell types).

The spatial gene expression data come from the coronal series of the in situ hybridization (ISH)-based Allen Gene Expression Atlas (AGEA) (Lein et al, 2007, Nature). While the sagittal atlas has better gene coverage, we chose to use the coronal atlas because of its superior spatial coverage, which provides an isotropic resolution of 200 um per voxel. Furthermore, MISS uses a feature selection algorithm to remove uninformative and noisy genes, partly mitigating the effect of the reduced gene coverage. We performed unweighted averaging on genes for which multiple probes were available, resulting in a dataset of 4083 unique genes. Lastly, we removed the 320 genes that were not present in both the scRNAseq and ISH datasets, resulting in a final set of 3763 genes.

Tauopathy experiments

We queried five studies to obtain twelve individual mouse tauopathy datasets (which we refer to interchangeably as "experiments"):

BoludaCBD and BoludaDSAD (Boluda et al, 2015, Acta Neuropathol.)

DS4, DS6, DS7, DS9, DS6 110, DS9 110 (Kaufman et al, 2016, Neuron)

Hurtado (Hurtado et al, 2010, Am J Pathol.)

IbaHippInj and IbaStrInj (Iba et al, 2013, J Neurosci.)

IbaP301S (Iba et al, 2015, Acta Neuropathol.)

We selected these studies for their spatial coverage (>40 regions quantified across both hemispheres) and the fact that they all utilized the same mouse tauopathy model (PS19), which contains a P301S tau transgene on a C57BL/6 background. The only exception is the Hurtado experiment, which contained an additional mutation in the amyloid precursor protein (APP) gene.

Alzheimer's disease risk gene selection

We selected our 24 AD risk genes by finding the intersection set between the list given by the Alzheimer's Disease Sequencing Project (ADSP) (Bellenguez et al, 2022, Nature Genetics; Kunkle et al, 2019, Nature Genetics) and the AGEA (Lein et al, 2007, Nature)

Gene annotations were obtained from the UniProt database (Bateman et al, 2023, Nucleic Acids Res.) unless otherwise noted.

Matrix Inversion and Subset Selection (MISS)

We applied the MISS algorithm to the Yao, et al. scRNAseq dataset (Yao et al, 2021, Cell) and the AGEA ISH dataset (Lein et al, 2007, Nature) as described previously (Mezias et al, 2022, PNAS).

Briefly, MISS involves two steps: 1) subset selection, which utilizes a feature selection algorithm to remove low-information genes that add noise to the final prediction of cell-type density; and 2) matrix inversion, where the gene-subset spatial ISH-based gene expression matrix is regressed on the gene-subset scRNAseq-based gene expression matrix voxel-by-voxel to obtain cell-type densities.