henhao
Sign in
NeurosciencebioRxiv
Preprint

Mapping Alzheimer's neuropathology signatures to the whole brain transcriptome using machine learning data fusion

Anwesha BhattacharyaiD, Chloé SavignaciD, L. M. HodgsoniD, Jack StanleyiD, Guy WolfiD, Smita KrishnaswamyiD, David A. A BennettiD, Elizabeth Binder, Danilo BzdokiD

Abstract

In Alzheimer's disease (AD), misfolded proteins emerge across the entire brain in structured, yet not rigid, spatiotemporal patterns. Yet, a systematic bias of single-cell genomics toward sampling mostly cortical tissue limits our understanding of the whole-brain transcriptomic vulnerability to AD. Here, we develop a machine learning method to extrapolate local AD neuropathology signatures to the whole brain. By analyzing gene expression profiles of over two million cortical cells from 427 humans spanning the AD-pathology spectrum, we derive transcriptomic estimators of AD neuropathology. After extensive validations on datasets with known ground truth, we apply this framework to three million cells from 108 brain regions in the Siletti whole human brain atlas and derive an anticipated brain map of transcriptomic signatures indexing AD neuropathology. This interrogation of regions spanning the cortical, subcortical, and brainstem structures uncovers transcriptomic signatures associated with hyperphosphorylated tau in the medulla oblongata, dorsal raphe nucleus, and the tuberal and mammillary regions of the hypothalamus. At the cellular level, assessments of these signatures across 31 cell populations identify VGLUT1/2 expressing neurons, astrocytes, and microglia as key neuropathology-resembling populations. Within the hippocampus, pathology signatures surface in the rostral cornu ammonis (CA) subfields, particularly in the CA1 pyramidal neurons and dentate granule cells. {beta}-amyloid-like signatures localize to the neocortex with laminar selectivity--most prominently in upper layer somatostatin+ intratelencephalic neurons (L2-L3), but also in deep layer intratelencephalic and corticothalamic neurons (L5-L6). Neocortical astrocytes and microglia exhibiting disease associated signatures similarly demonstrate a unique laminar preference. Together, this study provides the first whole human brain map of AD pathology-associated transcriptomic signals, and exposes cell type, region, and cortex layer specific vulnerabilities.

PDFDOIData
Full-text digest ↓

02 /Specification & Classification

Click any tag to browse matching papers
Citations
0
OpenAlex Topics

In-depth summary

✦ Multi-agent verified full-text synthesis

Claude Opus 5.5 · xhigh effort · 8 Nov 2026 · from the full text of version 2 (26 Aug 2026) · CC BY 4.0

Cell type-specific logistic regression estimators trained on 427 ROSMAP prefrontal donors project Braak- and β-amyloid-like transcriptomic signatures onto the 108-region Siletti atlas, flagging brainstem, hypothalamic and rostral hippocampal regions.

Single-cell studies of Alzheimer's disease (AD) have mostly sampled cortex, and the only whole-adult-brain human snRNA-seq atlas (Siletti) has no neuropathology data, so the subcortical and brainstem sites that staging histology implicates early are largely unprofiled at cellular resolution. This work transfers donor-level pathology labels from cortical AD cohorts to every cell in the atlas, producing maps resolved by region, cell type and cortical layer without new tissue sampling.

Key results

  1. ShownForty-five cell type-specific estimators (5 neuropathology markers × 9 cell types) told apart cells from high- and low-pathology ROSMAP prefrontal donors better than chance. Their gene weights overlapped little with those of control models that predict age.

    On held-out donors (nested, donor-stratified cross-validation), the models reached above-chance receiver operating characteristic AUC (ROC-AUC), up to 0.87 for high vs low Braak stage in astrocytes. Gene weights correlated weakly with 9 age-estimator controls (maximum average Kendall τb 0.06, TDP-43 vs age). Performance depended on the cohort's sex mix, so cross-dataset work used males only. · Modelling neuropathology at single-cell…; Pathology inference model; Model level gene contribution

  2. ShownBraak estimators trained on prefrontal cortex transferred to SEA-AD middle temporal gyrus cells in every inferred cell type. β-amyloid (Abeta) estimators varied by cell type, and subtype-level Ex3 models lifted the weak Ex3 Abeta transfer above chance.

    Braak ROC-AUC ranged from 0.698±0.017 (vascular) to 0.931±0.002 (Ex2). The best Abeta model scored 0.606±0.004 (astrocytes); Ex3, oligodendrocyte and vascular Abeta models were below chance. Donor-mean scores vs the continuous pseudo-progression score: τb=0.44. All models beat a 1000-permutation label-shuffled null, and every Ex3 subtype model beat chance on ROC-AUC and area under the precision-recall curve (AUPRC). · Neocortex-to-neocortex validation of…; Carrying over pretrained pathology estimation…; Subtype specific pathology estimators

  3. ShownAcross six regions of the ROSMAP multi-region dataset, including entorhinal cortex, hippocampus and thalamus, Braak-proxy scores tracked region-specific reference pathology in most regions and cell types. Abeta-proxy scores did so only in prefrontal, midtemporal and entorhinal cortex.

    Braak: maximum Kendall τb 0.54 (immune cells); the exception was Ex2 in thalamus (−0.2). Abeta: weak or negative in angular gyrus, hippocampus and thalamus (τb<0 for astrocyte, Ex1 and inhibitory models). SHAP genes overlapped the differentially expressed genes reported earlier for this dataset in all six regions (hypergeometric p < 0.05). Prefrontal vs midtemporal SHAP concordance: adjusted R^2 0.85 (Braak). · Neocortex to non-neocortical validations of…; Gene-level contributions to neuropathology cell…; SEA-AD and ROSMAP multi-region dataset

  4. SuggestedAcross ~3 million cells from 108 Siletti regions, Braak-proxy signal concentrated in brainstem, hypothalamic, limbic thalamic and extended amygdala nuclei and rostral CA1-3, low in cerebellum and neocortex. Abeta-proxy also peaked in rostral CA1-3 but, unlike Braak, was high in cortex and low subcortically.

    FDR p<0.05 regions included periaqueductal gray, dorsal raphe, medullary nuclei, tuberal and mammillary hypothalamus, and mediodorsal and reuniens thalamus. 90th-percentile Braak scores peaked at 0.94 (spinal cord) and 0.75 (midbrain) in inhibitory neurons. Cerebral cortex maxima: 0.42 Braak (Opc) vs 0.84 Abeta (Ex3). Donor vs aggregate cosine similarity 0.63±0.05. Age-proxy control: similar across regions. · Mapping Alzheimer’s neuropathology across the…; Braak-pathology transcriptomic vulnerability map…; Transcriptomic signatures of β-amyloid plaque…

  5. SuggestedBraak-proxy high scorers were enriched for VGLUT2+ splatter neurons and astrocytes in subcortical territories, and for VGLUT1+ granule and CA1 pyramidal cells in hippocampus. In inferred neocortical layers, the Abeta-proxy peaked in L2 and L5-L6; astrocytes and microglia were also layer-biased.

    VGLUT2+ splatter neurons were 53% of spinal cord high scorers but 36% of splatter neurons there. Hippocampal high scorers were 15% granule cells and 13% CA1 pyramidal cells. Weakest Abeta layers: L1 and L4. Pseudo-vulnerable astrocytes: 35% in L2; microglia: 21% in L1 and 20% in white matter. 133 SHAP genes overlapped early-AD cerebrospinal fluid proteins (hypergeometric p = 2.5×10^−9). · Non-neuronal populations; Subcortical, deep brain neuronal populations; Hippocampal neuronal populations; Transcriptomic signatures of β-amyloid plaque…

Methods, data and reproducibility

Per-cell-type L2-regularized logistic regression models, trained on donor-binarized pathology labels in ROSMAP prefrontal snRNA-seq with nested donor-stratified cross-validation, were checked on SEA-AD and ROSMAP multi-region data, then applied to the Siletti atlas after classifier-based cell type and layer inference; interventional SHAP and pySCENIC regulons nominated driver genes and regulators.

  • ROSMAP-PFC: prefrontal cortex snRNA-seq with donor neuropathology (Braak, β-amyloid, CERAD, TDP-43, Lewy body); training set · human · n = 427 donors (212 male, 215 female); 2,359,994 nuclei
  • SEA-AD middle temporal gyrus snRNA-seq with reference pathology; external validation · human · n = 89 donors, 1,230,092 cells (validation: >500,000 cells, 36 donors)
  • ROSMAP-MR: six-region snRNA-seq (prefrontal, midtemporal, angular gyrus, entorhinal cortex, hippocampus, thalamus); cross-region validation · human · n = 48 donors, ~1.35 million nuclei; 679,252 cells analysed (24 male donors)
  • Siletti whole adult human brain snRNA-seq atlas (31 superclusters, 461 clusters); projection target · human · n = 3 donors, 3.369 million nuclei, 108 regions; 2.89 million cells retained
  • Jorstad layer-annotated cortical snRNA-seq (SMART-seq v4 and 10x Chromium); six-layer classifier training · human · n = 5 donors, 49,417 nuclei
  • Maynard annotated snRNA-seq; white- vs grey-matter classifier training
  • Published cerebrospinal fluid proteomics cohort (preclinical AD with elevated tau vs controls); comparison set for SHAP genes · human · n = 1,760 participants; 2,535 proteins measured (323 implicated)
  • Allen Human Reference Atlas 3D (2020) NIfTI volumes for brain-map visualisation · human (Allen Human Reference Atlas) · n = 84 atlas regions
Methods & tools
scikit-learn v1.7.2 (LogisticRegression, saga solver), scanpy v1.11.2, SHAP (interventional, median-centred), pySCENIC v0.12.1 (GRNBoost2, AUCell), Aerts lab human motif database v9, gseapy v1.1.9 / Enrichr (GO 2023, KEGG_2021_Human, SynGO_2024), SciPy v1.15.2 (weightedtau), Nilearn v0.12.1, Allen Human Reference Atlas 3D, Cell Ranger v3.0.2 (ROSMAP-PFC alignment, original dataset processing), MapMyCells
Data & accessions
https://download.alleninstitute.org/informatics-archive/allen_human_reference_atlas_3d_2020/version_1/examples/flip_annotation/https://resources.aertslab.org/cistarget/tf_lists/https://resources.aertslab.org/cistarget/motif2tf/https://maayanlab.cloud/Enrichr/#libraries
Reagents & environment
Versions stated: scanpy v1.11.2, scikit-learn v1.7.2, SciPy v1.15.2, pySCENIC v0.12.1, gseapy v1.1.9 (Enrichr accessed November 2025), Nilearn v0.12.1, motif database v9. No container or requirements file is described.
Reproducibility gaps
Code is not yet public: it is promised on GitHub upon publication and is available from the corresponding author on reasonable request. No accessions or download locations are given for the snRNA-seq datasets used.
Limitations (5)
  • Scores measure how similar a cell's transcriptome is to cells from high-pathology donors, not observed tangles or plaques. The authors also note that molecular vulnerability need not mean early degeneration. · stated by the authors
  • How well Abeta estimators transferred depended on cell type and region: several cell types were below chance in SEA-AD, and results were weak or negative in angular gyrus, hippocampus and thalamus. Vascular-cell models were dropped because they underperformed consistently. · stated by the authors
  • The authors describe the alternative, composition-adjusted region ranking as more restrictive: it may remove real regional differences along with supercluster composition. · stated by the authors
  • The map rests on three male Siletti donors with no matched pathology. Its region and cell rankings are supported by consistency across donors, an alternative composition-adjusted ranking and agreement with prior literature, not by pathology measured in those brains. Female brains are not covered. · noted in this summary
  • The laminar findings depend on a two-step layer classifier. Its white- vs grey-matter step averaged 0.79±0.34 accuracy across Jorstad subclasses (a wide spread), and the six-layer step's leave-one-donor-out performance appears only in a supplementary figure, not in the text. · noted in this summary

Who should read:Read this if you study regional or cell type vulnerability in Alzheimer's disease and want an interpretable way to score pathology-like transcriptomic states in snRNA-seq atlases that have no neuropathology data.

Reproducibility ledger

Nobody has reported trying to reproduce this paper yet. Did you replicate an experiment or run the code? Your report saves the next lab a week.

Sign in to file a reproducibility report.

Related papers

Similarity Map
SelectedSame fieldCross-field
THISChen '22Zhao '26Wang '21Kwon '23Zeng '23Dávila-Veld '21
74% semantic matchNeuroscience
Shuo Chen, Yuzhou Chang, Liangping Li+17 more·Acta Neuropathologica Communications·Dec 21, 2022· 143 citations