Volumetric analysis of imaging data
The brain MRI scans used in this study were obtained from two separate datasets. The first was from the AMBR study16. A second dataset from the CBA study17 was added to improve rigour and reproducibility. Both studies obtained 0.8-mm isotropic T1 magnetization-prepared rapid gradient echo and T2 sampling perfection with application-optimized contrasts using different flip-angle evolution sequences for segmentation and volumetric analysis of specific hypothalamic nuclei as described in ref. 15. Although several other high-resolution brain MRI datasets exist, the AMBR and CBA studies also obtained assessments of brain amyloid-β deposition and cognitive impairment. Differences in other aspects of the AMBR and CBA studies, such as cohort characteristics, were seen as valuable to improve the robustness and generalizability of this study. Test–retest reliability of the segmentation pipeline was assessed using the ICC ((2,1); two-way random effects, absolute agreement, single measurement) in eight participants with repeat scans in the CBA study. Following ref. 61, we classified ICC < 0.40 as poor and 0.40–0.59 as fair; we note that other schemes, such as ref. 62, place the poor–moderate boundary at 0.50. For structures with poor ICC, we later corroborate the volumetric findings with a complementary molecular modality. All associations reported here are based on bilateral (combined left and right) volumes; hemisphere-specific measures, which are inherently noisier, were examined only descriptively and did not form the basis of our conclusions.
To specifically assess the effect of AD on subregion volumes, participants were categorized on the basis of their amyloid-β and cognitive status. Amyloid status was determined from amyloid positron emission tomography imaging or CSF biomarker assessments when available using study-specific criteria: in AMBR, labelling was based on Knight Alzheimer Disease Research Center (ADRC) cut-offs for [carbon-11]-Pittsburgh compound B (C-11 PiB) standardized uptake value ratio (SUVR), [fluorine-18]-florbetapir (F-18 AV-45) SUVR, by visual read when SUVR data were not available, or by CSF test results. In the CBA study, labelling was based on provided centiloid values thresholded at 16.4, corresponding to the Knight ADRC C-11 PiB mean cortical SUVR cut-off of 1.42 (ref. 63). If amyloid-β status was positive any time before or up to 1 year after the MRI, it was labelled as positive. Conversely if amyloid-β status was negative any time after or up to 1 year before the MRI, it was labelled as negative. Otherwise amyloid-β status was considered not available (NA).
Cognitive status was labelled as ‘impaired’ or ‘unimpaired’ using study-specific criteria. In the AMBR study, labelling was based on the Clinical Dementia Rating scale (CDR), with CDR = 0 indicating cognitively unimpaired and CDR > 0 indicating cognitive impairment. A subject was labelled as impaired only if assessed as impaired within ±1 year of the MRI, and as unimpaired only if assessed as unimpaired at any time after the MRI or within 1 year before (otherwise coded as NA). In the CBA study, labelling was based on the timepoint-specific diagnosis code, with code 1 indicating cognitively unimpaired; codes 2, 4, 5 and 6 indicating evidence of cognitive impairment and code 3 (subjective impairment) coded as NA.
In both studies, individuals meeting criteria for both cognitive impairment and positive amyloid-β status were designated as the AD group. Conversely, participants without cognitive deficits and with negative amyloid-β status were classified as the non-AD (control) group. To minimize potential confounding from intermediate pathology profiles, only individuals clearly falling into these AD or non-AD categories were included in these specific comparative analyses. The cohort had a mean age of 63.4 ± 11.6 years (mean ± standard deviation) and included 60.5% women. Moreover, our dataset includes information directly relevant to AD, including 30.6% with abnormal cognitive status and 13.8% who were amyloid-β positive (Extended Data Fig. 1a). The mean total hypothalamic volume for this cohort was 736.2 ± 88.2 mm3.
Given that the dataset did not have predefined group structures for all desired comparisons, we used a GLM to assess the associations between hypothalamic subregion volumes (summed across left and right hemispheres) and various demographic and clinical variables.
Specifically, in the AMBR study (n = 202), we modelled (denoted as ‘~’) the natural logarithm of each subregion’s volume (log[subregion volume]) as a function of scaled age at MRI, sex, cognitive (cogn.) status, amyloid-β deposition status and other relevant covariates, where ‘+’ denotes additive main effects, ‘:’ denotes an interaction, ‘×’ denotes crossed terms (main effects plus their interaction), and ‘scale()’ denotes z-score standardization of a covariate:
$$\begin{array}{c}\log [{\rm{v}}{\rm{o}}{\rm{l}}{\rm{u}}{\rm{m}}{\rm{e}}]\sim {\rm{s}}{\rm{c}}{\rm{a}}{\rm{l}}{\rm{e}}({\rm{a}}{\rm{g}}{\rm{e}}\,{\rm{a}}{\rm{t}}\,{\rm{M}}{\rm{R}}{\rm{I}})+{\rm{s}}{\rm{e}}{\rm{x}}+{\rm{c}}{\rm{o}}{\rm{g}}{\rm{n}}.\,{\rm{s}}{\rm{t}}{\rm{a}}{\rm{t}}{\rm{u}}{\rm{s}}+{\rm{a}}{\rm{m}}{\rm{y}}{\rm{l}}{\rm{o}}{\rm{i}}{\rm{d}}.\,{\rm{s}}{\rm{t}}{\rm{a}}{\rm{t}}{\rm{u}}{\rm{s}}\\ +{\rm{c}}{\rm{o}}{\rm{g}}{\rm{n}}.\,{\rm{s}}{\rm{t}}{\rm{a}}{\rm{t}}{\rm{u}}{\rm{s}}:{\rm{a}}{\rm{m}}{\rm{y}}{\rm{l}}{\rm{o}}{\rm{i}}{\rm{d}}.\,{\rm{s}}{\rm{t}}{\rm{a}}{\rm{t}}{\rm{u}}{\rm{s}}+{\rm{s}}{\rm{c}}{\rm{a}}{\rm{l}}{\rm{e}}({\rm{B}}{\rm{M}}{\rm{I}})\\ +{\rm{s}}{\rm{c}}{\rm{a}}{\rm{l}}{\rm{e}}({\rm{H}}{\rm{g}}{\rm{b}}{\rm{A}}1{\rm{c}})+{\rm{s}}{\rm{c}}{\rm{a}}{\rm{l}}{\rm{e}}({\rm{E}}{\rm{p}}{\rm{w}}{\rm{o}}{\rm{r}}{\rm{t}}{\rm{h}}{\rm{\_}}{\rm{S}}{\rm{c}}{\rm{o}}{\rm{r}}{\rm{e}})+{\rm{s}}{\rm{c}}{\rm{a}}{\rm{l}}{\rm{e}}({\rm{I}}{\rm{C}}{\rm{V}})\end{array}$$
P values for individual variables from the GLM analyses were corrected for multiple comparisons using the Benjamini–Hochberg procedure to control the false discovery rate (FDR).
In the CBA study (n = 256), the metabolic and sleep covariates (body mass index (BMI)), HgbA1c and Epworth Sleepiness Score) were not available; we therefore modelled the natural logarithm of each subregion’s volume (log[subregion volume]) as a function of scaled age at MRI, sex, cognitive status, amyloid-β deposition status, their interaction and ICV:
$$\begin{array}{c}\log [{\rm{v}}{\rm{o}}{\rm{l}}{\rm{u}}{\rm{m}}{\rm{e}}]\sim {\rm{s}}{\rm{c}}{\rm{a}}{\rm{l}}{\rm{e}}({\rm{a}}{\rm{g}}{\rm{e}}\,{\rm{a}}{\rm{t}}\,{\rm{M}}{\rm{R}}{\rm{I}})+{\rm{s}}{\rm{e}}{\rm{x}}+{\rm{c}}{\rm{o}}{\rm{g}}{\rm{n}}.\,{\rm{s}}{\rm{t}}{\rm{a}}{\rm{t}}{\rm{u}}{\rm{s}}+{\rm{a}}{\rm{m}}{\rm{y}}{\rm{l}}{\rm{o}}{\rm{i}}{\rm{d}}.\,{\rm{s}}{\rm{t}}{\rm{a}}{\rm{t}}{\rm{u}}{\rm{s}}\\ +{\rm{c}}{\rm{o}}{\rm{g}}{\rm{n}}.\,{\rm{s}}{\rm{t}}{\rm{a}}{\rm{t}}{\rm{u}}{\rm{s}}:{\rm{a}}{\rm{m}}{\rm{y}}{\rm{l}}{\rm{o}}{\rm{i}}{\rm{d}}.\,{\rm{s}}{\rm{t}}{\rm{a}}{\rm{t}}{\rm{u}}{\rm{s}}+{\rm{s}}{\rm{c}}{\rm{a}}{\rm{l}}{\rm{e}}({\rm{I}}{\rm{C}}{\rm{V}})\end{array}$$
In the harmonized two-study pooled cohort (AMBR + CBA, n = 458), we modelled the natural logarithm of each subregion’s volume as a function of scaled age at MRI, sex, cognitive status, amyloid-β deposition status, their interaction, ICV and study of origin. The metabolic and sleep covariates above were not included here as they were not available across both studies:
$$\begin{array}{c}\log [{\rm{v}}{\rm{o}}{\rm{l}}{\rm{u}}{\rm{m}}{\rm{e}}]\sim {\rm{s}}{\rm{c}}{\rm{a}}{\rm{l}}{\rm{e}}({\rm{a}}{\rm{g}}{\rm{e}}\,{\rm{a}}{\rm{t}}\,{\rm{M}}{\rm{R}}{\rm{I}})+{\rm{s}}{\rm{e}}{\rm{x}}+{\rm{c}}{\rm{o}}{\rm{g}}{\rm{n}}.\,{\rm{s}}{\rm{t}}{\rm{a}}{\rm{t}}{\rm{u}}{\rm{s}}+{\rm{a}}{\rm{m}}{\rm{y}}{\rm{l}}{\rm{o}}{\rm{i}}{\rm{d}}.\,{\rm{s}}{\rm{t}}{\rm{a}}{\rm{t}}{\rm{u}}{\rm{s}}\\ +{\rm{c}}{\rm{o}}{\rm{g}}{\rm{n}}.\,{\rm{s}}{\rm{t}}{\rm{a}}{\rm{t}}{\rm{u}}{\rm{s}}:{\rm{a}}{\rm{m}}{\rm{y}}{\rm{l}}{\rm{o}}{\rm{i}}{\rm{d}}.\,{\rm{s}}{\rm{t}}{\rm{a}}{\rm{t}}{\rm{u}}{\rm{s}}+{\rm{s}}{\rm{c}}{\rm{a}}{\rm{l}}{\rm{e}}({\rm{I}}{\rm{C}}{\rm{V}})+{\rm{s}}{\rm{t}}{\rm{u}}{\rm{d}}{\rm{y}}\end{array}$$
To test whether the association between age and subregion volume differed by disease status, we fitted a separate GLM that included an age × group interaction term, where group distinguished cognitively normal, amyloid-negative individuals from all others:
$$\log [{\rm{volume}}]\sim {\rm{scale}}({\rm{age}}\,{\rm{at}}\,{\rm{MRI}})\times {\rm{AD}}{\rm{\_}}{\rm{group}}+{\rm{sex}}+{\rm{scale}}({\rm{ICV}})+{\rm{study}}$$
To compare subregion volumes between patients with AD and non-AD individuals while balancing demographic differences, we constructed a logistic regression model to estimate each participant’s propensity score for AD-group membership, using sex and age at MRI as covariates. Propensity-score matching was then performed using a nearest-neighbour algorithm, generating matched pairs of patients with AD and non-AD individuals with similar distributions of sex and age. This balanced these key covariates between groups for the subsequent volumetric comparisons.
To evaluate the predictive power of hypothalamic volumes in identifying AD and cognitive status, we conducted a series of binary classification tasks using receiver-operating-characteristic analysis. Three primary target conditions were defined: (1) cognitive status (impaired ‘A’ versus normal ‘N’), (2) amyloid status (positive ‘P’ versus negative ‘N’) and (3) a composite AD group (A+P+ versus all others). For each condition, we compared the diagnostic performance of two distinct feature sets: a multivariate model comprising 13 individual hypothalamic subregions (ARC, anterior hypothalamic area, dorsal periventricular nucleus, DMH, lateral hypothalamus, medial preoptic nucleusMPO, paraventricular nucleus, posterior hypothalamus, periventricular nucleus, SCN, supraoptic nucleus, tuberomammillary nucleus, ventromedial nucleus) and a univariate model using only the total hypothalamic volume. Classification was performed using logistic regression with balanced class weights to account for potential group size discrepancies. Model performance was validated using a fivefold stratified cross-validation approach. Features were standardized (z-score normalized) within each training fold before model fitting. The resulting area under the curve was calculated as the mean across cross-validation folds, with the standard deviation reported as a measure of model stability.
Given the central role of the hypothalamus in energy homeostasis and metabolic regulation, we examined whether BMI was associated with hypothalamic volume at the total and subregion level. BMI was not significantly associated with total hypothalamic volume (estimate ± s.e. = −0.005 ± 0.007, adjusted P = 0.505, Extended Data Fig. 2e and Supplementary Table 4) or with any bilateral subregion volume after FDR correction (all adjusted P > 0.086). In lateralized analyses, the left DMH showed a significant negative association with BMI (left DMH −0.042 ± 0.014, adjusted P = 0.014), although this was not observed on the right side (right DMH −0.010 ± 0.018, adjusted P = 0.664, Extended Data Fig. 2e and Supplementary Table 4).
Post-mortem human brain tissues
Unfixed, flash-frozen half-hypothalamus tissues were obtained from the Netherlands Brain Bank (NBB), Netherlands Institute for Neuroscience, Amsterdam (open access at www.brainbank.nl). All material has been collected from donors for or from whom a written informed consent for a brain autopsy and the use of the material and clinical information for research purposes had been obtained by the NBB.
This consent procedure was approved by the Ethical Committee of the VU University Medical Center (Vrije Universiteit, Amsterdam, the Netherlands), independently and before this study. Tissue blocks were stored at −80 °C.
The cohort comprised 14 donors: 7 control participants without dementia and 7 patients with AD (6 sporadic and 1 familial). Comorbidities were common in both groups and predominantly cardiovascular; control participants also included donors with malignancy (colorectal carcinoma, multiple myeloma, melanoma), whereas the AD group included one donor each with chronic obstructive pulmonary disease and bladder carcinoma. Causes of death included acute cardiac events and pulmonary embolism (control participants), euthanasia (2 control participants), end-stage malignancy (control participants) and respiratory insufficiency, aspiration pneumonia, cachexia, ileus and dehydration due to swallowing problems (AD patients). No donor had a documented intensive care unit admission; all deaths occurred in community, nursing home or palliative care settings. All key determinants of post-mortem gene expression quality (age, post-mortem interval, brain pH, RNA integrity number and agonal state estimated from clinical records) were comparable between groups (all P > 0.1 with Mann–Whitney U test). All samples showed high quality, with short post-mortem intervals of less than 9 h and RNA integrity number scores above 6.5. Full donor-level clinical and neuropathological details can be requested from the NBB using the NBB ID, which is available on request.
To overcome the input cell number limits of droplet-based single-cell sequencing (which typically captures tens of thousands of cells per sample) relative to the vast cellularity of the entire hypothalamus, we performed targeted macrodissection. In this initial study, we prioritized in-depth profiling of three subregions already implicated in ageing and AD over a shallower survey of the whole hypothalamus: the ARC, a hub for energy homeostasis housing pro-opiomelanocortin and agouti-related peptide neurons that integrate leptin and insulin signals64,65; the SCN, the master circadian pacemaker whose disruption underlies sleep–wake disturbances among the earliest non-cognitive symptoms of AD66,67; and the mammillary nucleus, a component of the Papez memory circuit disproportionately vulnerable to neurodegeneration in AD models10. Together, these subregions span the anterior–posterior extent of the hypothalamus and represent three of its main functional domains: metabolic control, circadian regulation, and memory circuitry. Focusing on three primary regions of interest, the ARC, SCN and mammillary nucleus, prevented the severe dilution of rare neuronal populations, ensuring sufficient sampling depth for downstream differential expression analysis. Macrodissection was performed on hemisected hypothalami within a Leica cryostat at the lowest temperature setting, with the tissue surrounded by dry ice. To maintain RNA integrity and prevent cross-contamination, the blade and all dissection tools were sanitized and treated with RNaseZap RNase Decontamination Solution between each sample.
Animals
Wild-type and heterozygous (8-month-old) 5xFAD-C57BL/6 female mice were obtained from The Jackson Laboratory. Mice were housed and used according to protocols approved by Brown University Institutional Animal Care and Use Committee and in accordance with institutional and national guidelines. Animals were housed at 21 ± 1 °C (70 ± 2 °F) with humidity from 50% to 70%. Animals were fed ad libitum Lab Diet 5010 chow and water. The light cycle was 12 h on 12 h off and lights were on between 7:00 and 19:00. Animals were exposed to male bedding 3 d before being euthanized to synchronize the oestrous cycle. Animals were euthanized at Zeitgeber time ZT2–ZT3. No statistical methods were used to predetermine sample size. Animals were not randomly allocated to experimental groups. Investigators were not blinded to group allocation during the experiments or outcome assessment.
Statistics and reproducibility
No statistical methods were used to predetermine sample sizes, but our donor numbers are comparable to those of established snRNA-seq studies of the human brain in AD, which typically profiled 3–10 donors per group68,69,70. Individual nuclei were excluded from analysis on the basis of quality control metrics (feature count and mitochondrial read count). Data collection and analysis were not performed blind to the conditions of the experiments.
snRNA-seq
Nuclei extraction was performed using the Nuclei PURE Prep kit (Millipore Sigma) according to the manufacturer’s instructions. All procedures were performed in the cold room before the 45-min spin down. Tissue was transferred using a transfer pipette into a refrigerated Dounce homogenizer with 5 ml of lysis solution following kit instructions. Tissue was homogenized with the Dounce B and the lysate was transferred into a 15-ml Falcon tube through a 40-μm filter. The sucrose purification step was performed with the modifications that half the volume of all reagents was used to account for the small tissue sample sizes, an SW34 rotor was used and samples were spun for 45 min at 30,000g (13,000 r.p.m.) at 4 °C.
Nuclei were counted using a hemocytometer and 10,000 nuclei per sample were loaded onto the Chromium Single Cell 3′ Chip (10x Genomics) and processed with the Chromium Controller (10x Genomics). Samples were sequenced at GENEWIZ on an Illumina NovaSeq, with a target of 50,000 reads per sample.
Quality control, data processing and analysis
Sequence alignment for human snRNA-seq data (including the dataset from ref. 22) was performed against the GRCh38 (2020) reference genome. These alignments were executed using Cell Ranger (10x Genomics; v.7.1.0 for human) with the –include-introns flag, essential for capturing nuclear RNA.
Following alignment, systematic background noise, such as ambient RNA, was removed from the count matrices using CellBender (v.0.3.0)71. For quality control, outlier cells were identified on the basis of established criteria. Specifically, cells were flagged as outliers if they exceeded 5 median absolute deviations for any of the following quality control metrics: log1p_total_counts, log1p_n_genes_by_counts or pct_counts_in_top_20_genes. In addition, cells with a mitochondrial transcript percentage of 10% or more were excluded.
Potential doublets (two or more cells captured as a single droplet) were then identified and filtered out using scDblFinder72. After these quality control steps, data from different individuals were integrated using single-cell variational inference23 to correct for batch effects. Subsequent data visualization and downstream analyses were performed using tools from the Seurat73 and Scanpy74 ecosystems.
Subregional samples were never pooled. Each donor contributed three tissue samples obtained by macrodissection enriching for distinct hypothalamic subregions (ARC, SCN, MN), yielding 42 samples across the 14 older donors. Each sample was collected into a separate tube and processed through an independent 10x Chromium channel, producing independent library preparations and therefore constituting a biologically and technically independent observation capturing distinct cellular compositions and independent sources of technical variation.
Differential expression analysis was performed using a pseudobulk approach with the LRT in edgeR75. Genes were significant if the Benjamini–Hochberg-adjusted P value was below 0.05. The three young reference donors22 were included only to anchor cell-type annotation and the microglial ageing trajectory; all AD-versus-control differentially expression analysis was restricted to the HypoAD cohort, excluding the familial AD (AD_44) donor.
GSEA was performed using the fgsea package76 with the hallmark gene-set lists and Kyoto Encyclopedia of Genes and Genomes (KEGG) and Reactome gene-set list from Molecular Signatures Database76,77. For GSEA input, genes within each cluster were ranked by their log2 fold change (log2[FC]) values, determined from edgeR LRT differentially expressed analysis. The GSEA was then performed using the fgseaMultilevel() command with default settings. Gene sets were considered significantly enriched if the adjusted P < 0.1.
Embedding analysis with UCE
To contextualize our data within a broader reference, we used a universal cell embedding (UCE) foundation model. UCE projects single-cell gene expression profiles into a universal embedding space that is agnostic to technical covariates such as batch effects or protocol differences78. We implemented the 33-layer version of UCE as specified in its official repository (https://github.com/snap-stanford/UCE), using Python (v.3.9.0) and PyTorch (v.2.1.1), with the publicly available pretrained model weights.
We then used UCE to project our control and AD snRNA-seq datasets into a shared embedding space alongside a publicly available, community-curated human brain atlas containing more than 25 million cells78. The resulting embeddings were visualized in a two-dimensional UMAP space to benchmark our newly generated data against this large-scale public reference.
Model training and feature ranking
To identify predictive features for the classification task distinguishing cells from AD versus control individuals, we developed and applied a machine learning pipeline using Python (v.3.10.9) and the scikit-learn library (v.1.6.1). The full pipeline is available through GitHub (https://github.com/Webb-Laboratory/HypoAD).
Data subsetting
For cell-type-specific models, the full dataset was filtered to include only cells corresponding to a single target cell type (for example, Neuron, Astrocyte), on the basis of predefined annotations.
Nested cross-validation
We used a leave-one-batch-out nested cross-validation strategy to ensure robust hyperparameter tuning, prevent data leakage and give an unbiased performance evaluation. For each of the seven outer folds, an inner sixfold cross-validation was used to tune the model’s hyperparameters. This approach generated seven distinct models, from which we ranked gene-associated weights to identify the most predictive features (Fig. 3a). Specifically, the data for all cells, as well as for each of the top five most abundant cell types, were split into seven outer folds on the basis of the sample batch of origin (leave-one-batch-out). Each batch comprised two donors (one control and one AD), and no donor appeared in more than one batch. For each outer fold, the test set consisted of nuclei from a single held-out batch, such that both donors in that batch were entirely unseen during training. The remaining six batches were used for an inner leave-one-batch-out cross-validation to select the optimal hyperparameters for an Elastic Net logistic regression model. The 5 × 5 hyperparameter grid included a logarithmic search for the regularization parameter C (from 10−2 to 102) and a linear search for the L1 ratio (from 0 to 1). The model with the best average inner-fold AUROC was then retrained on all six training batches and evaluated on the held-out test batch. We first trained a model on all cells, which achieved predictive performance (average AUROC = 0.75). We then trained models specific to the five most abundant cell types. All cell-type-specific models also performed well, with the microglia model showing the highest accuracy (average AUROC = 0.91). The comparatively lower performance for neuronal models may be attributable to the complexity and heterogeneity of neurons in the hypothalamus (Fig. 3b).
Feature weight extraction and ranking
After completing the outer cross-validation, we extracted the feature weights (coefficients) from the seven models, each trained on one outer fold. Two feature ranking lists were generated for each cell type. The ‘raw-weight’ rank was created by ranking all features on the basis of their mean weight across the seven folds, preserving the sign to indicate positive or negative prediction association. The ‘absolute-weight’ rank was created by ranking features on the basis of the mean of their absolute weights to measure overall predictive importance regardless of direction.
Shared and cell-type-specific features
To identify features consistently important across all cell types, we defined a ‘shared score’ for each gene as the mean of its absolute-weight rank across the five cell types analysed, and coloured by its direction (that is, whether the feature is on the top or bottom of the raw-weight rank list). Cell-type-specific ranks for features were defined by inquiry_rank/mean_rank_others.
GSEA
To identify biological pathways associated with our model’s predictions, we performed preranked GSEA using the gseapy (v.1.1.9) Python package. For each cell type, the full list of features was ranked using the mean raw weight. This ranked list was tested for enrichment against the KEGG pathway database (v.2021), Hallmark (v.2020) and Gene Ontology Biological Process (2021) from the Molecular Signatures Database. Significance was determined using a permutation-based test (1,000 permutations), and pathways with FDR q < 0.05 were considered significantly enriched.
Trajectory analysis using Monocle3
To infer the ageing process for the microglia cluster (n = 51,420 nuclei) generated in Seurat, we applied Monocle3 (refs. 43,44). Monocle3 typically uses dimensionality reduction to position cells in a lower-dimensional space, can correct for batch effects using mutual nearest-neighbour alignment and subsequently connects cells to construct trajectories in a semi-supervised manner.
For our analysis of the microglia–macrophage cluster, we directly used the integrated Seurat object, bypassing any further batch correction or dimensionality reduction steps within Monocle3 itself, as these had been addressed in the previous Seurat processing. After subsetting the microglia cluster, the root of the trajectory was programmatically defined by selecting the principal graph node most enriched for cells originating from young individuals. The horizontal axis in Fig. 4e represents a data-driven transcriptomic pseudotime learned from the single-cell data, not a chronological timeline, and individual donors do not necessarily follow this ordering. Trajectory-based expression analysis along this inferred trajectory was then performed using Moran’s I test, as implemented in Monocle3. Genes with a q value (FDR-adjusted P value) of less than 0.01 were considered significantly trajectory dependent, resulting in the identification of 13,104 such genes.
Functional enrichment analysis
Functional enrichment analysis of the 13,104 t-DEGs, grouped into individual expression groups identified through co-expression analysis, was performed using Enrichr (v.3.0)79. This analysis used the GO_Biological_Process_2018 gene-set database. Statistically significantly enriched Gene Ontology terms were identified on the basis of a Benjamini–Hochberg-adjusted P value (Padj) < 0.01. For visualization purposes, the top ten significantly enriched terms per group were shown in a dot plot.
To decode and visualize the dynamic expression patterns of the t-DEGs (n = 13,104) along the inferred pseudotime trajectory, we used the RVAgene package (v.1.0)48 in Python (v.3.9.6) with PyTorch (v.1.9.0). Before input into RVAgene, expression values for each gene were averaged within discrete pseudotime bins and subsequently rescaled to a range of [−1, 1]. The RVAgene neural network was configured with a symmetrical architecture comprising two hidden layers (48 nodes per layer) and two latent variable dimensions. The smoothed, reconstructed expression trajectories of the t-DEGs generated by RVAgene were then used to create the summary heatmap.
mTOR and ECM module scores
Two gene sets were assembled to score mTOR(PI3K-AKT-mTOR) signalling and ECM (matrisome) activity. Each set began from a curated list of canonical genes: for mTOR, core pathway components and downstream lipid–sterol effectors; for ECM, matrisome constituents, including collagens, laminins, proteoglycans, matrix metalloproteinases and their inhibitors (MMPs or TIMPs) and ADAMTS proteases. Each seed list was expanded by gene-symbol pattern matching (for example, COL[0–9] + A[0–9], MMP[0–9], LAM[ABC][0–9]) and filtered to remove unrelated ‘look-alike’ families (for example, mitochondrial import genes TIMM/TOMM, TIMELESS), then restricted to genes detected in the dataset, yielding 70 mTOR and 157 ECM genes (Supplementary Table 9, tabs mTOR_set and ECM_set).
Counts were normalized with Seurat’s NormalizeData (log-normalization, scale factor 10,000). Per-cell module scores were computed with Seurat’s AddModuleScore (default parameters: 24 expression bins, control pools of 100 randomly sampled genes per set gene), which subtracts the mean expression of a randomly selected control gene set from the mean expression of the target set. mTOR and ECM scores were calculated jointly (Score1 = mTOR, Score2 = ECM). For statistical testing, donors were grouped as young, control or AD. To avoid pseudoreplication, per-cell scores were aggregated to the sample level by taking the mean score across all cells of a given cell type within each sample, retaining only sample × cell-type combinations with ten or more cells. For each cell type and gene set, module scores were compared between control and AD samples using a two-sided Wilcoxon rank-sum test, and P values were adjusted for multiple comparisons across all cell type × gene-set tests using the Benjamini–Hochberg procedure. Analyses were performed in R with Seurat and rstatix.
Spatial mapping with cell2location
Cell2location is a Bayesian model that uses snRNA-seq cell-type signatures to infer cell types in Visium spatial transcriptomics by decomposing messenger RNA counts in each Visium voxel into cell types. We performed the three main steps in the cell2location workflow: estimating reference expression signatures of cell types using our dataset, mapping the learned cell-type signatures onto the slides, and performing downstream analysis. The code, model parameters and training evaluations can be found in the Jupyter notebooks in our GitHub repository. In brief, we used the default parameters to train the cell2location model.
Proteomic validation
Lumbar CSF samples from a cohort of 280 participants (including healthy control, AD and other conditions) were analysed by means of liquid chromatography–tandem mass spectrometry. Samples were processed on a Thermo EASY-nLC 1200 system coupled to a Thermo Orbitrap Eclipse Tribrid mass spectrometer. Detailed chromatography and data acquisition parameters follow the methods described in ref. 80. Quality control was maintained using inter-batch and inter-experiment samples across 96-well plates to evaluate and correct for technical effects. Protein quantification was performed by mapping raw spectra to an empirically corrected Prosit81 library (generated by means of gas-phase fractionated, narrow-window injections) using EncyclopeDIA, with peak integration and protein mapping finalized in Skyline.
Differential abundance analysis was performed to compare patients with AD (n = 19) against the healthy control group (n = 205) using the limma and eBayes framework82 on the preprocessed and log2 normalized protein quantification table in R. Statistical significance for differential protein abundance was defined by Padj < 0.05.
To validate cell-type-specific transcriptomic signatures at the protein level, significant CSF proteins were intersected with significant DEGs from single-cell pseudobulk analysis (Padj < 0.05). To address nomenclature inconsistencies between modalities, a ‘fuzzy-join’ logic was implemented. This approach matched protein symbols to RNA symbols on the basis of exact matches, prefix alignment and antisense transcript identification.
Validation of nine ageing-related and AD-related markers from the microglia pseudotime analysis (CSF1R, CX3CR1, P2RY12, APOE, B2M, CTSB, AXL, CLEC7A and TIMP2 in Fig. 4f) was performed by extracting their corresponding features from the CSF proteomics expression matrix using case-insensitive substring matching. To isolate age-related versus disease-related protein trajectories, the healthy control cohort was further stratified into ‘young’ (less than 65 years of age) and ‘aged’ (65 years or older) subgroups and visualized accordingly. Statistical analysis did not show significant differences between the groups, despite the protein trajectories’ showing trends similar to the gene expression data.
External validation in the PsychAD cohort
AD-associated transcriptomic signatures were validated against the PsychAD Consortium single-nucleus RNA-seq resource of the human dorsolateral prefrontal cortex (Mount Sinai Brain Bank cohort), obtained from CELLxGENE Discover (https://cellxgene.cziscience.com/collections/84ce6837-548d-4a1f-919f-0bc0d9a3952f). Two releases were used: an AD-enriched cohort for disease comparisons and a neurologically normal cohort spanning adulthood for the ageing analysis. For all analyses, Ensembl gene IDs were mapped to HUGO Gene Nomenclature Committee symbols using the feature_name field supplied with the dataset, and raw counts were normalized to 10,000 counts per cell and log1p-transformed.
For cross-cohort differential expression concordance, PsychAD cell types were mapped to their HypoAD counterparts at the main cell-type level. Non-neuronal populations (astrocytes, microglia, oligodendrocytes, OPCs and endothelial cells) were assigned direct one-to-one correspondences. Because neuronal subtypes are largely region-specific and the cortical subtypes annotated in PsychAD do not correspond one-to-one to hypothalamic populations, the finer PsychAD neuronal subtypes were aggregated into glutamatergic and GABAergic (GABA) lineages; for each gene, the subtype result with the lowest FDR was taken as the representative statistic for the lineage. Pseudobulk differential expression (AD versus control) was computed per cell type with edgeR and compared across datasets by plotting the paired log2[FC] of shared genes. Directional concordance of AD-associated changes was quantified with Spearman’s rank correlation (ρ) on the subset of genes significantly differentially expressed (FDR < 0.05) in both datasets.
For microglial DAM module scoring, module scores for homeostatic microglia (HEXB, CST3, CX3CR1, CSF1R, CTSS, TMSB4X, P2RY12, C1QB), DAM1 (TYROBP, CTSB, APOE, B2M, FTH1, TREM2) and DAM2 (TREM2, AXL, CST7, CTSL, LPL, CD9, CSF1, ITGAX, CLEC7A, LILRB4, TIMP2) were computed with scanpy.tl.score_genes (analogous to Seurat’s AddModuleScore). In the ageing cohort, scores were averaged per donor and then per age bin, correlated with age by Spearman’s ρ and shown with locally weighted scatterplot smoothing (bandwidth 0.6); individual representative genes were shown as scatter plots against donor age (locally weighted scatterplot smoothing bandwidth 0.5) and as violin plots comparing donors less than 65 years of age versus those 65 years of age or older. In the AD cohort, the same gene-level visualizations were produced across the four donor groups (young, control, AD and familial AD).
Ethics statement
AMBR-MRIs were obtained from studies involving positron emission tomography imaging that were approved by the Radioactive Drug Research Committee and the Institutional Review Board at Washington University in St. Louis. All participants provided written consent. CBA-MRI studies involving human participants were reviewed and approved by the Human Research Protection Office at the University of Pittsburgh. All participants provided written informed consent to participate17. All procedures and research protocols for post-mortem tissue were approved by the Institutional Review Board at Brown University. All research was performed in accordance with relevant guidelines and regulations. Informed consent for the use of post-mortem tissue was obtained from all donors or their next of kin.
Reporting summary
Further information on research design is available in the Nature Portfolio Reporting Summary linked to this article.