Personalized single-cell transcriptomics reveals molecular diversity in Alzheimer’s disease

Abstract

Alzheimer’s disease (AD) is highly heterogeneous and driven by diverse molecular and cellular mechanisms. Functional genomics investigates these mechanisms from genetic variants to gene expression and regulation. We performed personalized functional genomics analysis on population-scale single-nucleus RNA-seq data, with cross-cohort validation across multiple cohorts comprising over 1900 individual brains, capturing donor-level cell type interactions and gene regulatory networks. Using a knowledge-guided graph neural network, we learned latent representations of each donor’s functional genomics that accurately classified AD phenotypes, identified molecularly defined subpopulations, and traced disease progression trajectories. Our importance scores, derived from graph attentions, identified significant inter-donor differences and prioritized personalized cell type genes and regulatory networks. Finally, we identified gene regulatory QTLs (grQTLs) linking genetic variants to donor-level regulatory changes, providing insights into gene regulatory relationships beyond traditional eQTLs. All results are summarized into a personalized functional genomics atlas for AD, including an open-source framework, iBrainMap, for general use.

Personalized functional genomics atlas for Alzheimer’s disease that uses knowledge-guided graph neural networks to analyze donor-level functional genomics, identify disease subpopulations and trajectories, and link genetic variants to gene regulation.

Introduction

The human brain exhibits higher transcriptomic complexity than most other tissues^1,2, making it susceptible to a wide range of neurodegenerative and neuropsychiatric diseases. These diseases are heterogeneous, with a broad spectrum of symptoms, varying severity levels, and phenotypic differences^3,4 that stem from diverse cellular and molecular mechanisms^5,6. In the context of Alzheimer’s disease (AD), large-scale single-cell RNA-seq datasets across individuals (e.g., ROSMAP^7,8 and SEA-AD⁹) have led to the study of gene expression variation at the cell type (CT) level^7,8,10. These studies have expanded beyond binary case-control comparisons by incorporating detailed AD phenotypic stratification (e.g., resilience, pathology-cognition) and individual-level metrics. Complementing these studies, functional genomics has emerged to investigate gene regulatory mechanisms, linking genetic variants to changes in gene expression and downstream cellular functions underlying these diseases^2,11,12. For instance, gene regulatory networks (GRNs), which can model interactions between genes (transcription factor to target genes) at a CT level, have identified key genes and functions associated with AD phenotypes^13–15. Similarly, several population studies have shown that cell–cell communication is an essential component in AD (e.g., via CT interactions, such as astrocyte-microglia crosstalk in β-amyloid pathology¹⁶ and neuroinflammation¹⁷). However, how these interactions vary between individuals remains unclear. While many studies have pooled cells across individuals to identify regulatory mechanisms, fewer have aimed to construct individualized functional genomic models for AD. Here, we define “personalized functional genomics” as the modeling of cell type interactions and gene regulatory networks at the individual donor level, informed by each donor’s single-cell transcriptomic profile and known disease-related priors.

Large amounts of individual data are essential to capture the complexity and variability of functional genomics present across populations in AD. Recent developments in single-cell sequencing technology have enabled us to characterize cell diversity, reveal the genomic landscape, and understand disease heterogeneity of the human brain at single-cell resolution^18–20. To explore neurodegenerative and neuropsychiatric disease mechanisms, the PsychAD Consortium has generated population-level single-nucleus RNA sequencing (snRNA-seq) data from 1494 donors across 6.3 million nuclei from the human dorsolateral prefrontal cortex (DLPFC) brain region^21,22 (Supplementary Note 1). The dataset includes donors with neurodegenerative and neuropsychiatric disorders, such as AD, schizophrenia (SCZ), and diffuse Lewy body disease (DLBD), either as standalone diagnoses or in combinations of multiple diagnoses. For a subset of donors, detailed information on neuropsychiatric symptoms (NPSs), such as depression, is available. Additionally, quantitative measurements of AD-related phenotypes, including BRAAK stages (neurofibrillary tangles) and cognitive impairment, are available for all donors diagnosed with AD and the majority of neurotypical controls as well.

Analyzing such large-scale population data typically requires emerging computational approaches to discover personalized functional genomic information that traditional methods cannot capture. Several personalized analyses in brain diseases have used brain imaging and gene co-expression networks to capture individual-level variation^23–28. For example, a recent approach called Dozer²⁹ identified individual variations in immune response among AD using personalized gene co-expression network analysis. Another study identified personalized network patterns and driver genes in various cancers by building individual-specific networks from gene expression data²³. Other individual-specific network inference methods use a leave-one-out approach to model each sample’s contribution to an aggregate population network and generate sample-specific interactions^30,31. However, these approaches rely on correlation-based measures to infer co-expression patterns that may overlook gene regulatory interactions and lack the resolution to capture cell-type-specific interactions driving individual variation in AD. While traditional machine learning methods have provided valuable insights in identifying key genes and biomarkers and reconstructing cellular trajectories underlying AD and neurodegeneration using single-cell data^19,32,33, recent graph learning-based approaches have shown promise in modeling these complex interactions. For instance, approaches such as SIMBA³⁴, scGNN³⁵, and GLUE³⁶ learn cell embeddings which are then used for cell clustering^34,35, gene regulation inference^36,37, and multimodal integration^34,36,37. These approaches mainly focus on pooling cells from the population to learn embeddings and do not preserve intra-individual functional genomic interactions, such as cell-cell communication and gene regulatory networks.

To address this, we investigated population-scale single-cell transcriptomic data, primarily from the PsychAD cohort, and developed a graph learning framework to analyze personalized functional genomics. This approach advances the biological interpretability of prior network-based models mentioned above by capturing donor-level functional genomic patterns and prioritizing CTs, genes, and their interactions for AD and clinical phenotypes. We demonstrated the robustness of our results using independent cohorts. We also identified genetic variants associated with cell-type-level gene regulation for 27 emerging brain cell subclasses²¹. We provide a computational framework, iBrainMap, for general use, and our results are consolidated into a personalized functional genomic atlas accessible online using a webapp.

Results

Personalized functional genomic atlas for brain diseases

We present iBrainMap, a computational framework for analyzing and capturing functional genomic variations across donors by modeling CT interactions and GRNs from population-scale snRNA-seq data (e.g., PsychAD). This offers a refined alternative to pooled or group-averaged network approaches for phenotype classification, population subtyping, and prioritization of functional genomic information at the donor level (Fig. 1a). The iBrainMap framework first constructs a personalized functional genomic graph (PFG) for each donor, integrating CT interactions and GRNs. Each PFG is a directed graph, with nodes representing CTs, transcription factor genes (TFs), and target genes (TGs), and edges denoting CT interactions and CT regulatory links (TF to TG). We then enhance each PFG by incorporating prior biological knowledge of known disease genes (e.g., AD and SCZ genes^38,39) to model potential disease-relevant interactions, resulting in bio-diffused PFGs where edges that connect to disease gene nodes are assigned higher weights (Supplementary Note 2, Supplementary Fig 1). These bio-diffused PFGs are then used as input to our knowledge-guided graph neural network (KG-GNN) to classify AD vs. control and generate two key outputs for each donor: latent representations of functional genomics (graph embeddings) and importance scores for PFG nodes and edges (Fig. 1b, Supplementary Note 3, Supplementary Fig 2).

Fig. 1. Personalized functional genomic atlas and analysis for brain disease phenotypes.

a The iBrainMap framework uses the PsychAD21 snRNA-seq data of 1,494 human brain donors to construct personalized functional genomic graphs (PFGs), enabling donor-level analysis of CT interactions and gene regulation associated with brain diseases, including Alzheimer’s Disease (AD) and Schizophrenia (SCZ). The framework icon was modified from https://www.flaticon.com/free-icon/mind-map_2830784?related_id=2830912. b The iBrainMap framework first constructs a PFG and then applies graph diffusion to integrate known disease genes to generate bio-diffused PFGs for each donor. The KG-GNN, a multi-head graph attention model, uses the bio-diffused PFGs of each donor to learn graph embeddings for AD vs. control classification, while also prioritizing nodes and edges for AD (through learned AD prior, SCZ prior, and data-driven). Genotype data are used to associate genetic variants with gene regulation and to impute donor graph embeddings. These imputed embeddings can then be used in cohorts that only have genetic information for phenotype predictions. More details in Methods, Supplementary Notes 2-3, Supplementary Figs. 1, 2. c Using graph embeddings, we can classify donors across disease phenotypes and stratify them into potential disease subtypes. Using prioritized edges, we can prioritize phenotype-associated personalized networks, identify potential disease genes, and identify cell-type-level gene-regulation QTLs (grQTLs) that link associated SNPs with gene regulatory links. Box plots show the median (center line), the interquartile range (box limits correspond to the 25th and 75th percentiles), and whiskers extending to the smallest and largest data points within 1.5 × the interquartile range. Points beyond the whiskers are plotted as individual observations. Created in BioRender. Lab, D. (2026) https://BioRender.com/bksrjwv.

Using graph embeddings, we can classify donors according to disease-related phenotypes, including case-control status, pathology capturing different stages of AD, and presence of NPSs. Additionally, we can perform population subtyping to identify disease subgroups capturing different disease stages. With pseudotime analyses of the graph embeddings, we can uncover several population trajectories associated with AD pathology, cognitive status, and NPSs (Fig. 1c, left). The importance scores can help prioritize disease-associated functional genomics for each donor. These prioritized nodes and edges can then be used to identify significant subnetworks and uncover potential biomarkers associated with AD phenotypes, including CTs and regulatory elements (TFs, TGs). Additionally, the genetic variants can be associated with CT gene regulatory network changes across donors, i.e., gene regulatory QTLs (grQTLs), providing functional genomic insights linking SNPs to TF-TG regulatory mechanisms as opposed to linking SNPs to genes in traditional QTL analysis (Fig. 1c, right).

Finally, we used the PsychAD data to build the framework and independently validated our results using external datasets, ROSMAP⁸ and SEA-AD⁹. The PsychAD donors were obtained from Mount Sinai NIH Brain Bank and Tissue Repository (MSSM, n = 1042 donors), NIMH-IRP Human Brain Collection Core (HBCC, n = 300 donors), and Rush AD Center (RADC, n = 152 donors). The phenotypes of these donors were categorized into three levels: disease vs. control, AD progression, and NPS (Supplementary Note 4, Fig. 2a). An illustrative example of a personalized functional graph (PFG) is shown in Fig. 2b. Summary statistics of nodes and edges across all donors in PsychAD are presented in Supplementary Fig 3. The median number of nodes and edges per node remains consistent across all CT GRNs, except for microglia and oligodendrocytes, and CT-CT interactions (Supplementary Fig 3a). Notably, approximately 10% of each donor’s edges were unique, while nearly 80% were shared among more than eight donors (Supplementary Fig 3b).

Fig. 2. Multi-cohort snRNA-seq data for personalized functional genomic analyses.

a Donors from multi-cohort AD studies: PsychAD (Mount Sinai Brain Bank/MSSM (M) (n = 1042), Rush Alzheimer’s Disease Center/RADC (R) (n = 152), Seattle Alzheimer’s Disease Brain Cell Atlas/SEA-AD (S) (n = 80), and NIMH-IRP Human Brain Collection Core/HBCC (H) (n = 300). Data summary of phenotypic donors used in this study divided into three levels: disease vs. control (AD vs. control, SCZ vs. control, AD-DLBD vs. control, AD-resilient vs. AD-strict vs. Control), Disease Progression (BRAAK stages, CERAD, Cogdx), and Neuropsychiatric symptoms (Dysphoria, DecInt (Anhedonia), Sleep/Weight Gain/Guilt/Suicide (S/WG/G/S), Depression/Mood (D/M)); each horizontal bar is a phenotype contrast. b Part of a personalized functional genomic graph (PFG) for a donor from the Mount Sinai Brain Bank/MSSM (M) cohort. The donor is female with European ancestry, age >80 years old, diagnosed with AD, showing pathology for BRAAK stage 6 and CERAD score 4 (definite AD), and diagnosed with Dementia (Supplementary Note 4). Nodes can be cell types, transcription factors (TFs), and target genes (TGs); edges represent cell type interactions and cell-type-specific regulatory links (TF to TG).

Classifying and subtyping phenotypes via functional genomics

We first trained the KG-GNN model for AD vs. control classification using donor PFGs by learning their graph embeddings. Here, we used five-fold cross-validation to identify the optimal hyperparameter configuration (Supplementary Fig 4a, b, Supplementary Table 1). We trained the final model using the optimal hyperparameters on a subset of donors from the AD (n = 438) vs. control (n = 179) contrast in MSSM cohort (Supplementary Note 5). Our KG-GNN model achieved high classification performance, with an area under the curve (AUC) of 0.913 on the held-out donors from MSSM (AD: n = 62 vs. control: n = 30), and a combined AUC of 0.805 on independent donors from RADC and SEA-AD cohorts (AD: n = 93 vs. control: n = 68) (Fig. 3a, Supplementary Fig 5a). We also tested our model on independent ROSMAP data (Supplementary Fig 5c). Our model classified early (n = 59) vs. late (n = 84) BRAAK with an AUC of 0.788 (Supplementary Fig 5d). To further assess the robustness of our PFGs, we performed a permutation test by shuffling the class label assigned to the graphs and retrianing the model, repeating this procedure ten times. The KG-GNN model consistently outperformed the randomized models, achieving a significantly higher classification accuracy (BACC: 0.87 vs. 0.5035; p < 1.46e-9, one-sample t-test).

Fig. 3. Graph embeddings for personalized functional genomics enable phenotype classification and subtyping.

a ROC curves for classifying AD vs. control using KG-GNN graph embeddings from the MSSM held-out: AD (n = 62) vs. control (n = 30) and the independent validation datasets: RADC + SEA-AD: AD (n = 93) vs. control (n = 68) and ROSMAP: ROSMAP: Early (n = 59) vs late BRAAK (n = 84). b Donor PFGs from all cohorts are input into the pre-trained KG-GNN model to extract learned graph embeddings, which are then used for classifying additional phenotypes, either as binary or multi-class classification tasks. The iBrainMap framework icon was modified from https://www.flaticon.com/free-icon/mind-map_2830784?related_id=2830912. c Average five-fold cross-validated ROC for classification of KG-GNN graph embeddings across phenotype contrasts; DecInt: Anhedonia, S/WG/G/S: Sleep/Weight Gain/Guilt/Suicide, D/M: Depression/Mood. Box plots show the median (center line), the interquartile range (box limits correspond to the 25th and 75th percentiles), and whiskers extending to the smallest and largest data points within 1.5 × the interquartile range. Points beyond the whiskers are plotted as individual observations. d UMAP based on graph embeddings (Left: colored by AD vs. controls; Right: colored by clusters from unsupervised clustering/subtyping of graph embeddings). These embeddings reflect personalized patterns derived from functional genomic graphs, capturing latent biological variation not evident in bulk or pooled analyses. e Heatmaps depict phenotype enrichments for clusters from (c) across AD progressions: AD pathology (BRAAK and CERAD) and cognitive status (clinical dementia rating), showing an increasing trend from c1-c5, using hypergeometric test. All dataset details are described in Fig. 2a and Supplementary Note 4.

We benchmarked our KG-GNN model against traditional machine learning models, state-of-the-art graph learning models (GAT and GCN), and two recent graph embedding approaches (SIMBA³⁴ and scGNN³⁵, Supplementary Note 6). We outperformed traditional machine learning models, which were trained on average CT gene expression (Supplementary Fig 6a). KG-GNN also outperformed other graph learning methods, achieving a high AUC score (Supplementary Fig 4c) and recent graph-embedding methods (Supplementary Fig 6b). These results highlight the importance of modeling the complex relationships between CTs and genes for disease phenotype prediction. We further evaluated the key components of the KG-GNN model through ablation study (Supplementary Note 7). Varying the number of genes for GRN construction showed that larger gene sets improved held-out performance but reduced generalization, while the gene set used in the original model achieved the best overall performance (Supplementary Fig 7a). In addition, varying the number of genes (highly variable genes) used as node features showed that performance improved as more genes were included (Supplementary Fig 7b). Incorporating knowledge-guided graph diffusion also improved accuracy compared to no diffusion and random diffusion models (Supplementary Fig 7c). We then compared our PFGs with CT coexpression networks constructed using Dozer²⁹ and CT protein–protein interaction (PPI) networks derived by intersecting CT co-expression networks with known PPIs from STRINGdb⁴⁰ and BioGRID⁴¹. Our PFGs consistently showed higher performance on both held-out and independent datasets (Supplementary Fig 7d). We also evaluated the importance of different biological links. We modified the model by selectively removing specific edge types. Specifically, we retained the model with the following modification: (1) Only CCI edges, (2) CCIs and CT-TF links, and (3) CCIs, CT-TF, and CT-TG links. In all cases, model performance declined on held-out samples (Supplementary Fig 8).

Next, we applied the pre-trained KG-GNN model to PFGs from donors spanning different phenotypes beyond AD vs. control to extract graph embeddings (Fig. 3b, Fig. 2a). We found that these embeddings achieved good classification performance on additional phenotypes and NPSs, using standard machine learning models for both binary (disease vs. controls and NPSs) and multi-class tasks (disease progression and AD-resilience) (Fig. 3c, Supplementary Note 8, Supplementary Fig 5b). For binary classification, SCZ showed high performance (AUC = 0.83), while NPS contrasts, such as S/WG/G/S (AUC = 0.67) and D/M (AUC = 0.61), achieved moderate accuracy. Other NPS contrasts yielded relatively lower yet comparable results. For multi-class classification, performance was above baseline across all four contrasts, including AD-resilience (AUC = 0.87), BRAAK (AUC = 0.79), CERAD (AUC = 0.75), and Cogdx (AUC = 0.79), though not as high as AD classification (Fig. 3c, Supplementary Fig 9a). We also evaluated SCZ classification on an independent cohort from HBCC (SCZ: n = 50 vs. Control: n = 250), where performance was lower than the earlier AD and SCZ classifiers, suggesting the need for training disease-specific models (Supplementary Fig 10, Discussion).

Our graph embeddings distinguished AD and control donors more effectively than gene expression, other state-of-the-art methods, and original PFGs (Supplementary Fig 11). We further explored potential population subtypes using our graph embeddings through unsupervised clustering and identified five distinct clusters (c1-c5, Fig. 3d). These clusters showed clear trends across AD-related phenotypes, including BRAAK stage, cognition (Clinical Dementia Rating or CDRscore), and CERAD (Fig. 3e). For example, c1-c2 were enriched for early BRAAK stages (0-2), controls in CDRscore, and ‘No AD’ based on CERAD; while c4-c5 were enriched for late BRAAK stages (5-6), Dementia (CDRscore), and presence of AD (CERAD). This suggests that our pre-trained model can identify meaningful donor subpopulations beyond AD vs. control. To assess the robustness of these trends, we varied the number of clusters and observed consistent phenotype enrichment patterns across different settings, supported by clustering quality metrics, such as Davies-Bouldin and Silhouette scores (Supplementary Fig 12).

Population trajectories for AD progression and symptoms

AD progression exhibits heterogeneity across donors, particularly in age of onset, rate of cognitive decline, and worsening of disease pathology^42–44. Capturing personalized disease mechanisms can therefore provide insights into disease trajectories at a population level. In this study, we used donor graph embeddings, extracted from our pre-trained KG-GNN model, to infer population-level trajectories for AD phenotypes by computing phenotypic pseudotimes across donors (Fig. 4a, Methods). Specifically, these graph embeddings allowed us to temporally order donors by assigning pseudotimes based on their phenotypes (e.g., AD, BRAAK, CERAD).

Fig. 4. Population-level pseudotime analysis uncovers phenotypic pseudotimes for AD progression, cognition, and NPS using pre-trained KG-GNN model.

a Graph embeddings were extracted for donors with different phenotypic information using the pre-trained KG-GNN model. Then, pseudotime analysis of graph embeddings is performed to compute phenotypic pseudotimes. The iBrainMap framework icon was modified from https://www.flaticon.com/free-icon/mind-map_2830784?related_id=2830912. b AD phenotypic trajectory captured for graph embeddings of donors with AD vs. controls, c Boxplot comparing pseudotimes from (b). shows controls (n = 179) appearing earlier compared to donors with AD (n = 438) having a later occurrence (p < 3.17e-127, two-sided Mann-Whitney), Box plots show the median (center line), the interquartile range (box limits correspond to the 25 and 75th percentiles), and whiskers extending to the smallest and largest data points within 1.5 × the interquartile range. Points beyond the whiskers are plotted as individual observations. d Regression plots comparing AD trajectory pseudotimes (from (a)) with AD pathology (Plaque, r2 = 0.684, p < 7.01e-106, two-sided t-test) and cognition (Clinical Dementia Rating score, r2 = 0.856, p < 1.81e-181), computed using ordinary least squares (OLS), e Boxplots showing pseudotimes computed from graph embeddings for donors with AD pathology and cognition reveal an increasing trend with stage progression; left: BRAAK stages = early vs. mid vs. late (p < 2.26e-16, n = 541), middle: CDRScore = Control vs. Mild Cognitive Impairment (MCI) vs. Dementia (p < 2.26e-16, n = 517), right: CERAD=No AD vs. Possible vs. Probable vs. Definite AD (p < 2.26e-16, n = 541); inset plots: UMAPs based on graph embeddings colored by pseudotime. Statistical significance (p-values) was calculated using two-sided Mann-Kendall. Box plots show the median (center line), the interquartile range (box limits correspond to the 25 and 75th percentiles), and whiskers extending to the smallest and largest data points within 1.5 × the interquartile range. Points beyond the whiskers are plotted as individual observations. f Density plots showing the distribution of neuropsychiatric symptoms (NPS) across donors diagnosed with NPS, including Dysphoria (p < 4.21e-3, n = 316), Early Insomnia (p < 1.30e-3, n = 233), and Middle Insomnia (p < 4.21e-3, n = 232) across NPS phenotypic pseudotimes. Statistical significance (p-values) was calculated using unpaired two-sided t-test.

We first inferred a trajectory for all donors from the MSSM AD vs. control contrast, assigning an AD pseudotime to each donor (Fig. 4b). AD donors were assigned later pseudotimes compared to the controls (p < 3.17e-3, two-sided Mann-Whitney, Fig. 4c). Furthermore, when comparing this AD pseudotime against Plaque and CDR using ordinary least squares (OLS), we observed a clear progression trend, with increasing plaque levels (r2 = 0.684, p < 7.01e-106) and higher CDRscores (r2 = 0.856, p < 1.81e-181) associated with later pseudotimes (Fig. 4d). To further explore AD phenotypic pseudotimes, we inferred trajectories for progression and cognition contrasts like BRAAK, CERAD, and CDRScore (Fig. 4e). BRAAK staging scores range from 0-6 and were grouped into three stages: early (0-2), mid (3-4), and late (5-6). Further details of the grouping are available in Supplementary Note 4. Across all three phenotypes, we observed a consistent trend of increasing pseudotime associated with more advanced disease stages (p < 2.22e-16, p < 2.22e-16, p < 2.22e-16, two-sided Mann-Kendall, respectively). This suggests that our donor graph embeddings can effectively capture phenotypic trajectories, as shown by the alignment of increasing AD pseudotime with more advanced stages of AD.

NPSs often co-occur in AD donors, with different symptoms manifesting at distinct stages of the disease⁴⁵. For instance, dysphoria is reported to be more prevalent during the early stages of cognitive impairment^45,46. Consistent with previous findings, our analysis revealed that pseudotimes inferred from graph embeddings of AD donors diagnosed with dysphoria were closer to zero or initial stages of AD progression (Fig. 4f:left). Similarly, pseudotime analysis of AD donors diagnosed with early- and mid-insomnia revealed that these sleep disturbances occur along different stages as AD progresses (Fig. 4f:mid, right, Supplementary Fig 13).

We validated our phenotypic trajectories by extending our analysis to a sufficiently large independent SEA-AD dataset. We first constructed the bio-diffused PFGs for 80 donors from SEA-AD and extracted their graph embeddings using our pre-trained KG-GNN model (Fig. 5a). We then compared the pseudotimes inferred from these embeddings with the continuous pseudoprogression score (CPS)²⁰, which used a machine-learning model on quantitative neuropathology measurements and immunohistochemical stains. Interestingly, the two measures had strong concordance, with a Jensen-Shannon similarity of 0.77, defined as 1 – Jensen-Shannon Distance⁴⁷ (Fig. 5b). Furthermore, the pseudotimes effectively captured disease severity stages in AD progression (Thal⁴⁸, CERAD) and cognition (Cogdx) (Fig. 5c–e).

Fig. 5. Independent validation of phenotypic trajectories by SEA-AD cohort.

a Graph embeddings were extracted for donors with diverse phenotypes using a pre-trained KG-GNN model to compute phenotypic pseudotimes. b Kernel density estimate (KDE) plot comparing pseudotime based on graph embeddings with Continuous Pseudoprogression Score (Jensen-Shannon similarity = 0.77). c–e Boxplots showing increasing pseudotime from (a) along AD pathology for Thal (p < 1.44e-2, two-sided Mann-Kendall, n = 80 donors) and CERAD (p < 1.17e-2, two-sided Mann-Kendall, n = 80 donors), and cognition (Cogdx, p < 1.61e-2, two-sided Mann-Kendall, n = 80 donors).

Prioritizing donor-level CTs, genes, and networks

Graph embeddings improved phenotype classification and provided insights into disease progression (e.g., population trajectories). To further understand how CTs, genes, and their interactions drive phenotype classification, we calculated node and edge importance scores for each donor (Methods). The edge importance scores are defined by the attention weights learned by the model for each CT interaction and gene regulatory link (TF to TG). The node importance scores are derived from the edge importance scores for each gene and CT (Supplementary Note 9). These importance scores indicate the contribution of nodes and edges to AD vs. control classification. Compared with gene expression, we found a significantly higher correlation between our importance score and various clinical phenotypes (Supplementary Table 2). For example, gene expression of the GAD2 gene in IN_VIP CT was not correlated with PRS, while importance scores showed a positive correlation among AD donors and a negative correlation among control donors (Supplementary Fig 14). Both GAD2 and IN_VIP have been reported to be downregulated in AD in previous studies⁴⁹.

We compared edge importance scores (based on AD-prior, SCZ-prior, and data-driven) of CT gene regulatory links between AD and control donors. While some significantly differential edges were consistently prioritized across all three priors, several edges were uniquely prioritized by each prior (Fig. 6a, Supplementary Fig 15). For example, TF IRF8 regulatory interactions in microglia had significantly higher data-driven importance scores in controls (IRF8 → RASGEF1C: p < 7e-3; IRF8 → LILRB1: p < 9e-4) but not for AD-prior and SCZ-prior. Changes in IRF8, a crucial microglial TF for AD, can cause abnormal expression of many AD-related genes⁵⁰. Similarly, some of the interactions for ZEB1 TF (ZEB1 → SOX5; OPC: ZEB1 → NTNG1) were only significantly differential for SCZ-prior. Several other interactions, such as those involving ZEB1 (OPC: ZEB1 → SEMA5A; OPC: ZEB1 → LAMA4) and TCF7L2 (Astro: TCF7L2 → BCL6; Astro: TCF7L2 → PDGFC), were consistently differential across all three importance scores. ZEB1, which had higher importance scores in AD donors, plays a major role in epigenetic regulation in AD⁵¹. TCF7L2 is associated with the Wnt/β-catenin signaling pathway, which plays a crucial role in AD⁵². A complete list of all the regulatory links with importance scores is available in Supplementary Data 1.

Fig. 6. Donor-level prioritization of cell type interactions, genes, and regulatory networks for AD.

a Heatmap showing differences in edge importance scores for different gene regulatory links (TF-TG) across AD and control donors (two-sided t-test; * p < 0.05, ** p < 0.01, *** p < 0.001) from different biological priors: AD, SCZ, and data-driven. b Plot showing cell type importance for AD (n = 438 donors). Dotted lines represent cell type fraction, and solid lines indicate cell type importance score. The plot is sorted by cell type fraction. The vertical lines represent the error bars indicating mean ± SEM. c Volcano plot of log-fold change against significance for cell type interactions. Blue: control; red: AD. Statistical significance (p-values) was calculated using unpaired two-sided t-test. d Prioritized subnetwork (top 1% edges across cell types within each donor) showcasing connectivity among cell types, TFs, and target genes for AD phenotype. Each circle represents a gene, shaded along a yellow to red gradient corresponding to the frequency of occurrence in prioritized nodes across donors. Edges represent regulatory links between TFs and target genes and are colored uniquely for each cell type as shown in the key below. Borders of genes linked to AD in the literature are shaded orange and labeled. Names of all other genes are hidden for ease of visualization.

We examined which CTs were prioritized among AD donors and found that neuronal subtypes consistently had the highest importance scores (Fig. 6b). For example, EN_L5_6_NP, EN_L6_IT_2, and IN_LAMP5_LHX6 were the top three CTs with the highest importance scores among AD donors. Previous transcriptomic studies have shown that the intratelencephalic neuron (IT) is associated with cognitive resilience in AD⁵³. The Lamp5 LHX6 inhibitory neuron (EN_LAMP5_LHX6) is affected by aging and associated with AD^54,55. We also observed that our model prioritized CTs even when they had relatively low cell fractions. Furthermore, differential analysis of the edge importance scores for CT interactions identified several AD-associated CT interactions. Notably, interactions originating from microglia were higher in control donors, whereas those involving IN_SST neurons as the source CT were higher in AD donors (Fig. 6c). Microglia is known to have many effects on AD pathology and therapeutic intervention, which corroborates their high importance⁵⁶. Somatostatin inhibitory neurons (IN_SST) are also linked to memory loss, and SST expression is reduced among AD patients⁵⁷.

We then combined the top-prioritized regulatory links from each donor (top 1% based on our edge importance scores) into a consensus functional genomic subnetwork (Fig. 6d). The subnetwork segregated the edges into three main clusters consisting mainly of oligodendrocytes, astrocytes, and excitatory neurons (Supplementary Fig 16), indicating that our model consistently prioritized regulatory links in these cells across donors. Oligodendrocytes are responsible for myelination, protecting neurons from damage, and dysfunction in oligodendrocytes leads to demyelination, a major factor in AD progression^58,59. Astrocytes play a critical role in AD pathogenesis by contributing to neuroinflammation, synaptic dysfunction, and amyloid-β (Aβ) plaque clearance^60–62. Excitatory neurons are particularly vulnerable in AD, exhibiting early synaptic dysfunction, loss of dendritic spines, and impaired glutamatergic signaling, which contribute to cognitive decline⁶³. Other single-cell transcriptomic analyses revealed distinct molecular alterations in excitatory neuron subtypes in AD¹⁸, astrocyte subclusters exhibiting varying response to neuropathology in AD progression^7,13, and AD-associated oligodendrocyte subpopulations with altered gene expression related to myelination and cellular responses⁶⁴. This subnetwork recapitulates several biologically relevant features of our model, with a statistically significant overrepresentation of known AD genes (p < 9.80e-4, hypergeometric test). For instance, PCDH11X, an aging-related gene associated with late-onset AD in GWAS⁶⁵, is prioritized in excitatory neurons in our model. Cross-verifying with an external independent dataset¹⁴ confirms PCDH11X predominant expression in excitatory neurons (Supplementary Fig 17). A previous study identified variants in the CR1 gene associated with AD⁶⁶. Our model suggests that FOSB mediates CR1 regulation, and this regulation is relevant for AD PVM cells. Overall, our model effectively prioritized a subnetwork that recapitulates known AD biology and suggests cell-type-specific regulation of several genes with prior evidence. Thus, other genes in the prioritized subnetwork are likely candidates that have remained elusive thus far.

Gene regulatory variations of phenotypes and subpopulations

As our importance scores identified several AD-related CTs, genes, and their interactions, we extended our analysis to additional phenotypes including AD progression, cognition, and ancestry. We examined differential gene importance scores for cross-disorder phenotypes (AD vs. SCZ) and AD progression (BRAAK). Several known SCZ TFs from inhibitory neurons (IN) and excitatory neurons (EN), like RBFOX2⁶⁷ (EN_L2_3_IT: p < 1.538e-07), ZMAT4⁶⁸ (IN_PVALB: p < 0.001), FOSB⁶⁹ (IN_SST: p < 0.018), and NR2F6⁷⁰ (IN_LAMP5_LHX6: p < 0.044) had significantly higher importance scores in SCZ donors compared to AD donors (Fig. 7a, Supplementary Data 2). These genes were also differentially expressed in PsychENCODE¹⁴. For BRAAK phenotype, we identified several TFs from IN CTs with higher scores in early BRAAK stages and TFs from astrocytes and microglia for late BRAAK stages (Fig. 7a). We also performed differential analysis of our gene importance scores between AD strict and AD resilience donors and found several known AD/neuropsychiatric genes that had significantly higher importance scores for AD resilience (e.g., PKNOX2: p < 0.009; TENM4: p < 0.012; ZMAT4: p < 0.024; DLGAP1: p < 0.032; CUX2: p < 0.033), which were also significantly differentially expressed for resilience in another study¹⁰.

Fig. 7. Subpopulation variation of prioritized transcription factors and target genes across brain diseases and clinical phenotypes using our importance scores.

a Differential gene importance scores between cross disorder (SCZ vs. AD) and AD progression (early vs. mid vs. late BRAAK). The x-axis represents cell types, and the y-axis represents the transcription factors. The dot size indicates the significance of the differential analysis (-log₁₀(p-value). The color of the dot is represented by log fold change (logFC). Red denotes positive log fold change and blue denotes negative fold change. Statistical significance (p-values) was calculated using unpaired two-sided t-test. b ANOVA comparison (two-sided Mann-Whitney-Wilcoxon test) of edge importance score distributions for select gene regulatory links for donors with AD progression phenotypes (n = 510 donors). Violin plots contain miniature boxplots indicating the mean (white circle), first to third quartile (black line), and whiskers with lengths 1.5 times the interquartile range. Significance is notated by *: p < 5.00e-2, **: p < 1.00e-2, ***: p < 1.00e-3, ****: p < 1.00e-4. c Histogram of high-scoring edges across the whole population. Bar plots for corresponding GO-term enrichment of genes in the resultant highly conserved (frequency > 50, orange), conserved (frequency = 23-–50, purple), and lowly conserved (frequency = 15–22, green) edge groups. The remaining edges are not conserved (frequency <15, gray). Only the top 20% of the histogram is shown. Statistical significance (p-values) was generated using a hypergeometric test with benjamini hochberg correction on 300 sampled genes from each edge group using Metascape74. d Functional enrichment of prioritized gene sets from EUR (n = 792 donors) and non-EUR (AFR, n = 120 donors, and AMR, n = 99 donors) ancestries (Methods).

We then examined differences in edge importance score distribution across AD progression phenotypes, particularly BRAAK stages, CDR score, and CERAD score (Fig. 7b). We found several significant CT regulatory links, like NFATc2 → IKZF1 in microglia, which displayed significant differences in distribution across all BRAAK stages (e.g., early vs. late with p < 1.79e-6). NFATc2 has been linked to AD pathology in mouse models through modulation of microglial activation⁷¹. Additionally, TCF7L1 → BOC edge in astrocytes showed significant differences across CERAD scores, particularly for No AD vs. AD-definite (p < 2.31e-5). TCF7L1 is known to be a highly astrocyte-specific regulator in AD pathology⁷². We then extracted top correlating genes for BRAAK phenotype (p < 5e-2) and compared the functional enrichment of positively and negatively correlated gene sets. Supplementary Data 3 contains the list of all significantly correlated genes for BRAAK phenotype. The positively correlated gene set was enriched for several developmental and signaling pathways, such as cell fate commitment and Nerve growth factor (NGF)-stimulated transcription (Supplementary Fig 18). NGF is known to be dysregulated in AD⁷³.

We extended our analysis beyond disease phenotypes to understand the conservation of prioritized gene regulatory links across all donors in PsychAD. As the population includes donors with different disease phenotypes beyond AD, we used data-driven importance scores. While ~81% of edges were unique or shared by fewer than four donors, a subset of edges were shared across many donors (Fig. 7c, Supplementary Fig 19). We separated these edges into groups based on their frequency and performed functional enrichment to identify pathways associated with each group. Specifically, we first computed the frequency with which each edge was high-scoring, i.e., scored above the 95th percentile of all data-driven importance scores, then grouped these edges into three categories: highly conserved (top 2%), conserved (top 2-4%), and lowly conserved (top 4–6%) edges based on empirical observation. We extracted unique genes from the edges in each group and performed functional enrichment on the resultant gene sets (Fig. 7c, Supplementary Data 4). Among the enriched terms in the highly conserved category, we observed relationships with regulation of leukocyte differentiation. Low leukocyte counts are associated with increased risk of developing AD⁷⁵. We also found dendritic cell differentiation to be enriched among conserved edges. Dendritic cells have been shown to be associated with AD progression and depression⁷⁶.

Differing subpopulations, such as ancestries, have been shown to exhibit significant AD risk disparities^77,78. We prioritized genes for European (EUR), African (AFR), and admixed American (AMR) ancestries using our importance scores (Methods). Permutation analysis showed the number of prioritized genes unique to EUR and non-EUR (shared between AFR and AMR) was significantly higher than random sampling (EUR: p < 1.00e-4, non-EUR: p < 1.00e-4, Methods, Supplementary Fig 20). We then performed functional enrichment on those genes unique to EUR and non-EUR (Fig. 7d). Regulation of phagocytosis had high significance among non-EUR. Phagocytosis of amyloid-beta (Aβ) plaques, particularly by microglia, is strongly related to AD severity⁷⁵. Connective tissue development was significant for both EUR and non-EUR gene sets. Expression of connective tissue growth factor is a regulator of Aβ plaque, which directly pertains to AD pathology⁷⁹. Functional enrichments for all ancestries are available in Supplementary Fig 21. We also performed gene set enrichment on the constituent genes from the top 2% of prioritized edges for our population subtypes (Supplementary Fig 22a), identifying several cluster-specific and shared AD-relevant pathways (Supplementary Fig 22b). For example, clusters c2-c5 were enriched with chromatin remodeling, which contributes to cognitive decline via HDAC dysregulation⁸⁰. Cluster c5 was enriched with brain morphogenesis, where neurogenesis in the adult hippocampus is believed to play a role in AD development⁸¹. These analyses demonstrate that our importance scores can offer subpopulation-level mechanistic insights for brain diseases and clinical phenotypes.

Genotype association and gene regulatory QTL analysis

Genetic variation affects gene expression and regulation at the CT level in brain diseases and disease progression^39,82,83. We first explored the relationship between polygenic risk score (PRS)⁸⁴ and our importance scores (Fig. 8a). We observed several edges significantly correlated with AD PRS scores (e.g., BCL6 → CDH23: r = -0.861, p < 3.26e-4; Immune → SPI1: r = -0.811, p < 4.41e-3). CDH23 has altered RNA expression in AD⁸⁵, and BCL6 overexpression mediates Aβ-induced neuronal damage and promotes SH-SY5Y cell survival⁸⁶. SPI1 has been associated with AD⁸⁷ and the age of Alzheimer’s onset⁸⁸. A positive correlation with disease PRS suggests that donors with high importance scores may have an increased disease risk from genetic variants, whereas a negative correlation implies a lower disease risk. Similarly, edges like EN_L3_5_IT_1 → BATF and IN_LAMP5_RELN → ARID5B were correlated with SCZ PRS. BATF has been associated with immune dysfunction in SCZ⁸⁹.

Fig. 8. Genotype association with gene regulatory networks.

a Spearman correlation (two-sided) between polygenic risk score and importance score for each regulatory link (transcription factor to target gene). The x-axis indicates the polygenic risk score and the y-axis represents AD-prior (top) and SCZ-prior (bottom) edge importance scores. Each dot in the plot represents a donor. The gray line is a monotonically increasing/decreasing best-fit line, included for clarity. b Upset plot showing the number of grQTLs that are cell-type-specific and common in at least two cell types. (top) Variation of the number of cell type grQTLs. Blue indicates cisTF-grQTLs and orange indicates cisTG-grQTLs. c grQTLs intersected with recent brain cell type eQTLs14. d the plot shows -log10(p) from grQTLs (p < 3.61e-21, top) and AD GWAS (p < 2.39e-8, bottom) from grQTLs and GWAS results for AD90 with posterior probability (PP) = 0.99 (colocalization test).

To further explore how genetic variants relate to gene regulation, we associated SNP genotypes with variations in CT gene regulatory links (TF–TG) using quantitative trait loci (QTL) analysis (i.e., gene regulatory QTLs or grQTLs), extending beyond traditional eQTLs that link SNPs to single gene expression (Methods). We call the SNPs associated with the TF region as cisTF-grQTL and those associated with the target gene region as cisTG-grQTL. In total, we identified 465,396 cisTF-grQTLs with 15,656 lead cisTF-grQTLs SNPs and 573,366 cisTF-grQTLs with 17,458 lead cisTG-grQTLs SNPs across 27 CTs at a false discovery rate corrected p-value cut-off of 1e-5 (Supplementary Table 3, Supplementary Data 5).

Among the lead SNPs, the number of unique grQTLs varied between CTs, with the highest count of 3105 grQTLs in SMC and the lowest count of 291 in PVM (Fig. 8b). We also saw that most grQTLs are cell-type-specific, with only 22% grQTLs shared with at least two CTs. The number of grQTLs showed high Pearson correlation with the number of genes (TF + TG) per CT (r = 0.95, p < 7.79e-14, Supplementary Fig 23). This suggests that CTs such as SMC, VLMC, and inhibitory neurons exhibit higher gene regulatory activity and thus have a higher number of grQTLs. Most CTs had similar numbers of cisTF- and cisTG- grQTLs with CTs like PVM, microglia, and oligodendrocytes being the exceptions, containing more cisTG-grQTLs (Fig. 8b).

We then compared the grQTLs with recent brain CT eQTLs¹⁴ and found several CTs with high intersection of the SNP-gene pairs (e.g., n = 1,559 for EN_L2_3_IT, n = 916 for EN_L6_IT_1, and n = 918 for oligodendrocytes, Fig. 8c). We then used Coloc⁹¹ to colocalize grQTLs with AD risk loci identified through GWAS⁹⁰. Co-localization analysis resulted in 30 co-localized loci (with posterior probability (PP) > 0.7). Among these, we identified several previously known AD TFs like RUNX1^20,92 and ZNF9⁹³ in excitatory and inhibitory subtypes. We also found two AD-associated target genes with colocalized signals, including JAZF1⁹³ and PLCG2⁹⁴. The top result was the SNP rs9648346 linked to JAZF1 (IKZF1 → JAZF1) with PP = 0.99 (Fig. 8d). JAZF1 is a protein-coding gene linked to tau phosphorylation in AD^93,95. These results provide insights into how genetic variation influences gene regulation in different brain CTs, potentially contributing to the heterogeneity observed in AD and other neuropsychiatric disorders. We also overlapped our grQTLs with 14 CRISPR-validated AD SNPs⁹⁶. We found 31 grQTLs within 5Kbp of 4 of the 14 AD SNPs, mainly consisting of 11 EN_L6_IT_1 grQTLs, 4 VLMC grQTLs, and 3 SMC grQTLs.

Beyond genotype association, we explored whether genotype data can be used to impute graph embeddings and whether these imputed embeddings could classify donor phenotypes (Supplementary Fig 24a, Supplementary Note 10). We trained our imputation pipeline using the PsychAD data and validated the phenotype classification on two independent datasets, ROSMAP⁹⁷ and ADNI⁹⁸, where only genotype data was available. In the ROSMAP cohort, we classified donors into early vs. late BRAAK stages, while in ADNI, we classified donors into AD vs. control. We evaluated several cross-modality imputation methods (CMOT⁹⁹, JAMIE¹⁰⁰, MOFA + ¹⁰¹, and autoencoders) and observed that the best performing model achieved an AUC of 0.57 for phenotype classification on both cohorts. Although modest, this performance was above random and comparable to results reported in prior PRS studies¹⁰² (Supplementary Fig. 24b, Supplementary Table 4). We next clustered the imputed donor graph embeddings and identified seven potential subtypes (Supplementary Fig 24c). We found that clusters c2-c6 were enriched with CERAD stages, which progressed in severity from ‘No AD’ to ‘Definite’ (Supplementary Fig 24d). These findings highlight how SNPs relate to CT gene regulation changes and open up new possibilities for using genotype data to classify disease phenotypes and discover subtypes, even when transcriptomic data is not available.

Discussion

In this study, we present what is, to our knowledge, the first personalized functional genomics analysis of AD using population-scale snRNA-seq data. Using the iBrainMap framework, we constructed PFGs and applied graph learning to classify AD phenotypes, stratify donors, and identify disease subtypes. Specifically, we trained a KG-GNN model to first classify AD vs. control and then used the pre-trained model to extract graph embeddings for all donors with diverse phenotypes. The KG-GNN model effectively classified several phenotypes using the AD graph embeddings. Our model also highlighted several CT interactions and gene regulatory links related to different AD phenotypes. We validated our findings using external snRNA-seq datasets, including SEA-AD and ROSMAP, demonstrating the framework’s robustness and generalizability. Beyond these analyses, our framework also serves as a publicly available resource and tool: a functional genomic atlas of AD featuring PFGs, a pre-trained KG-GNN model, and cell-type-level grQTLs, all accessible via an interactive webapp. In this study, the iBrainMap framework is applied to post-mortem human brain data, but it was broadly designed for population-level single-cell datasets. It can be readily trained and adapted to living patient samples (e.g., blood) and to other genetically complex disorders when large-scale functional genomics data are available.

Our approach has the potential to drive the development of tailored therapeutic strategies by capturing inter-individual variability of CT genes and gene regulatory interactions, which helps in identifying upregulated or downregulated genes that can serve as potential therapeutic targets. In this context, while many regulatory edges are person-specific and reflect inter-individual regulatory states, conserved edges highlight shared, cell-type-specific control points that are more suitable for therapeutic development. We compared our prioritized genes with group-level differentially expressed genes from PsychAD²¹. We observed an increase in overlap as we aggregate across donors and examine genes that are more frequently prioritization (Supplementary Fig 25a, b). Notably, while higher-ranked features increasingly overlap with differential expression markers, a substantial fraction of prioritized genes and regulatory edges show minimal or no differential expression, including conserved disease-associated programs, such as TYROBP-centered microglial networks¹⁰³ and GABAergic neuron differentiation¹⁰⁴ pathways, which would not be recovered by standard differential expression analyses (Supplementary Fig 25c). Prior research has shown that CT-specific networks can enhance the prediction of potential drug targets by identifying AD-related genes, which is valuable for predicting AD clinical phenotypes and for drug repurposing¹⁰⁵. AD therapeutic response can also vary based on genotype. For example, donepezil was shown to have a beneficial pharmacogenomic effect for female patients with MCI who carry the BCHE K variant¹⁰⁶. Additionally, a computational drug-repurposing screen identified bumetanide as a potential treatment to lower AD risk in donors carrying the APOE ε4 variant¹⁰⁷. As an exploratory analysis, we compared the top microglial genes prioritized by iBrainMap with amyloid-β clearance genes (“limited” and “extensive” clearance) reported by Olst et al. ¹⁰⁸ to investigate potential links with therapeutic outlook (Supplementary Fig 22c). We observed varying degrees of overlap across donor clusters, and these patterns may reflect diversity in microglial function.

We identified a substantial number of significant cisTF- and cisTG- grQTLs across 27 CTs. These findings were consistent with recent brain CT eQTL studies, which have reported comparable numbers, such as ~7500 eQTLs across eight major CTs¹⁰⁹, 26,597 eQTLs across 14 CTs¹¹⁰, and ~1.4 million single-cell eQTLs across 28 CTs¹⁴. Unlike eQTLs, which associate SNPs with gene expression, our grQTLs capture SNP changes in regulatory interactions, offering a complementary and potentially more mechanistic view of genetic influence on gene regulation.

Our work has several limitations. First, the PFGs are constructed by selecting TFs and TGs based on the RSS score, focusing on top regulatory links for each CT. Future work involves enhancing personalized graphs by incorporating additional biological networks, e.g., protein-protein interactions, to provide a comprehensive view of an individual’s regulatory mechanisms. Second, our graph neural network model currently treats all node and edge types as homogeneous, which overlooks distinctions between different types of nodes (e.g., CTs, TFs, TGs) and edges. To address this, future efforts could explore more sophisticated models, such as heterogeneous graph neural networks⁶⁴, that can distinguish different node and edge types. Third, while KG-GNN had high overall performance across various phenotypes, its performance was lower for AD comorbid conditions, such as NPSs (Supplementary Fig 5b) and AD-DLBD (Supplementary Fig 9b). These graph embeddings, derived primarily from AD vs. control comparisons, may not fully capture the complexity of comorbid conditions. Moreover, the schizophrenia importance scores for nodes and edges have higher variance than AD (Supplementary Fig 26). This suggests that retraining disease-specific KG-GNN models when sufficient samples become available in the future might help improve classification performance. Fourth, we incorporated genotype data to identify SNPs associated with variations in CT gene regulatory links (grQTLs), thereby extending beyond traditional eQTL analyses that link SNPs to single-gene expression. In addition, we explored the use of genotype data to impute graph embeddings, which were then used to predict disease phenotype. By developing this imputation model, we enable the prediction of graph embeddings for new donors who have only genotype data, without requiring gene expression. This imputation supports donor-level functional inference from genotype, though it does not yet imply direct clinical application or therapeutic stratification. While the classification performance was modest, it aligned with previous PRS studies¹⁰². However, a comprehensive assessment of disease stages requires additional information, such as environmental, epigenetic, and transcriptomic profiles^6,111. Thus, extending our analyses to emerging single-cell multimodal data, such as epigenomics, proteomics, and spatial data, could provide a more holistic view of disease mechanisms.

In summary, our framework leverages single-nucleus transcriptomics to construct personalized functional genomic graphs, capturing nuanced, biologically meaningful variation across donors. While not yet a tool for clinical precision medicine, iBrainMap offers a scalable strategy for hypothesis generation in molecular neurobiology and may inform future efforts toward individualized disease modeling.

Methods

Datasets and data preprocessing

Dataset

Our analysis is focused on the population-level snRNA-seq data from the PsychAD consortium covering the dorsolateral prefrontal cortex (DLPFC) brain region from 1494 donors, derived from three cohorts: NIH NeuroBioBank at the Mount Sinai Brain Bank/MSSM (M), Rush AD Center/RADC (R), and the NIMH-IRP Human Brain Collection Core/HBCC (H). The PsychAD data description is available in the “PsychAD dataset” (Supplementary Information). Our analyses utilized disease diagnosis (e.g., AD, SCZ, DLBD), quantitative assessment of the disease stages for donors with AD (e.g., BRAAK or CERAD), and diagnosis of neuropsychiatric symptoms (e.g., depression or weight loss). Accordingly, these donors were categorized into three levels: disease vs. control (AD, SCZ, DLBD, AD Resilience), AD progression (BRAAK stages, Cogdx, CERAD), and NPSs (Dysphoria, DecInt (Anhedonia), Sleep/Weight Gain/Guilt/Suicide (S/W/G/S), Depression/Mood (D/M)) (Fig. 2a, Supplementary Note 4). We mainly used the MSSM (M) cohort for all the training and used RADC (R) as well as external datasets from Religious Orders Study/Memory and Aging Project (ROSMAP)⁸ and Seattle AD Brain Cell Atlas/SEA-AD (S)⁹ for independent validation.

Data processing and feature selection

PsychAD snRNA-seq data was preprocessed as described in Lee et al. ²¹ and Fullard et al²². Briefly, the preprocessing involved identifying 5000 highly variable genes (HVGs). We selected the protein-coding genes from these 5000 HVGs and intersected them with the genes from the SEA-AD dataset. We ended up with 2,766 HVGs which were used as features in our analysis. We applied standard Scanpy (v1.9.3) functions to preprocess SEA-AD data. For SEA-AD, we selected DLPFC regions of 39 donors with dementia and 39 without dementia to retain individual balance. For ROSMAP, we first followed the PsychAD preprocessing guidelines to normalize and scale the raw counts. We then used scanpy.tl.ingest() function to perform label transfer to match the broad CTs (n = 12) with the refined cell subclasses from PsychAD (n = 27). For independent validation, we constructed bio-diffused PFGs for the donors and included the same set of 2766 HVG genes as features as the PsychAD dataset.

Construction of personalized functional genomic graph

For each donor, we define his/her personalized functional genomic graph (PFG) as a directed graph with two major node types: CTs and genes. Gene nodes include cell-type-specific transcription factors (TFs) and target genes (TGs). Each edge in the PFG conveys distinct information depending on the type of nodes. For example, CT to CT edge captures CT interactions, TF-TG edge captures gene regulatory links, and CT-TF edge captures cell-type-specific relationship of TFs. As some of the donors had fewer cells for the CT interactions and gene regulatory network algorithms to capture (Supplementary Fig 27), we were able to construct PFGs for 1478 out of 1494 donors. Below, we provide details on constructing each component of PFG.

Inferring cell type interactions

For each donor, the CT to CT links were constructed using CellChat¹¹² to quantify the cell-cell communications. CellChat provides a consolidated interaction probability for each CT to CT interaction by adding all the probabilities of ligand-receptor interactions between two CTs. We used snRNA-seq data from each donor as input to CellChat with 27 subclass information as the label. We used the default settings in CellChat to extract the probabilities using the function computeCommunProb. As the consolidated interaction probability can exceed 1, we normalized the probability between 0 and 1 and retained all links with the normalized score above a 0.5 threshold to construct CT to CT links for each donor.

Extracting CT gene regulatory links

We first used GRNBoost2 on the snRNA-seq data for each donor to deduce gene co-expression networks across all CTs. SCENIC¹¹³ was then used to infer CT gene regulatory links. Specifically, we applied runscENIC_1_coexNetwork2modules and runscENIC_2_createRegulons functions in Python with default settings (constraining the TF search to 10 kbps radius around the TSS or 500 bp upstream) to identify regulons. Regulons with at least ten genes were scored in each cell using runscENIC_3_scoreCells. We then computed AUCell enrichment based on the top 1% of genes detected per cell. We incorporated the regulon specificity score (RSS)¹¹⁴ to evaluate the cell-type-specificity of the regulon. For each CT, we select the top 10% of the regulons based on the RSS score as the CT regulons. We then keep the top 10% of the TGs for each CT regulon. Together, these comprise the CT gene regulatory links for each donor.

Definition of PFGs

Let $G_i=\left(V_i,E_i\right)$ be a PFG (Fig. 2b) for donor $i$, with $N_i=\mid V_i\mid$ nodes and $E_i$ representing the edges. Then, an edge $e_{i,st}\in E_i$ connecting the source node $s$ to a target node $t$ represents one of the following relationships:

$$e_{i,st}=\left\{\begin{array}{c}\hfill 1,ifsinteractswitht,wheresandtarecelltypes;\hfill \\ \hfill 1,ifsregulatest,wheresisatTF,tisaTG;\hfill \\ \hfill 1,iftbelongstothecelltyperegulonofs,wheresisacelltype,tisaTF;\hfill \\ \hfill 1,iftbelongstothecelltyperegulonofs,wheresisacelltype,tisaTG;\hfill \\ \hfill 0,Otherwise\hfill \end{array}\right)$$

The node features used for each PFG depend on the type of node. We used the average gene expression of 2,766 highly variable genes (HVGs) from all cells as the features for CT nodes and the co-expression of these HVGs with the gene expression for the TF/TG nodes.

Integrating prior biological knowledge into PFGs via network diffusion

We then applied network diffusion to propagate the influence of known disease genes on the PFGs (Supplementary Note 2). This diffusion process allows us to incorporate known disease-related information into our graph structure, potentially enhancing the model’s ability to capture disease-relevant patterns. We currently use well-known disease genes for AD and SCZ extracted from the DisGeNet³⁸ database and filtered based on manually curated sources with “CTD_human” or “GWASCAT” identifiers. Additionally, we downloaded high-confidence SCZ genes from the PsychENCODE project³⁹. In total, we identified 361 AD genes and 945 SCZ genes. The resulting matrices are called “bio-diffused PFGs” and are used to train our graph neural network model. While we focus on AD and SCZ in this paper, this method can be applied to any gene-of-interest (GOI) list.

Graph classification using Knowledge-guided Graph Neural Network (KG-GNN)

The KG-GNN model is based on the multi-head Graph Attention Networks (GAT)¹¹⁵ that learns from arbitrary bio-diffused PFGs as inputs for classifying AD vs. control. It is an interpretable machine learning model that integrates prior biological knowledge, personalized functional genomics, and biologically driven multi-head graph attention networks to derive insights about disease and underlying disease mechanisms at a donor level. The details of training procedure, validation, and benchmarking are provided in Supplementary Notes 4-5. The input to the KG-GNN mode is the PFGs. The trained model outputs graph embeddings and personalized AD importance scores for nodes and edges.

Population subtyping and enrichment based on graph embeddings

We clustered graph embeddings extracted for donors with AD vs. controls using the function sc.tl.louvain() from the Python package scanpy¹¹⁶. Before clustering, we identified the nearest neighbors using the sc.pp.neighbors() function and computed the UMAP embeddings using the sc.tl.umap() function. Then we performed Louvain clustering using the function sc.tl.louvain() and set the resolution = 0.35 to get 5 clusters. We clustered the graph embeddings to identify potential molecular subtypes and performed enrichment analysis using a hypergeometric test to see if any clusters were enriched with AD progression stages. We used the stats.hypergeom.cdf() from the Python package Scipy¹¹⁷ for this analysis and reported the -log₁₀(p-value) of phenotype enrichment in each cluster.

Phenotypic population trajectory inference

We inferred phenotypic trajectories across AD phenotypes based on donor graph embeddings extracted from our pre-trained KG-GNN model. Here we used a diffusion-based algorithm¹¹⁸ within the Scanpy package in Python to infer trajectories. Pseudotime was computed using sc.tl.diffmap() and sc.tl.dpt() functions from the Scanpy package. To determine phenotypic pseudotime for donors with a given phenotype, we selected a starting point or root by randomly choosing a donor from those assigned to the earliest phenotype stages. For AD & BRAAK, this was a donor from stage 0; for CERAD, a donor with a CERAD score of 0 (No AD); and for Cogdx, a donor with a cognitive diagnosis of 1 (control). To infer the pseudotimes for different NPS, we again set a random donor with a CERAD score of 2 as the root. This is because all donors diagnosed with NPS have a CERAD score equal to at least 2.

To infer the trajectory for SEA-AD donors, we first extracted their graph embeddings from the pre-trained KG-GNN model. Then using similar functions as above (sc.tl.diffmap(), sc.tl.dpt()), we computed a trajectory by randomly selecting a donor with CERAD score = “Absent” as the root.

Calculation of importance scores

The trained KG-GNN model outputs edge attentions which are used as edge importance scores. We get three different edge importance scores: AD-prior, SCZ-prior, and data-driven based on prior biological knowledge from known AD-genes and SCZ-genes. We derive the node importance scores using the edge importance scores. We use a combination of both incoming and outgoing edge importance as the basis to calculate these scores. Further details about node importance calculation are available in Supplementary Note 9. We averaged the importance scores across all three priors for each node and edge per donor.

Subpopulation-level edge conservation and gene prioritization analysis

To determine edge conservation across donors, we first counted instances where each edge was scored in the 95th percentile or above based on the attention scores of the selected head. We then sorted the edges by frequency and established three groups: Highly conserved edges (top 2%, frequency > 50 donors), conserved edges (top 2–4%, frequency = 23–50 donors), and lowly conserved edges (top 4-6%, frequency = 15–22 donors). The grouping of the edges was decided empirically, as the frequency with which highly scored edges appeared in the PFG graphs declined rapidly from the 100th to 94th percentiles.

We repeated this edge conservation analysis for AFR (n = 120 donors), AMR (n = 99 donors), and EUR (n = 792 donors) ancestries. Using the highly conserved edges, we sampled 300 prioritized genes for each ancestry, resulting in 587 unique genes which were used for downstream enrichment analyses. We analyzed the uniqueness of these prioritized genes for each ancestry using a permutation test as follows. First, we randomly constructed new gene sets for AFR, AMR, and EUR, matching their original sizes and recorded the number of overlapping and unique genes for each ancestry. Second, we repeated this random sampling 10,000 times to provide a baseline overlap distribution. Finally, we estimated one-sided p-values for gene set overlapping by comparing the number of overlaps and unique genes in our prioritized genes to our baseline overlap distribution.

Differential importance score analysis

We performed differential importance score analyses on nodes and edges for different phenotypes using Mann-Whitney U test in R. We adjusted p-values for multiple testing using the p.adjust() function with the Benjamini–Hochberg method, and features with a false discovery rate (FDR)–adjusted p-value < 0.05 were considered statistically significant.

Gene set enrichment analysis

We performed gene set enrichment using Metascape⁷⁴. Unique genes from the PsychAD data were used as background. Using a cutoff of p < 0.05, we extracted significant gene ontology terms for use in our analysis.

Polygenic risk score (PRS) analysis

We used AD GWAS⁹⁰ and SCZ GWAS¹¹⁹ to compute the PRS scores within the paper. Both were calculated for donors using PRS-CS¹²⁰ and PLINK 2.0¹²¹. The PRS-CS-auto method was utilized, which employs continuous shrinkage priors to refine effect sizes derived from summary statistics. We used an LD reference panel from the creators of PRS-CS, based on the 1000 Genomes Project data¹²² (available at https://github.com/getian107/PRScs). Default settings were maintained for PRS-CS, including the γ-γ prior parameters a = 1 and b = 0.5, 1,000 MCMC iterations, 500 burn-in iterations, and a thinning factor of 5. The global shrinkage parameter $\varphi$ was estimated using a fully Bayesian approach. PRS calculations at the donor level were performed using PLINK 2.0.

Gene-regulation QTL (grQTL) analysis

We associated genotype with the TF-TG edge scores per CT and called it grQTLs. To do this, we mapped cis-eQTLs within a 1-Mb window region (including the gene body) of both TF and TG genes with the TF-TG regulatory score from SCENIC for 27 CTs using QTLtools¹²³. We call the SNPs associated with the TF region cisTF-grQTL and the TG region cisTG-grQTL. We used the first three genotyping PCs, age, gender, and 50 PEER factors as covariates. Using the associated grQTLs, we then performed FDR correction and used 1e-5 as the cut-off to identify the most significant CT-specific cisTF-grQTLs and cisTG-grQTLs.

Statistics and reproducibility

The study does not involve biological experiments. No statistical method was used to predetermine sample size. No data were excluded from the analyses. The experiments were not randomized. The Investigators were not blinded to allocation during experiments and outcome assessment.

Ethics declarations

All procedures and research protocols were approved by the respective ethical committees of our collaborators’ institutions. The Ethics committee/Institutional Review Board (IRB) of Mount Sinai gave ethical approval for this work. The Ethics committee/IRB of the James J. Peters Department of the Veterans Affairs Medical Center gave ethical approval for this work. The Ethics committee/IRB of Rush AD Center gave ethical approval for this work. The Ethics committee/IRB of the National Institute of Mental Health Human Brain Collection Core gave ethical approval for this work.

Reporting summary

Further information on research design is available in the Nature Portfolio Reporting Summary linked to this article.

Author contributions

Conceptualization: D.W. and P.R. Methodology: P.B.C., S.A., T.J. and D.W. Software: P.B.C., S.A. and T.J. Formal analysis: P.B.C., S.A., N.C.K., T.J. and D.W. Investigation: X.H., S.L., A.L., K.G., G.V., G.E.H., J.B., J.F.F., D.L., P.R. and PsychAD Consortium. Resources: D.L., J.B., K.G., G.V., G.E.H., J.F.F., P.R. and PsychAD Consortium. Data Curation: D.L., J.B., K.G., G.V., G.E.H. and J.F.F. Writing: P.B.C., S.A., N.C.K., T.J., X.H., D.W., and P.R. with support from all co-authors. Visualization: S.A., P.B.C., N.C.K., T.J., C.G., and R.B. Supervision: DW, PR Funding acquisition: DW, PR All authors read and approved the final draft of the paper.

Peer review

Peer review information

Nature Communications thanks the anonymous reviewers for their contribution to the peer review of this work. A peer review file is available.

Data availability

The PsychAD single nucleus transcriptomics data is available via the AD Knowledge Portal (https://adknowledgeportal.org). The data, analyses, and tools are shared early in the research cycle without a publication embargo on secondary use. Data is available for general research use according to the following requirements for data access and data attribution (https://adknowledgeportal.synapse.org/Data%20Access). The data is available under controlled use conditions set by human privacy regulations through https://doi.org/10.7303/syn60084804. To access the data, a data use agreement is needed. The registration is in place solely to ensure the anonymity of the study participants. Interactive visualization of the single-cell data is available from CELLxGENE (https://cellxgene.cziscience.com/collections/84ce6837-548d-4a1f-919f-0bc0d9a3952f). Genotype data is available under controlled access and can be provided upon request. The Seattle AD Brain Cell Atlas (SEA-AD) can be accessed using the accession code syn52146347. The personalized functional genomic graphs, gene-regulation QTLs (grQTLs), and all the results from our analysis can be interactively visualized at https://daifengwanglab.shinyapps.io/iBrainMap/. All other data are included in the main paper or Supplementary Information.

Code availability

The iBrainMap framework and the pre-trained model, which takes personalized snRNA-seq data as input for classifying AD and generating embeddings and importance scores, together with tutorials and demos, can be accessed at https://github.com/daifengwanglab/iBrainMap. All code and data used for generating figures can be accessed at https://zenodo.org/records/15498102. All graphs were visualized using cytoscape¹²⁴ and panels of all figures were combined using BioRender (https://www.biorender.com/).

Competing interests

The authors declare no competing interests.

Supplementary information

The online version contains supplementary material available at https://doi.org/10.1038/s41467-026-72310-1.