Single-Cell Graph Learning Maps Personalized Gene Regulatory Networks


A computational framework combining population-scale single-nucleus transcriptomics with graph neural network modeling has mapped personalized functional genomic diversity across more than 1,400 human brains, revealing how cell type interactions and gene regulatory networks vary among individual Alzheimer’s disease donors. The study, analyzing post-mortem tissue from the dorsolateral prefrontal cortex across cohorts assembled by the PsychAD Consortium, addresses a fundamental limitation in neurodegenerative disease research: the tendency of conventional genomic studies to pool cellular profiles across large patient groups, thereby obscuring inter-individual variations in molecular regulation. By constructing individualized functional genomic graphs for each participant, the researchers demonstrated that donor-specific regulatory architectures capture distinct disease trajectories, identify previously hidden molecular subtypes, and map specific genetic variants directly to cell-type-level gene regulatory relationships beyond traditional expression quantitative trait locus mapping. The open-source platform, named iBrainMap, offers an analytical atlas for exploring the heterogeneous mechanisms governing cognitive decline, neuropathological burden, and co-occurring neuropsychiatric symptoms in Alzheimer’s disease.

To capture population-level complexity at cellular resolution, the investigators analyzed single-nucleus RNA sequencing data spanning 6.3 million nuclei derived from the dorsolateral prefrontal cortex of 1,494 human donors across three primary cohorts: the Mount Sinai NIH Brain Bank and Tissue Repository comprising 1,042 donors, the NIMH-IRP Human Brain Collection Core comprising 300 donors, and the Rush Alzheimer’s Disease Center comprising 152 donors. The expansive cohort encompassed neurotypical control individuals as well as patients diagnosed with Alzheimer’s disease, schizophrenia, and diffuse Lewy body disease, alongside documented assessments of cognitive status, Braak neurofibrillary tangle stages, CERAD neuritic plaque scores, and neuropsychiatric symptoms such as depression and anhedonia. Feature selection isolated 2,766 highly variable protein-coding genes shared with independent profiling cohorts, enabling consistent modeling across 27 refined brain cell subclasses. Out of the total cohort, 1,478 donors had sufficient cell coverage across designated cell types to build donor-level networks, providing an empirical foundation for investigating how individual cellular compositions and transcription factor interactions diverge across disease spectra.

The iBrainMap framework translates each donor’s single-nucleus transcriptomic data into a directed personalized functional genomic graph consisting of cell types, transcription factor genes, and downstream target genes connected by directed biological relationships. For every donor, cell-to-cell communication links were quantified through ligand-receptor interaction probabilities computed with CellChat, retaining links exceeding a normalized probability threshold of 0.5 across the 27 cell subclasses. Concurrently, cell-type-specific gene regulatory networks were inferred by coupling co-expression analysis from GRNBoost2 with regulon identification in SCENIC, prioritizing the top 10 percent of regulons by regulon specificity score and the top 10 percent of target genes per regulon. Node feature vectors integrated average expression across the 2,766 highly variable genes for cell-type nodes and pairwise co-expression vectors for regulatory gene nodes. Summary evaluations revealed that approximately 10 percent of network edges in each donor graph were entirely unique to that individual, whereas nearly 80 percent of edges were shared among more than eight donors, illustrating a balance between conserved baseline architecture and donor-specific regulatory variation.

To incorporate established pathological knowledge into the network architecture without biasing data-driven discovery, the team implemented a knowledge-guided graph diffusion step prior to deep learning classification. The diffusion procedure propagated biological signals from curated disease genes identified in databases such as DisGeNet and PsychENCODE, incorporating 361 known Alzheimer’s disease genes and 945 high-confidence schizophrenia genes to weight adjacent graph connections. The resulting bio-diffused personalized functional genomic graphs were then supplied to a knowledge-guided graph neural network built upon a multi-head graph attention network architecture. This model was trained to classify Alzheimer’s disease cases versus control individuals while projecting the high-dimensional personalized graphs into compact latent graph embeddings. Simultaneously, the multi-head attention mechanisms yielded importance scores for all graph nodes and edges, quantifying how much individual cell types, transcription factors, target genes, and intercellular links contributed to the classification of disease status across disparate clinical phenotypes.

When trained on an optimal hyperparameter configuration established through five-fold cross-validation on an initial subset of 438 Alzheimer’s disease donors and 179 control donors from Mount Sinai, the knowledge-guided model demonstrated robust diagnostic classification. On a held-out test cohort of 62 Alzheimer’s cases and 30 controls from Mount Sinai, the model achieved an area under the receiver operating characteristic curve of 0.913. To assess generalizability across separate clinical sites, the pre-trained architecture was evaluated on independent donors from the Rush Alzheimer’s Disease Center and the Seattle Alzheimer’s Disease Brain Cell Atlas, yielding a combined validation area under the curve of 0.805 across 93 disease cases and 68 controls. Application to independent data from the Religious Orders Study and Memory and Aging Project classified early versus late Braak stages with an area under the curve of 0.788. Label-shuffled permutation tests repeated across ten iterations yielded a balanced accuracy of 0.87 for the true model compared to 0.5035 for randomized graphs, demonstrating that performance stemmed from genuine functional network topology.

Beyond binary disease classification, unsupervised Louvain clustering of the learned graph embeddings separated the donor population into five distinct clusters, designated c1 through c5, reflecting underlying molecular stages of neuropathology. Hypergeometric enrichment testing revealed that clusters c1 and c2 were significantly enriched for early Braak stages zero through two, absence of dementia, and control designations under CERAD criteria, whereas clusters c4 and c5 were predominantly occupied by donors exhibiting late Braak stages five to six, severe clinical dementia, and definite Alzheimer’s disease pathology. Extracted graph embeddings also supported multi-class evaluations across intermediate phenotypes, classifying resilience status with an area under the curve of 0.87, Braak staging with 0.79, CERAD scoring with 0.75, and cognitive diagnosis classifications with 0.79. Binary evaluations in cross-disorder settings achieved an area under the curve of 0.83 for schizophrenia classification, while evaluations of neuropsychiatric symptoms achieved moderate accuracy, including 0.67 for sleep disturbances and guilt and 0.61 for depressive mood, indicating that latent graph embeddings capture nuanced clinical manifestations beyond coarse diagnostic boundaries.

To investigate how molecular networks shift over time across diverse patients, the authors extracted donor graph embeddings to compute continuous phenotypic pseudotimes using diffusion map algorithms. In the primary cohort, inferred Alzheimer’s disease pseudotimes separated disease cases from control donors with high statistical significance, showing strong positive alignments with quantitative plaque measurements and Clinical Dementia Rating scores via ordinary least squares regression. Similar monotonic progressions were observed across categorized Braak stages, CERAD scores, and cognitive impairment tiers. Furthermore, pseudotime trajectories for neuropsychiatric symptoms revealed that dysphoria manifested at earlier pseudotimes near the initial onset of cognitive impairment, whereas early and middle insomnia were distributed across distinct intermediate and advanced stages of disease progression. Independent validation within 80 donors from the Seattle Alzheimer’s Disease Brain Cell Atlas corroborated these trajectories, yielding a Jensen-Shannon similarity of 0.77 between graph embedding pseudotimes and continuous neuropathology pseudoprogression scores derived from physical immunohistochemical stains.

Model-derived attention weights enabled donor-level prioritization of specific regulatory links, illuminating which cellular compartments and molecular circuits drive disease states. Neuronal subtypes consistently received the highest importance scores among Alzheimer’s disease donors, with layer five and six near-projecting excitatory neurons, layer six intratelencephalic neurons, and Lamp5-Lhx6 inhibitory interneurons emerging as the top three prioritized cell types. Analysis of intercellular communications showed that interactions originating from microglia were elevated in control donors, whereas communications driven by somatostatin-expressing inhibitory interneurons were enriched in Alzheimer’s cases. When aggregating the top one percent of prioritized edges into a consensus subnetwork, the model isolated dense regulatory modules focused in oligodendrocytes, astrocytes, and excitatory neurons that exhibited significant enrichment for known disease risk genes. Within these modules, specific transcription factor interactions were prioritized across clinical stages, including the regulatory link between NFATc2 and IKZF1 in microglia across Braak stages, and the TCF7L1-to-BOC regulatory edge in astrocytes across CERAD neuropathology scores.

Extending functional genomics to genetic regulation, the investigators conducted gene regulatory quantitative trait locus analyses, termed grQTLs, correlating single-nucleotide polymorphisms with individual variations in cell-type-specific regulatory links. Across 27 cell subclasses, the analysis identified 465,396 cis-transcription factor grQTLs featuring 15,656 lead variants and 573,366 cis-target gene grQTLs featuring 17,458 lead variants at a false discovery rate threshold of 1e-5. Approximately 78 percent of these regulatory associations were strictly cell-type-specific, with smooth muscle cells displaying the highest count at 3,105 unique associations and perivascular macrophages displaying the lowest at 291. Colocalization analyses with Alzheimer’s disease genome-wide association study summary statistics identified 30 colocalized loci with posterior probabilities exceeding 0.70. Cross-referencing against CRISPR-validated disease variants uncovered 31 grQTLs within five kilobases of four validated risk variants, linking non-coding genetic variants to explicit regulatory disruptions rather than isolated gene expression alterations.

Subpopulation analyses across European, African, and admixed American ancestries demonstrated that conserved regulatory networks capture shared and lineage-specific functional pathways. Highly conserved regulatory links shared across multiple donors were enriched for leukocyte differentiation and dendritic cell pathways, while ancestry-stratified analyses revealed significant overrepresentation of phagocytic clearance mechanisms among non-European donors. The authors noted several technical limitations, emphasizing that current personalized graphs rely on regulon specificity scores and treat graph nodes and edges homogeneously rather than through heterogeneous graph neural network architectures. Furthermore, cross-modality imputation models predicting graph embeddings solely from genotype data achieved modest classification performance, with areas under the curve of 0.57 in external validation cohorts, confirming that genetic profiles alone cannot substitute for direct single-cell transcriptomic measurements. Nevertheless, by organizing complex single-nucleus datasets into interpretable donor-specific networks, the iBrainMap framework establishes a structured computational foundation for exploring inter-individual heterogeneity and prioritizing cell-type-specific mechanisms in neurodegenerative disease.

Subject of Research: Personalized single-nucleus functional genomics and graph neural network modeling in Alzheimer’s disease

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

Article References: Chandrashekar, P. B., Alatkar, S. A., Cohen Kalafut, N., Jin, T., Gupta, C., Burczak, R. C., Huang, X., Liu, S., Li, A. Z., PsychAD Consortium, Hong, A., Zeng, B., He, C., Dillard, C., Porras, C., Casey, C., McClung, C. A., Spencer, C., Bennett, D. A., … Wang, D. (2026). Personalized single-cell transcriptomics reveals molecular diversity in Alzheimer’s disease. Nature Communications, 17(1), Article 9596. https://doi.org/10.1038/s41467-026-72310-1

Image Credits: AI Generated

DOI: 10.1038/s41467-026-72310-1

Keywords: Alzheimer’s disease, single-nucleus RNA sequencing, functional genomics, graph neural networks, iBrainMap, PsychAD Consortium, gene regulatory networks, cell-cell communication, dorsolateral prefrontal cortex, phenotypic trajectories, pseudotime analysis, grQTLs, neuropsychiatric symptoms, microglia, inter-donor heterogeneity, subpopulation subtyping, cross-cohort validation, colocalization analysis

Leave a Reply

Your email address will not be published. Required fields are marked *