Single‐Nucleus Multiomic Analysis Reveals Immune–Metabolic Reprogramming Consistent With Maladaptive Trained Immunity in HIV‐Associated Neurocognitive Disorders

ABSTRACT

HIV‐associated neurocognitive disorders (HAND) remain prevalent in people with HIV (PWH) despite effective antiretroviral therapy, suggesting that persistent immune activation contributes to ongoing neurological dysfunction. Maladaptive trained immunity (TRIM), long‐term innate immune reprogramming characterized by sustained inflammatory responses, metabolic rewiring, and epigenetic remodeling, has been proposed as a sustaining mechanism. We performed single‐nucleus RNA‐seq and ATAC‐seq on post‐mortem brain from PWH with HIV‐associated dementia (HAD) or asymptomatic neurocognitive impairment (ANI) and compared these data with published HIV‐uninfected (PWoH) datasets. In PWH versus PWoH, glia, led by microglia, showed enrichment of innate immune signaling and upregulation of inflammatory mediators including NLRP3, TLR2, and TLR4. In parallel, glia showed coordinated cholesterol remodeling, with efflux transporter and apolipoprotein upregulation alongside LDLR downregulation, and partial microglial glycolytic reprogramming (PFKFB3, HK2, PGK1 upregulation). Chromatin accessibility profiling showed concordant gains at inflammatory, cholesterol regulatory (notably RXRA and APOE), and glycolytic loci. The HAD versus ANI comparison revealed selective reorganization rather than uniform amplification, with increased lipid scavenger receptor accessibility in oligodendrocytes and reduced accessibility at cholesterol efflux and glycolytic loci. Neurons exhibited predominantly bystander epigenetic changes. Three features distinguish this pattern from chronic inflammation alone and align it with trained immunity hallmarks: concordant transcriptional and chromatin‐level priming at inflammatory loci, parallel rewiring of glycolytic and cholesterol metabolism, and persistence despite long‐term viral suppression. These multiomic data are consistent with maladaptive trained immunity sustaining neuroinflammation in HAND, though functional validation is required.

Single‐nucleus multiomic profiling of postmortem brain reveals coordinated immune, metabolic, and epigenomic reprogramming in glia of people with HIV consistent with maladaptive trained immunity sustaining neuroinflammation in HAND.

Abbreviations

ANI

asymptomatic neurocognitive impairment

ART

anti‐retroviral treatment

EV

extracellular vesicle

GEO

gene expression omnibus

HAD

HIV‐associated dementia

HAND

HIV‐associated neurocognitive disorders

HIV

Human immunodeficiency virus

MND

mild neurocognitive disorder

NB

neuroblasts

NNRTI

non‐nucleoside reverse transcriptase inhibitor

NNTC

national neuroAIDS tissue consortium

NRTI

nucleoside reverse transcriptase inhibitor

PI

protease inhibitor

PWH

people living with HIV

snATAC‐seq

single nucleus assay for transposase‐accessible chromatin using sequencing

snRNA‐seq

single‐nucleus RNA sequencing

TRIM

trained immunity

UMAP

uniform manifold approximation and projection

1. Introduction

HIV‐associated neurocognitive disorders (HAND) remain a major clinical concern despite the widespread success of combination antiretroviral therapy (ART) in suppressing viral replication. Although the incidence of severe HAND (HIV‐associated dementia or HAD) has declined in the ART era, milder cognitive impairments (mild neurocognitive disorder or MND and asymptomatic neurocognitive impairment or ANI) affecting memory, executive function, and psychomotor speed continue to affect up to 50% of people living with HIV (PWH) (Thompson et al. 2024). The persistence of these deficits in individuals with undetectable viral loads suggests that HAND is not solely a result of direct viral neurotoxicity, but rather a multifactorial condition involving chronic neuroinflammation and innate immune dysregulation.

HAND preferentially affects specific brain regions. The middle temporal gyrus (MTG), part of the temporal neocortex, is a region that shows consistent neuropathological changes in HAND, including loss of neuronal density, reactive gliosis, synaptic damage, and altered white matter integrity (Thompson et al. 2024). Postmortem brain samples from PWH analyzed in this study were derived from the MTG. Inflammatory pathway dysregulation in the MTG is particularly relevant to the cognitive deficits characteristic of HAND, including impairments in memory, processing speed, and executive function, given the MTG’s roles in language processing, semantic memory, and multimodal sensory integration. The MTG shares substantial transcriptomic overlap with the prefrontal cortex (from which PWoH samples were derived), another neocortical region, including shared neuroimmune signatures and glial activation profiles (Gelman et al. 2012; Jorstad et al. 2023), providing justification for cross‐regional comparisons while acknowledging anatomical differences discussed in the Limitations.

Several mechanisms have been proposed to account for HAND pathogenesis. ART itself may contribute to HAND through off‐target effects on the central nervous system. Nucleoside reverse transcriptase inhibitors (NRTIs) can disrupt mitochondrial function and promote oxidative stress (Hung et al. 2017), whereas other antiretroviral classes, including non‐nucleoside reverse transcriptase inhibitors (NNRTIs) and protease inhibitors (PIs), have been implicated in neuronal damage through mechanisms such as endoplasmic reticulum stress and lipid dysregulation (Akay et al. 2014; Phulara et al. 2024).

Early models emphasized direct neurotoxicity mediated by viral proteins such as Tat, gp120, and Nef, which can impair synaptic integrity, mitochondrial function, and redox balance. However, in ART‐suppressed individuals, brain levels of these proteins are typically low and unlikely to produce sustained neurotoxic effects (Fields et al. 2019). A more compelling mechanism involves extracellular vesicles (EVs) released from HIV‐infected cells and carrying HIV‐derived factors, which circulate in the blood of ART‐treated PWH (DeMarino et al. 2024; Tang et al. 2024), and can carry these factors across the blood–brain barrier and into the central nervous system. Among these, Nef‐containing EVs are particularly relevant, as they are detectable in a substantial proportion of ART‐treated PWH with suppressed viral loads (Ferdin et al. 2018) and have been shown to provoke neuroinflammatory responses and impair myelination (Schenck et al. 2024). Chronic neuroinflammation remains a defining feature of HAND, characterized by sustained activation of microglia and astrocytes and persistent production of pro‐inflammatory cytokines, including IL‐6, TNF‐α, and IFN‐γ (Thompson et al. 2024). While HIV proteins such as Tat may contribute to this inflammatory milieu, Nef‐containing EVs appear to be more potent inducers of glial activation and inflammation in the ART era (Sviridov and Bukrinsky 2023).

Beyond classical inflammatory mechanisms, recent studies have implicated maladaptive trained immunity (TRIM) as a novel contributor to HAND (Capriotti and Klase 2024; Hajishengallis et al. 2025; Sviridov and Bukrinsky 2023). TRIM refers to the long‐term reprogramming of innate immune cells following an initial pathogenic exposure, leading to heightened inflammatory responses upon secondary stimulation. While originally characterized as a protective phenomenon, maladaptive TRIM can promote chronic inflammation and tissue damage. In the context of HAND, maladaptive TRIM may help explain the persistence of neuroinflammation in the absence of ongoing viral replication, sustaining glial activation and contributing to cumulative neuronal injury (Sviridov and Bukrinsky 2023).

HAND is a multifactorial condition shaped by a complex interplay of immune activation, metabolic dysregulation, and neurodegenerative processes. To fully capture this complexity, a systems‐level approach is essential. Multi‐omics technologies that integrate transcriptomic, epigenomic, proteomic, and metabolomic data provide a comprehensive framework for identifying novel disease mechanisms and therapeutic targets. In this study, we take an initial step toward such an integrated approach by analyzing multiomic single‐nucleus RNA sequencing (snRNA‐seq) and single‐nucleus ATAC sequencing (snATAC‐seq) data from post‐mortem brain tissue of PWH with HAD or ANI diagnosis, and comparing it to published data from PWoH controls.

We focused our analysis on inflammation, glycolysis, and cholesterol metabolism, aiming to identify dysregulated immune and metabolic pathways and characterize molecular signatures of trained immunity (TRIM), a form of innate immune memory that may underlie persistent microglial and glial activation. By defining cell‐specific inflammatory programs and chromatin accessibility patterns, this study seeks to reveal potential drivers of chronic neuroinflammation in HAND and lay the groundwork for future mechanistic and therapeutic investigations.

2. Materials and Methods

2.1. Brain Samples

Post‐mortem flash‐frozen brain samples from the middle temporal gyrus (MTG) were obtained from ART‐treated PWH through the National NeuroAIDS Tissue Consortium (NNTC), including two individuals diagnosed with HIV‐associated dementia (HAD), a severe form of HIV‐associated neurocognitive disorder (HAND), and two diagnosed with asymptomatic neurocognitive impairment (ANI), a milder HAND phenotype. For comparison, single‐nucleus RNA‐seq and ATAC‐seq data from four age‐matched PWoH individuals were obtained from the dorsolateral prefrontal cortex (DLPFC) from a previously published study (Ma et al. 2022). Donor characteristics for all samples are summarized in Table 1. The GWU IRB determined that the study is exempt from the IRB review.

TABLE 1.

Brain donors.

ID	Brain region	Sex	Age	HAND status	HIV status	VL g (copies/ml)	LOI h (years)	RNA‐seq (GEO)	ATAC‐seq (GEO)
7100107766 a	MTG c	M	38	HAD e	+	1843	12	GSE296943	GSE296943
7100616568 a	MTG	M	32	HAD	+	489,796	17	GSE296943	GSE296943
7100626868 a	MTG	M	43	ANI e	+	249	12	GSE296943	GSE296943
7102536771 a	MTG	F	40	ANI	+	157,009	4	GSE296943	GSE296943
HSB8050 b	DLPFC d	M	43	NA f	−	NA	NA	SRR19918320	SRR19918325
HSB6154 b	DLPFC	F	68	NA	−	NA	NA	SRR19918319	SRR19918324
HSB8073 b	DLPFC	F	51	NA	−	NA	NA	SRR19918318	SRR19918323
HSB6195 b	DLPFC	M	45	NA	−	NA	NA	SRR19918322	SRR19918327

^a NNTC ID.

^b Yale School of Medicine ID.

^c MTG, middle temporal gyrus.

^d DLPFC, dorsolateral prefrontal cortex.

^e HAV‐associated dementia.

^f Asymptomatic neurocognitive impairment.

^g Not applicable.

^h Viral load.

ⁱ Length of infection.

2.2. Single‐Nucleus Multiome Library Preparation and Sequencing

Single‐nucleus Multiome library preparation, sequencing, and analysis were conducted by Singulomics Corporation (https://singulomics.com/, Bronx, NY). Single‐nucleus gene expression and single‐nucleus ATAC libraries were constructed from nuclei isolated from flash‐frozen human brain tissue samples using the 10× Genomics Chromium System with the Chromium Next GEM Single Cell Multiome ATAC + Gene Expression Reagent Bundle. Flash‐frozen brain tissue was homogenized and subjected to lysis and permeabilization using Nonidet P40 Substitute, Digitonin, and Tween‐20. Isolated nuclei were purified, centrifuged, and resuspended in Diluted Nuclei Buffer before being loaded onto the Chromium X instrument (10× Genomics, Pleasanton, CA) for droplet encapsulation, following the manufacturer's protocol. For each sample, approximately 5000 nuclei were targeted for capture. Library preparation was performed according to the Chromium Next GEM Single Cell Multiome ATAC + Gene Expression User Guide. Amplified cDNAs and libraries were quantified using the Qubit dsDNA HS Assay (Thermo Fisher Scientific, Wilmington, DE) and quality‐checked using the Agilent TapeStation (Agilent Technologies, Santa Clara, CA). Libraries were sequenced on an Illumina NovaSeq X Plus instrument (Illumina, San Diego, CA), and raw sequencing reads were processed using the 10× Genomics Cell Ranger ARC pipeline (v2.0.2, cellranger‐arc count) (Satpathy et al. 2019; Zheng et al. 2017) with the human GRCh38 (2020‐A) reference genome to perform barcode processing, trimming, alignment, duplicate marking, peak calling and the creation of the peak‐barcode matrix. Comparative single‐nucleus RNA‐seq and ATAC‐seq data from four PWoH individuals were obtained from published report (Ma et al. 2022), and these datasets were processed and integrated following the same computational pipeline to ensure consistency across cohorts.

2.3. Data Analysis

The Seurat software package (v5.1.0) (Hao et al. 2024) and Signac (v 1.14.0) (Stuart et al. 2021) in R and their standard workflows were then used to perform quality control, clustering, integration, and differential gene expression analysis. For each sample (ANI, HAD, and PWoH controls), distributions of quality metrics were first examined in the unfiltered dataset. We then applied uniform thresholds across all groups to exclude low‐quality nuclei, defined as those with abnormally low UMI counts or detected genes, or with excessive mitochondrial read content, consistent with stressed or dying cells. The RNA and ATAC data were filtered by the following parameters: 200 < nFeature_RNA < 9500, percent.mt < 5 (stringent since this is single nuclei data), nCount_RNA < 60,000, nucleosome_signal < 2, and 2 < TSS.enrichment < 9. Doublets were also removed using established computational detection. The resulting filtered datasets demonstrated tighter distributions of quality metrics and improved comparability across samples. Pre‐ and post‐filtering QC metric distributions stratified by condition (ANI, HAD, PWoH) for ATAC‐seq (nCount_peaks, TSS enrichment, nucleosome signal) are shown in Figure S1A, and for RNA‐seq—nCount_RNA, nFeature_RNA, percent.mt—in Figure S1B. Pre‐filtering distributions show overlapping ranges across conditions with no evidence of systematic disease‐associated RNA degradation that would differentially bias filtering. HAD samples showed a marginally higher mean mitochondrial read proportion (3.1%) compared to ANI (2.4%) and PWoH controls (2.6%), but all samples remained well within the 5% threshold; post‐filtering QC distributions are comparable across groups. Importantly, filtering reduced variability attributable to technical noise while retaining sufficient numbers of high‐quality nuclei for each group, thereby ensuring the robustness of subsequent RNA‐seq and ATAC‐seq analyzes. The remaining cell barcodes were retained from the filter and were applied to subset the cells in two separate workflows, an snRNA sequencing workflow and an snATAC sequencing workflow.

2.4. snRNA‐Seq

In the snRNA sequencing data workflow, samples were merged, scaled, normalized, integrated, dimensionally reduced, and cell typed using the R packages Seurat and ScType (v2021) (Ianevski et al. 2022). A slightly modified gene list for the brain tissue type was used when implementing ScType in which Schwann cells were removed from this list because they were not expected to be found in the samples. Nuclei identified as unknown cells, endothelial cells, and tanycytes were removed from the analysis to focus on glia, and dopaminergic neurons and immature neurons were removed due to their low and imbalanced abundance across the samples. Integration was performed for the RNA samples utilizing the recommended workflow from Seurat for integration using SCTransform and CCAIntegration.

2.5. snATAC‐Seq

In the snATAC sequencing data workflow, the samples were read into Seurat and subsetted by the initial filter. We then employed the recommended Signac workflow (RunTFIDF, FindTopFeatures, RunSVD, and RunUMAP). Annotation was performed using GenomicRanges and EnsDb.Hsapiens.v86. Activity matrices were generated from the Seurat objects, merged, and scaled/normalized using SCTransform. Cell labels from the cell typed snRNA data were transferred to the ATAC data.

2.6. Differential Expression and Accessibility Analysis

Differential gene expression analysis was performed using a pseudo‐bulk approach with DESeq2. For each cell type, raw counts were aggregated per donor to generate donor‐level pseudo‐bulk profiles, which were then used as biological replicates (n = 4 per group for PWH vs. PWoH, HAD and ANI are combined for this comparison to maximize statistical power, consistent with established practice in the HAND field; n = 2 HAD versus n = 2 ANI, this comparison is underpowered and is treated as exploratory and hypothesis‐generating). Differential expression was modeled using a negative binomial framework with donor as the experimental unit, yielding results that account for within‐donor cell correlations. Genes with FDR‐adjusted p value ≤ 0.05 (Benjamini–Hochberg correction) and |logFC| ≥ 0.5 were considered differentially expressed. The same pseudo‐bulk DESeq2 framework was applied to ATAC activity matrices. A side‐by‐side comparison of single‐cell and pseudobulk logFC values with adjusted p value ≤ 0.05 for all genes analyzed across both RNA and ATAC modalities is provided in Table S1. Figure S2 shows the correlation of these two methods of differential gene expression testing for all genes and for the genes of interest pertaining to cholesterol metabolism, glycolysis, and inflammation and immunity—for ATAC‐seq (A) and RNA‐seq (B) data.

GO and KEGG gene set enrichment analysis (GSEA) (Subramanian et al. 2005) were conducted using the R package clusterProfiler (v4.12.6) (Wu et al. 2021). The gseGO and gseKEGG functions using all ontology options within the package were used to perform gene set enrichment analysis of ranked average log2FC differential gene expression, the FDR Benjamini–Hochberg correction, and a qvalue cutoff of 0.05. The results were filtered to include adjusted p values of less than or equal to 0.05, accounting for multiple testing, and were de‐duplicated if pathways were overlapping or similar across databases. Pathway‐level GSEA and gene‐level differential expression and accessibility analyzes are reported in parallel throughout the manuscript and provide complementary information. GSEA detects coordinated, rank‐ordered enrichment of changes distributed across all members of a defined gene set, whereas gene‐level analysis identifies individually significant loci that may or may not collectively reach pathway‐level enrichment thresholds. Consequently, a cell type may show individually significant gene‐level changes without GSEA pathway enrichment when accessibility or expression shifts are confined to a subset of pathway members or show mixed directionality across the set, and conversely may show GSEA pathway enrichment without large numbers of individually significant genes when changes are modest but coordinated across many pathway members.

2.7. Batch Correction and Integration Validation

To assess harmonization between the PWH dataset (middle temporal gyrus) and the published PWoH control dataset (dorsolateral prefrontal cortex), we computed the Local Inverse Simpson's Index (LISI) and the k‐nearest neighbor batch effect test (kBET) after CCA integration. LISI scores by batch and by sample demonstrate adequate data integration and are presented in Figure S3.

2.8. Cell‐Type Proportion Analysis

Proportions of each cell type per donor were calculated. Results are summarized in Table S2 and Figure S4.

2.9. Peak‐Gene Correlation Analysis

Peak‐gene comparisons between promoter ATAC peaks and corresponding normalized and scaled RNA expression were computed using Seurat and Signac within 500 kb windows (Figure S5). Visualizations were created using Python's matplotlib.

2.10. Data Availability Statement

The datasets of PWoH controls used in this manuscript can be accessed from GEO (SRR19918318, SRR19918319, SRR19918320, SRR19918322). The datasets of the four samples of PWH may be downloaded from GEO (GSE296943).

3. Results

3.1. Overview of Cell Type Composition and Disease‐State Clustering

To characterize global transcriptomic relationships across brain cell populations, we performed UMAP analysis on the snRNA‐seq data (Figure 1). Major neural and glial populations were annotated using canonical markers, including glutamatergic, GABAergic, and dopaminergic neurons, astrocytes, oligodendrocytes, oligodendrocyte precursor cells (OPCs), microglia, neuroblasts, and unclassified neurons. Cell clusters were well‐preserved across all groups, validating annotation quality and data integration.

FIGURE 1.

UMAP visualization of cell type‐ and disease state‐specific transcriptomic profiles. UMAP projection of single‐nucleus RNA‐seq data from PWH with ANI (left), PWH with HAD (center), and PWoH controls (right), colored by cell type annotated on the right. The UMAP projection was computed using the top principal components derived from highly variable genes and visualized in two dimensions. Cell clusters are labeled with numeric identifiers overlaid directly on the UMAP. Color coding is consistent across all three panels. “Unclassified neurons” denotes neurons that passed quality filters but did not meet expression thresholds for classical neurotransmitter‐associated genes (glutamatergic, GABAergic, or dopaminergic).

The following sections describe transcriptomic and epigenomic changes in glial cells and neurons in turn. For each cell type, pathway‐level and gene‐level RNA‐seq findings are presented first, followed by complementary ATAC‐seq chromatin accessibility data, allowing transcriptional and epigenetic changes to be interpreted together. Functional descriptions, cell type specificity, and trained immunity connections for all genes are provided in Table S3.

3.2. Glial Cells

3.2.1. Pathway‐Level Transcriptomic Changes in Glia

GSEA of snRNA‐seq data (Wilcoxon, FDR Benjamini‐Hochberg adjusted p value ≤ 0.05, |log2FC| ≥ 0.5) revealed cell type–specific dysregulation of metabolic and inflammatory pathways across glial populations in PWH versus PWoH comparisons. Cholesterol metabolism pathways (Figure 2A) showed differential regulation across glial subtypes: “cholesterol metabolism” was upregulated in oligodendrocytes and microglia, whereas astrocytes showed downregulation of related biosynthetic pathways. Glycolysis‐related pathways (Figure 2B) were similarly cell type–specific: “canonical glycolysis” was downregulated specifically in oligodendrocytes. Immune and inflammatory pathways (Figure 2C) showed the most pronounced changes: microglia and oligodendrocytes displayed broad upregulation, while astrocytes showed more limited or downregulated responses. A subset of pathways, notably “cytokine–cytokine receptor interaction” and “interleukin‐1 production,” were consistently upregulated across microglia and oligodendrocyte subtypes. HAD versus ANI comparisons revealed relatively few additional changes, the most notable being downregulation of the “chemokine production”; given the limited cohort size (n = 2 per group), these findings should be considered exploratory. Overall, these results demonstrate widespread but cell type–specific transcriptional dysregulation in glial cells from HIV‐infected brains, with microglia displaying the most robust pro‐inflammatory signature.

FIGURE 2.

snRNA‐seq gene set enrichment analysis (GSEA) of glial pathways. The figure shows the results of GSEA in glial cells for genes associated with cholesterol metabolism (A), glycolysis (B), and inflammation (C) comparing HAD vs. ANI, and PWH (HAD and ANI) vs. PWoH. Statistical test: GseGO and gseKEGG using a gene list ranked by average log2FC; FDR correction: Benjamini‐Hochberg; significance threshold: Adj. p ≤ 0.05; effect size filter: |log2FC| ≥ 0.5.

3.3. Gene‐Level Transcriptomic Changes in Glia

To complement pathway‐level findings, we analyzed differential gene expression in glial populations using dot plots focused on cholesterol metabolism, glycolysis, and inflammation pathways (Figure 3; all statistics are in Table S1). Full gene descriptions are provided in Table S3.

FIGURE 3.

snRNA‐seq differential gene expression (DGE) analysis in glia. The figure depicts the main differentially expressed genes in glia between the same groups as in Figure 2. Panel labels: (A) cholesterol metabolism, (B) glycolysis, and (C) inflammation. Statistical test: Single‐cell Wilcoxon; FDR correction: Benjamini‐Hochberg; significance threshold: Adj. p ≤ 0.05; effect size: |log2FC| ≥ 0.5.

Cholesterol homeostasis (Figure 3A). In PWH versus PWoH, ABCA1 (Villa et al. 2024), ABCG1 (Villa et al. 2024), and APOE (Chen et al. 2026) were generally upregulated across astrocytes, microglia, and oligodendrocytes. Conversely, LDLR (Nowacka et al. 2025) was strongly downregulated in astrocytes and microglia. Pseudobulk analysis additionally identified downregulation of HMGCR and HMGCS1, key enzymes in de novo cholesterol biosynthesis, in astrocytes in the PWH versus PWoH comparison (Table S1), further supporting suppression of endogenous cholesterol synthesis alongside increased expression of cholesterol efflux mediators. Together, the combination of increased expression of efflux transporters and apolipoproteins with reduced expression of cholesterol uptake and biosynthetic components suggests substantial remodeling of cholesterol homeostasis in PWH glia. These changes are consistent with compensatory responses to increased intracellular cholesterol and may have downstream consequences for membrane composition, lipid raft organization, and inflammatory signaling (Ho et al. 2022). HAD vs. ANI comparison revealed downregulation of APOE in astrocytes and ABCA1 in oligodendrocytes, and upregulation of LDLR in oligos. Together with the broader increase in cholesterol efflux and reduction in cholesterol biosynthetic pathways observed in PWH glia, these HAD‐associated changes may reflect a compensatory shift toward increased intracellular cholesterol retention and uptake in oligodendrocytes in response to myelin stress at the HAD stage.

Glycolysis (Figure 3B). In PWH versus PWoH, GAPDH expression was reduced in astrocytes, microglia, and oligodendrocytes (Donnelly and Finlay 2015; Schmalhausen et al. 2024). In contrast, PFKFB3 (Keating et al. 2020) was upregulated in microglia and oligodendrocytes, while HK2 (Jia et al. 2025) and PGK1 (Kokotos et al. 2024) were increased in microglia. HAD versus ANI differences were minimal, suggesting that HIV infection itself accounts for most of the glycolytic alterations, with relatively limited additional remodeling associated with neurocognitive disease severity. Overall, these findings indicate partial and cell type–specific glycolytic reprogramming in PWH glia rather than the coordinated induction of the full glycolytic machinery typically observed in canonical trained immunity.

Immunity and inflammation (Figure  3C). In PWH versus PWoH, microglia showed increased expression of innate immune sensors and inflammatory signaling mediators including IL18 (Mohammad et al. 2025), IRAK3 (Pereira and Gazzinelli 2023), NLRP3 (Lee et al. 2023), IL6ST (D. Lin et al. 2023), JAK1 (Qin et al. 2024), TLR2, and TLR4 (Alexopoulou and Irla 2025). TLR4 was also upregulated in astrocytes (Henneberger and Steinhauser 2016), consistent with secondary astrocytic activation driven by microglia‐derived DAMPs and inflammatory mediators. Notably, astrocytic TLR4 upregulation was observed across all pairwise pseudobulk comparisons (PWH vs. PWoH, ANI vs. PWoH, and HAD vs. PWoH; Table S1), making it one of the most consistently supported inflammatory findings in the dataset. Astrocytes and oligodendrocytes additionally showed increased expression of HMGB1 (Rajkovic et al. 2025; Vuscan et al. 2024) and IL6ST, together with reduced IL1RAP (Zarezadeh Mehrabadi et al. 2024), indicating broad remodeling of inflammatory signaling pathways across glial populations. In the HAD versus ANI comparison, microglia showed reduced expression of IL18, IL6ST, and TLR2, potentially reflecting partial counter‐regulatory adaptation or transition away from the highly activated inflammatory state established during HIV infection. In contrast, astrocytes and oligodendrocytes showed reduced JAK1 but increased HMGB1 expression, suggesting further reinforcement of DAMP‐associated inflammatory signaling in non‐microglial glia at the HAD stage.

These findings reveal cell type–specific transcriptional reprogramming in PWH glia consistent with maladaptive trained immunity: microglia showed the most complete signature, with upregulation of inflammasome sensors, pattern recognition receptors, and cytokine signal transducers, while oligodendrocytes and astrocytes showed convergent but partial reprogramming marked by HMGB1 upregulation and JAK1 and IL1RAP suppression. HAD‐associated changes were subtler but suggested further reinforcement of this configuration in HAD relative to ANI.

Taken together, gene‐level transcriptomic analysis of PWH glia reveals a coordinated pattern of dysregulation across cholesterol homeostasis, glycolytic metabolism, and inflammatory signaling collectively recapitulating the core molecular hallmarks of maladaptive trained immunity. Microglia displayed the most complete convergence of all three programs, while astrocytes and oligodendrocytes showed partial and cell type–specific reprogramming. HAD‐associated transcriptional shifts in glia involved selective reinforcement across all three arms, suggesting that trained immunity‐related configuration deepens and reorganizes in HAD, rather than qualitatively alters the state established by HIV infection itself.

3.4. Chromatin Accessibility Changes in Glial Cells

To examine the epigenetic basis of glial dysfunction in HAND, we analyzed snATAC‐seq data using dot plots at the pathway (Figure 4) and gene (Figure 5) levels focused on cholesterol metabolism, glycolysis, and inflammation, supplemented by unbiased volcano plot analysis (Figures [Link], [Link]). Full descriptions of related genes are provided in Table S3.

FIGURE 4.

snATAC‐seq GSEA in glia. The figure presents chromatin accessibility changes and the resulting pathway analysis in glia comparing the same groups as in Figure 2. Panel labels: (A) cholesterol metabolism, (B) glycolysis, and (C) inflammation. Statistical test: GseGO and gseKEGG using a gene list ranked by average log2FC; FDR correction: Benjamini‐Hochberg; significance threshold: Adj. p ≤ 0.05; effect size: |log2FC| ≥ 0.5.

FIGURE 5.

snATAC‐seq differentially accessible regions (DARs) in glia. The figure shows differentially accessible regions (DARs) in glia between the same groups as in Figure 2. Panel labels: Cholesterol metabolism (A), glycolysis (B), and inflammation (C). Statistical test: Single cell Wilcoxon; FDR correction: Benjamini‐Hochberg; significance threshold: Adj. p ≤ 0.05; effect size: |log2FC| ≥ 0.5.

Cholesterol metabolism (Figures 4A, 5A). Across glial cell types, PWH versus PWoH comparisons revealed widespread increased chromatin accessibility at cholesterol metabolism‐related pathways: “cholesterol metabolism” and “steroid hormone biosynthesis” in astrocytes and microglia, “bile secretion” in microglia, and “steroid biosynthesis” in oligodendrocytes (Figure 4A). In gene‐level analysis (Figure 5A), all glial cell types showed increased accessibility at RXRA (Natrajan et al. 2015; Nunez et al. 2010), and astrocytes and microglia—at APOE, concordant with transcriptional upregulation of these genes (Figure 3A).

In the HAD versus ANI comparison, additional cholesterol‐related loci demonstrated increased chromatin accessibility, including MSR1 (Husemann et al. 2002; Idiiatullina and Parker 2025) across all glial populations and CD36 (Zhang et al. 2026) in astrocytes and oligodendrocytes. Both genes represent major scavenger receptors involved in lipid uptake and foam cell–like inflammatory phenotypes and are well‐established components of trained myeloid cell reprogramming, supporting convergent cross‐glial epigenetic priming of scavenger receptor–mediated cholesterol uptake. Both loci were additionally supported by pseudobulk analysis (Table S1).

In contrast, RXRA accessibility was reduced across all glial populations, together with NPC1 in microglia and oligodendrocytes, LRP1 in microglia, and CYP46A1 and ABCG1 in oligodendrocytes. These coordinated changes suggest broad disruption of cholesterol trafficking, efflux, and regulatory pathways at the stage when myelin maintenance and cholesterol redistribution become increasingly stressed. In this configuration, chromatin accessibility at lipid scavenger receptors remains preserved or enhanced, whereas multiple components of the homeostatic cholesterol regulatory network become epigenetically suppressed. HAD‐associated transcriptional shifts in glia demonstrated selective reinforcement across all three functional arms, suggesting that the trained immunity‐related configuration established during HIV infection becomes further amplified and reorganized in HAD, rather than qualitatively transformed into a distinct state.

Glycolysis (Figures 4B, 5B). PWH astrocytes showed limited pathway‐level GSEA enrichment in glycolysis gene accessibility (Figure 4B), consistent with the limited glycolytic pathway changes observed in astrocyte RNA‐seq (Figure 2B), with only the “pentose phosphate” and “starch and sucrose metabolism” pathways showing increased chromatin accessibility. Gene‐level analysis in astrocytes (Figure 5B) identified partially increased accessibility only at HK1, indicative of weak and gene‐specific epigenetic priming rather than a coordinated pathway‐level response. Glycolytic chromatin remodeling was also observed in PWH microglia and oligodendrocytes. Microglia showed partially increased accessibility at HK2 (Jia et al. 2025), consistent with epigenetic priming of glucose‐trapping glycolysis previously described in trained myeloid cells. Oligodendrocytes showed increased accessibility at GAPDH, PFKFB3, PRKAA1 (T. Sun et al. 2024; Y. Zhang et al. 2022), and SLC2A1 (Figure 5B).

In HAD versus ANI comparison, PFKFB3 accessibility was reduced across all glial populations, together with HK2 in microglia and SLC2A1, ALDOC (Fujita et al. 2014; Liu et al. 2020), ENO1 (Liang et al. 2023), and GAPDH in oligodendrocytes. In contrast, accessibility was increased in oligodendrocytes at HK1, LDHB (Spate et al. 2024), and PRKAA2 (Shariq et al. 2024). These findings suggest progressive remodeling of glycolytic regulation away from coordinated glycolytic activation toward a more restricted metabolic configuration characterized by preserved glucose phosphorylation and lactate metabolism but reduced accessibility of multiple intermediate glycolytic components. Microglial glycolytic loci exhibited relatively limited differences between HAD and ANI, suggesting that much of the glycolytic epigenetic remodeling is already established during HIV infection and remains largely preserved across neurocognitive disease severity states.

Inflammation (Figures 4C, 5C). In the PWH versus PWoH comparison, microglia showed robust GSEA enrichment of inflammation‐related pathways (Figure 4C), together with increased chromatin accessibility at inflammatory loci including AIF1 (De Leon‐Oliva et al. 2023), IL10 (Porro et al. 2020), NLRP3, and TLR7 (Suvieri et al. 2024) (Figure 5C), spanning inflammatory sensing, cytokine production, phagocytic activation, and pattern recognition receptor programs. In striking contrast, no significant inflammation‐related chromatin accessibility changes were detected in astrocytes or oligodendrocytes in the PWH versus PWoH comparison, consistent with the limited inflammatory pathway changes observed by RNA‐seq (Figure 2C). This chromatin‐level stability may reflect relative restriction of inflammatory epigenetic remodeling to microglia, with astrocytes and oligodendrocytes instead exhibiting predominantly metabolic and cholesterol‐associated alterations.

In the HAD versus ANI comparison at the gene level (Figure 5C), HAD microglia showed no consistent additional gains at inflammatory loci already opened in PWH, suggesting that HIV infection itself establishes much of the microglial inflammatory epigenetic priming, with relatively limited further remodeling associated with neurocognitive disease severity. Astrocytes likewise showed minimal inflammatory accessibility changes, reinforcing the conclusion that the astrocytic response to HIV and HAND is primarily metabolic and cholesterol‐associated rather than inflammatory. In oligodendrocytes, increased accessibility was observed at IL1R1 (Kim and Lee 2024) and NLRP3 (Figure 5C), indicating stage‐associated epigenetic licensing of discrete inflammatory loci that may render oligodendrocytes more susceptible to inflammasome‐associated injury at the HAD stage.

Unbiased analysis (Figures [Link], [Link]) identified additional disease‐associated chromatin remodeling not captured by the analysis focusing on cholesterol, glycolysis, and inflammation. In PWH versus PWoH astrocytes (Figure S6A), reduced accessibility was observed at ISG20, an interferon‐inducible antiviral exonuclease involved in innate immune defense (Deymier et al. 2022), together with HIST1H2AE and HIST1H2BG, histone genes associated with nucleosome assembly and chromatin organization (Garciaz et al. 2019). These changes suggest altered antiviral signaling and epigenetic remodeling in astrocytes during HIV infection. However, astrocytes showed little evidence of increased accessibility at canonical trained‐immunity inflammatory loci, supporting the conclusion that the astrocytic response to HIV is driven primarily by metabolic and noncanonical epigenetic alterations rather than robust inflammatory chromatin priming.

In HAD versus ANI astrocytes (Figure S6B), reduced accessibility was observed at CEBPD, a stress‐ and inflammation‐responsive transcription factor induced by IL‐1β, TNFα, and TLR signaling that regulates reactive astrocyte activation and inflammatory gene expression (Wang et al. 2018). Decreased accessibility was also observed at MTHFR, a regulator of one‐carbon metabolism and methylation capacity linked to redox balance and inflammatory regulation (Jadavji et al. 2018). These changes suggest suppression of selected astrocytic inflammatory and metabolic regulatory programs in advanced HAND. However, the overall scarcity of canonical trained‐immunity inflammatory loci in astrocytes further supports the conclusion that astrocytic epigenetic remodeling during HIV infection and HAND is driven predominantly by metabolic and homeostatic alterations rather than robust inflammatory chromatin priming.

In PWH versus PWoH comparison of microglia (Figure S7A), increased accessibility was observed at NCF1 (Gao et al. 2024), consistent with epigenetic priming of oxidative burst capacity that simultaneously amplifies inflammatory effector function and reinforces the trained epigenetic state, SRGAP2B (Schmidt et al. 2019), and FCGR1A (Li et al. 2025; Yu et al. 2025), whose FcγR‐mediated activation is a recognized inducer of trained immunity in macrophages.

In the HAD versus ANI comparison of microglia (Figure S7B), further increased accessibility at ASB17 (Wan et al. 2022), consistent with further opening of NF‐κB‐dependent trained immunity loci in HAD relative to ANI, was accompanied by reduced accessibility at TOMM40 (Honea et al. 2025) and MZF1 (Luo et al. 2009). Mitochondrial TCA‐derived metabolites are required to sustain and expand trained immunity epigenomes upon restimulation; TOMM40 closure may therefore limit the capacity of maladaptively trained microglia to further propagate their inflammatory epigenetic programs in response to subsequent challenges, a potentially self‐limiting constraint that may be paradoxically beneficial in the most severely affected brains. MZF1 reduced accessibility parallels this TOMM40 contraction and may reflect a broader epigenetic retraction of the transcriptional infrastructure required to sustain and propagate the maladaptive trained immunity program. Together, these HAD‐specific changes indicate HAD‐associated suppression of mitochondrial epigenetic metabolite production alongside continued NF‐κB pathway activation, a configuration consistent with intrinsic constraints on further microglial inflammatory reprogramming in end‐stage HAND.

In the PWH versus PWoH comparison of oligodendrocytes (Figure S8A), increased accessibility was observed at ICOSLG, an immune regulatory ligand involved in inflammatory cell–cell signaling (Holst et al. 2021), whereas reduced accessibility was detected at CEBPB, a transcription factor regulating inflammatory activation, stress responses, and oligodendrocyte differentiation programs (Sha et al. 2025). These changes suggest selective remodeling of inflammatory and glial maintenance pathways rather than broad activation of canonical trained‐immunity programs. However, neither CEBPB nor ICOSLG is canonically expressed in mature oligodendrocytes, and their reduced accessibility may reflect lineage‐appropriate chromatin closure rather than disease‐driven repression. We therefore report these observations as descriptive findings from unbiased volcano analysis without assigning mechanistic significance for oligodendrocyte biology.

In the HAD versus ANI comparison (Figure S8B), increased accessibility was observed at ARG1, an immunometabolic regulator linked to arginine metabolism, tissue repair, and remyelination‐associated glial responses (Khoja et al. 2022), while ICOSLG accessibility was reduced (Holst et al. 2021). Increased ARG1 accessibility may reflect a compensatory response aimed at supporting repair and limiting inflammatory injury in the setting of HAD‐stage neurodegeneration. Overall, oligodendrocytes demonstrated limited but detectable inflammatory epigenetic remodeling coupled to pathways involved in myelin maintenance and repair, consistent with secondary adaptation to chronic neuroinflammatory stress rather than establishment of a robust canonical trained‐immunity state.

Peak‐gene correlation analysis confirmed concordance between chromatin accessibility and gene expression at key loci (Figure S5).

Together, glial ATAC‐seq findings reveal extensive epigenomic remodeling in PWH that both parallels and extends the transcriptional alterations identified by RNA‐seq. HIV infection established widespread chromatin accessibility changes at cholesterol metabolism, glycolytic, and inflammatory loci, with microglia showing the strongest inflammatory epigenetic priming, oligodendrocytes demonstrating coordinated metabolic and myelin‐associated remodeling, and astrocytes exhibiting predominantly metabolic alterations with relatively limited inflammatory chromatin activation.

Differences between ANI and HAD were characterized less by uniform amplification of inflammatory accessibility and more by selective reorganization of glial metabolic and inflammatory programs. Oligodendrocytes showed more pronounced disruption of cholesterol efflux in HAD, glycolytic coordination, and myelin‐supportive regulatory pathways together with increased accessibility at scavenger receptor and inflammasome‐associated loci, consistent with mounting metabolic stress and compensatory repair responses. Astrocytes demonstrated further suppression of inflammatory and metabolic regulatory loci without evidence of broad inflammatory epigenetic activation, reinforcing the conclusion that astrocytic remodeling in HAND is primarily noncanonical and homeostatic rather than classically inflammatory. Microglia showed relatively limited additional inflammatory accessibility gains in HAD, suggesting that much of the microglial inflammatory epigenetic priming is established early during HIV infection rather than progressively acquired during dementia‐stage disease.

Overall, these findings support a model in which chronic HIV infection induces persistent glial epigenetic remodeling involving inflammatory, metabolic, and cholesterol‐associated pathways, with distinct cell type–specific configurations. While the data are consistent with maladaptive innate immune reprogramming, particularly in microglia and oligodendrocytes, the epigenetic changes observed in advanced HAND appear to reflect progressive dysregulation and compensatory remodeling of glial homeostatic functions rather than simple escalation of canonical trained inflammatory programs.

3.5. Neurons

3.5.1. Pathway‐Level Transcriptomic Changes in Neurons

GSEA of neuronal RNA‐seq data revealed a pattern strikingly different from that observed in glia (Figure 6). In PWH versus PWoH comparisons, glycolytic and cholesterol metabolism pathways showed only minimal changes. By contrast, HAD versus ANI comparisons revealed a reversal of this pattern: prominent upregulation of inflammation‐associated pathways including “cytokine‐cytokine receptor interaction,” “JAK–STAT signaling,” and “MAPK signaling” across all neuronal subtypes, along with glycolytic pathway upregulation in unclassified neurons.

FIGURE 6.

snRNA‐seq GSEA of neuronal pathways. The figure shows the results of GSEA in neurons for pathways associated with cholesterol metabolism (A), glycolysis (B), and inflammation (C). Comparisons were done as in Figure 2. Statistical test: GseGO and gseKEGG using a gene list ranked by average log2FC; FDR correction: Benjamini‐Hochberg; significance threshold: Adj. p ≤ 0.05; effect size: |log2FC| ≥ 0.5.

3.5.2. Gene‐Level Transcriptomic Changes in Neurons

Targeted (Figure 7) and unbiased (Figures S9 and S10, Table S1) analyzes of neuronal gene expression provided resolution of the individual genes driving these pathway patterns. Full gene descriptions are provided in Table S3. Across both glutamatergic and GABAergic neurons in PWH versus PWoH, increased expression of mitochondrial transcripts MT‐ND1, MT‐ND2, MT‐ND3, MT‐ND6 (X. Lin et al. 2024), and MT‐ATP6 (Torregrosa‐Munumer et al. 2025) was observed (Figure S9). However, because nuclear‐encoded oxidative phosphorylation genes were simultaneously downregulated in both neuronal populations (e.g., ATP5F1E, NDUFA13, UQCRC1, COX4I1 in GABAergic neurons, and NDUFS2, ATP5F1E, COX4I1 in glutamatergic neurons, Table S1) this mitochondrial transcript enrichment likely reflects a technical or normalization‐related artifact rather than a coordinated compensatory response to mitochondrial stress, despite the application of standardized filtering, scaling and normalization across all samples. Glutamatergic neurons showed downregulation of NPAS4 (Fu et al. 2020), NR4A1 (Han et al. 2022), and BCL6 (Bonnefont et al. 2019), while GABAergic neurons showed downregulation of HSPA6 (Deane and Brown 2018) and NR4A3 (Han et al. 2022), collectively reflecting suppression of experience‐regulated neuroprotective, stress‐response, and synaptic maintenance programs across neuronal subtypes.

FIGURE 7.

snRNA‐seq DGE in neuronal samples. The figure presents the key differentially expressed genes related to cholesterol metabolism (A), glycolysis (B), and inflammation (C) in neurons, comparing the same groups as in Figure 2. Statistical test: Single cell Wilcoxon; FDR correction: Benjamini‐Hochberg; significance threshold: Adj. p ≤ 0.05; effect size: |log2FC| ≥ 0.5.

Although GSEA identified enrichment of immune pathways in HAD neurons, many individual pathway genes were downregulated or did not reach significance, likely reflecting modest but coordinated changes captured by GSEA's ranking approach but falling below individual differential expression thresholds. In the HAD versus ANI comparison (Figure 7), glutamatergic neurons showed upregulation of PRLR (Mellado et al. 2022) and TUBA1C (Gui et al. 2021), likely representing compensatory but insufficient neuroprotective responses, alongside downregulation of TRIM36 (Mascaro et al. 2022). GABAergic neurons exhibited upregulation of DUSP8 (Baumann et al. 2019), SEZ6L (Ong‐Palsson et al. 2022), and BID (Plesnila et al. 2001), alongside downregulation of ANOS1 (Garcia‐Gonzalez et al. 2016; Hara and Tanegashima 2014), CNTNAP3B (Ji et al. 2025), and VAMP2 (Yan et al. 2022).

These findings reveal a two‐phase transcriptional response in HAND neurons. In PWH, the dominant signature is transcriptional silencing of neuroprotective and synaptic maintenance programs, impairing synaptic plasticity and cellular resilience at a stage preceding overt dementia. At the HAD stage, this suppressive baseline is accompanied by activation of stress kinase, complement‐mediated synaptic pruning, and pro‐apoptotic programs in GABAergic neurons alongside compensatory neuroprotective responses in glutamatergic neurons, with concurrent repression of inhibitory circuit organization, synaptic vesicle release, and axon guidance genes, indicating coordinated deterioration of both excitatory and inhibitory circuit function concordant with the glial neuroinflammatory signature observed in HAD.

3.5.3. Chromatin Accessibility Changes in Neurons

Neuronal ATAC‐seq data (Figures 8 and 9; Figure S11) revealed that epigenetic changes in neurons were generally more limited in magnitude than in glia but showed disease stage–specific and subtype‐specific patterns that complement the transcriptional findings. Full gene descriptions are provided in Table S3.

FIGURE 8.

snATAC‐seq GSEA in neurons. The figure shows the results of chromatin accessibility pathway analysis (GSEA) in neurons comparing the same groups as in Figure 2. Panel labels: (A) cholesterol metabolism, (B) glycolysis, (C) inflammation. Statistical test: GseGO and gseKEGG using a gene list ranked by average log2FC; FDR correction: Benjamini‐Hochberg; significance threshold: Adj. p ≤ 0.05; effect size: |log2FC| ≥ 0.5.

FIGURE 9.

snATAC‐seq DARs in neurons. The figure displays key differentially accessible regions (DARs) in neurons between the same groups as in Figure 2. Panel labels: (A) cholesterol metabolism, (B) glycolysis, (C) inflammation. Statistical test: Single cell Wilcoxon; FDR correction: Benjamini‐Hochberg; significance threshold: Adj. p ≤ 0.05; effect size: |log2FC| ≥ 0.5; biological replicates.

In glycolytic and bioenergetic pathways (Figure 8B), “ATP synthesis coupled electron transport” was associated with reduced chromatin accessibility in GABAergic, glutamatergic, and unclassified neurons in PWH versus PWoH, suggesting a broad epigenetic suppression of neuronal oxidative phosphorylation and mitochondrial energy production programs. “Inositol phosphate metabolism” showed increased accessibility in GABAergic and glutamatergic neurons, alongside “pentose and glucuronate interconversions” in GABAergic neurons. Gene‐level analysis (Figure 9B) revealed a mixed pattern: reduced accessibility at GAPDH, LDHA (Frame et al. 2024), PDHB (Jiang et al. 2023), PGK1 (Kokotos et al. 2024), and SLC2A1 (this reduced accessibility in PWH neurons contrasts with the increased accessibility observed in PWH oligodendrocytes and SLC2A1 upregulation in oligodendrocyte RNA‐seq, suggesting cell type–specific epigenetic divergence in the regulation of glucose import consistent with glial prioritization of glucose resources) in unclassified neurons coexisted with increased accessibility at GPI (Knight et al. 2014), SRC (Gallo 2024), and TRAP1 (a mitochondrial chaperone suppressing oxidative phosphorylation, whose increased accessibility alongside reduced ETC loci suggests convergent epigenetic suppression of oxidative phosphorylation from two directions), suggesting dysregulated rather than uniformly suppressed glycolytic organization. PFKFB3 appeared among differentially accessible loci only in neuroblasts, consistent with its known constitutive proteasomal degradation in mature neurons. HAD versus ANI changes were limited to reduced PFKP accessibility in glutamatergic neurons.

At the gene level (Figure 9A), accessibility was increased at LRP8 (Passarella et al. 2022) and NCOA1 (Z. Sun and Xu 2020). Pseudobulk ATAC analysis provided additional support for accessibility gains at VAT1L and NCOR1 (cholesterol metabolism) in unclassified neurons in the PWH versus PWoH comparison, and at MYO9A (inflammatory signaling) in unclassified neurons in the HAD versus PWoH comparison; these additional loci are reported in Table S1. HAD versus ANI differences in these pathways were minor, suggesting that neuronal cholesterol epigenetic remodeling is established early in HIV infection and may not accumulate further with cognitive decline.

In inflammatory pathways (Figure 8C), “cytokine‐cytokine receptor interaction,” and “JAK–STAT signaling” showed increased pathway‐level accessibility in both comparisons, more pronounced in HAD versus ANI. Surprisingly, closed chromatin was found at JAK1 (mirroring JAK1 downregulation in astrocytes and oligodendrocytes in RNA‐seq HAD vs. ANI comparison, suggesting convergent cross‐cell‐type epigenetic impairment of cytokine regulatory signaling), PARK7 (Bao et al. 2025), PIN1 (S. C. Wang et al. 2023), PRKCA (X. Wu et al. 2025), RIPK1 (Pajulas et al. 2025), and TBK1 (Ahmad et al. 2016), while CAMK4 (Madhi et al. 2021) showed increased accessibility. This discordance between pathway enrichment and predominantly closed gene‐level chromatin is consistent with dysregulated rather than coordinated inflammatory pathway engagement, in which epigenetic silencing of both pro‐ and anti‐inflammatory regulators impairs the neuron's capacity to mount or resolve a coordinated inflammatory response.

Unbiased HAD versus ANI volcano analysis (Figure S11) revealed subtype‐specific remodeling: glutamatergic neurons showed reduced accessibility at HTR1B (Tadic et al. 2009), NEUROD2 (Runge et al. 2021), consistent with epigenetic erosion of the transcriptional program maintaining excitatory neuron identity, paralleling downregulation of other synaptic structural genes in glutamatergic RNA‐seq, and NECAB2 (Bueno et al. 2023). GABAergic neurons showed reduced accessibility at EPHB3 (Perez et al. 2016) and PCDHGC5/3 (Mancia Leon et al. 2020), suggesting that GABAergic neurons in HAD are particularly vulnerable to stress‐induced apoptotic priming and loss of inhibitory circuit stability.

These findings reveal a two‐phase epigenomic response in HAND neurons concordant with and extending the transcriptional findings. In PWH, the dominant signature is epigenetic constraint at mitochondrial and glycolytic flux loci alongside closure of neuroprotective and neuromodulatory regulatory genes, consistent with bystander epigenetic reprogramming of neurons by the surrounding glial neuroinflammatory microenvironment rather than cell‐autonomous trained immunity induction, though the two mechanisms are not mutually exclusive. At the HAD stage, this suppressive baseline is reorganized into subtype‐specific chromatin remodeling: GABAergic neurons show concordant epigenetic and transcriptional priming of stress kinase, complement‐mediated pruning, and pro‐apoptotic programs alongside closure of inhibitory circuit adhesion and synapse specificity loci, while glutamatergic neurons show more pronounced epigenetic erosion of excitatory neuron identity and calcium regulatory programs, together indicating coordinated epigenetic deterioration of both excitatory and inhibitory neuronal circuit maintenance programs concordant with the glial neuroinflammatory epigenome observed in HAD.

4. Discussion

4.1. Neuroinflammation in HAND: Maladaptive Trained Immunity Across Glial Cell Types

This study reveals that HIV infection triggers widespread but cell type–specific transcriptional and epigenomic reprogramming in the human brain, consistent with, though not proof of, a model of maladaptive trained immunity in which glial cells acquire durable epigenetic configurations that sustain and amplify neuroinflammatory programs beyond the resolution of acute infection. Across all glial cell types examined, the pattern of chromatin remodeling was concordant with transcriptional changes and organized into two distinct conditions: a baseline reprogramming associated with HIV infection itself and a selective reorganization, rather than uniform amplification, associated with differences in neurocognitive disease severity.

Microglia displayed the most complete signature consistent with maladaptive trained immunity, with transcriptional upregulation of innate immune sensors and signaling mediators (including the inflammasome sensor NLRP3, toll‐like receptors TLR2 and TLR4, and cytokine signal transducers) supported by increased chromatin accessibility at inflammatory loci in the targeted analysis (notably AIF1, IL10, NLRP3, and TLR7) and at oxidative burst, cytokine production, and phagocytic activation loci in unbiased volcano analysis—together spanning the three defining features of trained immunity described in peripheral macrophages (Hajishengallis et al. 2025). Critically, microglial epigenetic priming at these loci was established uniformly across the HAND severity spectrum, with HAD and ANI microglia showing no consistent differences in accessibility at the inflammatory loci opened in PWH, suggesting that microglial epigenetic reprogramming is a consequence of HIV infection itself rather than an accumulating correlate of cognitive decline. This interpretation is consistent with the classical trained immunity model, in which epigenetic changes are established by the initial stimulus and persist independently of ongoing transcriptional activation (Fanucchi et al. 2021).

The metabolic arm of this signature shows a two‐stage pattern that further refines the trained‐immunity interpretation (Table S1). At the PWH baseline, microglia exhibit broad epigenetic licensing of the glycolytic‐to‐lactate flux that underlies trained myeloid metabolism (Arts et al. 2016; Cheng et al. 2014), with increased chromatin accessibility across multiple glycolytic, lactate‐handling, and AMPK‐regulatory loci (including HK1, HK2, PFKFB4, ENO3, ACSS1, SLC16A3, and PRKAB1) accompanied by concordant transcriptional upregulation of PFKFB3, HK2, and PGK1 and remodeling of cholesterol regulatory programs (APOE, ABCA1, ABCG1 upregulated; LDLR, HMGCR, and HMGCS1 downregulated). Notably, this glycolytic chromatin opening is not accompanied by coordinated closure of mitochondrial loci at the PWH baseline; to the contrary, multiple nuclear‐encoded OXPHOS genes (including NDUFS2, NDUFA2, NDUFA11, COX14, and ATP6V0C) show increased rather than decreased accessibility, indicating that the OXPHOS chromatin landscape remains permissive at the HIV‐infected baseline. This dual‐permissive configuration is consistent with the metabolic flexibility required to sustain trained‐immunity effector functions, in which glycolytic flux supports rapid ATP generation and the oxidative burst (supported in our data by NCF1 chromatin opening) while mitochondrial TCA‐cycle intermediates continue to supply the succinate, fumarate, α‐ketoglutarate, and acetyl‐CoA required as cofactors and substrates by the histone‐modifying enzymes that maintain the trained epigenome (Dominguez‐Andres and Netea 2019). At the dementia stage, this dual‐permissive landscape contracts: HAD microglia show coordinated reduction of accessibility at glycolytic and lactate‐handling loci (including HK2, PFKFB3, ACSS1, ACSS2, ALDOA, and SLC16A3) alongside closure of the mitochondrial‐import locus TOMM40 and of the myeloid transcription factor MZF1, with continued NF‐κB‐associated accessibility gain at ASB17. This bilateral contraction of both glycolytic and mitochondrial‐import arms of energy metabolism would be expected to limit both the flux required for inflammatory effector function and the mitochondrial metabolite supply required to refresh the trained epigenome upon restimulation, providing a metabolic substrate for the trained‐immunity contraction model—a potentially self‐limiting feature of the microglial inflammatory program in end‐stage HAND. Whether this contraction reflects homeostatic counter‐regulation, loss of capacity to sustain the trained state, or a compensatory neuroprotective response cannot be resolved from cross‐sectional epigenomic data and will require longitudinal sampling and functional metabolic assays in patient‐derived or iPSC‐derived microglial systems.

In astrocytes, inflammatory chromatin remodeling was strikingly limited in the PWH versus PWoH comparison despite transcriptional upregulation of TLR4 and HMGB1 together with suppression of JAK1 and IL1RAP, distinguishing the astrocytic epigenetic response from the extensive inflammatory locus remodeling characteristic of microglial trained immunity. This chromatin‐level stability at inflammatory loci may reflect an already permissive astrocytic inflammatory epigenome, or alternatively that TLR4‐driven astrocytic inflammatory signaling does not require extensive chromatin remodeling at this disease stage. The dominant astrocytic ATAC signal was instead at cholesterol regulatory loci, with increased accessibility at RXRA and APOE concordant with transcriptional upregulation of APOE, ABCA1, and ABCG1, downregulation of LDLR, and pseudobulk downregulation of the cholesterol biosynthetic enzymes HMGCR and HMGCS1, consistent with epigenetic licensing of the LXR‐RXRα cholesterol regulatory axis and lipoprotein handling capacity as a bystander response to the microglial trained immunity‐driven lipid dysregulation program. Astrocytes also showed partially increased accessibility at HK1, the only glycolytic gene showing accessibility gains in this cell type, indicative of gene‐specific epigenetic priming rather than a coordinated pathway‐level response.

Oligodendrocytes showed coordinated reprogramming across cholesterol metabolism, glycolytic energy production, and inflammatory gene priming. Transcriptional changes included upregulation of HMGB1 alongside suppression of JAK1 and IL1RAP, consistent with partial trained immunity reprogramming. Chromatin‐level changes were most evident in the glycolytic pathway, with increased accessibility at GAPDH, PFKFB3, PRKAA1, and SLC2A1 in PWH versus PWoH, paralleling the metabolic reprogramming that underlies trained immunity. Increased accessibility at RXRA in PWH versus PWoH, together with concordant glial‐wide cholesterol regulatory transcriptional changes (APOE, ABCA1, ABCG1 upregulated; LDLR downregulated), points to coordinated licensing of the cholesterol regulatory axis. However, compared with ANI, oligodendrocytes from HAD cases showed reduced accessibility at several key cholesterol regulatory genes, including RXRA, NPC1, CYP46A1, and ABCG1, together with decreased accessibility at glycolytic loci such as ALDOC, ENO1, GAPDH, and SLC2A1. In contrast, accessibility increased at scavenger receptors (MSR1, CD36) and inflammasome‐related genes (IL1R1, NLRP3). Collectively, these changes suggest impaired cholesterol trafficking and efflux, disruption of coordinated glycolytic metabolism, and a shift toward a more pro‐inflammatory state in HAD oligodendrocytes. The combination of partial inflammatory gene priming, coordinated glycolytic epigenetic remodeling, and progressive cholesterol regulatory disruption directly threatens the two functions most essential to oligodendrocyte viability, myelin synthesis and maintenance. The coexistence of these changes with advancing disease suggests that oligodendrocyte epigenetic reprogramming contributes to rather than merely reflects demyelination in HAND.

4.2. Neuronal Epigenetic Reprogramming and Functional Implications

In neurons, chromatin remodeling was more limited in magnitude than in glia and more consistent with bystander epigenetic reprogramming driven by the surrounding glial neuroinflammatory microenvironment than with cell‐autonomous trained immunity induction, though the two mechanisms are not mutually exclusive. The dominant PWH signature was epigenetic constraint at glycolytic and mitochondrial flux loci—reduced accessibility at GAPDH, LDHA, PDHB, PGK1, and SLC2A1—coexisting with increased accessibility at upstream glycolytic regulatory nodes including GPI and SRC and at the mitochondrial chaperone TRAP1, a dysregulated rather than uniformly suppressed bioenergetic configuration. Concurrently, multiple inflammatory regulatory genes including PARK7, PIN1, PRKCA, RIPK1, TBK1, and JAK1 showed reduced accessibility, while CAMK4 alone showed increased accessibility—a discordance consistent with epigenetic silencing of both pro‐ and anti‐inflammatory regulators rather than coordinated inflammatory engagement. The parallel reduction of JAK1 accessibility in neurons and JAK1 transcript downregulation in astrocytes and oligodendrocytes points to convergent cross‐cell‐type epigenetic impairment of cytokine regulatory signaling.

In excitatory glutamatergic neurons, the combined reduction in bioenergetic flexibility and transcriptional downregulation of activity‐dependent neuroprotective genes, including NPAS4, NR4A1, and BCL6, suggests impaired coupling between neuronal activity and homeostatic survival programs in PWH. In HAD, this pattern was accompanied by reduced accessibility at neuronal identity regulators such as NEUROD2, HTR1B, and NECAB2, indicating additional disruption of the transcriptional programs that maintain excitatory neuronal identity and circuit integrity. In inhibitory GABAergic neurons, the concordant epigenetic and transcriptional priming of stress kinase signaling at DUSP8, complement‐mediated synaptic pruning at SEZ6L, and pro‐apoptotic programs at BID in HAD, alongside transcriptional downregulation of ANOS1, CNTNAP3B, and VAMP2 and HAD‐stage reduced accessibility at EPHB3 and PCDHGC5/3, suggest that GABAergic neurons are particularly vulnerable to stress‐induced apoptotic priming and inhibitory circuit destabilization in advanced dementia. The resulting imbalance between impaired excitatory maintenance and active inhibitory circuit deterioration provides a potential epigenomic basis for the excitatory‐inhibitory dysregulation and network instability that characterizes cognitive decline in HAND.

Neurons, traditionally viewed as less transcriptionally active in inflammatory processes than glial cells (Muller et al. 2025), nonetheless exhibited pathway‐level alterations in MAPK, PI3K‐AKT, and JAK–STAT signaling in HAD, suggesting secondary neuronal responses to the glial neuroinflammatory microenvironment. The apparent discordance between pathway‐level enrichment and predominantly closed chromatin at individual inflammatory gene loci including PARK7, PIN1, PRKCA, RIPK1, TBK1, and JAK1 is consistent with dysregulated rather than coordinated inflammatory pathway engagement, in which epigenetic silencing of both pro‐ and anti‐inflammatory regulators impairs the neuron's capacity to mount or resolve a coordinated inflammatory response. Together, these findings indicate that neuronal epigenetic reprogramming in HAND reflects bystander conditioning by the surrounding glial maladaptive trained immunity program, with functional consequences for circuit maintenance.

4.3. Trained Immunity as a Unifying Framework for HAND Pathogenesis

The convergence of transcriptional and epigenomic evidence across glial cell types supports, though does not prove, a model in which maladaptive trained immunity is a central mechanism of HAND pathogenesis. The simultaneous transcriptional upregulation of innate immune sensors and signaling mediators (IL18, IRAK3, NLRP3, IL6ST, JAK1, TLR2, and TLR4 in microglia; HMGB1 in astrocytes and oligodendrocytes), coordinated cholesterol regulatory and glycolytic transcriptional remodeling (APOE, ABCA1, ABCG1 upregulated; LDLR downregulated; GAPDH, PFKFB3, HK2, and PGK1 differentially expressed across glia), and chromatin accessibility changes at inflammatory, cholesterol, and glycolytic loci (AIF1, IL10, NLRP3, TLR7, NCF1, SRGAP2B, FCGR1A inflammatory; RXRA, APOE cholesterol; HK2, GAPDH, PFKFB3, PRKAA1, SLC2A1 glycolytic) in microglia and oligodendrocytes correspond to the three defining hallmarks of trained immunity described in peripheral macrophages (Hajishengallis et al. 2025). The persistence of microglial epigenetic priming uniformly across the HAND severity spectrum is consistent with the durable epigenetic memory that characterizes trained immunity and distinguishes it from classical transcriptional activation (Fanucchi et al. 2021). The partial and cell type–specific nature of non‐microglial reprogramming—astrocytes showing primarily cholesterol regulatory remodeling at RXRA and APOE with limited inflammatory chromatin opening, oligodendrocytes showing coordinated glycolytic accessibility gains (GAPDH, PFKFB3, PRKAA1, SLC2A1) alongside selective inflammatory transcript changes—is consistent with paracrine propagation of partial trained states from activated microglia, mediated by DAMPs, proinflammatory cytokines, and metabolic signals released by trained microglia into the CNS microenvironment.

Several findings from our postmortem dataset are supported by complementary iPSC‐based models of HIV neuropathogenesis. Sustained type I interferon signaling reported in iPSC‐derived microglia following HIV infection (Boreland et al. 2024) and in plasma of PWH (Mackelprang et al. 2023) is consistent with our observation of innate immune activation in PWH microglia, including upregulation of TLR2 and TLR4 and chromatin opening at multiple inflammatory effector loci. Persistent neuroinflammation and EIF2 pathway activation in hiPSC tri‐culture HIV models (Ryan et al. 2020) align with our transcriptional findings in PWH glia. Microglial activation observed in HIV‐infected cerebral organoids (Dos Reis et al. 2020; Gumbs et al. 2022; Min et al. 2023) provides mechanistic support for the inference that microglial epigenetic priming may be initiated by direct HIV exposure and persist in the ART era. Together, these convergent lines of evidence strengthen the interpretation that microglial epigenetic activation is a core feature of HIV neuropathogenesis, though the trained immunity framework requires functional validation as noted in the Limitations.

4.4. Stage‐Associated Chromatin and Transcriptional Differences Between ANI and HAD

A key observation of this study is that the relatively modest differences between HAD vs ANI and PWH vs PWoH across cell types suggest that many epigenetic and transcriptional alterations characteristic of advanced HAND may already be established early during HIV infection. The strongest HIV‐associated effects were observed in microglia, astrocytes, and oligodendrocytes, with additional transcriptional and epigenomic shifts in glutamatergic and GABAergic neurons. These findings are consistent with neuropathological evidence highlighting glial dysfunction and synaptic damage as central features of HAND (Irollo et al. 2021; Watson and Tang 2022).

The divergence between PWH‐versus‐PWoH and HAD‐versus‐ANI signatures takes distinct forms across cell types. In microglia, the epigenetic landscape is largely established by HIV infection and does not accumulate further with cognitive decline at the loci examined, with the exception of HAD‐specific epigenetic contraction at TOMM40 and MZF1 alongside continued accessibility gain at ASB17—a configuration that together suggests HAD‐associated suppression of mitochondrial epigenetic metabolite supply alongside continued NF‐κB pathway activation, consistent with intrinsic constraints on further microglial inflammatory reprogramming in end‐stage HAND rather than disease deepening per se. In neurons, HAD introduces subtype‐specific chromatin and transcriptional remodeling that is largely absent in ANI: glutamatergic neurons show reduced accessibility at HTR1B, NEUROD2, and NECAB2 alongside upregulation of PRLR and TUBA1C and downregulation of TRIM36, while GABAergic neurons show upregulation of stress kinase, complement, and pro‐apoptotic genes (DUSP8, SEZ6L, BID), downregulation of inhibitory circuit genes (ANOS1, CNTNAP3B, VAMP2), and reduced accessibility at EPHB3 and PCDHGC5/3, with GABAergic neurons showing the most pronounced vulnerability.

Regarding the mechanistic interpretation of trained immunity in this context, classical trained immunity requires that cells return toward a pre‐activation baseline after the initial stimulus is removed, such that secondary stimulation elicits a heightened response relative to a naïve cell. In the context of ART‐treated PWH, whether microglia fully return to a naïve baseline is uncertain, as residual antigen exposure from Nef‐containing extracellular vesicles or low‐level microbial translocation may persist. Three non‐mutually exclusive models are therefore proposed: a classical trained immunity model, in which initial HIV‐associated microglial activation induces durable epigenetic changes that sensitize microglia to exaggerated responses upon subsequent low‐grade antigen exposure; a persistent priming model, in which microglia never fully return to a naïve state but maintain an epigenetically sensitized profile that constitutively amplifies inflammatory responses; and a trained immunity contraction model, in which prolonged maladaptive trained immunity in end‐stage HAND imposes intrinsic metabolic and transcriptional constraints, reflected in TOMM40 and MZF1 epigenetic closure, on further inflammatory epigenetic propagation, representing a potential self‐limiting feature of the microglial trained state at the most advanced disease stage. Our cross‐sectional data cannot distinguish between these models; longitudinal studies and secondary challenge experiments in iPSC‐derived microglial systems will be required to resolve this question.

5. Limitations

This study has several limitations that should be considered when interpreting its findings. First, the overall sample size was modest, reflecting the high cost of single‐nucleus multiomic profiling. The PWH group comprised four individuals, of whom two had HAD and two had ANI, meaning that all cell type–specific comparisons are based on very small group sizes. This severely limits statistical power, increases the risk that individual sample‐level biological variation drives apparent group differences, and constrains the generalizability of findings, particularly for cell type–specific and gene‐level results. All findings should therefore be interpreted as hypothesis‐generating rather than definitive, and validation in larger, independently collected cohorts will be essential. A side‐by‐side comparison of single‐cell and pseudobulk DESeq2 log2FC values for all genes analyzed across both RNA and ATAC modalities is provided in Table S1. Single‐cell and pseudobulk estimates showed consistent direction of effect across the dataset, with pseudobulk values generally attenuated in magnitude relative to single‐cell values, as expected from donor‐level aggregation. This concordance, spanning glial and neuronal cell types, provides an additional layer of statistical confidence in the gene‐level findings and partially mitigates concerns about pseudoreplication (i.e., the inflation of statistical significance that arises when individual cells from the same donor are treated as independent observations), though it does not substitute for validation in larger cohorts. Additionally, despite standard ambient‐RNA‐aware filtering, residual ambient signal cannot be fully excluded for transcripts with strongly cell‐type‐restricted expression. Apparent differential expression of such genes in non‐source cell types—for example, microglia‐enriched transcripts such as PLCG2 appearing among neuronal differentially expressed genes in Table S1—should be interpreted cautiously, as the signal may reflect ambient RNA contamination during nuclei isolation rather than bona fide expression in the annotated cell type. Relatedly, formal sub‐clustering of microglia into canonical activation substates (e.g., homeostatic, disease‐associated, and interferon‐responsive) and statistical testing of per‐substate proportions across HAND severity groups were not performed, as the per‐group sample size (n = 2 HAD vs. n = 2 ANI) provides insufficient donor‐level power to support reliable inference on substate frequencies. Larger, cognitively characterized cohorts will be required to formally quantify microglial substate composition and its relationship to HAND severity.

Second, the PWH versus PWoH comparison is subject to potential confounds arising from differences in brain region, tissue collection conditions, and processing batch between the in‐house PWH specimens and the external PWoH control dataset (Ma et al. 2022). PWH samples were derived from the middle temporal gyrus while PWoH samples consisted of dorsolateral prefrontal cortex (DLPFC), introducing potential anatomical variability. Although prior transcriptomic studies have demonstrated substantial overlap in neuroimmune signatures and glial activation profiles between these cortical regions (Gelman et al. 2012; Jorstad et al. 2023), and although stringent batch correction and integration methods were applied, residual technical confounding cannot be fully excluded. The in‐house HAD versus ANI comparison avoids these batch and regional confounds but, as noted above, comprises only two samples per group, meaning that both comparisons are subject to important but distinct limitations that preclude designating either as fully reliable. Additionally, the PWoH donors from the Ma et al. (2022) reference dataset were not subjected to formal neuropsychometric testing, so subclinical cognitive impairment in this group cannot be excluded. Future studies should include cognitively characterized HIV‐negative controls processed in parallel with HIV‐positive specimens under identical tissue collection, nuclei isolation, and library preparation protocols to enable rigorous evaluation of cohort effects and of the relationship between molecular signatures and cognitive status.

Third, the cross‐sectional design of this study means that the temporal ordering of the transcriptional and epigenomic changes reported here cannot be established from these data alone. The two‐stage model of HIV‐associated baseline reprogramming followed by HAD‐specific reorganization is inferred from cross‐sectional group differences rather than from longitudinal observation of individual subjects, and the direction of causality between epigenomic and transcriptional changes cannot be determined. Longitudinal studies with repeated sampling across the HAND severity spectrum will be required to establish the temporal sequence of molecular events during HAND progression.

Fourth, the trained immunity framework proposed here rests entirely on correlative epigenomic and transcriptomic evidence. The defining functional features of trained immunity, including enhanced cytokine production upon secondary stimulation, increased glycolytic flux, and deposition of permissive histone marks including H3K4me3 and H3K27ac at inflammatory gene promoters, have not been directly measured in this dataset. In the absence of functional assays such as metabolic flux measurements, ChIP‐seq for trained immunity‐associated histone modifications, or ex vivo secondary challenge experiments in patient‐derived or iPSC‐derived cells, the trained immunity interpretation remains a mechanistically consistent but unproven framework. Future studies employing such approaches will be required to confirm or refute maladaptive trained immunity as a functional mechanism in HAND pathogenesis.

6. Conclusions

Our multiomic analysis reveals cell type–specific transcriptional and epigenomic reprogramming in the HIV‐infected brain that is organized along two disease stages and is consistent with, though not proof of, a model of maladaptive trained immunity. Microglia display the most complete reprogramming signature spanning inflammatory sensing, cytokine production, glycolytic rewiring, and cholesterol dysregulation, established uniformly by HIV infection and persistent across the HAND severity spectrum, with nascent signs of self‐limiting epigenetic contraction in end‐stage dementia. Astrocytes and oligodendrocytes show partial bystander epigenetic reprogramming consistent with paracrine propagation of the trained state from activated microglia, with dementia‐stage HAND marked by deepening scavenger receptor accessibility, paradoxical loss of cholesterol efflux and metabolic flexibility in oligodendrocytes, and further epigenetic silencing of neuroprotective and homeostatic functions in astrocytes in HAD. In neurons, bystander epigenetic reprogramming produces a two‐stage pattern of mitochondrial and glycolytic dysregulation followed by subtype‐specific deterioration of inhibitory and excitatory circuit maintenance programs concordant with the transcriptional findings, with neuronal epigenetic constraint at multiple bioenergetic and inflammatory regulatory loci potentially compounding the surrounding glial neuroinflammatory program. Together, these findings support a model in which HIV infection establishes a durable maladaptive microglial trained immunity program that drives epigenetic reprogramming of surrounding glia and neurons that is more pronounced at the HAD stage, with dementia‐stage HAND reflecting entrenchment of this configuration alongside further loss of accessibility at neuroprotective regulatory loci in HAD, a legacy that may persist despite viral suppression and represents a potential target for therapeutic intervention. These conclusions are based on correlative multiomic evidence in a small cross‐sectional cohort and require validation in larger longitudinal studies employing functional trained immunity assays.

Author Contributions

Conceptualization: Michael Bukrinsky and Dmitri Sviridov. Methodology, Michael Bukrinsky, Siera Martinez, and Anelia Horvath; Software, Siera Martinez and Anelia Horvath; Formal analysis, Michael Bukrinsky. Data curation, Siera Martinez; Resources, Tatiana Pushkarsky and Larisa Dubrovsky; Visualization, Anelia Horvath and Luke Johnson. Writing – original draft, Michael Bukrinsky. Writing – review and editing, Daria Starosyla and Michael Bukrinsky. Funding acquisition, Michael Bukrinsky. All authors have read and agreed to the published version of the manuscript.

Funding

This work was supported by the National Institutes of Health (R01 NS124477, R01 MH134776, R21 NS137986, P30 AI117970).

Ethics Statement

The authors have nothing to report.

Conflicts of Interest

The authors declare no conflicts of interest.

Supporting information

Figure S1: Quality control metrics before and after filtering of single‐nuclei datasets. Quality control plots are shown for each sample from individuals with asymptomatic neurocognitive impairment (ANI 1–2) (n = 2), HIV‐associated dementia (HAD 1–2) (n = 2), and PWoH controls (n = 4). For each sample, unfiltered data (top row) represent all captured nuclei prior to QC, whereas filtered data (bottom row) show the remaining nuclei after exclusion of low‐quality cells based on standardized thresholds for total UMI counts, number of detected genes, and percentage of mitochondrial reads. Doublets were also computationally removed. These curated datasets were used for all subsequent single‐nucleus RNA‐seq and ATAC‐seq analyzes.

Figure S2: Correlation plots of single cell Wilcoxon and pseudobulk differential gene expression testing.The correlation of pseudobulk and single cell differential gene expression analysis is displayed for snRNA‐seq (A) and snATAC‐seq (B) data separately. For single nucleus differential expression analysis, Wilcoxon Rank Sum Test was performed using Seurat, FDR Benjamini‐Hochberg, and an adjusted p value filter of ≤ 0.05. Pseudobulk differential expression analysis was performed using Seurat, AggregateExpression, Deseq2, FDR Benjamini‐Hochberg, an adjusted p value filter of ≤ 0.05 and two replicates. Only genes with an adjusted p value ≤ 0.05 are displayed as points and genes of interest pertaining to cholesterol metabolism, glycolysis and inflammation are highlighted in orange. The Pearson correlation coefficient is calculated and displayed as a line.

Figure S3: Evaluation of dataset integration following CCA correction. Integration quality was assessed using the Local Inverse Simpson's Index (LISI) and the k‐nearest neighbor batch effect test (kBET) following canonical correlation analysis (CCA)‐based integration. LISI scores exceeded 2.0 for all major cell types except unclassified neurons when batches were defined by diagnostic group (ANI, HAD, or PWoH; theoretical maximum = 3.0), indicating effective mixing across disease categories after integration. When batches were defined by individual sample identity, average LISI scores were approximately 5.0 across most cell types (theoretical maximum = 8.0 for eight samples), consistent with substantial reduction of sample‐specific batch effects while preserving biologically meaningful cellular structure.

Figure S4: Cell type annotation and sample composition across experimental conditions. Bar plots show the total number and relative proportion of cells across filtered cell types and experimental conditions following quality control and integration. Cell composition is displayed for each diagnostic group and individual sample, allowing assessment of cell type representation and sample balance across the dataset.

Figure S5: Comparison of chromatin accessibility and transcriptional changes. Post quality control filtered cells and cell types are compared by their scaled and normalized gene expression and their chromatin accessibility (in promoters only). These comparisons are made for astrocytes, microglia and oligodendrocytes, and selected cholesterol metabolism genes (A), glycolysis genes (B) and inflammation genes (C). Scaled and normalized gene expression is calculated using SCTransform (Seurat) and the mean of this value per gene and per cell type is calculated. Chromatin accessibility is the percentage of cells with greater than zero open areas of chromatin within the given cell type.

Figure S6: Analysis of differential chromatin accessibility in astrocytes. Volcano plots depict differentially accessible regions (DARs) for comparisons of PWH vs. PWoH (A) and HAD versus ANI (B) brains.

Figure S7: Analysis of differential chromatin accessibility in microglia. Volcano plots depict DARs for comparisons of PWH vs. PWoH (A) and HAD vs. ANI (B) brains.

Figure S8: Differential chromatin accessibility analysis in oligodendrocytes. Volcano plots depict DARs identified in oligodendrocytes for comparisons of PWH versus PWoH (A) and HAD versus ANI (B).

Figure S9: Differential gene expression analysis in neuronal populations from PWH versus PWoH samples. Volcano plots show differential gene expression (DGE) identified by snRNA‐seq analysis in glutamatergic neurons (A) and GABAergic neurons (B) comparing PWH and PWoH.

Figure S10: Unbiased analysis of DGE (HAD vs ANI) in neurons. Volcano plots show DGE obtained from snRNA‐seq analysis for glutamatergic (A) and GABAergic (B) neurons.

Figure S11: Unbiased analysis of differential chromatin accessibility (comparing HAD to ANI) in neurons. Volcano plots depict DARs obtained from snATAC‐seq analysis for glutamatergic (A) and GABAergic (B) neurons.

Table S1: Notes on interpretation.

Table S2: Cell‐type proportion analysis. Proportions of each cell type across donor groups, determined from single‐nucleus RNA‐seq cell type annotations. Values represent the percentage contribution of each donor group (HAD, ANI, PWoH) to the total nuclei assigned to each cell type. Abbreviations: ANI, asymptomatic neurocognitive impairment; HAD, HIV‐associated dementia; OPCs, oligodendrocyte precursor cells; PWoH, people without HIV; Unclassified Neu., unclassified neurons.

Table S3: Gene function descriptions and trained immunity connections. Functional descriptions, cell type specificity, direction of change, trained immunity connections, and manuscript figure locations for all genes shown in Results figures. Genes are listed alphabetically. Analysis: ATAC‐seq, chromatin accessibility; RNA‐seq, transcriptomic. Comparison: HAD versus ANI, HIV‐associated dementia versus asymptomatic neurocognitive impairment. Direction: Closed, decreased chromatin accessibility; Down, downregulated; Open, increased chromatin accessibility; PWH vs PWoH, people with HIV versus people without HIV; Up, upregulated.

Data Availability Statement

The original data presented in the study are openly available in GEO, accession number GSE296943.