Mapping genetic regulation of gene expression to cellular contexts identifies long non-coding RNAs associated with brain disorders

Abstract

How genetic variation regulates long non-coding RNA (lncRNA) expression across brain cell types remains poorly understood. A major barrier is the resolution–power trade-off between bulk and single-nucleus expression quantitative trait locus (eQTL) studies. Here, we quantify gene expression from cortical RNA-seq of 2443 individuals using an expanded transcriptome annotation and perform transcriptome-wide interaction eQTL mapping with deconvolution-derived cell-type proportions. Among 17,541 analyzed lncRNAs, we identify 3763 lncRNAs with cellular-context-dependent effects, of which 2,783 (74%) are not annotated by GENCODE. Colocalization with genome-wide association studies (GWAS) of brain-related traits identifies 118 lncRNA–trait colocalization events, approximately two-thirds of which are detectable only after modeling cellular context. Our study establishes a map of cellular-context-dependent genetic regulation in the human brain, providing a basis for nominating lncRNAs with potential roles in mediating genetic risk for brain disorders.

Genetic regulation of long non-coding RNAs (lncRNAs) in the human brain remains poorly understood. Here the authors map cell-type-dependent genetic regulation of lncRNAs, highlighting their potential roles in brain disorders.

Introduction

Expression quantitative trait locus (eQTL) mapping offers an effective route for functionally interpreting noncoding genetic variants identified by genome-wide association studies (GWAS)^1–5. Bulk transcriptomic analyses of the brain have cataloged maps of genetic variants regulating gene expression at the tissue level^6–11. The advance of single-cell transcriptomics has revealed pervasive cellular heterogeneity in brain gene expression^12,13, and single-cell eQTL analyses have demonstrated that genetic regulatory effects can vary across cell types and states, thereby extending the functional interpretation of variants beyond tissue-averaged signals to cellular contexts^14–16.

Despite these advances, long non-coding RNAs (lncRNAs) remain a particularly important yet understudied class. These transcripts are increasingly recognized as components of regulatory programs operating within defined neuronal and glial contexts, suggesting that genetic effects on their expression may be strongly cell-type-dependent^17,18. Most efforts to date to characterize this regulatory layer have relied on bulk brain eQTL studies, where tissue-averaged effects can obscure cell-type-specific regulatory signals^7,19. However, current single-cell eQTL study designs face limited sample sizes and reduced statistical power for detecting effects on lowly expressed transcripts, including lncRNAs^20,21. As a result, the genetic regulatory landscape of lncRNAs in brain cell types remains insufficiently characterized, limiting the interpretation of noncoding GWAS signals for complex brain disorders.

In this study, to identify genetic variants regulating lncRNA expression in brain cell types, we used interaction eQTL (ieQTL) mapping, which models genotype-by-cell-type-proportion interactions in RNA sequencing (RNA-seq) data^22–25. We integrated and rigorously harmonized genotype and dorsolateral prefrontal cortex (DLPFC) RNA-seq data from 2,443 individuals, used a previously developed expanded transcriptome model that incorporates lncRNA transcripts beyond GENCODE annotations for expression quantification¹⁹, and applied BayesPrism²⁶ to infer cellular composition after systematic benchmarking of deconvolution methods. We identified numerous lncRNAs whose genetic effects showed cellular-context dependence, including many not annotated by GENCODE. We subsequently assessed the reproducibility of ieQTLs and their functional enrichment in cell-type-specific regulatory annotations. Integration with brain-related GWAS additionally identified a substantial number of cellular-context-dependent colocalization signals that were not detectable in bulk eQTL analyses, highlighting the additional resolution afforded by modeling cellular context. Collectively, this work expands the cell-resolved genetic regulatory landscape of the human brain by incorporating lncRNAs, filling an important gap in brain eQTL studies. This resource (https://brain.hitxqtl.org.cn/Brain-ieQTLAtlas/#/) will help elucidate the molecular mechanisms underlying neurological diseases.

Results

Study design and data overview

The overall design of this study is shown in Fig. 1, and it comprises three major components: benchmarking of cellular deconvolution, ieQTL mapping, and GWAS integration. Investigating genotype–cellular context interactions requires accurate quantification of cellular composition across individuals²⁷. To this end, we evaluated the robustness and reproducibility of computational deconvolution in complex brain tissues based on both simulated and real data (Fig. 1a). We collected genotype and RNA-seq data from seven cohorts within the PsychENCODE²⁸ and AMP-AD²⁹ consortia (Fig. 1b and Supplementary Data 1). After cohort harmonization and quality control, we retained RNA-seq data from 2443 unrelated individuals and $~$12 million common genetic variants with minor allele frequency (MAF) > 0.01 (Methods). We then used the expanded transcriptome model to quantify 17,541 lncRNAs, along with 14,850 protein-coding genes (PCGs) used to support cell-type deconvolution and comparative analyses. Using a single-nucleus RNA-seq (snRNA-seq) reference profile derived from representative cortical samples (referred to as PFC48)¹³, we estimated cellular composition in the RNA-seq data and performed ieQTL mapping. Finally, we integrated GWAS of brain-related traits to evaluate whether cellular-context-dependent lncRNA-ieQTLs contribute to the interpretation of disease-associated loci (Fig. 1c).

Fig. 1. Overview of the study design.

a We established a benchmarking pipeline to evaluate the robustness of four transcriptome-based deconvolution methods (CIBERSORTx, MuSiC, BayesPrism, and Bisque) under diverse conditions. Evaluation included (1) artificial mixtures generated by recombining single-cell profiles within PFC48¹³ for within-dataset benchmarking; (2) pseudo-bulk mixtures constructed by aggregating cells from different donors in PFC427³³ for cross-dataset evaluation; (3) FACS-purified cortical populations with defined cell identities; and (4) matched bulk RNA-seq, immunohistochemistry (IHC)-based cell quantification, and clinical phenotype–linked data from the ROSMAP cohort. b We applied BayesPrism to infer cell-type composition from RNA-seq profiles of 2443 individuals. Integrating cell abundance estimates with matched genotypes and transcript quantification of 14,850 PCGs and 17,541 lncRNAs, including both GENCODE-annotated and unannotated transcripts, we performed transcriptome-wide ieQTL mapping. c Finally, we integrated lncRNA- and PCG-ieQTLs with GWAS summary statistics to compare their associations with complex brain traits and disorders and to nominate lncRNAs with putative roles in mediating genetic risk. Created in BioRender. Jia, Y. (2026) https://BioRender.com/ccn6jan.

Benchmarking cellular deconvolution in brain transcriptomes

To test the accuracy and stability of cell-type abundance estimation in brain transcriptomes, we conducted a benchmarking analysis of four widely used deconvolution methods, including MuSiC³⁰, CIBERSORTx³¹, BayesPrism²⁶, and Bisque³². The deconvolution accuracy was measured by Spearman’s rho and mean absolute error (MAE) between the inferred abundance and the ground truth. Our simulation was based on the previously described PFC48 dataset¹³, with cell-type annotations obtained from the original publication. We split PFC48 into two portions, using one portion to generate synthetic bulk mixtures based on a predefined method and the other as the reference for deconvolution (Methods). Performance was evaluated against the known cellular composition. Across simulated mixtures, BayesPrism showed the highest accuracy (median correlation = 0.98; median MAE = 0.01), compared with CIBERSORTx (0.96; 0.03), MuSiC (0.95; 0.04), and Bisque (0.94; 0.04) (Fig. 2a, b), with cross-cell-type benchmarking showing the strongest agreement between matched estimated and reference cell types (Supplementary Fig. 1). To better approximate realistic transcriptomic heterogeneity, we further assessed cross-dataset performance using a large-scale snRNA-seq atlas comprising 427 cortical donors³³ (referred to as PFC427). Pseudo-bulk profiles were constructed by aggregating cells at the individual level, thereby explicitly incorporating inter-individual variation and dataset-specific heterogeneity. Under this setting, deconvolution performance was more stable for higher-abundance cell populations, whereas performance was reduced for lower-abundance populations (Supplementary Fig. 2a). These observations underscore inherent challenges in accurately resolving rare cell types in realistic biological settings, where reference and target profiles may differ in cellular composition, technical origin, and biological context.

Fig. 2. Evaluation of deconvolution performance under diverse benchmarking conditions.

a, b Boxplots showing the Spearman’s rho and mean absolute error (MAE) between estimated and true cell-type proportions across 5000 artificial mixtures, comparing four deconvolution methods: BayesPrism, MuSiC, CIBERSORTx, and Bisque. Each data point represents Spearman’s rho or MAE for one cell type, calculated across 5000 artificial mixtures. c Deconvolution accuracy in FACS-purified samples with known cell identities (Ex, In, Oli; n = 4), using PFC48 as the reference. Each data point represents one independent FACS-purified sample. Bars indicate mean values, and error bars indicate the standard error of the mean (SEM). Bisque and MuSiC exhibited greater sensitivity to reference noise than CIBERSORTx and BayesPrism. d Estimated cell-type proportions across bulk RNA-seq samples (n = 2443) using BayesPrism with the PFC48 reference. Each data point represents one cortical sample. For all boxplots (a, b and d), the box indicates the interquartile range (IQR), the line within the box represents the median, and the whiskers extend to the most extreme data points within 1.5 times the IQR. Points beyond the whiskers indicate outliers. A threshold of 0.05 (dashed line) was applied to define cell types for downstream ieQTL mapping. e Pairwise correlation between estimated cell-type proportions. Cell type abbreviations: Ast astrocytes, End endothelial cells, Ex excitatory neurons, In inhibitory neurons, Mic microglia, Oli oligodendrocytes, OPC oligodendrocyte precursor cells, Per pericytes.

We further evaluated deconvolution performance using fluorescence-activated cell sorting (FACS)-purified, lineage-specific expression profiles from the SuperAgerEpiMap³⁴ and EpiGABA³⁵ projects (Methods). This validation setting leveraged independent brain cell-type-enriched profiles with known target lineages to evaluate whether deconvolution methods recovered the expected dominant cell-type compartments. In the $Neu{N}^{+}$ and $Neu{N}^{-}$ samples (Supplementary Fig. 2b), BayesPrism, MuSiC, and CIBERSORTx recovered the expected dominant neuronal and non-neuronal compartments, whereas Bisque showed weaker concordance with the expected compartment structure. Similar patterns were observed in additional lineage-specific samples (Fig. 2c), where BayesPrism and CIBERSORTx generally aligned with the known target lineages, whereas MuSiC and Bisque showed less consistent recovery of the expected dominant cell type. In addition, we compared deconvolution estimates with matched immunohistochemistry (IHC)-measured cell-type proportions in ROSMAP samples³⁶. BayesPrism showed the highest overall concordance with IHC measurements among the evaluated methods (Spearman’s rho = 0.92, compared with 0.85–0.90 for the other methods), although exact cell-type proportions remain subject to uncertainty owing to the limited scale of the IHC experiment and the distinct molecular readouts captured by IHC and RNA-seq (Supplementary Fig. 3)^6,37,38. Based on the complementary evaluations described above, we selected BayesPrism for cell-type abundance estimation in ieQTL analyses.

Robust estimation of brain cell-type abundance

We applied BayesPrism to infer cell-type composition from RNA-seq data, resolving eight brain cell types, including excitatory neurons, inhibitory neurons, astrocytes, oligodendrocytes, oligodendrocyte precursor cells, microglia, endothelial cells and pericytes. Sensitivity analysis showed that BayesPrism-based deconvolution was robust to the gene set used for abundance estimation (Supplementary Fig. 4). As expected, excitatory neurons represented the most abundant population (Fig. 2d), while estimated cell-type proportions varied markedly among cohorts (Supplementary Fig. 5). Pairwise correlations between inferred cell types were generally low to moderate (median Spearman’s rho = 0.26; Fig. 2e), with excitatory neurons showing negative correlations with all other cell types, consistent with their dominant representation in the cortical transcriptome. In samples annotated with Braak stage, we additionally observed a progressive decline in neuronal proportions with increasing pathology, consistent with expected neurodegenerative trajectories³⁹ (Supplementary Fig. 3d). Given the limited reliability of deconvolution estimates for rare cell populations, downstream analyses were restricted to cell types with a median estimated proportion above 0.05, namely excitatory neurons, inhibitory neurons, astrocytes, and oligodendrocytes²². Although deconvolution provides relative rather than absolute estimates, these high-abundance cell types exhibited strong concordance when the reference profile was replaced with PFC427 (Supplementary Fig. 6; average Spearman’s rho = 0.94), confirming the robustness of proportion inference for these major cell types.

Mapping and validation of ieQTLs

To increase statistical power, we integrated RNA-seq and genotype data across cohorts (Supplementary Fig. 7) and performed ieQTL mapping within a ± Mb cis window (Methods), after which the interaction effects were classified according to how allelic effects varied with cellular context into positive, negative, and uncertain categories (Fig. 3a). At a 5% false discovery rate (FDR), we identified 9306 genes (ieGenes) with at least one significant ieQTL (Fig. 3b and Supplementary Data 2). We noticed that negative-direction ieQTLs in one cell type were often in linkage disequilibrium (LD) with positive-direction ieQTLs in other cell types (Supplementary Fig. 8a–c). Together with the correlation structure among estimated cell-type abundances, this observation prompted us to further assess whether the originally assigned cell-type interaction effects remained stable in a joint model including interaction terms for all four cell types simultaneously (Supplementary Fig. 8d). Under this more stringent model, positive-direction ieQTLs showed higher rates of significance retention for their originally assigned interaction terms than negative-direction ieQTLs. This result suggests that negative-direction ieQTLs may have less stable cell-type attribution than positive-direction ieQTLs. Approximately 15–21% of ieQTLs showed uncertain interaction effect directions (Fig. 3b), characterized by lower minor allele frequencies (Supplementary Fig. 9a) and greater distances from transcription start sites (TSSs) (Supplementary Fig. 9b). The proportion of uncertain-direction ieQTLs decreased progressively as FDR thresholds were increasingly stringent (Supplementary Fig. 9c, d), suggesting that many uncertain-direction ieQTLs likely represent less robust signals²². We therefore excluded uncertain-direction ieQTLs from subsequent analyses.

Fig. 3. Discovery of lncRNA-ieQTLs.

a Conceptual illustration of ieQTL, including no interaction, positive direction, negative direction, and uncertain direction. b Number of significant ieQTLs (FDR < 0.05) identified from an interaction linear model fitted with tensorQTL⁶⁴, with tests performed separately for each of the four cell-type proportions. Results are stratified by gene class and ieQTL directionality. Bar labels indicate the total number of ieQTLs in each category. c ${\pi }_1$ statistics estimating replication of ieQTLs using single-nucleus eQTL datasets from Fujita et al.¹⁵ (left) and Haglund et al.¹⁶ (right). Rows and columns are clustered using Euclidean distance and complete linkage. d Enrichment analyses comparing the genomic locations of lncRNA- and PCG-ieQTLs based on ANNOVAR-defined categories⁶⁷. Circle color represents the ${\log }_2$ odds ratio, and circle size indicates the $-{\log }_{10}$ P value from two-sided Fisher’s exact tests. e (Top) Number of ieLncRNAs with significant positive-direction ieQTLs across cell types, stratified by lncRNA biotype, including intronic, antisense, and intergenic lncRNAs. (Bottom) Among intronic and antisense lncRNAs with significant positive-direction ieQTLs, the fraction whose host PCGs are also identified as ieGenes in the same cell type. f Pairwise sharing of ieQTL effects between cell types, estimated by mashr⁴¹. Pairwise sharing is defined as the proportion of significant ieQTLs (LFSR < 0.05) with concordant effect direction and similar magnitude (within a twofold range) between two cell types.

We identified 3,763 interaction-effect lncRNAs (ieLncRNAs) and 3,445 PCGs (iePCGs) with at least one positive- or negative-direction ieQTL (Supplementary Fig. 10). Notably, 74% (2,783 out of 3,763) of the ieLncRNAs were previously unannotated in GENCODE (Supplementary Fig. 11a). Comparison with cis-eQTLs mapped in bulk tissue showed that the vast majority of ieGenes (93.5%) corresponded to genes harboring at least one cis-eQTL (hereafter referred to as eGenes; Supplementary Fig. 11b). A fraction of ieQTLs (19.2%) exhibited weak LD (coefficient of determination $R^2<0.2$) with any of the corresponding eGene’s conditionally independent bulk eQTLs (Supplementary Fig. 11c), based on an empirical threshold commonly used to distinguish strongly from weakly correlated regulatory variants⁴⁰. This observation is consistent with previous studies showing that ieQTL mapping can uncover cell-type-dependent regulatory effects not fully resolved by bulk eQTL analyses²². Among ieGenes exhibiting positive or negative ieQTL effects in at least one cell type, 6,480 (85.4%) were not previously reported in brain ieQTL studies (Supplementary Fig. 11d), likely reflecting advances in lncRNA annotation and genome-wide mapping⁶. We next assessed the concordance of ieQTLs using published single-nucleus eQTL datasets from human cortex, including the Fujita et al. study¹⁵ (n = 424, partially overlapping with the present cohort) and the Haglund et al. study¹⁶ (n = 391, independent). For positive-direction ieQTLs, consistent replication was observed in matched cell (sub)types across both datasets (Fig. 3c and Supplementary Fig. 11e), with median ${\pi }_1$ estimates of 0.82 (from 0.68 to 0.90) in the Fujita et al. dataset and 0.76 (from 0.54 to 0.84) in the Haglund et al. dataset. When evaluated separately by gene category, lncRNA-ieQTLs also replicated in matched cellular contexts, although their replication estimates were slightly lower than those of PCG-ieQTLs (Supplementary Fig. 11f). Cross-cell-type replication patterns reflected the expected mixture of shared and cell-type-enriched effects, as well as (anti)correlations with relative cell type abundances⁴⁰. In contrast, negative-direction ieQTLs showed a more dispersed replication pattern across cell types, with high replication observed in cell types different from the originally tested context (Supplementary Fig. 11g, h). This pattern is consistent with observations described earlier, suggesting that negative-direction ieQTLs may not primarily reflect regulatory effects in the tested cell type. Overall, the expanded lncRNA catalog and the increased statistical power from the large sample size allowed us to identify 3262 lncRNA–cell type pairs that had not been reported in previous studies (Supplementary Fig. 11i).

Characteristics of lncRNA-ieQTLs

We observed that lncRNA-ieQTLs were enriched near transcription start sites (TSSs) (Supplementary Fig. 12a). Compared with PCG-ieQTLs, lncRNA-ieQTLs were located closer to TSSs (median lncRNA-ieQTL distance = 30.1 kb; median PCG-ieQTL distance = 33.4 kb; $P=1.10{\times 10}^{-5}$, Wilcoxon test; Supplementary Fig. 12b) and exhibited a more compact cis-regulatory architecture (Fig. 3d), consistent with previous reports of more proximal regulatory effects on lncRNA expression⁹. Among intronic and antisense lncRNAs with positive-direction ieQTLs, only 6–48% had significant ieQTL signals for their host PCGs within the same cellular context (Fig. 3e), suggesting that many lncRNA-ieQTL effects are not simply explained by corresponding ieQTL effects on their host PCGs. In addition, the distribution of lncRNA-ieQTLs across cell types was highly uneven, with neuronal cell types harboring approximately twice as many lncRNAs with positive-direction ieQTLs as glial cell types (Fig. 3e). This pattern may partially reflect higher neuronal abundance and greater inter-individual variability, and may also be influenced by cell-type differences in deconvolution estimation accuracy and power for detecting genotype–cell-type interaction effects⁴⁰.

To characterize cell-type specificity and sharing patterns of ieQTLs, we used multivariate adaptive shrinkage (mashr), a statistical method for analyzing effect sizes across multiple conditions^23,41,42 (Methods). We identified 142 lncRNA-ieQTLs and 139 PCG-ieQTLs with cell-type-specific effects across the four major brain cell types analyzed, defined by a local false sign rate⁴³ (LFSR) < 0.05 in only one cell type. The cell-type-specific lncRNA-ieQTLs had effect sizes comparable to those of cell-type-specific PCG-ieQTLs (Supplementary Fig. 12c). LncRNA-ieQTL effect sizes decreased with increasing cross-cell-type sharing, consistent with the trend observed for PCG-ieQTLs (Kruskal–Wallis test; PCG-ieQTLs, $P=1.3\times {10}^{-10}$; lncRNA-ieQTLs, $P=5.6\times {10}^{-15}$; Supplementary Fig. 12d). The inverse relationship between effect size and cross-cell-type sharing remained consistent after restricting the analysis to ieQTLs passing a more stringent significance threshold ($P_{ieQTL}<5\times {10}^{-8}$; Supplementary Fig. 12e). Pairwise sharing was defined as the presence of effects that were significant in both cell types, had consistent effect directions, and had mashr-estimated effect sizes within a twofold range⁴¹. Both lncRNA-ieQTLs and PCG-ieQTLs showed high sharing between astrocytes and oligodendrocytes, whereas sharing between excitatory and inhibitory neurons was lower for lncRNA-ieQTLs than for PCG-ieQTLs (Fig. 3f).

Epigenomic annotation

To gain insight into the functional properties of ieQTLs, we annotated ieQTLs with $Neu{N}^{+}$ and $Neu{N}^{-}$ chromatin peaks identified by assay for transposase-accessible chromatin using sequencing (ATAC-seq) of FACS-sorted nuclei⁴⁴. ieQTLs were enriched in these regulatory annotations, with positive-direction ieQTLs showing preferential enrichment in lineage-matched peaks compared with negative-direction ieQTLs (Fig. 4a). This direction-dependent enrichment pattern is consistent with the differential interpretation of positive- and negative-direction interaction effects described above.

Fig. 4. Epigenomic annotation of lncRNA-ieQTLs.

a Enrichment of ieQTLs in non-neuron ($Neu{N}^{-}$) and neuron ($Neu{N}^{+}$) lineage ATAC-seq peaks, stratified by ieQTL direction. The x-axis indicates the fold enrichment, and the y-axis denotes the corresponding cell types. Each circle represents fold enrichment. Error bars represent the 95% CI computed using permutations (n = 1000). b, c Enrichment of positive-direction ieQTLs in cell-type–specific regulatory regions. Each square represents the fold enrichment of positive-direction ieQTLs, stratified by FDR thresholds (x-axis), within ATAC-seq peaks (b) and active regulatory regions marked by H3K4me3 and H3K27ac (ChIP-seq, c) in the corresponding brain cell type. Colors indicate gene categories; error bars denote 95% confidence intervals derived from permutation testing (n = 1000). d Chromatin landscape surrounding the excitatory neuron ieQTL rs3746746. The x-axis denotes the genomic coordinates on chromosome 20 encompassing HAR1A and nearby loci. From top to bottom: ChIP-seq tracks for H3K27ac (4 cell types), H3K4me3 (4 cell types), and ATAC-seq (4 cell types), followed by PLAC-seq tracks highlighting chromatin contacts across corresponding cell types.

Using cell-type-specific accessibility peaks derived from single-nucleus ATAC-seq (snATAC-seq)¹², we found that positive-direction ieQTLs for glial cell types, including astrocytes (PCG-ieQTLs fold enrichment = 2.20, 95% CI 1.82–2.57; lncRNA-ieQTLs fold enrichment = 1.80, 95% CI 1.40–2.20) and oligodendrocytes (PCG-ieQTLs fold enrichment = 3.25, 95% CI 2.92–3.58; lncRNA-ieQTLs fold enrichment = 2.33, 95% CI 1.99–2.68), were strongly enriched within their respective accessible chromatin regions, showing progressively stronger enrichment with increasing significance stringency. In contrast, positive-direction ieQTLs for excitatory neurons (PCG-ieQTLs fold enrichment = 1.78, 95% CI 1.60–1.96; lncRNA-ieQTLs fold enrichment = 1.51, 95% CI 1.33–1.68) and inhibitory neurons (PCG-ieQTLs fold enrichment = 1.36, 95% CI 1.18–1.54; lncRNA-ieQTLs fold enrichment = 1.22, 95% CI 1.03–1.41) showed comparatively modest enrichment (Fig. 4b). This overall pattern was similarly observed when using cell-type-specific chromatin immunoprecipitation followed by sequencing (ChIP-seq) marks⁴⁵ (H3K4me3 and H3K27ac; Fig. 4c), suggesting that neuronal ieQTLs may exhibit weaker cell-type specificity than glial ieQTLs, consistent with previous studies^14,46.

Figure 4d illustrates the chromatin landscape surrounding the lncRNA locus HAR1A, which contains the excitatory neuron-specific ieQTL rs3746746. This SNP resides in an exon of HAR1A and, despite lacking a strong ATAC-seq signal, shows extensive overlap with neuron-specific ChIP-seq marks, indicating that it is located within an active regulatory environment. The HAR1A locus is located within a human accelerated region (HAR), a class of noncoding regions that often function as enhancers in the nervous system⁴⁷, regulating neuronal development and maturation, and have been linked to central nervous system disorders such as Huntington’s disease^48–50. In addition, PLAC-seq data indicate a chromatin interaction between the HAR1A-associated regulatory region and the downstream MIR124-3 locus, which encodes a key neuronal regulator⁵¹. These features implicate rs3746746 as part of a neuron-specific regulatory architecture, potentially modulating transcriptional activity at the HAR1A locus.

Colocalization with disease susceptibility

To integrate lncRNA-ieQTLs with brain-related genetic risk, we compiled GWAS summary statistics for 36 phenotypes spanning four major categories: neuropsychiatric disorders, neurodegenerative diseases, cognitive traits, and brain morphological or developmental features (Supplementary Data 3 and Methods). We observed that lncRNA-ieQTLs were enriched for GWAS-associated variants (Supplementary Fig. 13). This observation supports the potential functional relevance of lncRNA-ieQTLs. We used coloc⁵² to investigate whether genetic associations for the brain-related traits described above may act through lncRNA-ieQTLs^22,40 (FDR < 0.05). For comparison, we also performed colocalization analyses using PCG-ieQTLs as well as matched lncRNA and PCG eQTLs. We identified 99 ieLncRNAs with evidence of colocalization (PP4 > 0.8), most frequently for traits such as schizophrenia, neuroticism, and educational attainment, whereas no colocalization signals were detected for multiple sclerosis or autism spectrum disorder (Supplementary Data 4). Notably, 85.8% (85 out of 99) of ieLncRNAs were absent from current GENCODE annotations, a pattern consistently observed under both more lenient and more stringent thresholds (Supplementary Fig. 14).

In total, 47.5% (77 out of 162) of lncRNA–trait colocalization events were uniquely identified through lncRNA-ieQTLs, whereas an additional 25.3% (41 out of 162) were supported by both lncRNA-ieQTLs and matched lncRNA eQTLs (Fig. 5a). Of the lncRNA-ieQTL-supported colocalization events, 64.4% (76 out of 118) were supported by positive-direction ieQTLs (Supplementary Data 4). These results underscore the added value of cellular-context-dependent regulatory mapping in detecting trait-relevant lncRNA loci and refining their likely cellular origins.

Fig. 5. Integrating lncRNA regulation in cellular context refines GWAS functional interpretation.

a Comparison of colocalization supported by ieQTLs versus eQTLs, showing the additional colocalization signals revealed when incorporating cellular context (PP4 > 0.8). b Expression levels of lncRNA AC096667.1 in excitatory neurons (Ex), oligodendrocytes (Oli) and inhibitory neurons (In) from EPIGABA samples (n = 4). The box indicates the interquartile range (IQR), the line within the box represents the median, and the whiskers extend to the most extreme data points within 1.5 times the IQR. Points beyond the whiskers indicate outliers. Statistical significance was assessed using two-sided Wilcoxon rank-sum tests: Ex versus Oli, P = 0.01; Ex versus In, P = 0.02. c Locus plot showing colocalization of lncRNA AC096667.1 in inhibitory neurons and at the bulk-tissue level with the neuroticism GWAS signal. Association P values ($-{\log }_{10}$ scale) are shown for inhibitory neuron ieQTLs (middle) and bulk eQTLs (bottom). d Across the 13 traits with significant SMR results, the number of loci prioritized by SMR and the subset with concordant evidence from COLOC in the same cellular context. Bars are stratified by evidence type: light color represents loci uniquely prioritized by SMR, while dark color indicates loci supported by both SMR and COLOC. e Locus plot showing colocalization between the SCZ GWAS locus rs12293670, NRGN, and the unannotated antisense lncRNA NRGN-AS1 in excitatory neuron ieQTLs. f ieQTLs of NRGN-AS1 in excitatory (Top) and inhibitory neurons (Bottom). Dots are colored on the basis of the genotype of rs12293670. For (c, e and f), ieQTL association P values were obtained from two-sided linear regression tests of the genotype-by-cell-type-proportion interaction term; eQTL P values shown in c were obtained from two-sided tests of the genotype term.

As shown in Fig. 5b, the GENCODE-annotated lncRNA AC096667.1 (chr2: 192,711,265–192,908,909; lead variant rs1019706) is specifically expressed in inhibitory neurons and exhibits strong colocalization with neuroticism risk loci in the inhibitory neuron ieQTLs (PP4 = 0.99; Fig. 5c), whereas no colocalization signal was detected using standard bulk eQTL analysis (PP4 = 0.00). Supplementary Figs. 15 and 16 provide additional examples of lncRNA-ieQTLs colocalizing with brain-related trait associations. These cases illustrate how lncRNA-ieQTLs can refine the cellular interpretation of GWAS associations and reveal regulatory signals that are masked in tissue-averaged analyses.

As colocalization analysis alone cannot distinguish true pleiotropy from confounding due to LD, we additionally applied summary-based Mendelian randomization⁵³ (SMR) with heterogeneity testing (HEIDI) to refine causal inference. We identified 134 loci (Supplementary Data 5), of which 52 were also supported by COLOC within the same cellular context (Fig. 5d). One illustrative locus involves the schizophrenia risk variant rs12293670^54,55, previously reported as a region-dependent eQTL in the prefrontal cortex affecting expression of Neurogranin⁵⁶ (NRGN), a neuron-specific gene and direct thyroid hormone target implicated in synaptic plasticity⁵⁷. lncRNA-ieQTL revealed that the SCZ GWAS signal colocalizes with the unannotated antisense lncRNA NRGN-AS1 (chr11:124,610,987–124,616,233; lead variant rs12293670; PP4 = 0.95; Fig. 5e), rather than with NRGN (PP4 = 0.00). The allelic effect of rs12293670 increased with excitatory neuron abundance and decreased with inhibitory neuron abundance. Epigenomic annotations further showed that rs12293670 resides within a neuron-specific enhancer marked by H3K27ac and H3K4me1 (Supplementary Fig. 17). Notably, this specificity was observed despite highly correlated expression of NRGN and NRGN-AS1 in excitatory neurons (Pearson’s r = 0.71; Supplementary Fig. 17). SMR evidence further supported an association between schizophrenia risk and NRGN-AS1 expression ($P_{SMR}=4.99\times {10}^{-9}$; $P_{HEIDI}=0.03$). SMR yielded a negative gene-trait effect estimate for NRGN-AS1 (${\beta }_{SMR}=-0.21$), indicating that higher expression levels of this lncRNA are associated with lower SCZ risk. Given the observed downregulation of NRGN in SCZ patients⁵⁸, we hypothesize that NRGN-AS1 may function to stabilize NRGN expression, mitigating synaptic vulnerability linked to disease risk. These analyses demonstrate how lncRNA-ieQTLs can nominate previously uncharacterized non-coding transcripts as candidate regulatory targets and inform hypotheses about the cell-type specificity of disease-associated genetic effects.

Discussion

Despite extensive efforts to characterize cellular-context-dependent genetic regulation in the human brain, lncRNAs remain a substantially understudied component. Here, leveraging a large-scale RNA-seq dataset (n = 2443) constructed through the integration of multiple cohorts and an expanded lncRNA transcript annotation, we established a cell-type-resolved genetic atlas of lncRNAs, identifying 3763 ieLncRNAs with cell-type-dependent genetic effects, including many absent from current GENCODE annotations. We also developed the Brain-ieQTLAtlas online resource (https://brain.hitxqtl.org.cn/Brain-ieQTLAtlas/#/) to facilitate exploration and reuse.

Current brain single-nucleus eQTL studies remain limited in their ability to characterize genetic regulation of lncRNA expression, owing to both limited sample sizes and technical constraints of prevailing protocols. In particular, the widespread reliance on oligo(dT)-primed poly(A) capture and intact 3′ transcript ends renders transcript detection highly sensitive to RNA integrity, posing additional challenges for frozen or partially degraded brain tissue and further reducing recovery of noncoding transcripts. In this context, ieQTL analysis offers a complementary approach by leveraging large bulk cohorts and inferred cellular composition to test whether genetic effects on lncRNA expression vary with cellular context. Our analysis detected many lncRNA–cell-type pairs that were not captured by single-nucleus eQTL studies, underscoring that current resources still miss a substantial fraction of cellular-context-dependent regulation of lncRNA expression. Nevertheless, deconvolution-based ieQTLs require cautious interpretation and should not be directly equated with cell-type-specific eQTLs. In our direction-stratified analyses, positive-direction ieQTLs showed stronger concordance with cell-type-matched single-cell eQTLs and preferential enrichment in regulatory annotations for the corresponding cell type, supporting more direct cellular-context attribution for this subset. By contrast, negative-direction ieQTLs showed weaker concordance with evidence supporting the inferred interacting cell type, indicating that their cellular attribution requires additional molecular evidence for interpretation.

We further integrated lncRNA-ieQTLs (FDR < 0.05) with brain-related GWAS datasets, identifying 99 ieLncRNAs with colocalization evidence at disease risk loci (PP4 > 0.8) across 12 phenotypes. A substantial proportion of lncRNA-ieQTL–supported colocalization signals were not detected using bulk eQTLs, indicating that, although ieQTLs do not necessarily pinpoint the precise cellular source of regulatory activity, some disease-relevant regulatory associations may be attenuated or obscured in tissue-level analyses. Furthermore, colocalization analyses prioritized lncRNAs over nearby PCGs, underscoring that previous efforts have yet to fully account for regulatory contributions from noncoding transcripts (Fig. 5e and Supplementary Fig. 15). These observations position lncRNA-ieQTLs as a critical complement to existing regulatory maps for interpreting disease-associated genetic variation in the human brain, while simultaneously expanding the catalog of functionally relevant lncRNAs with potential roles in disease biology.

Our findings should be interpreted within the limitations of our study. Cellular contexts in this study were inferred from RNA-seq data rather than measured directly at single-cell resolution, and our analysis may miss genetic regulatory effects involving rare cell populations or finer cell states. In addition, leveraging an expanded lncRNA transcript set broadened the scope of mapping genetic regulation of lncRNA expression, but unannotated lncRNAs still require continued experimental validation and careful curation. Furthermore, our discovery dataset consisted of DLPFC samples. Although the DLPFC is highly relevant to cognitive function and neuropsychiatric disease, extending this analysis to additional brain regions will be important for obtaining a more comprehensive understanding of genetic regulation of lncRNA expression in the human brain.

In summary, by extending the brain eQTL landscape to lncRNAs that are underrepresented in current single-nucleus eQTL resources, our study supports the investigation of non-coding regulatory mechanisms in psychiatric and neurological disorders. Efforts to incorporate larger and more diverse cohorts, additional brain regions, multi-omic evidence, and experimental validation will further enhance the utility and biological interpretability of this resource.

Methods

Ethical approval

This study received ethical approval from the Ethics Committee of Westlake University, under approval number 20200722YJ001.

Data collection and description

Bulk RNA-seq resources

We integrated genotype and bulk RNA-seq data from publicly available cohorts contributed by the PsychENCODE (https://nda.nih.gov/pec, BrainGVEx, LIBD, CMC, HBCC)²⁸ and AMP-AD (https://adknowledgeportal.synapse.org/)²⁹ (MSBB, Mayo, ROSMAP) consortia. The initial dataset comprised 2865 RNA-seq samples across all available cohorts (Supplementary Data 1). For the discovery analysis, we retained cortical samples to preserve sufficient power for ieQTL mapping. For donors with multiple cortical samples, we retained a single sample, prioritizing the DLPFC or closely related prefrontal cortical regions. This yielded a DLPFC-centered cortical discovery dataset comprising RNA-seq data from 2443 donors. Across cohorts, RNA-seq FASTQ files were processed using a uniform pipeline, including read-level quality assessment with FastQC, adapter and low-quality base trimming with Cutadapt⁵⁹, and alignment to GRCh37 using STAR (v2.7.8a) in two-pass mode⁶⁰. Gene-level expression of lncRNAs and PCGs was quantified using RNA-SeQC (v2.3.5)⁶¹, based on a transcript model generated by integrating our cohort-derived lncRNA catalog with GENCODE v37¹⁹. For downstream analyses, we retained only lncRNAs that were detected (TPM > 0.5) in all cohorts.

snRNA-seq resources

PFC48 (https://www.synapse.org/#!Synapse:syn18485175)¹³: The PFC48 dataset generated by Mathys et al. comprises prefrontal cortex (PFC) samples from 48 individuals in the ROSMAP cohort, including 24 neuropathologically normal or minimally affected controls and 24 age-matched individuals with varying degrees of Alzheimer’s disease–related pathology. In total, 80,660 nuclei were profiled, of which 70,634 passed quality control and were retained. Nuclei were annotated into eight major brain cell types based on transcriptomic profiles: astrocytes (n = 3392), endothelial cells (n = 121), excitatory neurons (n = 34,976), inhibitory neurons (n = 9196), microglia (n = 1920), oligodendrocytes (n = 18,235), oligodendrocyte precursor cells (n = 2627), and pericytes (n = 167). Cell-type annotations were taken directly from the original publication.

PFC427 (https://www.synapse.org/#!Synapse:syn52293417)³³: The PFC427 dataset, also obtained from Mathys et al., comprises approximately 2.3 million nuclei isolated from PFC samples of 427 participants in the ROSMAP cohort. Participants spanned a spectrum of Alzheimer’s disease (AD) pathology, ranging from non-AD and early-stage (Braak 0–II) to intermediate (III–IV) and late-stage AD (V–VI). After quality control and doublet removal, 2,359,994 nuclei were retained for analysis. The original study identified 54 transcriptionally defined cell types grouped into 12 major classes, including excitatory neurons (n = 1,043,230), inhibitory neurons (n = 329,699), oligodendrocytes (n = 645,142), oligodendrocyte precursor cells (n = 90,502), astrocytes (n = 149,558), five immune cell types—including microglia, CNS-associated macrophages (CAMs), and T cells (n = 83,889)—and several vascular cell types, including endothelial cells, smooth muscle cells, fibroblasts, and pericytes (n = 17,974).

FACS resources

EPIGABA (https://www.synapse.org/Synapse:syn4588488)³⁵. Gene expression profiles were generated from human PFC samples by Kozlenkov et al. using a FACS protocol. Nuclei were isolated and enriched for inhibitory GABAergic interneurons (${SOX}^{+}$, n = 4), excitatory glutamatergic neurons (${GLU}^{+}$, n = 4), and oligodendrocytes (${OLIG}^{+}$, n = 4).

SuperAgerEpiMap (https://www.synapse.org/#!Synapse:syn25672226)³⁴. FACS-based RNA-seq data generated by Dong et al. from the human cortex. Nuclei were sorted into neuronal ($Neu{N}^{+}$, n = 10) and non-neuronal ($Neu{N}^{-}$, n = 10) populations across five cortical regions (Brodmann areas 10, 17, 22, 36, and 44) from ten neurotypical adult donors.

IHC quantification data

IHC-based cell quantification data were obtained from ref. ³⁶. This dataset includes relative abundance estimates for neurons, astrocytes, microglia, oligodendrocytes, and endothelial cells from 49 individuals in the ROSMAP cohort. Samples with matched bulk RNA-seq data were used for deconvolution benchmarking.

Epigenomic resources

Cell-type-specific ChIP–seq peaks for H3K27ac and H3K4me3 were obtained from Nott et al.⁴⁵. Lineage-specific ATAC-seq peaks distinguishing $Neu{N}^{+}$ and $Neu{N}^{-}$ nuclei were obtained from Fullard et al.⁴⁴, whereas cell-type-specific chromatin accessibility peaks were derived from Emani et al.¹².

GWAS resources

We curated publicly available GWAS summary statistics for 36 brain-related phenotypes, classified into four major categories: (1) Neuropsychiatric disorders (e.g., schizophrenia and bipolar disorder); (2) Cognitive traits (e.g., intelligence and educational attainment); (3) Neurodegenerative disorders (e.g., Alzheimer’s disease and Parkinson’s disease); (4) Morphological and developmental traits (e.g., regional brain volume and age at menarche). Full details on included phenotypes, original publications, and GWAS sample sizes are provided in Supplementary Data 3.

Benchmarking of deconvolution methods

Basic settings

We systematically benchmarked four widely used cell-type deconvolution methods: MuSiC³⁰, CIBERSORTx³¹, BayesPrism²⁶, and Bisque³². For CIBERSORTx (https://cibersortx.stanford.edu/), we constructed a reference signature matrix from the PFC48 dataset by computing cell-type–average expression profiles, followed by counts-per-million (CPM) normalization and log transformation; this matrix was then submitted for deconvolution. For MuSiC, BayesPrism and Bisque, deconvolution was performed using each method’s R package with default parameters, using the raw UMI count matrix from PFC48 as the reference input. Method performance was evaluated using Spearman’s rank correlation coefficient to assess concordance between inferred and true cell-type proportions, and MAE to quantify estimation error.

Simulation evaluation settings

To generate synthetic bulk RNA-seq mixtures with predefined cell-type compositions, we first randomly divided the PFC48 dataset into two subsets: one serving as the source for constructing simulated mixtures and the other used as the reference for deconvolution. Each synthetic sample consisted of 500 nuclei drawn from a single donor in the source subset, thereby preserving inter-individual expression heterogeneity⁶². For each mixture, between five and all available cell types from that donor were included. One cell type was randomly chosen and assigned a free proportion $f_c\in \left[\mathrm{0,\; 1}\right]$. The remaining cell types were assigned unnormalized proportions $r_c^{\prime }$, which were then normalized and scaled so that the combined proportion equaled ${1-f}_c$, ensuring that total composition summed to 1:$f_c^{\prime }=\frac{r_c^{\prime }}{\sum r_c^{\prime }}\times \left(1-f_c\right)$. The number of nuclei contributed by each cell type was computed by multiplying the final proportions by 500. Expression profiles were generated by summing counts from nuclei sampled with replacement according to these proportions, yielding a total of 5000 simulated mixtures.

Real‑data evaluation settings

To complement simulation-based assessments, we further evaluated deconvolution accuracy using real datasets with known or high-confidence cell-type compositions. Specifically, we used FACS-purified samples from the SuperAgerEpiMap³⁴ and EpiGABA³⁵ projects as ground-truth benchmarks. Each deconvolution method was applied to the bulk RNA-seq profiles of these samples using PFC48 as the single-cell reference. Performance was evaluated based on whether the inferred cell-type abundances were predominantly assigned to the expected cell type. For example, a $Neu{N}^{+}$ sample should be almost entirely decomposed into neuronal components, with minimal contributions from glial or other non-target populations. As an additional cortical benchmark, IHC-based cell quantification data from ROSMAP individuals with matched bulk RNA-seq data were compared with deconvolution-derived cell-type proportions estimated using the PFC48 reference. IHC-derived abundances were normalized to sum to one across the measured cell types before comparison, and performance was assessed using Spearman’s rho and MAE.

Mapping ieQTLs

For ieQTL analysis, genotype quality control was performed using a standardized pipeline consistent with our previous work⁶³. SNPs were excluded if they had a genotype call rate below 95%, Hardy-Weinberg equilibrium p-value less than 1 × 10⁻⁸, MAF below 0.01, or imputation information score below 0.3. We further excluded individuals of non-European ancestry, as determined by principal component analysis, and one of each pair of related individuals (genetic relatedness > 0.05). The matched expression matrix was normalized for library size using the trimmed mean of M-values (TMM) method, followed by rank-based inverse normal transformation (RINT) to mitigate the influence of outliers and better satisfy linear model assumptions.

We employed tensorQTL⁶⁴ to perform ieQTL mapping. For each gene, genetic variants located within $\pm 1$ Mb of the TSS were tested. To minimize potential bias caused by extreme cell-type proportions, we applied a stratified filtering approach: only variants with MAF > 1% in both the lower and upper 50% of samples (based on cell-type abundance) were included. A linear interaction model (Y~G+E+G×E+C) was fitted for each gene–SNP pair, where $Y$ represents the RINT-transformed gene expression; $G$ is the genotype; $E$ is the RINT-transformed cell-type abundance estimated via BayesPrism; G × E is the genotype–cell abundance interaction term; $C$ denotes a set of covariates, including genotype principal components, sex, age at death, and probabilistic estimation of expression residuals (PEER)⁶⁵ factors. The number of PEER factors was dynamically adjusted based on sample size: 15 for n < 150; 30 for 150 ≤ n < 250; 45 for 250 ≤ n < 350; and 60 for n ≥ 350. To increase statistical power, we combined data across cohorts and used a unified linear regression model to remove batch effects prior to downstream analyses. Multiple testing correction was conducted in two stages: first, eigenMT⁶⁶ was used to correct for the effective number of tests per gene based on local LD structure; second, the Benjamini-Hochberg procedure was applied genome-wide to control the FDR, with significance defined as FDR < 0.05.

Based on the dynamic behavior of the genotype main effect (${\beta }_G$) across levels of cell-type abundance, we classified significant ieQTLs into three functional subtypes following a framework that has been extensively benchmarked and adopted by the GTEx Consortium²². Specifically, Positive direction: $∣{\beta }_G∣$ increases significantly in the high-abundance (top 25%) compared to the low-abundance (bottom 25%) group; Negative direction: $∣{\beta }_G∣$ decreases significantly in the high-abundance group; Uncertain direction: the sign of ${\beta }_G$ reverses between low- and high-abundance subgroups (${\beta }_{Glow}\times {\beta }_{Ghigh}<0$).

SNP annotation was performed using ANNOVAR⁶⁷ with the integrated transcriptome model established in our previous work¹⁹, which incorporates both GENCODE (v37) annotations and lncRNA transcripts absent from current GENCODE annotations.

Replication of ieQTLs

We assessed the reproducibility of ieQTLs by comparison with previously published single-nucleus eQTL datasets. Specifically, we evaluated reproducibility using datasets from ref. ¹⁵, which were derived from the ROSMAP cohort and mapped eQTLs across seven major brain cell types and 54 transcriptionally defined cell subtypes, and from ref. ¹⁶, which profiled eQTLs across eight major brain cell types in an independent cohort not included in this study. Replication was quantified using the ${\pi }_1$ statistic, which estimates the proportion of true positives⁶⁸. The ${\pi }_1$ statistic was calculated using the qvalue R package, with the tuning parameter $\lambda$ set to 0.5 when more than 20 tests were available and to 0.85 when more than 100 tests were available in the replication dataset⁴⁰.

Estimation of cell-type specificity of ieQTL effects

To estimate cell-type specificity and sharing of ieQTL effects across the four major brain cell types, we used multivariate adaptive shrinkage implemented in mashr⁴¹, a multivariate empirical Bayes method that jointly models effect estimates across conditions to obtain posterior effect-size estimates and uncertainty measures. The input to mashr consisted of genotype-by-cell-type interaction-effect estimates and their corresponding standard errors for each SNP–gene pair in each cell type. For each gene, we defined the lead variant as the tested variant with the smallest interaction P value observed across the four cell types. The resulting lead SNP–gene pairs were used as the strong-signal set for learning data-driven covariance matrices in mashr, with rows representing SNP–gene pairs and columns representing cell types. We additionally selected 250,000 tested SNP–gene pairs at random to estimate the correlation structure among effect estimates across cell types. The fitted mashr model was then applied to significant ieQTLs to obtain posterior effect-size estimates and LFSR for each SNP–gene pair in each cell type, with LFSR < 0.05 used to define significant effects.

Epigenomic enrichment

We assessed the functional enrichment of ieQTLs in cell-type-associated regulatory annotations. Let $N$ denote the total number of ieQTLs tested and $n_{\mathrm{obs}}$ the number overlapping a given regulatory annotation, yielding an observed overlap proportion $x=\frac{n_{\mathrm{obs}}}{N}$. To mitigate ascertainment bias, we constructed an empirical null distribution by randomly sampling $N$ control SNPs from the set of all SNPs tested in the analysis, matching MAF and distance to the transcription start site. This procedure was repeated 1000 times, and for each iteration $i$, the overlap proportion $y_i$ between the control variants and the same regulatory annotation was calculated. The expected overlap proportion was defined as $ȳ=\frac{1}{m}{\sum }_{i=1}^m{y}_i$. Functional enrichment was quantified as the fold enrichment $FE=\frac{x}{ȳ}$. The sampling variance of the fold enrichment estimate was approximated using the delta method^24,69. In general, the variance of the ratio $\frac{x}{ȳ}$ can be expressed as $\mathrm{var}\left(\frac{x}{ȳ}\right)\approx {\left(\frac{x}{ȳ}\right)}^2\left[\frac{\mathrm{var}\left(x\right)}{x^2}+\frac{\mathrm{var}\left(ȳ\right)}{{ȳ}^2}-\frac{2\mathrm{cov}\left(x,ȳ\right)}{xȳ}\right]$. Assuming that the covariance term $\mathrm{cov}\left(x,ȳ\right)$ is negligible and that $\mathrm{var}\left(x\right)$ can be approximated by the empirical variance of the control overlap proportions, the variance simplifies to $\mathrm{var}\left(\frac{x}{ȳ}\right)\approx {\left(\frac{x}{ȳ}\right)}^2\frac{\widehat{\mathrm{var}}\left(x\right)}{x^2}+\frac{\widehat{\mathrm{var}}\left(ȳ\right)}{m{ȳ}^2}$, where $\widehat{\mathrm{var}}\left(ȳ\right)$ denotes the sample variance of the control overlap proportions across resampling iterations.

Colocalization analysis

We integrated GWAS summary statistics with ieQTL signals using two complementary summary-level approaches. We performed Bayesian colocalization analysis using the coloc R package⁵² to evaluate whether an ieQTL signal and a brain disorder–associated GWAS signal were likely driven by the same causal variant. For each locus, coloc estimates posterior probabilities for five competing hypotheses, and colocalization was considered supported when the posterior probability for a shared causal variant (PP4) exceeded 0.8. We applied summary data–based Mendelian randomization (SMR; https://yanglab.westlake.edu.cn/software/smr/) analysis to further assess putative causal relationships between gene expression and disease risk⁵³. Statistical significance was determined using a Bonferroni-corrected threshold ($p_{SMR}<0.05/m$ where $m$ denotes the total number of tested genes). To distinguish true pleiotropic associations from confounding due to LD between independent causal variants, we further applied the heterogeneity in dependent instruments (HEIDI) test, excluding loci showing evidence of heterogeneity ($p_{HEIDI}\le 0.01$). For SMR and HEIDI analyses, LD estimates were derived from 8,652 unrelated individuals from the Health and Retirement Study (dbGaP: phs000428), which served as the reference panel (Supplementary Note). To validate cell type-specific gene expression, we quantified lncRNAs and PCGs in FACS-purified samples from the SuperAgerEpiMap³⁴ and EpiGABA³⁵ projects, applying the same inclusion criteria.

Reporting summary

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

Author contributions

Tianyi Zhao and Y.W. conceived and supervised the study. Y.J., L.C., and J.Y. designed the experiments. Y.J. and L.C. performed data collection, cleaning, and quality control. Y.J. conducted data analyses with support and guidance from L.S. and L.C. Y.J. and Tianyi Zheng drafted the initial manuscript. All authors contributed to manuscript review and revision and approved the final version.

Peer review

Peer review information

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

Funding

This work was supported by the National Key Research and Development Program of China (2025YFE0216700, 2024YFC3405800, 2024YFC3405802).

Data availability

The full summary statistics for PCG-ieQTLs and lncRNA-ieQTLs generated in this study are available through the Brain-ieQTL Atlas (https://brain.hitxqtl.org.cn/Brain-ieQTLAtlas/#/). PsychENCODE data: https://www.synapse.org/#!Synapse:syn4921369/files/. AMP-AD data: https://www.synapse.org/#!Synapse:syn5550382. Other publicly available datasets used in this study are listed below: Fujita et al. eQTL summary statistics: https://www.synapse.org/Synapse:syn52335807. Haglund et al. eQTL summary statistics: https://zenodo.org/records/13343729. MetaBrain ieQTL summary statistics: https://www.metabrain.nl. EPIGABA: https://www.synapse.org/Synapse:syn4588488. SuperAgerEpiMap: https://www.synapse.org/#!Synapse:syn25672226. HRS can be accessed through dbGaP under accession phs000428; GTEx can be accessed through dbGaP under accession phs000424. Source data are provided with this paper.

Code availability

Custom scripts used in this study are available at GitHub: https://github.com/yrjia1015/Brain-ieQTL.

Competing interests

The authors declare no competing interests.

Supplementary information

The online version contains supplementary material available at https://doi.org/10.1038/s41467-026-77304-7.