Early and late RNA eQTL are driven by different genetic mechanisms

Abstract

Understanding the genetic regulation of RNA abundance is essential for defining disease mechanisms. Conventional expression quantitative trait locus (eQTL) studies measure steady-state RNA and capture effects across the entire transcript lifecycle. While most eQTL likely affect transcription by altering promoter or enhancer function within the nucleus, others may act post-transcriptionally through RNA modification or stability in the cytosol. To distinguish these mechanisms, we compare eQTL from mature cellular RNA and recently transcribed nuclear RNA in brain and kidney. We identify distinct causal variants underlying cellular and nuclear eQTL at the same eGenes. Cellular eQTL are enriched in transcribed regions (P = 3.3×10⁻¹²⁶), suggesting post-transcriptional regulation, whereas nuclear eQTL are enriched in distal regulatory elements (P = 7.0×10⁻³²), consistent with transcriptional control. For example, stop-gain variants likely acting through nonsense-mediated decay appear only in cellular eQTL. Conversely, nuclear eQTL variants (e.g., TUBGCP4) within enhancers sometimes uniquely colocalize with disease loci (schizophrenia), revealing distinct regulatory mechanisms.

By comparing newly produced RNA in the nucleus with mature RNA in the cytosol, genetic variants that control gene expression showed striking differences in localization and biological mechanisms, helping explain how they contribute to disease.

Introduction

Across their entire lifecycle, RNA molecules are under tight temporal and spatial regulation. In eukaryotes, promoters and enhancers in genomic DNA regulate RNA transcription in the nucleus¹. The nascent transcript is then capped, spliced, and polyadenylated to form mature RNA. Mature RNA molecules are exported to the cytoplasm, where RNA-binding proteins (RBPs) make chemical modifications and modulate structural stability, subcellular localization, translation efficiency, and degradation^2,3. Safeguarding mechanisms for each step, such as nuclear retention in the nucleus and nonsense-mediated decay (NMD) in the cytosol, prevent aberrant RNA from being translated⁴. The breakdown of high-fidelity RNA processing may contribute to disease⁵.

The effect of complex disease alleles on RNA abundance has been extensively cataloged through expression quantitative loci (eQTL) studies^6,7. While eQTL studies have emerged as a valuable resource to interpret human disease alleles, most disease alleles from genome-wide association studies (GWAS) do not colocalize with eQTL loci^8–10. Notably, most previous eQTL studies used RNA-seq to quantify transcripts from the entire cell. A majority of causal eQTL variants fine-map to gene promoters, while a minority localize to enhancers^6,11,12. In either case, the emerging consensus is that eQTL variants regulate transcription from the DNA by altering the sequence of DNA regulatory elements.⁶ However, genetic variants may affect any process after transcription, including post-transcriptional modifications essential for mature RNA stability. Indeed, molecular QTL studies have identified many loci that affect post-transcriptional processes, such as splicing^13,14, alternative polyadenylation^15,16, and N⁶-methyladenosine (m6A)^17,18.

Here, we assert that single-nucleus (sn)RNA-seq predominantly measures relatively recently transcribed RNA molecules (early RNAs), whereas single-cell (sc)RNA-seq or bulk RNA-seq measures the entirety of RNA molecules in the cells. The majority (70–90%) of cellular RNA captured by bulk cellular RNA-seq comprised cytosolic RNAs¹⁹. These different RNA profiling technologies can be used to assess the presence of distinct genetic determinants of expression at different time points in the RNA lifecycle. Therefore, we conducted comparative analyses of eQTL variants from nuclear and cellular RNA. We obtained nuclear and cellular RNA-seq data in matched human tissues and conducted cis-eQTL analyses separately for each of the two modalities. We fine-mapped putative causal eQTL variants and identified systematic differences in the genomic locations of nuclear and cellular eQTL. We uncover how genetic regulation varies across RNA lifecycles and subcellular localizations.

Results

eQTL analyses with distinct RNA lifecycles

In this study (Fig. 1a), we analyzed two tissues: (i) brain samples from dorsolateral prefrontal cortex (DLPFC) from participants in the Religious Order Study (ROS) and the Memory Aging Project (MAP)^20,21 for discovery and (ii) kidney samples from lupus nephritis and control participants in the Accelerating Medicines Partnership (AMP) RA/SLE^22,23 for replication (Fig. 1b). The DLPFC study was larger, consisting of cellular bulk RNA-seq (n_samples = 1092) and snRNA-seq (n_samples = 424, n_nuclei = 1,562,635), while the kidney dataset offered direct comparison between scRNA-seq (n_samples = 179, n_cells = 547,796) and snRNA-seq (n_samples = 50, n_nuclei = 144,385). We first mapped cis-eQTL separately in both modalities and statistically fine-mapped causal variants for eGenes using approximate Bayesian factor (ABF) analysis^24,25 (Methods). For each shared eGene in both nuclear and cellular RNA eQTL (n-eQTL and c-eQTL), we assessed whether causal variants of n-eQTL and c-eQTL colocalize^26,27, and whether causal variant locations were systematically enriched within specific functional genomic annotations in one modality relative to the other (Methods).

Fig. 1. Overview of the analyses and datasets.

a Analytical framework. We mapped cis-eQTL of nuclear RNA (red) and cellular RNA (blue) in matched tissues, fine-mapped causal variants, and investigated systematic differences in causal variant location between nuclear eQTL (n-eQTL) and cellular eQTL (c-eQTL). b Discovery dataset in the brain and replication dataset in the kidney. The cell type column depicts UMAP representation of the snRNA-seq and/or scRNA-seq dataset, colored by annotated cell types in the original study. WGS whole genome sequencing, SNP single nucleotide polymorphism, OPCs oligodendrocyte progenitor cells.

In the discovery brain dataset, we identified 9618 eGenes in the nuclear snRNA-seq with FDR < 10%. We used 5632 eGenes that are shared with 9056 eGenes from eQTL summary statistics in cellular bulk RNA-seq (FDR < 5%) for downstream analyses. We confirmed high consistency of our n-eQTL effect estimates with the original publication (Pearson’s r = 0.98 in excitatory neurons; Supplementary Fig. 1). For the 5632 shared eGenes, the effect sizes for the n-eQTL lead SNP partly correlated with those for the c-eQTL (Pearson’s r = 0.72; Fig. 2a). We then tested if the n-eQTL and c-eQTL shared the same causal variant^26,27 (Methods). Surprisingly, 1835 of the shared eGenes (33%) had distinct causal variants (posterior probability (PP) for hypothesis 3 [H3] > 0.7 in coloc). In contrast, 3071 eGenes (55%) confidently had the same causal variants (PP(H4) > 0.7; Fig. 2b). In a comparison of two similarly sized bulk RNA-seq datasets (c-eQTLs) of the prefrontal cortex, PsychENCODE²⁸ (n_samples = 1387) and ROSMAP, we observed higher correlation of the lead SNP effect sizes (Pearson’s r = 0.81) and colocalizing eGenes (73% with PP(H4) > 0.7; Supplementary Fig. 2a, b). Fine-mapping of the causal variants by assuming multiple credible sets (see Methods) demonstrated that the majority of eGenes harbored a single credible set (Supplementary Fig. 3a, b). Further, more complex colocalization analyses assuming multiple credible sets still resulted in similar PP(H3) and PP(H4) with the primary analyses, with 26.4% eGenes having distinct causal variants between n-eQTL and c-eQTL (Supplementary Fig. 3c, d). This suggests that the distinct causal variants for n-eQTL and c-eQTL are more likely driven by biological differences in the regulation of nuclear and cytosolic RNA, rather than by platform-specific fluctuations or assumptions used in the statistical fine-mapping methods.

Fig. 2. Distinct location of causal variants between nuclear eQTL and cellular eQTL.

a Scatter plot for beta coefficients for the same eSNP-eGene pairs between c-eQTL (x-axis) and n-eQTL (y-axis) in the brain dataset. Each dot represents one eSNP-eGene pair, and eSNPs were selected based on n-eQTL statistics. b Histogram of coloc posterior colocalization probability for H3 (left; distinct causal variant between c-eQTL and n-eQTL) and H4 (right; shared causal variant between c-eQTL and n-eQTL). c $t_{pip}$ statistics comparing enrichment of c-eQTL and n-eQTL causal variant locations in the brain dataset. d Proportion of SNPs within each genomic bin among all eSNPs with PIP > 0.7 for n-eQTL (red) and c-eQTL (blue). Stars above the three bins indicate significant deviation from equal proportion between n-eQTL and c-eQTL (two-tailed chi-squared test P < 0.05; P = 0.046, 0.021, and 0.040 from left to right). Bars indicate 95% confidence intervals by bootstrapping variants. e Comparison of enrichment statistics differences in $t_{pip}$ analyses between the brain dataset (x-axis) and kidney dataset (y-axis). Circles represent mean differences of enrichment statistics in each dataset, and bars indicate 95% confidence intervals.

Distinct genomic locations of causal variants for nuclear and cellular eQTL

After fine-mapping, we examined the genomic features harboring eQTL causal variants. We derived the posterior inclusion probability (PIP) of variants in cis-regions of a shared eGene and calculated enrichment of variants within the following genomic annotations (Methods): promoter, candidate cis-regulatory elements (cCREs), exonic region, transcribed region, 3′UTR, 5′ UTR, RBP binding sites by eCLIP assay²⁹, and m6A modification sites¹⁷. Consistent with prior observations, we observed known enrichment in promoters, exonic regions, and 5′ UTRs³⁰ for both n-eQTL and c-eQTL (Supplementary Fig. 4).

We then assessed if there was differential enrichment between nuclear and cytosolic RNA causal variants. We defined a differential enrichment statistic $t_{pip}$ for each of the genomic annotations to test preferential enrichment for causal variants in n-eQTL or c-eQTL (Methods). Briefly, for each modality, we took the aggregate of PIPs for variants within each annotation for an eGene; then we used a paired T-test to compare these values between n-eQTL and c-eQTL across eGenes. A positive $t_{pip}$ indicates that the annotation has stronger enrichment for c-eQTL causal variants, whereas a negative $t_{pip}$ indicates enrichment for n-eQTL. We confirmed that this $t_{pip}$ metric for all genomic annotations is well calibrated and distributed around zero under the null simulation, in which we randomly permuted the PIPs in n- and c-eQTL (Supplementary Fig. 5).

Applying $t_{pip}$ to 5632 eGenes in the brain data, we found that c-eQTL causal variants are most strongly enriched for the transcribed regions ($t_{pip}$ = 24.5, P = 3.3 × 10⁻¹²⁶), followed by exonic and RBP binding sites by eCLIP (Fig. 2c). Given that chemical modification on the transcribed RNA and RBP binding sites affects folding, stability, and decay of RNA in the cytosol³¹, this enrichment likely reflects genetic mechanisms that disrupt post-transcriptional modifications. Consistent with this hypothesis, the significant c-eQTL enrichment in exons and RBP binding sites remained significant even after conditioning on observed enrichment within transcribed regions (Supplementary Fig. 6, Methods). Intronic regions, which in principle are spliced out and not exported into the cytosol, had instead enrichment for n-eQTL variants ($t_{pip}$ = −10.6, P = 7.9 × 10⁻²⁶ in conditional test; Supplementary Fig. 6). Investigation of each of the 168 RBPs individually was underpowered, but we found that the majority of RBPs had positive $t_{pip}$ indicating c-eQTL enrichment (Supplementary Fig. 7). RPS10 had significantly negative $t_{pip}$ enriched for n-eQTL, and DDX47 had nominally significant negative $t_{pip}$ (Supplementary Fig. 7). RPS10 and DDX47 had reported subcellular localization at granular component at nucleolus and nucleus, respectively, but not in the cytosol³². Among all tested annotations, we observed the most significant negative $t_{pip}$ at cCREs more than 250 kb distant from the transcription start site (TSS) ($t_{pip}$ = −11.8, P = 7.0 × 10⁻³²; Fig. 2c), reflecting n-eQTL enrichment in distal enhancers. This might reflect a weak enhancer’s regulatory effect at the transcriptional level within the nucleus, which is more likely to be detected in the nucleus but buffered in the cytosol.

To support these observations, we examined the location of high-confidence eQTL variants (PIP > 0.7) relative to the TSS. We observed significant differences in causal variant localization between n-eQTL and c-eQTL in the regions immediately downstream (3′) of the TSS corresponding to transcribed regions, with c-eQTL having more variants downstream of the TSS implicating transcribed regions (chi-squared P < 0.05, Fig. 2d; Methods). We confirmed this asymmetric distribution of causal variants (i.e., more causal variants downstream of TSS than upstream) in c-eQTL with a skewness statistic based on the third moment of the signed distances (Methods)^33,34. A positive third moment indicates a right-skewed distribution (downstream of TSS), and a negative third moment indicates the left skewed distribution (upstream of TSS). We observed the downstream skewness for c-eQTL (skewness = 0.535 in c-eQTL and −0.406 in n-eQTL).

We considered that the observed differential enrichment might be driven by tissue or dataset-specific effects. Hence, we replicated our findings in an independent kidney dataset with more matched profiling: snRNA-seq and scRNA-seq (Fig. 1b). In this study, the snRNA-seq samples are a subset of the scRNA-seq samples. Despite different tissues and a smaller sample size, we observed highly concordant $t_{pip}$ with the discovery brain dataset across 248 kidney eGenes, with the most positive $t_{pip}$ at the transcribed region ($t_{pip}$ = 11.7, P = 2.1 × 10⁻²⁵), and the most negative $t_{pip}$ at distal cCREs ($t_{pip}$ = −15.2, P = 3.5 × 10⁻³⁷; Supplementary Fig. 8a). The mean differential effect estimates were highly correlated (Pearson’s r = 0.95, P = 2.6 × 10⁻⁵; Fig. 2e).

Since the nuclear studies were smaller than the cytosolic studies, we considered the possibility that power differences in eQTL discovery and fine-mapping may influence the $t_{pip}$. We performed secondary analyses (i) by statistically simulating equally powered n- and c-eQTL for the brain dataset in which raw expression data is unavailable for bulk RNA-seq and (ii) by using fully matched individuals between snRNA-seq and scRNA-seq for the kidney dataset. We replicated the same trend in $t_{pip}$ across genomic annotations for both approaches (Supplementary Fig. 8b, c). We confirmed that matching cell type proportions between snRNA-seq and scRNA-seq did not significantly change the observed enrichment (Supplementary Fig. 8d). We also confirmed that the enrichment trend was consistent when we used the FDR < 5% threshold for defining eGenes in the brain dataset (Supplementary Fig. 9a). The magnitude of differential effect estimates was much larger for non-colocalizing eGenes when compared with all eGenes (Supplementary Fig. 9b) and colocalizing eGenes (Supplementary Fig. 9c). We noted that the differential enrichment for colocalizing eGenes was not completely null possibly reflecting technical differences in accuracy of fine-mapping probabilities between two studies. The additional rank-normalization and removing noisy estimates of PIPs lead to almost no differential enrichment in the colocalizing eGenes, as expected (Supplementary Fig. 9d, e, see Methods). We concluded that the observed systematic differences in causal variant locations between n-eQTL and c-eQTL are pervasive across human tissues.

Distinct biological mechanisms regulating RNA abundance in nuclear and cellular eQTL

We investigated the regulatory mechanisms explaining differences in causal variant locations between n- and c-eQTL by focusing on specific eGenes. First, we observed five eQTL variants (PIP > 0.5) of stop-gain high-confidence loss-of-function (LoF) variants³⁵ causing decreased gene expression in the brain. These variants still yield a full transcript, but the resulting RNA with premature translation termination codons (PTCs) is degraded by NMD, preventing the production of harmful truncated proteins³⁶. For example, in the VWA3B locus, a highly likely causal variant rs62154921 is a high-confidence LoF variant creating a stop codon (PIP = 0.96). The stop-gain allele T significantly decreases the VWA3B expression in c-eQTL (beta = −0.91, nominal P = 7.4 × 10⁻³⁵), which is consistent with NMD. In contrast, the same variant rs62154921 showed low causal probability (PIP = 3.0 × 10⁻⁴) in n-eQTL (Fig. 3a, Supplementary Fig. 10a). After conditioning on the lead n-eQTL variant, this association became insignificant (P = 0.16; Supplementary Fig. 10b). The causal variant(s) was not shared between c- and n-eQTL (PP(H3) = 0.70 by ABF-coloc, PP(H3) = 0.86 and 0.99 by SuSiE-coloc). However, we also note that technical factors could affect colocalization probability, given that the n-eQTL and c-eQTL were performed in overlapping but not completely the same group of individuals. In the kidney dataset, there were no stop-gain variants in eGenes. There was one frameshift LoF insertion (rs66949844), again in c-eQTL with marginal probability of causality (PIP > 0.1), which decreased the expression of SIGLEC12. Of these six LoF variant-eGene pairs, only one variant-gene pair (rs35233100 in MADD) was identified as significant both in c-eQTL and n-eQTL, with the remaining pairs identified solely in c-eQTL. NMD events occur in newly synthesized mRNAs as they exit the nuclear pore complex and are simultaneously translated³⁷. A previous study clarified that NMD occurred within 5–56 s of entering the cytoplasm³⁸, which corresponds to the pioneer round of translation³⁹. An exception has been reported in limited genes, in which NMD occurs after being fully released into the cytoplasm⁴⁰, but none so far within the nucleoplasm. Therefore, our observation of NMD acting predominantly in c-eQTL is consistent with the previously demonstrated cytosolic localization of NMD, while one exception in n-eQTL might be explained by contamination of cytosolic RNA or ambient RNA⁴¹ into the extracted nuclei.

Fig. 3. Different causal variants between nuclear and cellular eQTL reveal distinct biology.

a Stop-gain causal eQTL variant only observed in c-eQTL in the VWA3B locus. The top plot (red) and the bottom plot (blue) depict regional association statistics for the n-eQTL and c-eQTL, respectively, at this locus. The middle plot depicts regional association statistics for the n-eQTL colored by linkage disequilibrium (LD) R² values as shown in the right upper color scale from the lead n-eQTL variant (rs74746113). eQTL P values are from two-sided linear regressions. b Distinct causal eQTL variant between n-eQTL and c-eQTL in the RMB20 locus. The top plot (red) and middle plot (blue) depict regional association statistics for the n-eQTL and c-eQTL, respectively. The bottom plot depicts regional association statistics for the c-eQTL colored by LD R² values, as shown in the right upper color scale from the lead c-eQTL variant. eQTL P values are from two-sided linear regressions. c Schematic representation of the conventional and new concept of multiple causal variants for eQTL. d Violin plots indicating the distribution of the number of variants with PIP > 0.05 within 95% credible sets for eGenes with causal variants within transcribed regions (right) and all the other eGenes (left). The box plots within violin plots demonstrate median (bars) and 25th and 75th percentile values (box edges) and ±1.5 interquartile ranges (IQRs) from 25th and 75th percentiles (whiskers). e Beta coefficients (circles; x-axis) and standard error (horizontal bars) from the association between the LD score (i.e., sum of the LD R² of cis-variants that had greater values of LD R² than a sliding LD threshold (y-axis); Methods) and whether the eSNP is in the transcribed region or not. Right-hand P values indicating the significance of associations between the LD score and whether the eSNP is in the transcribed region or not are from two-sided linear regressions (see Methods).

Intriguing examples of distinct causal variants between n-eQTL and c-eQTL highlighted transcriptional vs. post-transcriptional regulation. There were 90 eGenes in the brain dataset with different causal variants between n-eQTL and c-eQTL, with causal probability in transcribed regions <0.1 in n-eQTL but >0.7 in c-eQTL. Of them, 9 eGenes in n-eQTL had instead regulatory variants with causal probability in cCRE>0.7. RBM20 is one such example (Fig. 3b). In n-eQTL, we fine-mapped a single causal variant, rs72641325, within cCRE proximal to the TSS (PIP = 0.80). However, in c-eQTL, we identified a different 95% credible set of nine causal variants, all located in the transcribed region. None of these variants individually had PIP > 0.3, but together accounted for 95% of causal probability. Fine-mapping with SuSiE identified two credible sets, with one set harboring all of these nine variants with low PIP and the other with another variant, which is also different from the one in the n-eQTL (Supplementary Fig. 11). Four variants in the first credible set were located within the RBP binding sites (FUS, SAFB, RBFOX2, and KHDRBS1)²⁹, suggesting functions in post-transcriptional modification and RNA structure. Moreover, these nine variants were in tight LD (R² > 0.97; Fig. 3b), indicating a shared haplotype with functional consequences on RNA structure. Such examples motivated us to consider the possibility that, in certain instances, multiple causal variants may drive eQTL and complex trait effects.

Multiple causal eQTL variants driving cellular eQTL

Assuming a single causal variant or multiple causal variants in a locus has important implications for statistical fine-mapping approaches in eQTL and GWAS^42,43. One approach to identify multiple causal variants is to use iterative conditional analysis, where a lead variant is identified and adjusted for to find additional independent lead variants^44–49. More recently developed methods, including SuSiE⁵⁰, assume and create multiple “credible sets” that are independent, and each contains a subset of variants that explain each of the independent “single effects.” Both approaches implicitly assume that we need to nominate and distinguish a true biologically causal variant from those that are non-functional and merely tagged by LD, regardless of the number of causal variants or credible sets in the locus. Functional genomic annotation has been used to assist this effort^30,51–54. Our observation of eSNP haplotypes in the transcribed regions might propose a specific scenario that challenges this implicit simpler assumption: correlated multiple variants tagged by LD (Fig. 3c) may in concert influence RNA stability. In this scenario, we do not have to distinguish a single true causal variant among the variants constituting a single effect, but rather we need to consider them as a joint single haplotype effect (Fig. 3c). Previous studies had also proposed similar scenario in regulatory elements, in which multiple functional variants in a haplotype LD block simultaneously affect enhancer activity and are associated with complex traits, such as cancer and autoimmune diseases^55–57.

We asked whether such a scenario is pervasive for eSNPs within the transcribed regions. First, we observed that the number of candidate causal variants with PIP > 0.05 within the 95% credible set was significantly larger for eSNPs within the transcribed regions than those in other annotations such as promoters and enhancers (two-sided t-test P = 4.6 × 10⁻¹⁷; Fig. 3d). Second, we examined LD scores, which capture the weighted averaged number of SNPs in linkage with a lead SNP. We observed that eSNPs within the transcribed regions had significantly larger LD scores than promoters and enhancers (Fig. 3e; Methods). We replicated this trend by using fine-mapped variants of independent bulk-eQTL data from GTEx⁶ with greater statistical significance (Supplementary Fig. 12). These observations may support our proposed scenario that multiple causal variants in the transcribed regions within a haplotype tend to synergistically affect the target RNA’s structure, leading to differences in RNA abundance. In this scenario, all variants within a single effect can be causal, as opposed to variants within promoters or enhancers, which span only ~250 bp^58,59 and most likely harbor one functional variant.

Disease colocalization for nucleus- and cellular-specific

Finally, we investigated whether the eQTL localizing to certain subcellular compartments shared causal variants with disease GWAS, which helps us understand molecular mechanisms. Recently, there is growing discussion that GWAS loci are not sufficiently explained by putative causal variants from cis-eQTL studies. Mostafavi et al.⁶⁰ proposed that the limited overlap might reflect systematic differences in causal variant locations and their target genes; disease GWAS variants are more likely to be under selective pressure and hence distribute further from TSS in the genes with high selective constraint than eQTL variants. Motivated by our observation of distinct causal variant locations between n- and c-eQTL, we tested each of them for colocalization with disease GWAS. Across 11 neuropsychiatric disorders (Supplementary Data 1; Methods), we identified 156 unique disease-eGene pairs that share the same causal variants (Supplementary Data 2). Among them, 42 and 77 disease-eGene pairs specifically colocalized with n-eQTL and c-eQTL, respectively, while 37 disease colocalizations were shared among both eQTL. For example, we identified n-eQTL in tubulin gamma complex component 4 (TUBGCP4) that colocalized with schizophrenia GWAS (lead SNP P = 1.4 × 10⁻¹⁴, PP(H4_GWAS) = 0.90; Fig. 4a, Supplementary Fig. 13a). This locus had distinct causal variants at more powered c-eQTL (PP(H3_eQTL) = 0.94) with less significance (lead SNP P = 4.2 × 10⁻⁷) that did not colocalize with schizophrenia (PP(H4_GWAS) = 5.0 × 10⁻⁴), suggesting that bulk eQTL had not sufficiently explained this locus. In another example, c-eQTL in the calcium voltage-gated channel subunit alpha1 I (CACNA1I) shared causal variants with schizophrenia GWAS with PP(H4_GWAS) = 0.97 (Fig. 4b, Supplementary Fig. 13b). The causal probability was most localized in the transcribed region for c-eQTL, suggesting an effect on post-transcriptional modifications. In this locus, there were no shared causal variants (PP(H4_GWAS) = 0.10) between n-eQTL and schizophrenia. For such loci, the brain snRNA-seq or multiome RNA/ATAC-seq datasets^61–64 might not adequately identify causal mechanisms. Taken together, we argue that subcellular localization of RNA matters in identifying molecular mechanisms of GWAS loci in addition to cell types, tissues, and conditions of eQTL.

Fig. 4. Subcellular-eQTL-specific disease colocalizations.

Regional plots for eQTL and GWAS statistics (schizophrenia) at TUBGCP4 (a) and CACNA1 at locus (b). For both panels, the top red plots show n-eQTL statistics, the middle blue plots show c-eQTL statistics, and the bottom green plots show GWAS summary statistics for schizophrenia. Posterior colocalization probability (H4: shared causal variant) was assessed between eQTL and GWAS. GWAS and eQTL P values are from two-sided logistic and linear regressions, respectively.

Discussion

In this study, we showed that the genetic variants affecting inter-individual differences in RNA abundance are, in many instances, entirely different between nuclear RNA and cellular RNA. In two different tissues, we argue that causal variants that are unique for n-eQTL are more likely to be driven by distal regulatory elements affecting transcriptional activity. In contrast, c-eQTL variants localized to transcribed regions and RBP binding sites, likely acting through altered post-transcriptional regulation.

Our data further argues that c-eQTL may be driven by multiple variants in tight linkage and located on the same haplotype in the transcribed regions. Generally, in causal variant fine-mapping, investigators have considered the possibility that multiple causal variants in a single locus might individually affect different molecular processes⁵⁰, and methods assume that these variants are unlinked by LD, enabling multi-variant fine-mapping. However, our c-eQTL cases suggest that it may be inappropriate to narrow down the list of causal variants in a single effect to disentangle the LD structure⁶⁵. All variants in a causal haplotype could, in concert, affect inter-individual differences in RNA structure, stability, or degradation rate. While we sought to statistically investigate if there is a haplotype’s effect beyond individual variants’ effect using the kidney dataset, given that the haplotypes are almost completely nested with the variants in the 95% credible sets, no significant haplotype’s effect was observed in the current sample size (Supplementary Data 3, see Methods).

We finally demonstrated that nuclear- or cytosolic-specific eQTL causal variants sometimes colocalize with disease variants. There has been a growing debate on the reasons why there has been limited overlap between currently available eQTL variants and GWAS variants for complex diseases^9,10. Major reasons discussed so far have included tissue- and cell-type-specific eQTL variants and context-dependent effects, such as cellular activation^66–69. Our observation highlights the importance of expanding our eQTL catalog in terms of (1) subcellular localization of RNA and (2) processes that affect RNA abundance after the transcriptional regulation.

We were intrigued by the instances when genetic disease risk variants sometimes colocalize with variants affecting nuclear RNA abundance but not cellular aggregated abundance. Recent genomic studies also suggest that genetic variants affecting epigenomic changes in the nucleus, such as chromatin accessibility, are more likely to colocalize with disease alleles even in the absence of corresponding changes in cellular RNA abundance^70,71. One possible explanation is that for those genes, the rate of RNA transcription in the nucleus is more associated with translation efficiency⁷², rather than the accumulated sum of RNA in the cytosol in a steady state. Nuclear eQTL variants may capture cells’ ability to dynamically respond to a changing environment. Indeed, it has been recently demonstrated that the transcription rate might be coupled with translation efficiency⁷³. Another study showed that for stress-response genes, newly synthesized RNA, as opposed to RNA pre-existing the stress condition, preferentially escapes translational repression as a mechanism to rapidly adapt to the environment⁷⁴. These recent findings are in line with the functional significance of n-eQTL alleles that affect transcription rate and their potential role in diseases.

Our study has several potential limitations. First, we compared nuclear and cellular RNA abundance to infer distinct genetic regulators for the nuclear and cytoplasmic subfraction of RNA. Multiple previous studies have confirmed that cytoplasmic RNA molecules comprise the majority of cellular RNA across human tissues¹⁹. However, minor contamination of nuclear RNA into the cellular RNA compartment (reported to be 10–30%¹⁹) could have reduced our power to detect distinct causal variants between n- and c-eQTLs. To assess the magnitude of this effect, we performed simulations in which RNA from one cell type was artificially contaminated into the other cell type at a locus with distinct causal variants between the two cell types. We observed that we could still recover sufficient power to discriminate distinct causal variants between the two cell types at contamination rates up to 60% (see Methods, Supplementary Fig. 14). Still, a more direct and powerful comparison might be achieved by extracting cytoplasmic RNA, if in the future such technologies are applied widely alongside matched nuclear RNA profiling. Second, all the datasets we analyzed used the 3′-end RNA-sequencing method, which does not allow us to perform transcript-level analyses. To partly address this, we tested colocalization of n-eQTL and c-eQTL with published sQTL from the DLPFC¹⁴. We observed that colocalizing probability with sQTL did not significantly vary between n-eQTL and c-eQTL (paired t-test P = 0.14; Methods). This is biologically plausible, since the splicing machinery acts before RNA molecules are exported out of the nucleus. We also identified putative causal variants (PIP > 0.7) localized to 5′ UTR regions, even with 3′-end RNA-sequencing data, since genotype covers the entire cis-region (Supplementary Data 4). If in the future 5′-end or full-transcript technologies are applied for both nuclear and cytoplasmic fractions, these datasets might reveal a more precise effect on alternative splicing, alternative start sites, and other post-transcriptional mechanisms. Third, the statistical power to detect eGenes are affected by technical factors including the number of cells in single-cell or nucleus studies (Supplementary Fig. 15) and the number of variants and genes analyzed (Supplementary Data 5), while our main conclusion remains stable with different FDR thresholds for defining eQTLs (Supplementary Fig. 9a). Fourth, to definitively prove the regulatory effects of the identified eQTL variants, especially on RNA structure and RBP binding capacity as a haplotype consisting of multiple putatively functional variants, we need experimental perturbation approaches accompanied by molecular profiling of RNAs. While this is beyond the scope of this paper, novel technologies for base-editing putative multiple causal SNPs simultaneously and detecting genomic and transcriptomic changes in single cells are emerging⁷⁵.

In summary, we demonstrated distinct regulatory mechanisms affecting RNA abundance that localize in the nucleus or in the cytosol, which encompass various post-transcriptional processes beyond conventional transcriptional regulation. These mechanisms can sometimes explain missing molecular causes of diseases, expanding our view of gene regulation and its association with pathogenic processes.

Methods

Ethics statement

This study was performed in accordance with protocols approved by the Brigham and Women’s Hospital. We used publicly available data of ROS/MAP data, which was approved by the Institutional Review Board of Rush University Medical Center. The Religious Orders Study (ROS) enrolls Catholic nuns, priests, and brothers from more than 40 groups across the United States. The Memory and Aging Project (MAP) study primarily enrolls residents of continuous care retirement communities throughout northeastern Illinois. All participants are without known dementia and agree to annual clinical evaluation and brain donation. The details of study approval and human participant recruitment are described previously²¹. We also used the AMP SLE study, whose protocol was approved by the institutional review boards and ethics committees of participating sites in adherence with the Declaration of Helsinki. Informed consent was obtained from all participants. The details of study approval and human participant recruitment are described previously^76,77.

Datasets

We used two independent sources of data for the main analyses in the paper. The first dataset is snRNA-seq and bulk RNA-seq from the DLPFC of the postmortem brain tissue samples from the ROSMAP cohort²⁰. We used this dataset as a discovery dataset since it had both RNA and genotype data from > 400 individuals. At the time of death, 34% of participants were cognitively nonimpaired, 26% were mildly impaired, and 40% had dementia. 68% were female, and 63% fulfilled a pathological diagnosis of Alzheimer’s disease by the National Institutes of Health (NIH) Reagan Criteria²⁰. We downloaded the raw count matrix of snRNA-seq and whole-genome sequence (WGS) genotype from 424 individuals²¹ from Synapse under agreement on Controlled Data Usage. Since raw data from bulk RNA-seq was not available at the time of our analysis, we downloaded the available summary statistics of eQTL from the original study²¹ from Synapse, instead of running eQTL mapping on our own. We also obtained a list of significant eGenes as defined in the study through communication with the authors of the paper.

The second dataset is snRNA-seq and scRNA-seq from biopsied kidney tissues from the AMP RA/SLE consortium²². We used this dataset as a replication dataset. In brief, as part of the AMP RA/SLE Phase 2 consortium patients, individuals >16 years old undergoing a clinically indicated kidney biopsy to evaluate proteinuria were enrolled if they met sufficient criteria of systemic lupus erythematosus (SLE) diagnosis based on the revised American College of Rheumatology or the Systemic Lupus Erythematosus International Collaborating Clinics classification criteria, along with healthy control individuals. We processed the scRNA and snRNA expression matrices and genotype data, the details of which will be described in the flagship paper (Gurajala et al.⁷⁸).

Genotype quality control

For the downloaded WGS data from the brain dataset, we selected variants that were (i) annotated as “PASS,” (ii) biallelic, (iii) variant call rate >0.95, (iv) minor allele frequency (MAF) > 0.01, and (v) P_HWE > 1.0 × 10⁻⁶. For the SNP-chip genotype data from the kidney dataset, we applied the same genotype quality control and imputation steps as described in previous studies from the AMP RA/SLE consortium^79,80. In brief, we genotyped donors by using the Illumina Multi-Ethnic Genotyping Array. We retained samples and variants with (i) sample call rate >0.99, (ii) variant call rate >0.99, (iii) MAF > 0.01, and (iv) P_HWE > 1.0 × 10⁻⁶. We performed haplotype phasing with SHAPEIT2 software and performed whole-genome imputation by using minimac3 software with a reference panel of 1000 Genomes Project phase 3⁸¹. After imputation, we selected variants with imputation Rsq > 0.7 and MAF > 0.05 as post-imputation QC.

eQTL mapping

For snRNA-seq and scRNA-seq of both the brain and the kidney datasets, we pseudo-bulked and normalized the expression matrix using post-QC cells from all the cell types to maximize the statistical power in eQTL discovery and fine-mapping. We note that this strategy was necessary in the brain snRNA-seq to be comparable with the brain bulk RNA-seq, and in the kidney dataset because the statistical power to detect cell-type-specific eQTL was limited (Supplementary Data 6). We aggregated and averaged the log-normalized expression values from a given donor for a given gene. We retained genes with non-zero expression sample rate >0.2 and performed inverse normal transformation of the aggregated expression across samples. We then performed PEER correction⁸² of the gene expression by including known covariates: Sex, age at death (the age of sample collection), study cohort, postmortem interval, and five first genotype principal component (PC) values for the brain dataset and sex, age, and five first genotype PC values for the kidney dataset. We confirmed that the PEER factors for correction are not collinear with each other.

We performed standard cis-eQTL mapping for the genes after QC by linear models between corrected gene expression and genotype dosages. We used FastQTL software (version 2.0)⁸³ with default settings. In addition to obtaining nominal association statistics from the linear models, we used FastQTL’s beta approximation to compute permutation P values from 10,000 permutations. Finally, to adjust for multiple hypothesis testing across genes and define significant eGenes, we obtained Storey’s q values using q value package (version 2.30.0) in R (version 4.2.3). We used q values < 0.10 as the level of FDR significant eGenes for these analyses for snRNA-seq and scRNA-seq. We summarized the number of genes and variants analyzed in the brain dataset in Supplementary Data 5.

Fine-mapping and colocalization with ABF

We used Coloc software (version 5.2.3)^26,27 to perform fine-mapping and colocalization analyses between n-eQTL and c-eQTL. The fine-mapped posterior probability for causal variants was calculated from the ABF analysis implemented in finemap.abf() function with default settings. Importantly, these coloc analyses assume that there is a single causal variant in the cis-region for a given gene. This setting has a good feature for the goal of our analyses, that the sum of PIP across the variants in the cis-region should always sum up to one when we analyze significant eGenes, which makes the analyses comparable across different genes. In addition, we discussed potential multiple causal variant scenarios when multiple variants had very similar PIPs in the Main text. While the finemap.abf() function outputs the posterior probability of the null model, where no SNP is causal in the locus. In addition to the posterior probability of each SNP being causal, we considered that the probability for the null model is negligible since we have only analyzed FDR-significant eGenes with at least one causal SNP being associated with the gene’s expression. Therefore, for the differential enrichment test that is going to be described, we adjusted raw PIPs by dividing them by the sum of PIPs across the cis-variants (but excluding the probability for the null) such that the sum of SNP’s PIP adds up to one. For the colocalization between n-eQTL and c-eQTL, we used coloc.abf() function with default settings and defined posterior probability (PP) for hypothesis 3 (H3) for loci with distinct causal variants between n-eQTL and c-eQTL, and PP for hypothesis 4 (H4) for loci with shared causal variants between n-eQTL and c-eQTL. We adjusted the PP for H3 and PP for H4 such that the sum of PP for H3 and PP for H4 added up to one, because we have only analyzed FDR-significant eGenes for both n-eQTL and c-eQTL, and therefore reasoned that PP for H0 (no causal variant), H1 (causal variant for the first trait only), and H2 (causal variant for the second trait only) should be negligible. We excluded any loci (eGenes) with fewer than 200 genetic variants shared between n-eQTL and c-eQTL. We defined non-colocalizing genes with distinct causal variants for those with PP for H3 > 0.7 in coloc. We defined colocalizing genes with shared causal variants for those with PP for H4 > 0.7 in coloc. The remaining genes with both 0.3 ≤ PP(H3) ≤ 0.7 and 0.3 ≤ PP(H4) ≤ 0.7 were categorized as ambiguous genes for colocalization of causal variants between n- and c-eQTL.

Fine-mapping and colocalization with SuSiE

As a secondary analysis in the brain dataset, we also performed fine-mapping and colocalization with SuSiE, which can assume multiple credible sets in a locus⁵⁰. For consistency of the analysis, we used summary statistics of eQTL for fine-mapping and used an in -sample LD for the n-eQTL and LD reference using ROSMAP samples for the c-eQTL. We removed monomorphic sites from the LD reference. We used the susieR package in R (version 0.12.35) with default settings and with specifying five as the number of the maximal number of credible sets in both eQTL studies. We then performed colocalization analysis between n- and c-eQTL using the coloc-SuSiE package in R (coloc version 5.2.3) with default settings.

Analyses with the PsychENCODE dataset

We sought to compare our observation of effect size correlation and colocalization between n-eQTL and c-eQTL by comparing two similarly sized c-eQTLs (bulk RNA-seq eQTLs) in the human brain prefrontal cortex. The first dataset was bulk RNA-seq eQTL from the ROSMAP cohort (n = 1092) that we had described in the previous section, and the second dataset was bulk RNA-seq eQTL from the PsychENCODE Consortium. The PsychENCODE²⁸ collected bulk RNA-seq and genotype data from 1387 individuals and released eQTL summary statistics.

We assessed the lead SNP effect size correlation between ROSMAP and PsychENCODE. We took 4289 shared eGenes from both cohorts, aligned the effect alleles, intersected cis-genetic variants for each eGene shared between the two cohorts, and calculated Pearson’s correlation R of the beta coefficients from the two cohorts by selecting one most significant genetic variant from the ROSMAP study for each of the 4289 shared eGenes. We next assessed colocalization of causal variants between ROSMAP c-eQTL and PsychENCODE c-eQTL. We ran coloc.abf() function for the shared eGenes between ROSMAP and PsychENCODE, and only retained loci with at least 200 genetic variants shared between ROSMAP and PsychENCODE. Similarly to the primary analysis, we ensured that the sum of PP for H3 and PP for H4 added up to one, since we had already selected loci to be analyzed for the significant eGenes for both studies.

Differential enrichment test

To quantify differential enrichment of causal eQTL variants between n-eQTL or c-eQTL within a specific genomic annotation or unit, we defined an enrichment statistic $t_{pip}$ for each of the genomic units. For the genomic annotations or units to be tested for enrichment, we examined promoters, cCREs with varying distances from TSS, exonic and transcribed regions, 3′ and 5′ UTRs, RBP binding sites, and m⁶A modification sites. We defined exonic regions by aggregating all possible exonic regions across all patterns of transcripts for the target gene registered at the RefGene database (GRCh37). The brain-specific cCREs were based on the Brain Open Chromatin Atlas⁸⁴ and PsychENCODE enhancer list for prefrontal cortex²⁸. The kidney-specific cCREs were based on cCRE annotation in the human kidney sample from ENCODE v3 (accession: ENCSR751YZB). We defined potential RBP binding sites by using peaks from the eCLIP assay released from the ENCODE consortium²⁹ for 168 unique RBPs for HepG2 and K562 cell lines, as well as adrenal gland tissue. We defined m⁶A modification sites from published m⁶A-seq data from 60 Yoruba lymphoblast cell lines.¹⁷ For RBP binding sites and m⁶A modification sites, we ensured that those regions are included in the transcribed region of the eGene, which we tested for enrichment, so that we have accurate insights into the post-transcriptional regulation for the target eGene but not for the other genes.

Based on fine-mapped PIPs for either n-eQTL or c-eQTL, we first took the sum of PIPs for all the SNPs that belong to a given genomic unit $c$ (e.g., promoter) for each eGene $g$ as an enrichment statistic for $\left\{\mathrm{nucleus}\mathrm{or}\mathrm{cell}\right\}$ eQTLs:

$$\begin{array}{c}{\mathrm{Enrichment}}_{g,c,\left\{nucleusorcell\right\}}=\sum _m{\mathrm{PIP}}_m\hfill \end{array}$$ 1

where $m\in \mathrm{genomic}\mathrm{unit}c$

$m$ indicates SNPs in the cis-region for eGene $g$ that belong to the genomic unit $c$.

We then compared ${\mathrm{Enrichment}}_{g,c,cell}$ and ${\mathrm{Enrichment}}_{g,c,nucleus}$ for each eGene $g$, and ran paired T-test across all the shared eGenes between n-eQTL or c-eQTL to calculate $t_{pip,c}$ for the genomic unit $c$.

$$\begin{array}{c}{d}_{g,c}={\mathrm{Enrichment}}_{g,c,cell}-{\mathrm{Enrichment}}_{g,c,nucleus}\hfill \end{array}$$ 2

$$\begin{array}{c}{t}_{pip,c}=\overline{d_c}/\mathrm{SE}\left(\overline{d_c}\right)\mathrm{across}\mathrm{all}\mathrm{eGenes}\hfill \end{array}$$ 3

Given the subtraction direction above, the positive $t_{pip,c}$ value indicates that the genomic unit c has stronger enrichment for c-eQTL causal variants, whereas the negative $t_{pip},c$ value indicates that it has stronger enrichment for n-eQTL causal variants. While we calculated the $t_{pip,c}$ across all the FDR-significant eGenes shared between n- and c-eQTL regardless of the colocalization of causal variants, we performed secondary analyses just focusing on the non-colocalizing eGenes with different causal variants and on the colocalizing eGenes with shared causal variants (Supplementary Fig. 9).

To demonstrate the calibration of the $t_{pip}$, we performed a permutation experiment, in which we permuted the variant’s PIP in the summary statistics of n- and c-eQTL 500 times. We calculated the $t_{pip}$ in each genomic annotation in each set of the permuted PIPs. We collected the $t_{pip}$ results from 500 permutations and confirmed the distribution (the mean and standard error) of the $t_{pip}$ in each genomic annotation (Supplementary Fig. 5).

We also performed a sensitivity analysis with a rank-normalization and thresholding of PIPs to have more similar distributions of PIPs between c- and n-eQTL which will be expected to yield conservative differential enrichment results. First, we intersected the cis-wide genetic variants in each gene shared between c- and n-eQTL. We then ranked the intersected cis-wide genetic variants in each gene based on the original PIP values from ABF fine-mapping from the greatest to the smallest, separately for the c- and n-eQTL studies. In the ranking, we randomly assigned unique ranks when the ties of PIPs arose between variants. We then assigned a new rank-normalized PIPs for the variants with the same rank for the same gene between c- and n-eQTL by calculating the averaged PIPs between c- and n-eQTL. We also removed any variants with very low normalized PIPs (normalized PIP < 0.01) to avoid the effect of noisy PIP estimates from inaccurate fine-mapping. We finally conducted the same $t_{pip}$ analyses between c- and n-eQTL for the remaining eGenes.

Test for asymmetric distribution of causal eQTL variants

We selected variants with PIP > 0.7 as highly probable causal variants for each eGene. We obtained a list of such variants and their distance to the TSS with accounting for the gene’s strand. We concatenated the list across all the eGenes separately for n-eQTL and c-eQTL. We then binned the cis-region (±1 Mb window from TSS) by a window size of 20 kb. Separately for n-eQTL and c-eQTL, we calculated the fraction of causal variants belonging to each bin as the proportion of SNPs belonging to the bin, similarly to ref. ⁶⁰. We also calculated the 95% confidence intervals (CIs) of these proportions by bootstrapping the variants in the list 1000 times. To statistically test the differences in the proportion between n-eQTL and c-eQTL for each bin, we performed a chi-squared test for a two-by-two table, with the number of causal variants within the bin and outside of the bin (columns) for n-eQTL and c-eQTL (rows).

We also tested the asymmetric distribution of causal variants relative to the TSS as the center of the distribution. We defined the skewness statistics^33,34 as follows using the 3rd moment of the signed distances between the causal variant and the TSS.

$$\begin{array}{c}\mathrm{skewness}=\frac{n}{\left(n-1\right)\left(n-2\right)}\sum _{i=1}^n{\left(\frac{x_i-x_{TSS}}{s}\right)}^3\hfill \end{array}$$ 4

where $n$ denotes the total number of causal variants. $x_i$ indicates the genomic position for SNP $i$, $x_{TSS}$ indicates the genomic position for TSS of the eGene, and $s$ indicates the standard deviation of the signed distances (i.e., differences in genomic positions) between the variant and the TSS. We flipped the sign of $x_i-x_{TSS}$ if the eGene is coded in the negative strand.

Baseline enrichment test

We also assessed the baseline distribution of the causal variants for the genomic units in n-eQTL and c-eQTL (Supplementary Fig. 4). To compare enrichment across the genomic units rather than within a unit, as we did in the $t_{pip}$ analyses, we needed to adjust for the length and the number of variants within the genomic units and across the cis-region.

$${Enrichment}_g=\frac{{\sum }_m{PIP}_m/\mathrm{\#}variantsintheunitc}{{\sum }_n{PIP}_n/\mathrm{\#}variantsacrossthecisregion}$$ 5

where $m\in \mathrm{causal}\mathrm{variants}\mathrm{in}\mathrm{genomic}\mathrm{unit}c$, and $n\in \mathrm{causal}\mathrm{variants}\mathrm{in}\mathrm{the}\mathrm{entire}\mathrm{cis}\mathrm{region}$. We defined the causal variants with variants with PIPs > {0.3, 0.5, 0.7}. We then calculated the overall baseline enrichment value by taking the mean of ${\mathrm{Enrichment}}_g$ across all $n$ eGenes,

$$\begin{array}{c}{\mathrm{Overall}}_{Enrichment}=\frac{1}{n}\sum _{i=1}^n{\mathrm{Enrichment}}_g\hfill \end{array}$$ 6

and derived the 95% CIs by bootstrapping genes 1000 times.

Conditional enrichment test

In assessing $t_{pip}$ in genomic annotations included in the transcribed regions in principle (i.e., UTRs, exons, introns, and m⁶A modification sites and RBP binding sites), we wanted to account for known strong enrichment within the transcribed regions. To this end, we performed conditional analyses for these genomic units (Supplementary Fig. 6). We first defined $\mathrm{Fractional}\_\mathrm{Enrichment}$ statistics by dividing sum of PIPs in a given genomic unit c by sum of PIPs in the transcribed regions as follows:

$$\begin{array}{c}{\mathrm{Fractional}\_Enrichment}_{g,c,\left\{nucleusorcell\right\}}=\frac{{\sum }_m{\mathrm{PIP}}_m}{{\sum }_t{\mathrm{PIP}}_t}\hfill \end{array}$$ 7

where $m\in \mathrm{genomic}\mathrm{unit}c$ and $t\in \mathrm{transcribed}\mathrm{region}$

We also ensured that all SNPs $m$ that belong to the genomic unit c are also within the transcribed region of the gene $g$. Similarly to the primary analyses, we then compared $\mathrm{Fractional}\_{\mathrm{Enrichment}}_{g,c,cell}$ and $\mathrm{Fractional}\_{\mathrm{Enrichment}}_{g,c,nucleus}$ for each eGene $g$, and ran paired T-test across all the shared eGenes between n-eQTL or c-eQTL to calculate $\mathrm{conditional}{\_t}_{pip,c}$ for the genomic unit $c$.

$${{\mathrm{fractional}}_d}_{g,c}={\mathrm{Fractional}\_Enrichment}_{g,c,cell}-{\mathrm{Fractional}\_Enrichment}_{g,c,nucleus}$$ 8

$${{\mathrm{conditional}}_t}_{pip,c}=\overline{{{\mathrm{fractional}}_d}_c}/\mathrm{SE}\left(\overline{{{\mathrm{fractional}}_d}_c}\right)\mathrm{across}\mathrm{all}\mathrm{eGenes}$$ 9

Credible set size

We assessed significant differences in the size of the fine-mapped credible set between eGenes with probable causal variants within transcribed regions or eGenes with probable causal variants within conventional promoters or cCREs. For each eGene $g$ in c-eQTL in the brain dataset, we assessed the genomic unit with the highest ${\mathrm{Enrichment}}_{g,c,cell}$ with assuring the ${\mathrm{Enrichment}}_{g,c,cell}>0.05$. Based on this, we categorized eGenes into promoter-driven, enhancer (cCREs)-driven, transcribed region-driven, and other ones. We constructed 95% credible sets for eGenes by sorting variants based on their PIPs in descending order and cumulatively summing and collecting them until the cumulative sum is 0.95. We used variants with at least PIP > 0.05 for this analysis (i.e., maximum credible set size is 19), and then we compared the number of variants within the 95% credible sets between transcribed region-driven eGenes and promoter- or enhancer-driven eGenes by T-test. This filter by PIP is introduced because we observed that the distribution of the raw number of variants in 95% credible sets is highly long-tailed (Supplementary Fig. 16), with some credible sets with thousands of uncertain variants with extremely low PIPs. We considered that such distributions are not appropriate for comparison using T-test as they are susceptible to being biased by outliers, and therefore we decided to focus on loci (eGenes) with credible variants with at least 5% of causality.

Tagged variants in LD

We also assessed significant differences in the number of genetic variants tagged by LD between probable causal variants in transcribed region-driven eGenes and those in promoter- or enhancer-driven eGenes. Using the WGS genotype data in the brain dataset, we comprehensively calculated the LD R² between a sentinel variant with maximum PIP (>0.05) for a given eGene and all other variants in the cis-region for the eGene within a 1 Mb window. We defined the LD scores for each eGene by summing the total LD R² of variants that had greater values of LD R² than a sliding LD threshold {0, 0.1, 0.3, 0.5, 0.7} to assess the effect of removing noisy R² from weakly tagged variants. We associated these LD scores from each LD threshold with the categorical variable whether the eGene is classified as transcribed region-driven or promoter- or enhancer-driven, including MAF of the sentinel variant as a covariate.

Analyses with the GTEx dataset

To examine if the observation on credible set size and tagged variants in our brain dataset is generalizable across other tissues’ eQTL, we used the published GTEx v8 dataset. We used the GTEx fine-mapping results by CAVIAR. Similar to our primary analysis in the brain dataset, we defined promoter-driven, enhancer (cCREs)-driven, and transcribed region-driven eGenes for each tissue. cCRE regions were defined using cCRE annotation from the ENCODE consortium⁸⁵ that covers many human cell lines and tissues. We ran the same test for comparison of the 95% credible set size between transcribed region-driven eGenes and promoter- or enhancer-driven eGenes by T-test. We also ran the same association test between the LD scores for the sentinel variant of an eGene and the categorical variable whether the eGene is classified as transcribed region-driven or promoter- or enhancer-driven, including MAF of the sentinel variant as a covariate. We used 1000 Genomes Project phase 3 genotype data to calculate the LD scores and MAF.

Disease colocalization

Using the brain dataset, we conducted colocalization analysis of causal variants from n-eQTL or c-eQTL and those from neuropsychiatric diseases. We downloaded publicly available GWAS summary statistics of 10 neuropsychiatric disorders from the Psychiatric Genomics Consortium (Alzheimer’s disease, anorexia nervosa, attention deficit hyperactivity disorder, autism spectrum disorder, bipolar disorder, major depressive disorder, obsessive compulsive symptoms, panic disorder, post-traumatic stress disorder, schizophrenia) and one from the GWAS catalog (Parkinson’s disease⁸⁶; Supplementary Data 1 for details). We defined loci for colocalization based on the cis-regions for all eGenes from either n-eQTL or c-eQTL in the brain dataset. For each disease GWAS, if the cis-region contains any variant with genome-wide significance (P_GWAS < 5 × 10⁻⁸), we proceeded on assessing colocalization of causal variants between eQTL and GWAS by using the coloc software with default settings. If they do not contain any genome-wide significant variants, we did not perform colocalization analyses. We also excluded any loci with fewer than 200 genetic variants shared between eQTL and GWAS.

sQTL colocalization

We obtained sQTL summary statistics from 450 DLPFC samples in ROSMAP through personal communication with the author of the original publication¹⁴. We selected 1313 genes and their transcripts that are significant in previously described n-eQTL and c-eQTL, as well as in the sQTL study. We tested colocalization of causal variants between sQTL and n-eQTL and between sQTL and c-eQTL by using coloc with the same procedures described in the previous sections. We finally tested whether there is a statistically significant difference in the posterior probability of shared causal variants (PP(H4) for sQTL and n-eQTL vs. PP(H4) for sQTL and c-eQTL) by paired t-test across 1313 genes.

Haplotype test

We sought to test whether there are loci that had a significant haplotypes’ effect beyond the individual SNPs’ effect. We conducted this experiment in the kidney dataset, in which we have availability for both the raw genotype and expression data. We first identified 27 loci with more than five variants in 95% credible set in the transcribed region. We constructed haplotypes that consisted of these SNPs within the 95% credible set and QCed out any rare haplotypes with frequency <5%. We then tested whether including haplotype information in the full model significantly improves the explained variance of the gene expression when compared with the reduced model only including the SNPs by comparing the full and reduced models with an ANOVA F test.

Contamination experiment

To indirectly test the magnitude of the effect of the potential inclusion of nuclear RNA compartment in cellular RNA compartment in identifying distinct eQTL causal variants between nuclear RNA and cytosolic RNA, we computationally mixed the RNAs between two different cell types that had clear distinct causal variants in a locus, with variable degrees of contamination. For this experiment, we chose astrocytes and oligodendrocytes in the brain dataset and examined the EGFR locus. For this locus, an independent study (Zeng et al.⁸⁷) reported two clearly distinct causal variants between astrocytes and oligodendrocytes (rs74504435 and rs7800174, respectively). We replicated these two causal variants in our brain snRNA-seq eQTL data (Supplementary Fig. 14a). The colocalization analysis resulted in PP(H3) = 1.0 and PP(H4) = 1.1e-11, highly likely with distinct causal variants between the two cell types. We then replaced astrocytes with oligodendrocytes with variable fractions between 5 and 80% and ran the same pseudo-bulk eQTL analysis in each fraction, retaining the oligodendrocyte eQTL results. We finally assessed the colocalization probability between contaminated astrocyte eQTL and the oligodendrocyte eQTL across contaminated fractions (Supplementary Fig. 14c). This will be considered as a simulation of the capability to detect distinct causal variants between n-eQTL and c-eQTL contaminated by nuclear RNA by around 10–30%¹⁹.

Reporting summary

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

Author contributions

S.S. and S.R. conceived the work and wrote the manuscript. S.S. conducted all analyses with inputs from S.R. Accelerating Medicines Partnership®: RA/SLE Network provided the kidney dataset.

Peer review

Peer review information

Nature Communications thanks Anelia Horvath and the other, anonymous, reviewer(s) for their contribution to the peer review of this work. A peer review file is available.

Data availability

The publicly available datasets from ROSMAP (https://www.synapse.org/#!Synapse:syn31512863) can be accessed via Synapse under the Data Usage Agreement. AMP SLE dataset with the processed count matrices is also available on Synapse under accession code syn64064827. Interactive visualization of marker genes in our data is available on our website: https://immunogenomics.io/ampsle2/app/?ds=b_pbmcs.

Code availability

The computational scripts related to this manuscript are available at https://github.com/immunogenomics/EarlyLate_RNA_eQTL (10.5281/zenodo.18962458).

Competing interests

S.R. is a founder of Mestag, Inc, a scientific advisor for Jannsen and Pfizer, and serves as a consultant for Sanofi, AbbVie, Nimbus, and Third Rock Ventures. S.S. declares no competing interests.

Supplementary information

The online version contains supplementary material available at 10.1038/s41467-026-72139-8.