Functional dissection of complex trait variants at single-nucleotide resolution

Siraj, Layla; Castro, Rodrigo I.; Dewey, Hannah B.; Kales, Susan; Butts, John C.; Nguyen, Thanh Thanh L.; Kanai, Masahiro; Berenzy, Daniel; Mouri, Kousuke; Wang, Qingbo S.; Fiziev, Petko P.; Tsuo, Kristin; McCaw, Zachary R.; Gosai, Sager J.; Aguet, François; Cui, Ran; Kassam, Irfahan; McRae, Jeremy; Vockley, Christopher M.; Lareau, Caleb A.; Abramov, Sergey; Boystov, Alexandr; Vierstra, Jeff; Okada, Yukinori; Gusev, Alexander; Jones, Thouis R.; Lander, Eric S.; Sabeti, Pardis C.; Finucane, Hilary K.; Reilly, Steven K.; Ulirsch, Jacob C.; Tewhey, Ryan

doi:10.1038/s41586-026-10121-6

Download PDF

Article
Open access
Published: 25 February 2026

Functional dissection of complex trait variants at single-nucleotide resolution

Nature (2026)Cite this article

18 Altmetric
Metrics details

Subjects

Abstract

Identifying the causal variants and mechanisms that drive complex traits and diseases remains a core problem in human genetics^1,2,3,4,5. Most of these variants individually have weak effects⁶ and lie in non-coding gene-regulatory elements^7,8,9,10, for which we lack a complete understanding of how single-nucleotide alterations modulate transcriptional processes to affect human phenotypes^{5,11,12,13,14,15}. To address this problem, we measured the activity of 221,412 fine-mapped trait-associated variants using a massively parallel reporter assay^{16,17,18,19,20} in 5 diverse cell types. We show that this assay effectively discriminates between likely causal variants and controls, and identified 13,121 regulatory variants with high precision. Although the effects of these variants largely agree with orthogonal measures of function, only 69% of them can plausibly be explained by the disruption of a known transcription factor binding motif. We investigated the mechanisms of 136 variants using saturation mutagenesis and assigned affected transcription factors for 91% of variants without a clear canonical mechanism. Finally, we detected regulatory epistasis at 11% of tested regulatory variants in close proximity and identified multiple functional variants on the same haplotype at a small, but important, subset of trait-associated loci. Overall, our study provides a systematic functional characterization of likely causal common variants that underlie complex and molecular human traits, enabling new insights into the regulatory grammar underlying disease risk.

Main

Genome-wide association studies (GWASs) have successfully linked tens of thousands of loci to complex human traits and diseases^{1,21,22,23,24}. A more difficult task has been to pinpoint the exact causal site^12,25,26, or variant, in each locus, which is a critical step for understanding how individual variants and the genes they affect contribute to genetic risk^2,10,27,28. Most trait-associated variants individually have small effect sizes⁶ and are located in non-coding cis-regulatory elements (CREs)^7,12. The identification of causal variants is further hindered by linkage disequilibrium, in which alleles of nearby variants are correlated with each other²⁹.

Genetic fine-mapping^25,26,30 can partly disentangle these correlations by assigning to each variant a posterior inclusion probability (PIP) that it is causal for the association. Currently, only around 10–20% of trait-associated loci can be resolved to a single variant using genotypic information^8,31,32. To help with the identification of causal variants at challenging loci, fine-mapping constructs 95% credible sets (CSs), which are minimal sets of variants that contain at least 95% of the cumulative PIP. Compared with alternative linkage disequilibrium windowing approaches, CSs are typically smaller and more experimentally tractable³¹.

Although large-scale experimental catalogues of CREs⁹ and other genomic annotations can help in identifying causal variants^4,33,34, most fine-mapped variants do not disrupt recognized units of regulatory syntax⁸. Direct genome editing of CREs containing these variants in vitro can uncover affected sequences and nucleotide-specific effects^13,14, but these methods are limited in scale and sensitivity.

To address these challenges, we used a massively parallel reporter assay (MPRA), which is a high-throughput approach that simultaneously evaluates the regulatory activity of thousands of DNA sequences^{16,17,19,20,35,36}. Designed to measure CREs from both promoters and distal elements, MPRAs effectively detect responses to a diverse range of transcription factors (TFs)^35,37. Each sequence is uniquely tagged with barcodes embedded in the reporter transcript, enabling not only the measurement of transcriptional activity³⁷, but also testing of allele-specific activity by comparing barcode abundance in RNA transcripts with input plasmid DNA.

We developed the following systematic approach to evaluate trait-associated loci and measure the performance of our assay at the scale required by GWAS:

1.
We selected variants in 95% CSs from recent fine-mapping studies of complex traits and expression quantitative trait loci (eQTLs; molecular traits). We included variants with high PIP (greater than 0.9) and low PIP (less than 0.01) as proxies for causal and non-causal control variants.
2.
We synthesized elements of 200 base pairs (bp) containing each variant, placing the reference or alternative allele in the centre, cloned these elements upstream of a minimal promoter, and transfected the libraries into five diverse cell lines^38,39. For variant pairs within 150 bp and in a CRE, we tested all four haplotype combinations to assess epistatic (non-additive) interactions.
3.
Elements were defined as active if one or both alleles appreciably increased or decreased transcriptional output in at least one cell line (|log₂-transformed fold change [log₂FC]| > 1, Bonferroni-adjusted P < 0.01). Variants in active elements were defined as expression-modulating variants (emVars) if there was a significant difference in activity between alleles (false discovery rate [FDR] < 0.1).
4.
To assess the ability to discriminate between likely causal and non-causal variants, we evaluated how well emVars could separate high-PIP variants from controls.

High-PIP variants were enriched among emVars and other genomic annotations, such as CREs. Here, we defined CREs as genomic regions that have evidence of both accessible chromatin and an activating histone modification in at least one cell type, which we obtained by compiling cell atlases of DNase I hypersensitivity sites (DHSs), transposase-accessible chromatin (ATAC) and H3K27ac (Methods). Consistent with our previous work¹⁸, emVars that reside in endogenous CREs best identified the likely causal regulatory variants. We call these 13,121 high-confidence variants trait-associated regulatory variants (TARVs).

To understand how molecular and complex TARVs affect expression, we used a systematic approach to nominate TF sites and other functional features affected by TARVs. We used this information to select 136 TARVs, representing a diverse range of traits and predicted regulatory effects, on which we performed deep mutational scanning (saturation mutagenesis). By using the MPRA, we scanned entire elements containing either the reference or the alternative allele and measured the impact of changing each nucleotide to every other possible nucleotide. We developed a method to identify short, contiguous DNA sequences, termed activity blocks (ABs), that contribute to element activity, enabling us to directly nominate TF sites affected by TARVs, including cases missed by conventional approaches. Finally, we leveraged these sequence-to-function maps to explore crucial aspects of regulatory grammar and demonstrate how these maps can be used to understand not only common but also rare variant effects.

In summary, our approach systematically measures the regulatory effects of hundreds of thousands of trait-associated variants, dissects their mechanisms, identifies their epistatic effects, and reveals the complex interplay between common regulatory variants and their sequence contexts.

Measuring variant effects at scale

We selected 221,412 fine-mapped variants, consisting of: 78,238 trait-associated variants for at least one of 48 complex human traits and diseases from European (the UK Biobank (UKBB)) and East Asian (Biobank Japan (BBJ)) populations; and 148,805 eQTL variants for at least one of 17,969 genes across 49 tissues from a diverse US population (the Genotype Tissue Expression Project, GTEx v.8), including 5,631 variants shared between cohorts (Fig. 1a, Methods and Supplementary Tables 1–3). In total, variants from 89,387 CSs (Methods) and 17,161 unique high-PIP variants were included. We further selected 86,064 control variants, consisting of 14,170 variants associated with a simulated null trait, 32,534 variants matched with a fine-mapped variant for genomic location, and 39,708 variants matched with a fine-mapped variant for genomic annotation (Fig. 1b, Methods, Extended Data Fig. 1a, Supplementary Fig. 1a and Supplementary Table 1).

**Fig. 1: Identification of functional genetic elements housing fine-mapped complex and molecular traits.**

To create MPRA libraries, 304,278 unique elements with either the reference or the alternative allele (608,556 total sequences) were synthesized and paired with 20-bp 3′ untranslated region (UTR) barcodes (median of 163 unique barcodes; Fig. 1c, Extended Data Fig. 1b and Methods). The MPRA libraries containing either complex trait variants (UKBB and BBJ) or eQTL variants (GTEx) were transfected into four cell types each, representing five distinct cell types overall: blood and myeloid (K562), liver (HepG2), brain (SK-N-SH) and colon (HCT116) or lung epithelial (A549) (Fig. 1c, Extended Data Fig. 1c,d and Methods). Overall, 30.4% of elements were active (92,560 of 304,278; Extended Data Fig. 1e), of which 96.6% enhanced rather than repressed transcription, consistent with the use of a minimal promoter in this assay (Supplementary Table 4 and Methods).

We sought to characterize the facets of regulatory grammar underpinning element activity in MPRAs. Active elements were more likely than non-active elements to be derived from CREs (odds ratio = 2.77 for promoters and 1.77 for distal CREs, P < 10⁻³⁰⁰; Extended Data Fig. 2a and Supplementary Fig. 2a) or to overlap with TF occupancy sites (Extended Data Fig. 2b, Supplementary Fig. 2b and Supplementary Table 5), using chromatin immunoprecipitation (ChIP) peaks from 1,741 TFs across 1,046 cell types^40,41 (Methods). Inactive elements were not enriched for variants with coding or splicing effects, but we observed an enrichment for variants in 3′ UTRs (P = 4.2 × 10⁻¹⁰; Extended Data Fig. 2c,d), indicating that assays measuring the effects of 3′ UTR variants could resolve additional variant effects.

Among active elements in accessible chromatin, we observed that element activity in our assay is largely correlated across the cell types tested (median Pearson r = 0.70), more strongly so for promoters than for distal CREs (median Pearson r = 0.82 versus 0.61), and is higher at cell-type-matched, cell-type-specific CREs (Extended Data Fig. 2a,g and Methods). We compared the maximum element activity in MPRAs across the cell types tested to the maximum chromatin accessibility in an atlas of 438 cell types⁹. The two were well correlated (Pearson r = 0.47), although notably more so at promoters than distal CREs (Pearson r = 0.61 versus 0.30, P < 10⁻³⁰⁰), probably reflecting differences in the complexity and composition of TFs at distal and proximal elements⁴² (Fig. 1d and Methods).

We modelled element activity in MPRAs as a function of TF binding site (motif) usage, based on position weight matrices (PWMs) for 1,210 TFs (Methods). For most TFs, another motif occurrence had similar effects on element activity across cell types (r = 0.88–0.94), including strong activators (such as JUN–FOS), repressors (SNAI and REST) and lineage-specific regulators (GATA, KLF, GFI1B, HNF and IRF) (Fig. 1e, Supplementary Table 6 and Methods).

Having assembled a credible catalogue of active elements, we examined whether the variant in each active element modulated transcription in an allele-specific manner; that is, whether the variant was an emVar. In active elements, 40.2% (37,249 of 92,560) of variants were detected as emVars (Fig. 1f, Extended Data Fig. 3a,b, Supplementary Figs. 3–5, Supplementary Table 4 and Methods). emVars typically had modest effects (median |∆log₂ activity| = 0.58), and most emVars had allelic effects in more than one tested cell type (complex traits, 75%; eQTLs, 69%; Supplementary Table 7 and Methods), reflecting many shared transcriptional components. However, we observed a small but significant increase in emVar detection for cell types most similar to the tissue in which the eQTL was identified (Extended Data Fig. 3d and Supplementary Note 1). Importantly, control variants were less likely to reside in active elements or to be emVars than the high-PIP variants (P = 9.0 × 10⁻²⁵ and 9.9 × 10⁻⁵⁰ for complex traits and eQTLs, respectively; Fig. 2a and Extended Data Fig. 3e,f).

**Fig. 2: Reporter assays recapitulate endogenous regulatory function.**

We assessed how well emVar allelic effects correlated with orthogonal measures of regulatory variant effects. Focusing on high-PIP eQTLs, we observed that emVar allelic effects modestly agreed with endogenous expression effects (Spearman’s ρ = 0.47 and 0.35 in promoters and distal CREs, respectively; Fig. 2b). Next, in cell type and tissue samples heterozygous for emVars, we found that the higher-activity alleles were associated with both higher chromatin accessibility (Pearson r = 0.65; Fig. 2c) and greater endogenous TF occupancy (Pearson r = 0.54; Extended Data Fig. 3g). Finally, we examined whether emVar allelic effects on MPRA activity were consistent with their predicted impact on TF binding (motif strength), finding a significant correlation between the predicted change in TF binding and MPRA allelic effects for 78% (479 of 614) of TFs with sufficient power in at least one cell type (FDR < 0.05; Fig. 2d, Supplementary Table 8 and Methods), including ETS1 (ρ = 0.86), CEBPG (ρ = 0.83), CTCF (ρ = 0.74), GATA1 (ρ = 0.63), SNAI2 (ρ = −0.78), GFI (ρ = −0.83) and BCL6 (ρ = −0.84).

Prioritizing causal regulatory variants

We investigated the ability of our assay to discriminate between high-PIP and low-PIP variants (Methods). We found that emVar status discriminated between high-PIP and low-PIP variants from balanced sets with a precision and recall of 0.70 and 0.25 for complex traits and 0.68 and 0.22 for eQTLs, respectively (Fig. 2e, Extended Data Fig. 4a,b and Supplementary Tables 9 and 10). Similar values were observed using alternative negative approximation sets (Methods). Comparing our assay with other genomic annotations, we found that CRE overlap had a similar ability to discriminate between high-PIP and low-PIP variant sets, consistent with previous observations¹⁸. Importantly, we found that requiring CRE overlap with allele-specific regulatory effects from our assay (emVar status) improved precision to 0.83 while maintaining a recall of 0.20 for complex traits, and 0.82 and 0.15 for eQTLs. By contrast, emVars outside CREs had no predictive power for high-PIP status. Of note, the exact precision and recall for a specific variant depends on the underlying trait, the resolution of fine-mapping, and the cell types used for each assay (Extended Data Figs. 3a–c and 4c–g, Supplementary Tables 10 and 11, and Discussion).

Overall, we identified 13,121 disease- and trait-associated regulatory variants (Supplementary Tables 12 and 13). The TARVs were observed in trait-relevant, cell-type-specific CREs, but a substantial minority reside in multi-tissue CREs and promoters⁴³ (Supplementary Fig. 6a and Supplementary Table 14)_. Of these, we identified 9,285 TARVs across 18,976 exclusively non-coding CSs (Methods), including a high-confidence subset of 2,368 TARVs across 4,274 CSs where no other CS variants had evidence of alternative function (Supplementary Table 15, Supplementary Note 2 and Methods). To contextualize TARVs, we provide putative target genes from multiple CRE-to-gene predictors (Supplementary Table 16 and Methods). Compared with other similar genome-wide approaches⁴⁴, we found that our targeted study provides improved trait-associated variant coverage, sensitivity and recall (Supplementary Fig. 6b–d).

Mechanisms of transcriptional regulation

After establishing a catalogue of TARVs, we explored how well existing approaches for characterizing regulatory variant mechanisms could separate high-confidence TARVs from other CRE variants. We compared TF motif strength between elements containing the reference or alternative alleles⁸ and observed that 69% of TARVs disrupt known motifs for one or more of 839 TFs. However, 52% of low-PIP CRE non-emVars also disrupt such motifs (Fig. 2f and Methods), which is consistent with the ubiquity of seemingly non-functional TF motifs across the genome^45,46. Requiring TF occupancy to corroborate motif disruptions better separates high-PIP TARVs from low-PIP CRE non-emVars. Specifically, restricting this to either TF footprints in DHS from 243 cell types⁴⁷ or overlaps with ChIP–seq peaks⁴¹ prioritizes functional binding sites for 37% and 43% of high-PIP TARVs, but only 12% and 10% of background CRE variants, respectively (P < 2.3 × 10⁻¹⁶¹; Supplementary Table 17). However, we note that no single TF explains more than 2% of trait-associated high-PIP TARVs (Supplementary Fig. 7a).

We explored potential mechanisms for the remaining 57% of high-PIP TARVs that do not disrupt the motif at an occupied TF site, and observed residual enrichment at the flanking regions of occupied TF sites (±10 bp), indicating that a subset of TARVs act outside catalogued TF binding sites^8,17 (Fig. 2f). We then applied Enformer⁴⁸, which is a transformer-based neural network that models variant effects on both chromatin accessibility and TF occupancy (Methods), and found that an ensembled Enformer score was best at distinguishing high-PIP TARVs from other CRE variants (62.6% versus 7.3%; Fig. 2f). Of the high-PIP TARVs not predicted to disrupt a TF motif, Enformer identified 51% as having an impact on chromatin accessibility or TF occupancy (16% of all high-PIP TARVs, P = 2.8 × 10⁻¹¹³; Supplementary Fig. 7b). Predicted variant effects were often correlated (average of 177 TFs affected per variant), obscuring the exact molecular mechanisms without further information. Taken together, our results nominate high-confidence molecular mechanisms for thousands of fine-mapped TARVs^8,17.

Multiple variant and epistatic effects

Allelic heterogeneity, in which multiple independent (different CS) but nearby (same locus) variants are associated with a trait, has been widely established for both complex and molecular traits⁴⁹. There is evidence of multiple causal variants in tight linkage disequilibrium (same CS) in humans^50,51 and other species⁵². To directly evaluate the cis-regulatory aspects of genetic architecture, we performed MPRA on variant pairs (all four diplotypes) in the same CREs in up to six sliding windows. We selected 1,377 pairs of variants in the same CS, 961 pairs of variants from different CSs, and 184 pairs of variants with PIP > 0.1 where a CS could not be constructed (Fig. 3a, Extended Data Fig. 5a–f, Supplementary Tables 18 and 19 and Methods). Leveraging our sliding-window approach (Methods), we showed that allelic effects between diplotypes are robust when approximately 75% of nucleotides are shared (Extended Data Fig. 5g–i).

**Fig. 3: Evidence for regulatory allelic heterogeneity and multiple causal variants.**

We next explored whether regulatory allelic heterogeneity could occur in the same CRE and identified 218 TARV pairs from different CSs. We subsequently investigated non-additive, epistatic effects, and found 180 such pairs (11%, FDR < 0.1; Fig. 3b, Extended Data Fig. 5j,k, Supplementary Tables 20 and 21 and Methods). Regulatory epistasis was driven in part by variant proximity (mean difference of 18 bp, P = 2.0 × 10⁻¹⁰; Fig. 3c), and the presence of both activity-increasing alleles most commonly had a smaller effect than expected based on their individual effects, resulting in dampened transcriptional output (139 of 180, binomial P = 1.1 × 10⁻¹³; Fig. 3d and Extended Data Fig. 5l). If these observations hold broadly across complex traits, our results provide evidence for linkage masking^53,54, in which variants escape negative selection by having opposing effects.

We highlight two examples of regulatory epistasis, in which each variant is in strong linkage disequilibrium with its pair. Both rs1936950 and rs1936951 are fine-mapped (PIP = 0.64 and 0.36, respectively) in cerebral-hemisphere tissue for effects on the ESS2 gene, which is linked to developmental disorders including DiGeorge syndrome⁵⁵. Both variants alter the same CTCF PWM and independently activate transcription (Δlog₂ activity of 0.51 and 1.41, respectively). However, their combined effect is greater than their additive contributions (Δlog₂ activity of 2.54 versus 1.92, FDR = 5.5 × 10⁻⁴; Fig. 3e); rs9294987 and rs9294988 are also in a single CS (PIP = 0.48 and 0.39, respectively) for systolic blood pressure at the THBS2 locus, a gene with proposed roles in cardiovascular function³. In contrast to the ESS2 locus, this pair can activate transcription significantly only when both C alleles are present to create a Jun motif (Δlog₂ activity of 1.68; Extended Data Fig. 5m). These examples underscore that TF binding can be a nonlinear function and support shared variation potentiation⁵⁶, in which effects on a shared regulatory element are not equally distributed but act on average to either repress or activate gene expression.

Looking beyond specific pairs of variants in a single CRE, we investigated whether we could observe an excess of TARVs in CSs containing at least one such variant (Methods). We found significant enrichments for additional TARVs in CSs across multiple CS sizes and linkage disequilibrium thresholds (risk-ratio range: 1.03–2.81; Fig. 3f and Supplementary Table 22). This observation was mainly due to TARVs in different CREs⁵¹ (range of 0.69–0.83 additional CREs per excess TARV), ultimately corresponding to an excess of TARVs in 0.1–3.0% of CSs. Our findings are consistent with previous studies based on CRE annotations⁵¹; however, we found that further conditioning on emVar status improved enrichments (CRE-only risk-ratio range: 1.17–1.60 using low-PIP controls; Supplementary Table 22). Importantly, we did not observe an enrichment for emVars without considering CRE annotation (risk-ratio range: 0.62–1.39), highlighting the importance of accounting for genomic context and background rate of emVar identification when assessing genetic architectures^16,17,50.

Saturation mutagenesis maps mechanisms

To investigate the molecular mechanisms for a subset of TARVs, we selected 128 CREs containing 136 TARVs with either a canonical mechanism (83 of 136) where the variant disrupts the known motif at an occupied TF site, or a non-canonical mechanism (53 of 136) where the variant does not (Fig. 4a, Supplementary Tables 23 and 24, and Methods). For each allele, we synthesized DNA oligonucleotides with every possible single-nucleotide substitution across the 200-bp element and measured their effects on transcription using MPRA in K562 and HepG2 cells (Fig. 4b). We also investigated the haplotype effects of 14 variant pairs, assaying across all 4 possible diplotypes (Supplementary Fig. 8a and Methods). In total, we recovered 99% of unique substitutions (more than 170,000) across 284 total elements, with excellent replication of emVar effects from the initial screen (Pearson r = 0.98, P < 10⁻¹⁵⁰; Supplementary Fig. 8b–d, Supplementary Tables 24, and 25, and Methods).

**Fig. 4: Saturation mutagenesis of 136 fine-mapped TARVs.**

To identify the short motif-like sequences that affect transcriptional activity, we used a Gaussian filtering approach to call ABs of functional nucleotides (Methods). We found that 98% of elements contained at least one AB, which were approximately the size of a TF binding motif (mean of 9 bp, range 5–33 bp), covered an average of 26% of each element, and were 6.2 bp closer to each other than expected by chance (P = 0.002). The ABs significantly overlapped known TF footprints (56% of ABs and 52% of footprints, odds ratio = 1.51, P < 10⁻³⁰⁰), indicating the identification of both shared and complementary regulatory functions. To nominate causative TFs for each AB, we derived PWMs from saturation mutagenesis across ABs and matched these to catalogued PWMs (Methods), assigning at least one TF for 89% of ABs (Supplementary Table 25).

Although all 128 elements drove transcriptional activity in at least one background and cell type, a surprising proportion of ABs acted as repressors (34%), increasing expression when disrupted. Saturation mutagenesis-derived PWMs matched known repressor TFs, including GFI1B and SNAI1/SNAI3, exclusively at repressive ABs (Supplementary Table 25). Overall, the direction and magnitude of TF motif effects at ABs was consistent with their inferred contributions from the initial screen (ρ = 0.68, P < 10⁻³⁰⁰; Fig. 4c). Several TARVs disrupted repressors to either reveal underlying activators or mediate switching between activating and repressive sites. For example, in K562 cells, rs536864738 (minor allele frequency (MAF) = 0.0015) ablates an evolutionarily constrained GFI1B repression site, revealing a PBX3 activator (Fig. 4d,e). Conversely, rs11571842 transforms an activating TFDP1 to a repressive ZNF343 site (Extended Data Fig. 6a,b), and rs11864973 converts an activating KLF1 to a SNAI1 site, notably within the MCS-R4 element of the α-globin locus enhancer cluster⁵⁷ (Fig. 4f,g).

Single-nucleotide effects and AB assignments were often correlated across cell types (Fig. 4h), but several elements demonstrated cell-type-specific effects largely explained by master transcriptional regulators. For example, rs536864738 ablated a GFI1B motif with repressive effects in K562 but not HepG2 cells (Fig. 4d,e and Extended Data Fig. 6c,d). At rs2529369, an eQTL for CDHR3, saturation mutagenesis identified two ABs matching SRY and SOX9 motifs only in HepG2 cells⁵⁸ (Fig. 4i,j). Because Y-linked SRY is present in XY HepG2 but not in XX K562 cells, and the activating A allele of rs2529369 creates an SRY and SOX9 motif, we tested whether this allele displayed sex-dependent effects on CDHR3 expression in GTEx, and observed higher expression in XY than in XX individuals with A allele(s) (P = 0.026; Fig. 4k).

Mechanisms of trait-associated variants

We investigated how well saturation mutagenesis and other methods explain the molecular mechanisms underlying TARVs. Overlap with an AB or matching motif explained 91% of TARVs (124 of 136), including 92% of canonical (76 of 83) and 91% of non-canonical (48 of 53) mechanisms. This outperformed all other methods, including Enformer⁴⁸ (P = 0.001), particularly for non-canonical TARVs (Fig. 5a and Supplementary Table 23). Notably, saturation mutagenesis nominated at least one TF for 47 of 53 non-canonical TARVs, whereas traditional PWM matching identified a TF disruption for only 5. For 26 of 42 variants in which only saturation mutagenesis identified the disrupted TF, a predicted PWM disruption could be matched after relaxing the calling thresholds (Methods). The remaining misses were attributed to motifs with low information content (short motifs) or variant disruption of low-information positions in the catalogued PWM. For example, at an eQTL for DDX1, saturation mutagenesis revealed that rs7953706 partly ablates the LIN54–E2F4 DREAM complex (Extended Data Fig. 6e,f). Although ChIP–seq provides evidence of occupancy by LIN54, E2F4 and 16 other TFs, all have either a sub-threshold match to the catalogued PWM or a sub-threshold predicted alteration (Methods). This demonstrates how saturation mutagenesis can identify variant mechanisms missed by conventional approaches.

**Fig. 5: Saturation mutagenesis uncovers variant mechanisms.**

After generating sequence-to-function maps for all possible substitutions across 128 elements, we asked whether measures of evolutionary conservation aligned with changes in transcriptional output. Although most (2,277, 80%) large-effect substitutions did not have evidence of evolutionary constraint across mammals, similar to other studies⁴, nucleotides under constraint (FDR < 5%) were more likely to have a large transcriptional impact than unconstrained positions (odds ratio = 7.8, P = 10⁻⁵⁰). This relationship was variable across elements, with most (80 of 128, 63%) displaying significant correlation between the magnitude of constraint and the impact of the position on transcriptional activity (FDR < 5%; Extended Data Fig. 6g,h, Supplementary Table 26 and Methods).

We next used these sequence-to-function maps to compare the spectrum of possible mutation with common, trait-associated variation. Across all substitutions tested by MPRA, a median of 9% of positions and 5% of substitutions had large effects on transcription (a more than 50% increase or decrease; Fig. 5b). Full ablation (log₂FC < 1) of activity was rare (1% of substitutions) and achievable in only 38% of elements. Compared with the largest-effect substitution, TARVs reduced expression less on average (a 47% reduction compared with 73%), although there was considerable heterogeneity across elements (Fig. 5c), including slightly larger reductions at elements with TARVs from complex traits (P = 0.03).

Elements harbouring larger transcriptional-effect substitutions than the TARV provide an intrinsic opportunity to test whether individuals with extant rare alleles have similar or more extreme phenotypic values. One example is rs191148279, which was fine-mapped (PIP = 0.92) for glycosylated haemoglobin A1c (HbA1c) levels and fully disrupts a GATA motif when tested by MPRA (Pearson r = 0.92; Fig. 5d,e). Base editing in K562 cells confirmed that the alternative allele reduced the expression of the nearest gene, C2orf88 (FDR = 3.10 × 10⁻⁶), with no changes at neighbouring expressed genes (Fig. 5f, Extended Data Fig. 7a–c and Methods). Although the variant has a weak effect in the reporter assay (∆log₂ activity = −0.49), two adjacent GATA motifs contribute substantially more to transcriptional activity (∆log₂ activity = −2.93 and −2.42 for a similar G>A substitution), probably because of adjacent binding by TAL1, an important cofactor for GATA1 function^59,60. In the UKBB cohort (n = 402,990), we identified 39 individuals carrying 1 of 5 rare mutations (an allele count of less than 100) at positions in the adjacent GATA and cofactor sites with stronger effects than the common variant (Fig. 5e). Carriers had higher variance (P = 4.7 × 10⁻⁶) in HbA1c and were more likely to decrease HbA1c by 1 s.d. or more (odds ratio = 3.8, P = 6.2 × 10⁻⁵; Fig. 5g). Motivated by this finding, we searched for more variants at higher allele frequencies that were underpowered in the initial UKBB fine-mapping. We identified an allele common to only individuals of African ancestries (MAF = 0.18 versus 5.6 × 10⁻⁴ in individuals of European ancestries) that was predicted to disrupt an adjacent GATA site (T>C). A meta-analysis across four cohorts showed that the C allele significantly reduced HbA1c levels (P = 6.4 × 10⁻¹⁰; Fig. 5h), consistent with the other genetic and functional studies across this element.

Next, we used saturation mutagenesis to examine 14 variant pairs (9 non-additive and 5 additive). Non-additive effects agreed with the effect directions (15 of 16) and effect sizes from the original experiment (Spearman’s ρ = 0.97, P = 9.1 × 10⁻¹⁰; Extended Data Fig. 8a). Saturation mutagenesis enabled us to differentiate between pairs in the same binding motif or independent motifs, confirming that rs1936950 and rs1936951 alter the same CTCF motif (Fig. 3e and Extended Data Fig. 8b,c). We characterized another non-additive interaction at a liver-specific eQTL for DOP1B in a CS containing three variants, all located in CREs at the third intron of DOP1B. Only rs7282770 (PIP = 0.08) was a TARV (Δlog₂ activity = 0.38 in HepG2) with saturation mutagenesis revealing that the minor allele (G) creates two adjacent HNF4-binding sites. Introducing the minor allele (C) of rs7282886 (PIP = 0.08), in perfect linkage disequilibrium (R² = 1.0) with rs7282770, between the two HNF4 sites, resulted in a non-additive increase (Δlog₂ activity of 2.25 versus 1.48; Fig. 5i,j), indicating the facilitation of dimeric HNF4 binding. Although we cannot exclude the possibility that the third CS variant, rs2246810, which had the highest PIP (0.8), also contributes to differences in DOP1B expression, this variant does not lie in an exon, UTR, CRE or splice site, and fine-mapping relies on haplotypes from only two individuals (R² = 0.98).

Another example is found in the promoter of ZNF329, where rs35081008 and rs34003091 (R² = 0.99) are TARVs associated with decreased ZNF329 expression (PIPs = 0.22) and decreased low-density lipoprotein (LDL) cholesterol (PIP = 0.50 and 0.43, respectively). Saturation mutagenesis revealed a shared mechanism for rs35081008, in which the T allele creates a HIC2 repressor site in both K562 and HepG2 cells, and rs34003091 disrupts an HNF1 activator in only HepG2 cells (Extended Data Fig. 8d,e). The role of rs35081008 has previously been validated using base editing, with rs34003091 showing directional concordance but not reaching significance¹³.

Finally, we investigated the interaction between the 114 centred emVars and all 114 × 597 possible substitutions at other positions. Non-additive effects occurred with 10% of substitutions (|∆activity | > 25%), of which 58% were dampening effects (Extended Data Fig. 8f). Most non-additive effects occurred in ABs or matched motifs (78%), and nearly 1 out of 5 substitutions in ABs had non-additive effects (18% versus 4%, P < 10⁻³⁰⁰). Notably, non-additive effects exhibited a strong positional dependence to the central emVar (35% lower odds per log(bp), P < 10⁻³⁰⁰), consistent with our observations from common variant pairs. Similarly, amplifying effects were 85% more likely to be observed in an AB (P = 2.4 × 10⁻¹¹⁹) and had stronger positional dependence than dampening effects (P = 5.4 × 10⁻³³). Although these results demonstrate that epistatic effects can frequently be observed between random substitutions, the extent to which selective forces exploit or avoid these alleles to shape complex and molecular traits remains an open question.

Discussion

We combined state-of-the-art statistical fine-mapping^30,31 with comprehensive experimental validation of non-coding, cis-regulatory variants to identify causal trait-associated regulatory variants and dissect their mechanisms. In total, we tested 221,412 fine-mapped variants associated with hundreds of complex human phenotypes, identifying 13,121 TARVs that alter the transcriptional activity of annotated CREs. Through careful comparisons with 86,064 unique controls, we conclude that MPRA combined with endogenous CRE annotations can confidently identify putative causal alleles at an important subset of trait-associated loci with high precision (0.82–0.83) and modest recall (0.15 for eQTLs and 0.20 for complex traits). Mechanistically, we found that 31% of probably causal (PIP > 0.9) TARVs are not confidently predicted to alter the binding of a known TF motif, and 39% are not predicted to alter TF occupancy.

Saturation mutagenesis of 136 TARVs uncovered sequence motifs directly associated with transcriptional outcomes and assigned a known TF for 91% of TARVs that initially had a non-canonical or unknown mechanism, expanding our mechanistic understanding of regulatory architecture. These data highlight a sensitivity–specificity trade-off in identifying functional TF binding sites using standard PWM-based approaches, as well as revealing some cryptic mechanisms, such as variants between TF dimer binding sites. These approaches provide sequence-to-function maps across the allelic spectrum and serve as essential validations for sequence-based models of regulatory function⁶¹.

Our study further quantifies the genetic architecture of transcriptional regulation, providing clear evidence for widespread allelic heterogeneity⁴⁹ and relatively infrequent but compelling examples of multiple functional variants underlying a single, fine-mapped association. We detected multiple functional variants in individual CSs for only 0.1–3.0% of relevant non-coding associations, which is much fewer than estimates from previous reports⁵⁰ that did not adjust for background effects. As recall of assays such as MPRA increases, we might find that the true proportion of CSs with multiple functional variants is higher (1–15% after ‘adjusting’ for a recall of around 20%). In general, we found that regulatory variants in the same CRE are capable of dampening, and occasionally amplifying, each other’s effects on transcription. We note that our findings might not generalize to all variants, traits or contexts.

Some non-coding loci were missed by our experimental approach by testing elements upstream of a minimal promoter on an episomal plasmid. Variants in these loci might modulate TFs that are inactive in the cell types tested, have regulatory effects below the detection limits, affect chromatin structure, act in isolated repressors, affect post-transcriptional mechanisms or be affected by fine-mapping miscalibration^35,62 (Extended Data Fig. 2e,f). We expect that MPRA in other cell types and contexts will increase the recall by 2.5–3.4% per additional cell type (Extended Data Fig. 4c,d and Supplementary Table 10). This gain is probably due to additive discovery power with each new cell line and access to untested TFs, although this will eventually plateau. Poorly represented cell types offer prioritization opportunities. For example, although TARVs represent DHSs from all the major organ systems, 14% of high-PIP yet MPRA-inactive elements overlap annotated CREs, with connective tissue representing a substantial proportion, and therefore being a promising candidate for further testing (Supplementary Fig. 9a,b, Supplementary Tables 27 and 28, and Supplementary Note 1). These results highlight the need for modified MPRA designs⁶³, further high-throughput functional-characterization methods and testing in more cell types to enable the identification of more TARVs.

We caution that conclusively demonstrating causality for a cellular or organismal phenotype requires genetic manipulation, such as endogenously modifying variants in their native context, or modification of the CRE harbouring the variant, combined with appropriate phenotypic read-outs¹⁵. Although MPRA experiments capture the local sequence context of regulatory elements, they do not identify their target gene(s). Furthermore, transcription is only the first step in the pathways to a cellular or organismal phenotype, and understanding the post-transcription process linking a gene to a phenotype is also essential.

Overall, we have demonstrated that large-scale reporter assays provide a tractable experimental system for nominating and dissecting the functions of TARVs across three large biobanks. Ultimately, the application of variant-centric approaches, including not only MPRA but also predictive models based on large perturbation datasets, like the one introduced here, will help to bridge the crucial gap between variant association and functional mechanisms.

Methods

Fine-mapping

We obtained fine-mapping summary data, including PIPs for each variant and 95% CSs, for 48 traits in the UKBB and BBJ from previous studies^31,66. Detailed methods are available from these studies, but we describe the general approach briefly here. The UKBB is a population-based cohort in the UK, of which 366,194 ‘white British’ individuals were selected for inclusion (https://github.com/Nealelab/UK_Biobank_GWAS). The BBJ is a large non-European hospital-based biobank from Japan, from which 178,726 individuals of Japanese descent were included in this study. We selected 48 traits across 12 domains (including an ‘other’ category) from the UKBB and BBJ (Supplementary Table 1). GWAS was performed on each using a generalized linear mixed model, as implemented in SAIGE⁶⁷ (for binary traits) or BOLT-LMM^68,69 (for quantitative traits), with the following covariates: age, sex, age², age × sex, age² × sex and the top 20 genetic principal components. Fine-mapping was then performed using SuSiE³⁰ with the GWAS summary statistics and in-sample dosage linkage disequilibrium in merged 1.5-Mb windows.

Similarly, we obtained fine-mapping results across 49 tissues from the GTEx v.8 from previous studies^34,66,70 and summarize the approach below. The GTEx project is a tissue-based biobank with genotypes (whole genome sequencing) and gene-expression data (RNA-seq) for 838 individuals across 49 tissues (Supplementary Table 1), which are grouped into 11 systems. We obtained cis-eQTL summary statistics, genotype principal components (PCs) and other associated data as input to fine-mapping, which was performed in a similar way to the fine-mapping of complex traits from the UKBB, with the main modification being that covariates (including genetic PCs) were projected out of the genotypes before the linkage disequilibrium calculation, because GTEx is a more ancestrally heterogeneous cohort³⁹. When collapsing PIPs across multiple tests into a single estimate, the maximum PIP across traits and/or tissues is used. As well as individual trait CSs, merged CSs across traits or across tissues were obtained, as previously described³¹.

Variant selection and oligonucleotide design

We designed 200-bp oligonucleotides centred around the reference and alternative alleles of fine-mapped eQTLs from GTEx v.8 and fine-mapped complex traits from the UKBB and BBJ cohorts^71,72 (Fig. 1a). Variants were selected for inclusion if they demonstrated a PIP > 0.5 or greater in any tissue or a PIP > 0.1 in any trait, or if they fell within a 95% CS with fewer than 25 variants (eQTLs), 30 variants (quantitative traits) or 75 variants (binary traits). Variants with PIP > 0.1 in any of 3 GTEx tissues approximately corresponding to cell types for MPRA were also included. That is, we selected additional fine-mapped eQTLs from liver (matched to HepG2), transverse colon (matched to HCT116) and any brain tissue (matched to SK-N-SH). CSs were considered to be successfully assayed if variants containing at least 80% of the probability had MPRA measurements. Most (92.5%) of the selected variants were single nucleotide polymorphisms. The remaining 22,951 variants were small insertions or deletions (indels), of which 92.2% affected fewer than 10 bp, and only 11 involved more than 50 bp. In total, we included 148,805 eQTL test variants and 78,238 complex trait test variants (Fig. 1b and Supplementary Tables 1–3), with 5,631 variants shared across the two sets. All test and control variants (see below) were split into 8 libraries (Supplementary Table 29).

To test for potential non-additive effects between trait-associated variants in the same element, we selected 2,522 pairs of variants that were within 150 bp of each other, were fine-mapped with PIP > 0.1 for at least one trait, and overlapped with an endogenous CRE in at least one cell type (Supplementary Table 18). To maximize the power to detect non-additive effects, oligonucleotides were designed for up to six possible windows. Centring on each variant in the pair, the variant was placed at approximately 50 bp, 100 bp or 150 bp (the exact location could change for variants causing insertions or deletions), and windows in which the second variant was at least 10 bp from the edge of the oligonucleotide were included (Fig. 3a). For each window, we designed oligonucleotides for all four possible diplotypes (Ref/Ref, Ref/Alt, Alt/Ref and Alt/Alt).

As well as trait-associated test variants, we designed several comprehensive sets of controls. We selected 32,534 ‘location-matched’ controls that were more than 150 bp but less than 500 bp away from test variants and were not significantly associated with a complex trait (P > 0.0001) or change in gene expression (P > 0.001) in our summary data. Next, we selected 39,708 ‘annotation-matched’ controls that were not in 95% CSs or significantly associated with a trait that were matched for MAF, linkage disequilibrium scores, standard genomic annotations (promoter, 3′ UTR, coding, 5′ UTR, intron, CRE, evolutionary constraint) as well as CRE annotations from relevant ENCODE cell types and tissues. Matching was done by computing propensity scores and selecting the closest score match using the MatchIt R package. An example of this matching procedure is shown in Extended Data Fig. 1a. Finally, we selected 14,710 ‘null’ trait controls, for which we simulated a typical GWAS phenotype and performed fine-mapping and variant selection in the same manner as was done for complex traits for UKBB and BBJ. To simulate these phenotypes⁷³, we drew causal effects from variants present in UKBB or BBJ, largely following a previous approach. That is, we assumed 20% trait heritability, either 1,000 (for binary traits) or 5,000 (for quantitative traits) total causal variants, and a MAF-dependent per-allele effect size consistent with previously reported complex-trait architecture⁷³ (α = −0.38). Unlike annotation- and location-matched controls, null control variants were not expected to be enriched at CREs. Instead, this set of controls was designed to reflect the linkage disequilibrium patterns of real complex traits and provide a true null set from the perspective of genomic annotation. For all three classes of control, matching was done across PIP bins, resulting in 37,437 controls for eQTLs and 48,880 for complex traits (Fig. 1b).

We also selected 128 elements containing variants from the original screen that overlap endogenous CREs on which to perform saturation mutagenesis. This set contained a total of 136 emVars, where 114 elements each contain 1 emVar, 8 elements contain 2 emVars, and 6 elements contain 1 emVar and 1 non-emVar (Supplementary Table 23). Three categories of variants were prioritized for saturation mutagenesis. First, we prioritized 83 emVars with ‘canonical regulatory mechanisms’, in which a fine-mapped emVar was predicted to disrupt the known motif at an occupied TF site (that is, the variant disrupts a TF motif and that variant resides within a ChIP–seq peak for that TF). Second, we prioritized 53 emVars with ‘non-canonical or unclear regulatory mechanisms’, in which a fine-mapped emVar did not disrupt the known motif of a TF with evidence of occupancy from ChIP–seq. Candidate variants were manually inspected and selected variants had at least one alternative measure of function, such as residing in a DHS footprint, lying in the flank (0–10 bp) of a motif at an occupied TF site, or having strong predicted allelic effects in sequence-based models of function, such as Enformer. Third, we prioritized 14 ‘same CRE variant pairs’, in which at least one variant in the pair was an emVar and the pair demonstrated a non-additive effect, or both variants were strong candidates for the first two categories. Unless otherwise indicated, the 22 emVars selected as variant pairs were also included in their respective canonical and non-canonical categories.

Overall, variants were chosen to survey a variety of complex and molecular traits, a range of PIPs to allow for both confirmation (PIP > 0.9) and exploration (PIP > 0.1 and PIP < 0.9), a variety of predicted effects on different TFs and in different cell types, and sufficient representation in all three categories. In total, we selected 83 variants from category 1 (canonical mechanism), 53 variants from category 2 (non-canonical or unclear regulatory mechanisms) and 14 variant pairs from category 3 (same CRE variant pairs). For each of the 114 single-variant elements, we designed 200-bp oligonucleotides using both the reference and the alternative background sequences, as well as all possible single-nucleotide substitutions for both backgrounds (3 alternative nucleotides × 200 positions × 2 backgrounds). For each of the 14 variant pairs, we designed mutagenesis oligonucleotides using each diplotype (Ref/Ref, Ref/Alt, Alt/Ref and Alt/Alt) as a background (3 alternative nucleotides × 200 positions × 4 backgrounds).

Finally, we also included technical controls to evaluate the quality of individual MPRA experiments. For each library, we included 91 activity controls, 96 emVar controls and 506 negative controls selected from previously published experiments^16,74 (Supplementary Table 30). Activity controls were chosen for their ubiquitous activity across multiple cell types and were selected to represent a broad range of enhancer activities. For the saturation-mutagenesis library, we included an additional 2,126 negative controls, selecting 1,876 ORF controls and 250 shuffled sequences from previously published work^75,76.

MPRA library construction

The MPRA libraries were constructed as previously described¹⁶ with several modifications to the original protocol. The 230-bp oligonucleotides were synthesized using Agilent Technologies HiFi libraries for complex traits (seven 60 K libraries) and eQTL (two 244 K libraries) or Twist Biosciences for saturation mutagenesis (one 300 K library) with the designed 200-bp oligonucleotide in the middle and 15-bp adaptor sequences on either end. Unique 20-bp barcodes were added by PCR using primers #82 and #202 (Supplementary Table 31). Each oligonucleotide amplification consisted of 16–32 50-μl reactions containing 25 μl Q5 2xMM Ultra II, 2.5 μl each of 10 μM primer and 0.5 μl oligonucleotide library using the following PCR conditions (98 °C for 30 s; 6× [98 °C for 10 s; 60 °C for 15 s; 65 °C for 45 s]; 72 °C for 5 min). The PCR products were purified using AMPure XP SPRI (Beckman Coulter, A63881) and incorporated into the SfiI digested pMPRAv3:∆luc:∆xbaI (Addgene #109035) backbone vector by Gibson assembly (50 μl 2 × NEB HiFi Assembly MM, 2.2 μg DNA oligonucleotide pool, 2.0 μg SfiI digested pMPRAv3:∆luc:∆xbaI in a 100-μl reaction incubated at 50 °C for 1 h). To achieve a library composition of 200–300 barcodes per oligonucleotide, a test transformation was performed using 1 μl of the 20 μl SPRI purified Gibson assembly mixture and 50 μl of 10-beta electrocompetent cells (NEB) to determine the optimal amount of Gibson assembly and 10-beta cells to achieve the desired colony-forming unit (CFU) count. Based on the results of the test transformation, a subset of a second identical transformation was taken directly after electroporation and split across ten 1-ml cultures with 10-beta recovery media (NEB), then incubated for 1 h at 37 °C. After 1 h, each culture was independently expanded in 20 ml of Luria broth (LB) with 100 μg ml⁻¹ carbenicillin. In parallel, CFU colony counting plates were created from four of the ten cultures. After 6.5 h of growth at 37 °C, the cultures were individually pelleted and frozen, and the desired number of cultures from ten expanded cultures was selected, based on the CFU counting plates, to reach an average of 200–300 CFUs (barcodes) per oligonucleotide (15 million to 21 million CFUs for complex traits, 49 million CFUs for eQTLs and 86 million for saturation mutagenesis). Cultures were combined and purified using a Qiagen Plasmid Plus Maxi or Midi kit. For each library, 16 colonies from the CFU plates were checked by colony PCR to determine the oligonucleotide insertion rate for the test transformation (insertion-rate range: 93–100%). The expanded purified pMPRAv3:∆GFP plasmid library was sequenced using Illumina 2 x 150 bp chemistry to acquire oligonucleotide–barcode pairings. To construct the final MPRA libraries, 10 μg of the pMPRAv3:∆GFP library was digested with AsiSI, and a GFP amplicon with a minimal TATA promoter (amplified from pMPRAv3:minP-GFP, Addgene 109035) was inserted using Gibson assembly (125 μl 2 × NEB HiFi Assembly MM, 5.28 μg GFP amplicon, 1.6 μg pMPRAv3:∆GFP in a 250-μl reaction incubated at 50 °C for 1.5 h). The resulting pMPRAv3 library includes the 200-bp oligonucleotide sequence positioned directly upstream of the minimal promoter and the 20-bp barcode falling in the 3′ UTR of GFP. The Gibson reaction was purified using 1.5× SPRI and eluted in 40 μl water before electroporation of 2–16 μl of purified plasmid library into 100–400 μl 10-beta cells. The electroporation was split across six 2-ml cultures and recovered at 37 °C for 1 h followed by expansion of each 2-ml culture in 500 ml of TB media with 100 μg ml⁻¹ carbenicillin. After 16 h of growth at 30 °C, plasmid was purified using Qiagen Plasmid Plus Giga kits. The CFU counting plates from 4 of the 500-ml cultures were made to monitor transformation efficiencies (23 million–185 million CFUs for complex traits, 285 million–1 billion CFUs for eQTLs, and 123 million–171 million CFUs for saturation mutagenesis). Then, 16 colonies from the CFU plates were checked by colony PCR to determine the GFP insertion rate for the test transformation, with all libraries having a minimum of 80% of plasmids containing a GFP insert.

MPRA library transfections

We used five cell types for MPRA transfections, representing five distinct tissue types. The complex trait libraries were tested in K562, SK-N-SH, HepG2 and A549 cells; the eQTL libraries were tested in K562, SK-N-SH, HepG2 and HCT116 cells; and the saturation-mutagenesis library was tested in K562 and HepG2 cells. All cell lines were acquired from ATCC and routinely tested for mycoplasma and other common contaminants by The Jackson Laboratory’s Molecular Diagnostic Laboratory. Each replicate experiment was expanded from a low-passage cryopreserved aliquot followed by consistent handling for each cell type across all libraries. Specifically, K562 cells were grown in RPMI (Life Technologies, 61870127) supplemented with 10% FBS (Life Technologies, 26140) maintaining a cell density of 0.5–1.0 million cells per ml; SK-N-SH cells were grown in DMEM (Life Technologies, 10566-024) supplemented with 10% FBS (Life Technologies, 26140); HepG2 cells were grown in DMEM (Life Technologies, 10566-024) supplemented with 10% FBS (Life Technologies, 26140); A549 cells were grown in Ham’s F-12K (Life Technologies, 21-127-030) supplemented with 10% FBS (Life Technologies, 26140); and HCT116 cells were grown in McCoy’s 5a (ThermoFisher, 16600108) supplemented with 10% FBS (Life Technologies, 26140). All transfections were done using a Neon transfection system (Life Technologies), transfecting 150 million (complex traits), 500 million (eQTL) or 700 million (saturation mutagenesis) cells for each replicate. Transfections were done using 100-μl tips with optimized settings for each cell type: K562, 10 million cells per 100 μl, 5 μg of plasmid, 3 pulses of 1,450 V for 10 ms; SK-N-SH, 10 million cells per 100 μl, 10 μg of plasmid, 3 pulses of 1,200 V for 20 ms; HepG2, 10 million cells per 100 μl, 5 μg of plasmid for complex traits and saturation mutagenesis or 10 μg of plasmid for eQTLs, 1 pulse of 1,200 V for 50 ms; A549, 10 million cells per 100 μl, 5 μg of plasmid, 2 pulses of 1,200 V for 30 ms; and HCT116, 10 million cells per 100 μl, 10 μg of plasmid, 2 pulses of 1,350 V for 20 ms. We did a minimum of 5 replicates for each library and collected each replicate 24 h after transfection by rinsing three times with PBS and collecting by centrifugation. Cell pellets were either frozen at −80 °C for later processing or homogenized immediately in RLT buffer (RNeasy Maxi kit) and dithiothreitol, before freezing at −80 °C. For each cell type and library, no more than two replicates were done on the same day and with parallel replicates processed from independently expanded batches of cells.

RNA isolation and MPRA RNA library generation

Total RNA was extracted from the cell homogenates using the Qiagen RNeasy Maxi kit followed by a 1-h Turbo DNase treatment (ThermoFisher) with 5 μl of DNase in 1,475 μl total volume for 1 h at 37 °C. To stop the DNase digestion, 15 µl of 10% SDS and 150 µl of 0.5 M EDTA were added, and the sample was incubated at 70 °C for 5 min. GFP mRNA was then pulled down using a mixture of three GFP-specific biotinylated primers at 0.5 nM (#120, #123 and #126; Supplementary Table 31) in 0.2× SSC and 33% formamide (MilliporeSigma 4650–500 ML) and incubated for 2.5 h at 65 °C. Biotin probes hybridized with GFP mRNA were then captured by adding 400 μl of pre-washed Sera Mag beads (Fisher Scientific) eluted in 500 μl of 20× SSX by agitation at room temperature for 15 min. The beads were captured on a magnet and washed once with 1× SSX and twice with 0.1× SSC. After removal of the last wash, 50 μl of water was added and the RNA was treated overnight with Turbo DNase at 37 °C. The beads were then collected by magnet and the supernatant removed and purified with 2× RNA SPRI. Complementary DNA was created using SuperScript III (Life Technologies) in a 100-μl reaction with a gene-specific primer for the GFP transcript (20 μM primer #19; Supplementary Table 30) and a modified elongation temperature (47 °C for 80 min). The GFP mRNA abundance was quantified by quantitative PCR to determine the cycle threshold for each replicate using a QuantStudio 5 qPCR instrument (Reaction mix: 5 μl Q5 Ultra II 2x, 0.5 μl of 10 μM primers #801 and #802, 1.66 μl 1:10,000 Sybrgreen 1, 1 μl cDNA or GFP mRNA in a 10 μl reaction, Cycle conditions: 98 °C for 30 s; 40× [98 °C for 10 s; 62 °C for 15 s; 72 °C for 30 s]; Supplementary Table 31). A standard curve with the final MPRA library (10 fg–1 ng) was run alongside 1 μl of cDNA and 1 μl of GFP mRNA for each replicate. Any samples showing amplification in the mRNA sample were discarded for having plasmid carry-over. Replicates within a cell type and library were diluted to approximately the same concentration, based on the quantitative PCR (qPCR) results, and first-round PCR (98 °C for 30 s; 6–13× [98 °C for 10 s; 62 °C for 15 s; 72 °C for 30 s]; 72 °C for 2 min) with primers #801 and #802 (Supplementary Table 31) were used to amplify barcodes associated with GFP cDNA sequences for each replicate. A second round of PCR (98 °C for 30 s; 6× [98 °C for 10 s; 68 °C for 15 s; 72 °C for 30 s]; 72 °C for 2 min) was used to add Illumina sequencing adaptors. The resulting MPRA barcode libraries were spiked with 0.01–1% PhiX and sequenced on an Illumina NextSeq 500 or NovaSeq SP.

MPRA data processing and main analysis

Data from MPRA were analysed as previously described¹⁶ using MPRAsuite, which includes MPRAmatch (barcode and oligonucleotide pairing), MPRAcount (tag assignment and counting) and MPRAmodel (activity and allelic effect estimates) (https://github.com/tewhey-lab/MPRASuite). In brief, oligonucleotide–barcode pairings were identified by merging paired-end reads into single amplicons using Flash (2.2.00)⁷⁷. Genomic DNA sequences and barcodes were extracted and the genomic DNA element was mapped back to the oligonucleotide design file using minimap2 (v.2.17-r941)⁷⁸. Alignments showing more than 5% error to the design files were discarded, leaving only high-quality alignments, which were compiled in a look-up table with their matching barcodes. After sequencing of the RNA and plasmid libraries, 20-bp barcode tags were assigned to oligonucleotides from the look-up table and oligonucleotide counts were aggregated by addition across all barcodes. Variants with fewer than 30 DNA counts or with exactly 0 RNA counts in oligonucleotides for either allele were excluded from all downstream analysis. In the diplotype analyses, a stricter filter was used, and variants with fewer than 100 DNA counts or fewer than 10 RNA counts across all 4 diplotypes were excluded. When variants were assayed in multiple libraries, the library with the highest sum of the square roots of plasmid counts in reference and alternative sequences was selected.

MPRAmodel uses DESeq2 (1.42.1) as the framework for normalization and statistical analysis. For each library, counts were normalized across samples using a ‘summit-shift’ approach, which centres the distribution of RNA/DNA ratios for each sample at 1. To determine significant differences between DNA plasmid count and RNA count, a negative binomial generalized linear model (GLM) was used in DESeq2 with independent dispersion estimates for each cell type and library. We estimated element activity in units of log₂(fold change) for RNA counts compared with DNA counts by including a contrast in the design matrix of DESeq2 to compare treatment (RNA versus DNA) and used Wald’s test to evaluate the significance of this estimate, correcting for multiple hypothesis testing in each cell type and library using Bonferroni’s method. We estimated an allelic effect in units of ∆ log₂(fold change) (or ∆log₂ activity) for the difference in RNA counts relative to DNA counts at the Alt oligonucleotide compared with the Ref oligonucleotide by including a contrast in the design matrix of DESeq2 to compare alleles (Alt versus Ref) and treatment (RNA versus DNA) and used Wald’s test to evaluate significance followed by FDR estimation using Benjamini and Hochberg’s method. For simplicity, we refer to P values obtained after adjusting for the FDR as ‘FDR’ rather than ‘q-value’ or ‘adjusted P-value’ in all relevant instances. Elements passing a specific effect size and significance threshold for RNA versus DNA are defined as active, and variants in active elements passing a specific significance threshold are defined as expression-modulating variants or emVars. By performing a grid search as described in the following section, we define elements as active if |log₂[fold change]| > 1 and Bonferroni-adjusted P value < 0.01, and variants as emVars if either the reference or alternative allele element is active and FDR < 10% for a non-zero allelic effect. For specific analyses requiring cell-type agnostic estimates, we performed fixed-effect meta-analyses of activity and allelic effects across cell types and used these estimates. In Supplementary Table 4, we report the best measurement across libraries (by plasmid coverage across variants) chosen per variant per cell type.

For saturation mutagenesis, we used the saturation-mutagenesis mode of MPRAmatch, which applies increased stringency requirements when pairing barcodes to oligonucleotide sequences requiring perfect alignment matches during barcode pairing. Because most of the saturation-mutagenesis library consisted of sequences with strong MPRA activity by design, we used a summit-shift normalization procedure where linear scaling factors were determined only from negative control sequences. Because the primary goal of the saturation-mutagenesis experiment was to compare the effects of substitutions across an element, rather than to detect individual emVars, we used an empirical Bayes approach to shrink the allelic effect sizes towards 0, based on how accurately they were measured (the number of RNA and DNA reads) (Supplementary Fig. 8c,d). First, we obtained an improved estimate of the baseline activity of the background sequence by pooling the log₂(fold change) estimates of oligonucleotides between the observed 25th and 75th percentiles of substitutions across this element, assuming that approximately 50% of substitutions will have little effect and the mean activity of these substitutions is an unbiased estimate of true baseline activity. Taking the mean and variance of this set as a normal prior, we used DESeq2 estimates of activity (log₂[fold change]) and their corresponding standard errors to define a normal likelihood, allowing for closed-form posterior inference, owing to conjugacy. We used this prior and likelihood to obtain a posterior distribution for baseline activity of the background sequence. Next, we used the mean and variance of activity estimates between the 5th and 95th percentiles of substitutions to form a less-stringent prior with higher variance, better reflecting the distribution of substitutions with true allelic effects. Using a similar Bayesian approach with this prior, we estimated the posterior activity distribution of each substitution. To obtain shrunken estimates of each allelic effect, we subtracted the posterior mean activity for each substitution from the posterior mean baseline activity for the background sequence.

Grid search for activity and allelic effect thresholds

To find the optimal thresholds for activity and for allelic effect, we did a grid search across Bonferroni-adjusted P value or FDR and effect magnitude, aiming to maximize precision and recall in a set with known positives and negatives. For activity, we used overlap with a CRE to approximate our positive set, and assigned all other oligonucleotides to the negative set, reasoning that sequences originating from accessible chromatin marked by H3K27ac should approximate active transcriptional regulatory elements. We selected sets of variants in CREs and not in CREs, and calculated the precision and recall across a grid of Bonferroni-adjusted P value and activity (log₂[fold change]) thresholds (Supplementary Fig. 3a,b). For activity grid searches, to compare cut-offs using different annotations to separate positive and negative controls, we did not scale the sets 1:1. Setting |log2[fold change]| > 1 and Bonferroni-adjusted P value < 0.01 optimized the precision–recall trade-off and is used as the definition of active throughout the manuscript. For allelic effects, we repeated our grid search, using sets of high PIP (more than 0.9) and low PIP (less than 0.02) variants as positive and negative sets. We scaled the larger, negative sets down to the size of the positive set using the false-positive rate. First requiring all elements to be active, we calculated precision and recall across a grid of FDR thresholds and allelic effect (∆log₂[fold change]) thresholds (Supplementary Fig. 4a,b). Based on these results, we defined emVars as variants residing in active elements with any magnitude of allelic effect and an FDR < 10% for a non-zero allelic effect.

Non-additive effect estimation and diplotype analyses

Because alternative and reference alleles for trait-associated variants are arbitrarily coded based on the reference genome, rather strictly by effect direction, disease risk or minor allele, we recoded the alleles to best aid the interpretation of MPRA effects. Using MPRA estimates of activity for each of the four diplotypes, we recoded alleles in order from lowest to highest activity, with lowest activity as the reference category. We set the reference category as the ‘decrease–decrease’ pair and the highest value as the ‘increase–increase’ pair, with ‘decrease-increase’ and ‘increase-decrease’ pairs in between. The additive allelic effect is defined as the sum of the individual allelic effect (decrease–increase + increase–decrease). The observed double-allele effect is the allelic effect for the increase–increase pair compared with the decrease–decrease pair. Interaction effects are the difference between the expected additive allelic effect and the double-allele effect, which are estimated in the DEseq2 model and recoded here. Interacting pairs of variants were detected as described below and classified into amplifying or dampening pairs (Fig. 3b,d, Extended Data Fig. 5i and Supplementary Tables 20 and 21).

To gain power to detect interaction effects between variant pairs, we performed several meta-analyses. First, we did fixed-effect meta-analyses to obtain both activity and allelic-effect estimates for each pair. Meta-analyses were done by aggregating across available windows (within a single cell type) or by aggregating across both windows and cell types. In these meta-analyses, activity and emVar definitions were the same as for the meta-analyses of single-variant elements but with the consideration of additional alleles; an element was considered active if any of the four elements (versus two elements for single-variant) were active. Interaction effects were also meta-analysed across windows within a single cell type and windows across all cell types using both a standard covariance matrix and an empirical covariance matrix.

For the meta-analyses with a standard covariance, the observed variance of each interaction-effect estimate was placed on the diagonal and zeros were placed elsewhere, encoding the hypothesis that windows are independent. We reasoned that windows might not be independent, and if we could capture this relationship in the covariance matrix, we would increase the power to detect significant interaction effects in meta-analyses. To estimate the relationship between windows, of which each variant pair could be tested in up to six (centring on variant 1 or variant 2; windows around the centred variant shifted left, middle or right), we calculated the expected correlation between these effects based on two factors that could influence these effects and how they covary: how much of the 200-bp oligonucleotide was shared, and the distance between the variants in the pair. Specifically, we split the haplotype-interaction effect measurements into a 7 × 9 grid of 7 bins of shared oligonucleotide percentage (0%, 25%, 50%, 60%, 70%, 80%, 95% and 100%) and 9 bins of variant distance in nucleotides (0, 10, 15, 20, 25, 50, 75, 100 and 200). In each bin, we then computed the correlation between observed interaction effects to obtain an empirical correlation matrix, and used a generalized additive model to smooth the correlation matrix across the grid (Extended Data Fig. 5g). The covariance matrix was obtained by scaling the correlation matrix by the variance vector, and fixed-effect meta-analysis was done using this covariance matrix.

Interaction emVars were defined as pairs in which at least one diplotype was an emVar (versus the Ref/Ref diplotype) and the FDR was less than 0.05 and the absolute value of the interaction effect was greater than 0.25 for one of the 4 meta-analyses (standard versus empirical covariance, across windows in cell types versus across windows and cell types). In the saturation-mutagenesis dataset, interaction-effect estimates were similarly obtained by subtraction of the double allele and additive variant posterior activity estimates, and interaction emVars were defined as having an interaction effect greater than 0.415 (a 25% change in activity). Analyses of the distance between pairs of variants was restricted to pairs in which variants were tested in all three windows (left, middle and right) because the number of windows a variant pair is tested in is dependent on the distance between the variants (Fig. 3c).

Multivariate adaptive shrinkage

To account for differences in power when estimating whether element activity or allelic effects are shared across cell types, we used the Multivariate Adaptive Shrinkage in R (mashr) package⁷⁹. Because MPRA was done in four cell types for complex trait variants a four cell types for molecular trait variants, we used only the canonical covariance matrices in the mixture model. We selected variants with PIP > 0.1 that are in CREs and fit the mash model separately for complex trait variants and eQTLs separately. Cell-type-specific effects were determined using the local false sign rate (lfsr) < 0.05 (Supplementary Table 7).

Coding and non-coding genomic annotations

Genomic annotations were obtained as described previously^31,66. In brief, we ran the Ensembl Variant Effect Predictor v.85 (ref. ⁸⁰), selecting the most severe consequence for each variant including coding (missense, loss of function and splice site) and UTR annotations. Non-coding variants were annotated as residing within a promoter (using annotations from the S-LDSC baseline model), a CRE or having evidence of neither. CREs were defined as accessible chromatin in at least one cell type and evidence of histone modification at H3K27ac in at least one cell type. We used the combined, normalized and quality-controlled chromatin accessibility and histone modification datasets described in ref. ⁶⁶, including DNase-seq, ATAC-seq, single-cell ATAC-seq and H3K27ac measurements from seven different atlases^{9,40,41,81,82,83,84,85}. This combined dataset consists of more than 1,000 cell types: ROADMAP Epigenomics DNase-seq across 39 cell types and H3K27ac across 98 cell types⁸¹; DNase-seq across 438 cell types⁹; single-cell ATAC-seq across 54 cell types⁸²; ATAC-seq data for 18 haematopoietic cell types⁸³; single-cell ATAC-seq across 24 brain cell types⁸⁴; ATAC-seq for 25 immune cell types⁸⁵; and ChIP-Atlas DNase-seq for 284 cell types and H3K27ac for 720 cell-types⁴¹ to identify CREs. Non-coding CSs were defined as 95% CSs where no variant was annotated as a missense, loss of function or splice-site variant.

For analyses comparing the strength of endogenous CRE activity with MPRA activity (Fig. 1d), only CREs from the dataset in ref. ⁹ are used and cell-type agnostic CRE activity is estimated as the maximum normalized DNase I hypersensitivity counts across cell types. To identify cell-type-specific CREs, we computed the euclidean norm of normalized counts from the dataset in ref. ⁹, selected the maximum value across biosamples representing similar cell types, and selected CREs with a euclidean norm greater than 0.2. Furthermore, we obtained representative tissue and organ system annotations provided in ref. ⁹. In CREs from this dataset⁹, variants were annotated for consensus DHS footprints across 243 cell types, which were downloaded from ref. ⁴⁷ and lifted over to hg19.

Precision and recall of MPRA

We calculated the precision and recall of different annotations (such as emVars) for separating trait-associated variants from background variants. We approximated the true causal (positive) set of variants by using non-coding, trait-associated variants with PIP > 0.9. Background non-coding variant (negative) sets were defined using either location-matched controls, annotation-matched controls or trait-associated variants with PIP < 0.01. For the eQTL dataset, variants with PIPs up to 0.02 were included as low-PIP to have sufficient matches. We scaled the larger, negative sets down to the size of the positive set using the false positive rate to create a 1:1 set. Precision is defined as the number of true positives divided by the number of variants selected as positive by the annotation. Recall is defined as the number of true positives divided by all positives. All three control sets produced similar estimates, and results using low-PIP variants as controls have the lowest standard errors and are used in the main analysis shown in Fig. 2e. As well as evaluating MPRA measurements and genomic annotations, we estimated precision and recall across CS size (Extended Data Fig. 4e and Supplementary Table 11) and specific cell types used in MPRA or for defining CREs (Extended Data Fig. 4f,g and Supplementary Table 10). For cell-type analyses, meta points were estimated as averages of cell-type combination, and linear regression was used to predict the impact of each additional cell type on precision and recall.

Comparison with genome-wide reporter assays

Measurements of allelic effects using the Survey of Regulatory Elements (SuRE) reporter assay for 5.9 million variants in K562 and HepG2 cells were downloaded from ref. ⁴⁴. To compare performance between the Survey of Regulatory Elements and MPRA, we again estimated precision and recall using PIP > 0.9 variants as positives and PIP < 0.05 variants as negatives, but now restricted to variants assayed by both approaches.

Regulatory quantitative trait loci

Chromatin accessibility quantitative trait loci and allele-specific TF occupancy quantitative trait loci were obtained from ENCODE^9,86. Variants with an FDR < 0.25 for allele-specific differences were included in comparisons with MPRA allelic effects.

TF binding motifs and occupancy

Positional motif matches on oligonucleotides were identified using a modified version of FIMO (MEME Suite 5.5.5)⁸⁷ implemented in motif liquidator⁸⁸ for PWMs from the Homo Sapiens Comprehensive Model Collection (HOCOMOCO, H12CORE⁸⁹) and JASPAR⁹⁰ (2022 release). Motifs shorter than 8 bp or with an information content of less than 12 were excluded. Sequence matches were called when the absolute percentage (scaled by the difference between the best and worst possible sequence matches scores) and the relative percentage (scaled by the difference between the best and random sequence matches scores) of each sequence was greater than 12, and the significance for this match was less than P = 0.0001. Disruption of a motif required a difference of more than 10% in the absolute percentage of each sequence match to the PWM. Weak calls with disruptive alleles required only an information content score of more than 8, no relative percentage filter, P < 0.001 and an allelic difference of only 5% in the absolute percentage of each sequence. Motifs were defined as flanking variants if the motif started or ended within 10 bp of the variant (Fig. 2f).

TF occupancy for tested elements was determined by ChIP–seq support. We obtained ChIP–seq peaks from ChIP-Atlas, which uniformly reprocessed peaks from 25,823 experiments in 1,046 human cell types or tissues for 1,741 TFs^40,41. TFs were retained in analyses if they met a minimum number of overlaps with variants or elements (as specified in the text). For each specific motif match or disruption, a corresponding TF was defined as occupying an element if that element overlapped with a TF peak for any TF that matched a highly similar motif (Pearson r > 0.9 for PWM similarity). For example, 1,513 TFs occupied at least 20 elements, and we observed that, of these, 1,355 were more likely to overlap active elements at a Bonferroni-adjusted P < 0.05 (Extended Data Fig. 2b). Furthermore, 54 ‘dark TFs’, or TFs that bind to closed chromatin, were obtained from ref. ⁹¹.

Motif contributions to the activity of elements derived from CREs containing trait-associated variants were estimated using linear regression, modelling the marginal effect of the number of motifs in each element on the activity in K562, HepG2 and SK-N-SH cells. Before fitting the linear regression, activity was quantile normalized across cell types and a generalized additive model using smoothing splines of 1-mer and 2-mer counts in each element to model activity (each cell-type measurement was an independent observation) was fit using the mgcv R package, and residual activity was used as the outcome for linear regression. P values for each motif were obtained from a Wald test and FDRs were estimated using Storey’s q value. Coefficients from these regressions for motifs present in more than 0.1% of elements are shown in Fig. 1e, although these estimates should not be considered causal effects. We also assessed whether changing the extent of PWM disruption for each motif was correlated with the extent of change in activity in each cell type. Restricting to variants that are emVars in at least one cell type and motifs represented by at least 10 variants, we computed Spearman’s correlations and estimated FDRs using Storey’s q value.

Regulatory sequence classes

We obtained regulatory sequence (Sei) classifications, including polycomb and heterochromatin annotations, and lifted the annotations over to hg19 (ref. ⁴³). Variants were excluded if they fell into multiple Sei annotations after liftover (less than 1%). The proportion of emVars for each complex trait, stratified by domain, is estimated and shown in Supplementary Fig. 6a.

Enformer

To capture more regulatory mechanisms than those mediated through TF motifs catalogued in JASPAR and HOCOMOCO, we scored each variant (after lifting over to hg38) using Enformer⁴⁸, a transformer-based neural network model trained to predict functional genomic data, including TF occupancy and chromatin accessibility. Allelic-effect predictions across 1,447 distinct tissue–TF pairs and 320 accessible chromatin predictions were z-score normalized to variants with non-emVars in CREs with PIP < 0.02. A combined score for each variant was computed by squaring its allelic effect for individual TF occupancy or accessible chromatin predictions and taking the sum. Thresholds were set separately for TF occupancy and accessible chromatin, each determined such that only 5% of non-emVars in CREs with PIP < 0.02 had higher scores. Thresholds for individual Enformer predictions for all variants were similarly obtained. A variant was considered disrupted by Enformer if it exceeded the threshold for either the TF occupancy or chromatin accessibility. Using these thresholds, 7.3% of non-emVars in CREs with PIP < 0.02 were classified as disrupted (Fig. 2f).

Comparison of eQTL and MPRA effect sizes

The MPRA allelic effects were compared with eQTL allelic effects by computing their Spearman’s correlation coefficients. Variants were restricted to emVars with PIP > 0.5 for at least one gene, and MPRA effect sizes were obtained from the meta-analyses across cell types. The eQTL effect sizes were obtained from the most significantly associated gene and tissue for each variant. To ensure a clear distinction between proximal and distal regulatory elements, distal CREs were further required to be more than 10 kb from the TSS of the associated gene, and proximal CREs were required to be less than 2 kb from the TSS of the associated gene. To compare agreement in sign, a binomial test that the proportion of variants agreed in sign more than expected (0.50) was performed (P_omnibus = 4.3 × 10⁻³⁰). To compare agreement in magnitude, the absolute value of each effect-size measurement was taken before computing correlations (P_omnibus = 7.1 × 10⁻¹⁵). To control for background at distal CREs, correlations were computed for a set of variants that differed only in CRE status (not in a CRE), and the obtained correlation was subtracted off to obtain the adjusted estimate.

Multiple causal variants

We designed a hypothesis test to evaluate whether our data provided evidence of multiple causal variants in independent signals of genetic association (the same CS). First, we restricted our analysis to non-coding 95% CSs where we obtained sufficient MPRA measurements on variants harbouring most of the CS probability (more than 90%). Then, we removed CSs without evidence of the functional annotation being tested (CRE emVars, CRE or emVars). We counted the total excess number of variants in each remaining CS (containing at least one variant with a functional annotation) and compared this with the background rate estimated from controls (location matched, annotation matched or low PIP), computing both risk ratios and risk differences. Because the emVar call rate has both biological and technical variance across experiments, we designed each library such that individual traits were fully contained in a single library along with all three types of matching control. Thus, risk ratios and risk differences were computed in each experiment and meta-analysed using a random-effects model implemented in the R package metafor. Meta-analysis was performed separately for complex traits and eQTLs across multiple CS sizes (5, 10 and 15) and r² thresholds (0.8, 0.9 and 0.99).

Single-TARV CS selection

Across all traits, we identified 95% CSs with only one TARV that lacked other ‘interesting’ variants (Supplementary Table 15). We reasoned that the TARV in these CSs were enriched for being the likely causal variant. To obtain this set, we first excluded CSs containing any coding variants, as well as those consisting of only a single variant. We then filtered to CSs containing a single TARV and removed CSs when any non-TARVs in that CS were not evolutionarily constrained, not in a 3′ or 5′ UTR. As an extra precaution, we also excluded CSs containing variants in promoters (not annotated as CREs) and variants that were emVars.

CRE-to-gene annotations

To better contextualize the variant effects identified in our MPRA, we annotated all autosomal TARVs using multiple CRE-to-gene prediction datasets:

PCHi-C in 27 diverse cell types⁹²
PCHi-C in 17 primary haematopoietic cell types⁹³
CRE–promoter correlation in diverse 127 cell types⁹⁴
CRE–promoter correlation in diverse 808 cell types⁹⁵
CRE–promoter correlation in 16 primary haematopoietic cell types²⁸
E2G models trained on CRISPRi perturbations (E2G)⁹⁶.

TARVs were overlapped with each dataset to generate a compendium of CRE–gene links, as and nearest gene annotations (Supplementary Table 16). These results include the following gene connections for TARVs highlighted in our study:

rs6536864738 (Fig. 4d,e): fine-mapped in UKBB for platelet count (PIP = 1.0). Target genes identified include the nearest gene KDM7A⁹⁶, as well as UBE3C⁹³, PARP12 (ref. ⁹⁶), SSBP1 (ref. ⁹³), MKRN1 (ref. ⁹⁶), RAB19 (ref. ^93,96), SLC37A3 (ref. ^93,96) and KLRG2 (ref. ⁹⁶).

rs11864973 (Fig. 4f,g): fine-mapped in BBJ for red blood cell count (PIP = 0.99). PCHi-C analysis identified 16 putative target genes, including the nearest gene NPRL3. Of these, several encode haemoglobin subunits: HBQ1 (refs. ⁹⁴,⁹⁵), HBA1 (ref. ²⁸), HBA2 (ref. ²⁸) and HBM^28,94,95.

rs2529369 (Fig. 4i,j): fine-mapped in GTEx tibial nerve (PIP = 1.0) and associated with CDHR3 expression. This target is confirmed by PCHi-C⁹³. Other targets include the nearest gene ATXN7L1 (ref. ⁹³) and EFCAB10 (ref. ⁹⁶).

rs191148279 (Fig. 5d–f): fine-mapped in UKBB for HbA1c levels (PIP = 0.919). PCHi-C identifies C2orf88 as a potential target^28,96, which we confirmed through endogenous editing experiments. Other putative targets include SLC40A1 (ref. ²⁸) and INPP1 (ref. ⁹⁶).

rs7282770 and rs7282886 (Fig. 5h,i): located in the same CS in GTEx, associated with DOP1B expression in liver (PIP = 0.08), the loci shows target gene contacts to DOP1B in the PCHi-C and E2G datasets. Other targets include CBR3 (ref. ⁹⁶) and CHAF1B⁹².

rs35081008 and rs34003091 (Extended Data Fig. 8d,e): located in the ZNF329 promoter and associated with ZNF329 expression in liver and LDL cholesterol levels. Both variants contact ZNF329 (refs. ^94,96) and 83 other genes (Supplementary Table 16).

Activity blocks

We developed a Gaussian filter approach to identify blocks of quasi-contiguous non-zero variant positions. Gaussian filters are widely used across fields such as image processing, in which they detect edges by reducing pixel noise; astronomy, in which they enhance features such as stars and galaxies against noisy backgrounds; and audio processing, in which they reduce high-frequency noise while preserving signal integrity. By smoothing data, Gaussian filters reduce noise while preserving localized features and highlighting important patterns in the signal. This approach emphasizes features by weighting nearby positions more heavily and provides a smooth transition between regions of high and low signal. Because of this, they are well suited to identify TF-binding sites and other regulatory sequence segments in which sets of contiguous variant positions all participate in transcriptional regulation.

To identify blocks, we used the crossover points between a threshold T ∈ ℝ and a smoothed signal S = [S_k ∈ ℝ] (at position k) of the ∆log₂(fold change) [Δ_k] of an element as follows. For any given sequence element, each Δ_k was obtained by averaging all three posterior allelic substitution effects at each position k. Then, S was obtained by applying a one-dimensional Gaussian filter (scipy.ndimage.gaussian_filter1d) to [Δ_k] with a kernel s.d. of σ = 1.15. To identify blocks of negative allelic effects, which indicate positive contributions to activity by the reference sequence, we found the crossover indices {c_i ∈ ℕ} between S and T = −0.2, and considered as salient those regions [c_i, c_i+1] such that:

i)
S_k ≤ T S_k ≤ T = −0.2 ∀ k ∈ [c_i, c_i+1]; that is, the signal S along the block is below the threshold;
ii)
c_i+1 − c_i + 1 ≥ 5; the length of the region is at least 5;
iii)
mean(Δ_k, k ∈ [c_i, c_i+1]) ≤ −0.15; the mean of the allelic effects has to be at most 0.15.

Similarly, we took T = 0.2 and mean(Δ_k, k ∈ [c_i, c_i+1]) ≥ 0.15 to identify blocks of positive allelic effects.

Motif matches in activity blocks

To find a TF motif match to a given activity block b = [b_start,b_stop], we constructed a PWM from the allelic effects as follows. Let δ^a_k denote the three allelic effects at position k ∈ [b_start,b_stop] between the reference allele r ∈ {A,C,G,T} and the alternative alleles {A,C,G,T} ∋ a ≠ r. We define a matrix M^t_k, where t ∈ {A,C,G,T} and k ∈ [b_start,b_stop], as M^t_k = δ^a_k if t = a, M^t_k = 0 if t = r. Then, we define PWM[b_start,b_stop] = 20 × M^t_k/M_max for t ∈ {A,C,G,T} and k ∈ [b_start,b_stop], as the PWM induced by the allelic effects in the activity block [b_start,b_stop], where 20 is a chosen temperature parameter and M_max = max_t(∑_k|M^t_k |). Using PWM[b_start,b_stop], we construct the block position probability matrix, PPM[b_start,b_stop], by applying the Softmax function to the values of PWM[b_start,b_stop] at each position. We also define the block information content matrix, ICM[b_start,b_stop], by mapping the values of PPM[b_start,b_stop] at each position into information bits with respect to the uniform background p_A = p_C = p_G = p_T = 0.25. When we had obtained the matrices PWM[b_start,b_stop] and ICM[b_start,b_stop] from the allelic effects in an activity block b, we sought to find their best match to a human TF motif in JASPAR⁹⁰ or HOCOMOCO⁸⁹. Let \({\mathcal{D}}\) represent the dataset of known human TF motifs in the union of JASPAR (2022 REDUNDANT) and HOCOMOCO (H12CORE). Let PWM_m, ICM_m and λ_m represent the position weight matrix, information content matrix and length, respectively, of a motif \(m\in {\mathcal{D}}\). We defined the best alignment offset o(b, m) ∈ ℤ between a block b = [b_start,b_stop] and a motif \(m\in {\mathcal{D}}\) as the integer that maximizes the sum of the element-wise product between ICM[b_start,b_stop] and PWM_m for all possible alignments of the nucleotide positions between the two matrices (with the appropriate zero padding to match the dimensions of the two matrices). Next, we computed the (block) Pearson correlation between ICM[b_start,b_stop] and ICM_m (with the appropriate zero padding) in their best alignment given by o(b, m) for every m in \({\mathcal{D}}\), and collected the motifs in the top 50 Pearson correlations. Then, for those top 50 motifs, we computed the (match) Pearson correlation between ICM_m and ICM[b_start + o(b, m), b_start + o(b, m) + λ_m]; that is, the correlation to the ICM allelic effects given the alignment induced by the alignment to the activity block. Finally, we selected the motif with the highest match_Pearson + 0.2 × block_Pearson. If the overlap between the selected TF match and the activity block leaves five or more contiguous positions in the block unassigned, we repeat the process above to the unassigned sub-block.

Rare-variant analysis

Whole genome sequencing for 402,990 unrelated (max kinship coefficient < 0.884) individuals without filtering on genetic continental ancestry in the UK Biobank was analysed on DNAnexus. Carriers were defined as single-nucleotide substitutions with a higher MPRA allelic effect size than rs191148279 from the saturation-mutagenesis experiments in K562 cells. Residual HbA1c phenotypes were obtained as previously described³¹, and an odds ratio using Firth’s correction for more than 1 s.d. change in the phenotype was computed as well as Levene’s test, which was run to test whether the variances were different between carriers and non-carriers.

GWAS meta-analysis

We selected 121,247 individuals of African continental ancestries from UKBB (n = 5,803), All of Us (n = 7,442), Million Veterans Project (n = 98,427) and Alliance for Genomic Discovery (n = 9,575) with genotype calls at rs785206 and HbA1c level measurements. In brief, linear regression in UKBB and Alliance for Genomic Discovery was done as previously described³¹; linear regression in All of Us was done controlling for age, sex and three genetic PCs; and linear regression results from Million Veterans Project were obtained from ref. ³² (HARE, A1C_Min_INT). Fixed-effect meta-analysis was done using the R metafor package. Bonferroni-adjusted 95% confidence intervals were obtained using the critical value corresponding to P < 5.0 × 10⁻⁸ (5.33).

GTEx interaction analysis

Sex-biased eQTLs were mapped as described in ref. ⁹⁷. In brief, a linear model with a genotype × biological sex interaction term was used to model normalized gene expression. Genotype PCs and expression PEER factors were included as covariates. The interaction P value was obtained from a Wald test.

Constraint

Nucleotide-level PhyloP scores for constraint in 427 mammals were downloaded from Zoonomia⁴. Saturation-mutagenesis sequences were lifted over to hg38 and matched to PhyloP scores. Specific positions were defined as constrained if the PhyloP FDR was less than 0.05, corresponding to a PhyloP score of more than 2.27. Before estimating the correlation with PhyloP, ∆log₂(fold changes) for all three substitutions at each position were averaged and inverted to obtain a contribution score, similar to the height of letters in the saturation-mutagenesis examples in Figs. 4 and 5. Correlations between either the signed contribution score or the |contribution score| and PhyloP using Pearson’s method and corresponding FDRs obtained using Storey’s q value are provided in Supplementary Table 26.

Endogenous base-editing

The K562 cells were cultured in RPMI 1640 medium, GlutaMAX Supplement (Life Technologies, 61870127) with 10% FBS (Life Technologies, 26140). For base-editing, 1 M cells were transfected in a 10-μl volume with 1 μg of a 3:1 mass ratio pool of base-editing plasmid, BE4 (Addgene, 100802)⁹⁸ and sgRNA-expressing plasmid, BPK1520 (Addgene, 65777)⁹⁹, using a Neon transfection system (Life Technologies) (3 pulses of 1,450 V for 10 ms). Two sgRNAs were used to target the centre GATA motif and transfected independently (AGTACTATCTTTGAGGATAG, AAGTACTATCTTTGAGGATA). Transfected cells were grown for 6 days followed by removal of 5 × 10⁵ cells for gDNA extraction and PCR amplification of the target site (rs191148279_gDNA_F, rs191148279_gDNA_R; Supplementary Table 31) using Q5 Polymerase (NEB) to confirm presence of edits by Sanger sequencing. Cells were grown for a further 96 h, then individual cells were isolated by limiting dilution cloning, plating at 0.3 cells per well in 384-well plates to yield single-cell colonies in approximately one-third of wells. Single cells were expanded followed by gDNA extraction on an aliquot of cells using QuickExtract DNA Extraction Solution (Biosearch Technologies, QE09050). DNA was amplified using Q5 Polymerase (NEB) and the same primers for bulk verification followed by Sanger sequencing. Nine clones containing the rs191148279 G>A knockout mutation with a second silent G>A mutation outside GATA and 11 unedited clones with the wild-type allele containing the GATA site were selected for downstream analyses.

Expression profiling of base-edited clones

RNA was extracted from around 10⁶ cells taken from K562 base-edited and unedited clones using Qiagen RNeasy Mini Columns (Qiagen, 74104). BRB-seq libraries were prepared as described previously¹⁰⁰. Specifically, 100 ng (quantified on Qubit) of RNA was incubated with the BU3_index primer (Supplementary Table 31) for 10 min at 65 °C and placed on ice before performing the room-temperature reaction (14 μl RNA + primer mixture, 4 μl 5× First Strand Buffer, 2 μl 0.1 M DTT and 0.25 μl Superscript II Enzyme (ThermoFisher, 18064014)) (42 °C for 50 min, 70 °C for 15 min). Individual samples were then pooled and purified (Monarch Spin PCR & DNA Cleanup Kit, NEB T1130) before eluting in 20 μl water. Samples were then treated with exonuclease I (20 μl cDNA, 1 μl exonuclease I, 2 μl 10× reaction buffer (NEB, MO293), 30 °C for 30 min, 80 °C for 20 min) before second-strand synthesis. Second-strand synthesis was done by nick translation in the following reaction: 23 μl exonuclease-treated cDNA, 2 μl RNase H (NEB, M0297), 1 μl Escherichia coli DNA ligase (NEB, M0205), 5 μl E. coli DNA polymerase (NEB, M0209), 1 μl 0.2 mM dNTP (ThermoFisher, R0192), 10 μl 5× second strand buffer (100 mM Tris-HCl, pH 6.9, 25 mM MgCl₂, 450 mM KCl, 0.8 mM β-NAD, 60 mM (NH₄)2S0₄) and 11 μl water; overnight incubation at 16 °C. Tagmentation was then performed with 50 ng second strand synthesized sample in the following reaction: 10 μl 2× tagmentation buffer (Diagenode, C01019043), 9 μl DNA, 1 μl Tagmentase (Diagenode, C01070012-30), incubation at 55 °C for 7 min. Tagmented samples were purified (Monarch Spin PCR & DNA Cleanup Kit, NEB T1130) and resuspended in 20 μl water before final library amplification: 20 μl tagmented DNA, 2.5 μl 5 μM P5 BRB Primer, 2.5 μL 5 μM BRB Idx7N5 primer, 25 μl NEBNext High-Fidelity 2× PCR Master Mix (NEB, M0541) with the following thermocycler conditions: 3 min at 72 °C, 30 s at 98 °C, 15×(10 s 98 °C, 30 s 63 °C, 30 s 72 °C), 5 min at 72 °C. Amplified libraries were double SPRIselect purified (0.5×, 0.8×, Beckman, B23319) to select for 200–800-bp fragments before confirmation on an Agilent TapeStation with D1000 High Sensitivity ScreenTape (Agilent, 5067-5584).

Libraries were sequenced on a NextSeq 2000 sequencer with a P3-100 cycle kit (Illumina) with 21-bp read 1 using a custom primer (BRB_Ilmn_read_1; Supplementary Table 31), and 71 bp for read 2 with the standard primer. Sequenced reads were aligned with STAR (v.2.7.11b, --outFilterMultimapNmax 1)¹⁰¹ and demultiplexed with BRB-seq Tools (v.1.6.1)¹⁰⁰. Count tables from BRB-seq Tools (CreateDGEMatrix) were used as input for differential expression analysis with DESeq2 (v.1.48.1)¹⁰² and significant genes (Wald test) were identified at an FDR threshold of 0.01 (Benjamini–Hochberg procedure). BRB-seq results for C2orf88 and HIBCH were confirmed using quantitative PCR. Total RNA (1–2.5 μg) from the same extractions was converted to cDNA using SuperScript III, and qPCR was done using Luna Universal Master Mix on a QuantStudio 5 with gene-specific primers for C2orf88 (C2orf88_F and C2orf88_R; Supplementary Table 31), HIBCH (HIBCH_F and HIBCH_R; Supplementary Table 31) and TBP as the normalizing control gene (TBP_F and TBP_R; Supplementary Table 31). Two independent experiments with four technical replicates each were done, and fold changes were calculated using the ΔΔCt method normalized to TBP and reference allele expression. Statistical differences were assessed using a two-sided Mann–Whitney U test (Extended Data Fig. 7c).

Data visualization and figure generation

All plots were generated in R using ggplot2 (v.3.5.0) and assembled with patchwork (v.1.2.0) and cowplot (v.1.1.3). Heatmaps were created using ComplexHeatmap (v.2.2.0). Colour palettes were from BuenColors (v.0.5.6) and PNWColors (v.0.1.0). Rasterization was done using ggrastr (v.1.0.2) and marginal distributions added using ggExtra (v.0.10.1).

Reporting summary

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

Data availability

ENCODE and NCBI GEO accession numbers for all the MPRA data generated for this study are available in Supplementary Table 29. Processed MPRA datasets, along with annotation files and code used for analysis, are available at Zenodo at https://doi.org/10.5281/zenodo.15297965 (ref. ¹⁰³). The UKBB analysis in this study was conducted though application number 31063.

Code availability

The code used to analyse and visualize data for this study is available at https://github.com/julirsch/finemapped_mpra/.

References

Loos, R. J. F. 15 years of genome-wide association studies and no signs of slowing down. Nat. Commun. 11, 5900 (2020).
Article ADS CAS PubMed PubMed Central Google Scholar
Weeks, E. M. et al. Leveraging polygenic enrichments of gene features to predict genes underlying complex traits and diseases. Nat. Genet. 55, 1267–1276 (2023).
Article CAS PubMed PubMed Central Google Scholar
Mountjoy, E. et al. An open approach to systematically prioritize causal variants and genes at all published human GWAS trait-associated loci. Nat. Genet. 53, 1527–1533 (2021).
Article CAS PubMed PubMed Central Google Scholar
Sullivan, P. F. et al. Leveraging base-pair mammalian constraint to understand genetic variation and human disease. Science 380, eabn2937 (2023).
Article CAS PubMed PubMed Central Google Scholar
MacArthur, J. et al. The new NHGRI-EBI Catalog of published genome-wide association studies (GWAS Catalog). Nucleic Acids Res. 45, D896–D901 (2017).
Article CAS PubMed Google Scholar
O’Connor, L. J. The distribution of common-variant effect sizes. Nat. Genet. 53, 1243–1249 (2021).
Article PubMed Google Scholar
Gusev, A. et al. Partitioning heritability of regulatory and cell-type-specific variants across 11 common diseases. Am. J. Hum. Genet. 95, 535–552 (2014).
Article CAS PubMed PubMed Central Google Scholar
Farh, K. K.-H. et al. Genetic and epigenetic fine mapping of causal autoimmune disease variants. Nature 518, 337–343 (2015).
Article ADS CAS PubMed Google Scholar
Meuleman, W. et al. Index and biological spectrum of human DNase I hypersensitive sites. Nature 584, 244–251 (2020).
Article ADS CAS PubMed PubMed Central Google Scholar
Nasser, J. et al. Genome-wide enhancer maps link risk variants to disease genes. Nature 593, 238–243 (2021).
Article ADS CAS PubMed PubMed Central Google Scholar
Flynn, E. D. & Lappalainen, T. Functional characterization of genetic variant effects on expression. Annu. Rev. Biomed. Data Sci. 5, 119–139 (2022).
Article PubMed Google Scholar
Cano-Gamez, E. & Trynka, G. From GWAS to function: using functional genomics to identify the mechanisms underlying complex diseases. Front. Genet. 11, 424 (2020).
Article CAS PubMed PubMed Central Google Scholar
Ryu, J. et al. Joint genotypic and phenotypic outcome modeling improves base editing variant effect quantification. Nat. Genet. 56, 925–937 (2024).
Article CAS PubMed PubMed Central Google Scholar
Chen, Z. et al. Integrative dissection of gene regulatory elements at base resolution. Cell Genom. 3, 100318 (2023).
Lin, J. & Musunuru, K. From genotype to phenotype: a primer on the functional follow-up of genome-wide association studies in cardiovascular disease. Circ. Genom. Precis. Med. 11, e001946 (2018).
Article PubMed PubMed Central Google Scholar
Tewhey, R. et al. Direct identification of hundreds of expression-modulating variants using a multiplexed reporter assay. Cell 165, 1519–1529 (2016).
Article CAS PubMed PubMed Central Google Scholar
Ulirsch, J. C. et al. Systematic functional dissection of common genetic variation affecting red blood cell traits. Cell 165, 1530–1545 (2016).
Article CAS PubMed PubMed Central Google Scholar
Mouri, K. et al. Prioritization of autoimmune disease-associated genetic variants that perturb regulatory element activity in T cells. Nat. Genet. 54, 603–612 (2022).
Article CAS PubMed PubMed Central Google Scholar
Lee, S. et al. Massively parallel reporter assay investigates shared genetic variants of eight psychiatric disorders. Cell 188, 1409–1424 (2025).
Article CAS PubMed PubMed Central Google Scholar
Cooper, Y. A. et al. Functional regulatory variants implicate distinct transcriptional networks in dementia. Science 377, eabi8654 (2022).
Article CAS PubMed Google Scholar
Cerezo, M. et al. The NHGRI-EBI GWAS Catalog: standards for reusability, sustainability and diversity. Nucleic Acids Res. 53, D998–D1005 (2025).
Article CAS PubMed PubMed Central Google Scholar
The Wellcome Trust Case Control Consortium. Genome-wide association study of 14,000 cases of seven common diseases and 3,000 shared controls. Nature 447, 661–678 (2007).
Article Google Scholar
Minikel, E. V., Painter, J. L., Dong, C. C. & Nelson, M. R. Refining the impact of genetic evidence on clinical success. Nature 629, 624–629 (2024).
Article ADS CAS PubMed PubMed Central Google Scholar
Ochoa, D. et al. Human genetics evidence supports two-thirds of the 2021 FDA-approved drugs. Nat. Rev. Drug Discov. 21, 551 (2022).
Article CAS PubMed Google Scholar
Lappalainen, T. & MacArthur, D. G. From variant to function in human disease genetics. Science 373, 1464–1468 (2021).
Article ADS CAS PubMed Google Scholar
The Wellcome Trust Case Control Consortium. Bayesian refinement of association signals for 14 loci in 3 common diseases. Nat. Genet. 44, 1294–1301 (2012).
Article Google Scholar
Weissbrod, O. et al. Leveraging fine-mapping and multipopulation training data to improve cross-population polygenic risk scores. Nat. Genet. 54, 450–458 (2022).
Article CAS PubMed PubMed Central Google Scholar
Ulirsch, J. C. et al. Interrogation of human hematopoiesis at single-cell and single-variant resolution. Nat. Genet. 51, 683–693 (2019).
Article CAS PubMed PubMed Central Google Scholar
Uffelmann, E. et al. Genome-wide association studies. Nat. Rev. Methods Primers 1, 59 (2021).
Article CAS Google Scholar
Wang, G., Sarkar, A., Carbonetto, P. & Stephens, M. A simple new approach to variable selection in regression, with application to genetic fine mapping. J. R. Stat. Soc. Series B Stat. Methodol. 82, 1273–1300 (2020).
Article MathSciNet PubMed PubMed Central Google Scholar
Kanai, M. et al. Insights from complex trait fine-mapping across diverse populations. Preprint at medRxiv https://doi.org/10.1101/2021.09.03.21262975 (2021).
Verma, A. et al. Diversity and scale: genetic architecture of 2068 traits in the VA Million Veteran Program. Science 385, eadj1182 (2024).
Weissbrod, O. et al. Functionally informed fine-mapping and polygenic localization of complex trait heritability. Nat. Genet. 52, 1355–1363 (2020).
Article CAS PubMed PubMed Central Google Scholar
Wang, Q. S. et al. Leveraging supervised learning for functionally informed fine-mapping of cis-eQTLs identifies an additional 20,913 putative causal eQTLs. Nat. Commun. 12, 3394 (2021).
Article ADS CAS PubMed PubMed Central Google Scholar
Klein, J. C. et al. A systematic evaluation of the design and context dependencies of massively parallel reporter assays. Nat. Methods 17, 1083–1091 (2020).
Article CAS PubMed PubMed Central Google Scholar
Muerdter, F. et al. Resolving systematic errors in widely used enhancer activity assays in human cells. Nat. Methods 15, 141–149 (2018).
Article CAS PubMed Google Scholar
Agarwal, V. et al. Massively parallel characterization of transcriptional regulatory elements. Nature 639, 411–420 (2025).
Article ADS CAS PubMed PubMed Central Google Scholar
Sakaue, S. et al. A cross-population atlas of genetic associations for 220 human phenotypes. Nat. Genet. 53, 1415–1424 (2021).
Article CAS PubMed PubMed Central Google Scholar
The GTEx Consortium. The GTEx Consortium atlas of genetic regulatory effects across human tissues. Science 369, 1318–1330 (2020).
Article Google Scholar
Zou, Z., Ohta, T., Miura, F. & Oki, S. ChIP-Atlas 2021 update: a data-mining suite for exploring epigenomic landscapes by fully integrating ChIP-seq, ATAC-seq and Bisulfite-seq data. Nucleic Acids Res. 50, W175–W182 (2022).
Article CAS PubMed PubMed Central Google Scholar
Oki, S. et al. ChIP-Atlas: a data-mining suite powered by full integration of public ChIP-seq data. EMBO Rep. 19, e46255 (2018).
Article PubMed PubMed Central Google Scholar
Nguyen, T. A. et al. High-throughput functional comparison of promoter and enhancer activities. Genome Res. 26, 1023–1033 (2016).
Article CAS PubMed PubMed Central Google Scholar
Chen, K. M., Wong, A. K., Troyanskaya, O. G. & Zhou, J. A sequence-based global map of regulatory activity for deciphering human genetics. Nat. Genet. 54, 940–949 (2022).
Article CAS PubMed PubMed Central Google Scholar
van Arensbergen, J. et al. High-throughput identification of human SNPs affecting regulatory element activity. Nat. Genet. 51, 1160–1169 (2019).
Article PubMed PubMed Central Google Scholar
Grossman, S. R. et al. Systematic dissection of genomic features determining transcription factor binding and enhancer function. Proc. Natl Acad. Sci. USA 114, E1291–E1300 (2017).
Article CAS PubMed PubMed Central Google Scholar
Wang, J. et al. Sequence features and chromatin structure around the genomic regions bound by 119 human transcription factors. Genome Res. 22, 1798–1812 (2012).
Article CAS PubMed PubMed Central Google Scholar
Vierstra, J. et al. Global reference mapping of human transcription factor footprints. Nature 583, 729–736 (2020).
Article ADS CAS PubMed PubMed Central Google Scholar
Avsec, Ž. et al. Effective gene expression prediction from sequence by integrating long-range interactions. Nat. Methods 18, 1196–1203 (2021).
Article CAS PubMed PubMed Central Google Scholar
Hormozdiari, F. et al. Widespread allelic heterogeneity in complex traits. Am. J. Hum. Genet. 100, 789–802 (2017).
Article CAS PubMed PubMed Central Google Scholar
Abell, N. S. et al. Multiple causal variants underlie genetic associations in humans. Science 375, 1247–1254 (2022).
Article ADS CAS PubMed PubMed Central Google Scholar
Corradin, O. et al. Combinatorial effects of multiple enhancer variants in linkage disequilibrium dictate levels of gene expression to confer susceptibility to common traits. Genome Res. 24, 1–13 (2014).
Article CAS PubMed Google Scholar
She, R. & Jarosz, D. F. Mapping causal variants with single-nucleotide resolution reveals biochemical drivers of phenotypic change. Cell 172, 478–490 (2018).
Article CAS PubMed PubMed Central Google Scholar
Brown, B. C., Price, A. L., Patsopoulos, N. A. & Zaitlen, N. Local joint testing improves power and identifies hidden heritability in association studies. Genetics 203, 1105–1116 (2016).
Article CAS PubMed PubMed Central Google Scholar
Zhang, M. J. et al. Pervasive correlations between causal disease effects of proximal SNPs vary with functional annotations and implicate stabilizing selection. Preprint at medRxiv https://doi.org/10.1101/2023.12.04.23299391 (2023).
Takada, I. et al. Transcriptional coregulator Ess2 controls survival of post-thymic CD4⁺ T cells through the Myc and IL-7 signaling pathways. J. Biol. Chem. 298, 102342 (2022).
Article CAS PubMed PubMed Central Google Scholar
Zhou, J. et al. Deep learning sequence-based ab initio prediction of variant effects on expression and disease risk. Nat. Genet. 50, 1171–1179 (2018).
Article CAS PubMed PubMed Central Google Scholar
Hay, D. et al. Genetic dissection of the α-globin super-enhancer in vivo. Nat. Genet. 48, 895–903 (2016).
Article CAS PubMed PubMed Central Google Scholar
Li, Y., Zheng, M. & Lau, Y.-F. C. The sex-determining factors SRY and SOX9 regulate similar target genes and promote testis cord formation during testicular differentiation. Cell Rep. 8, 723–733 (2014).
Article CAS PubMed Google Scholar
Wakabayashi, A. et al. Insight into GATA1 transcriptional activity through interrogation of cis elements disrupted in human erythroid disorders. Proc. Natl Acad. Sci. USA 113, 4434–4439 (2016).
Article ADS CAS PubMed PubMed Central Google Scholar
Han, G. C. et al. Genome-wide organization of GATA1 and TAL1 determined at high resolution. Mol. Cell. Biol. 36, 157–172 (2015).
Article PubMed PubMed Central Google Scholar
Kircher, M. et al. Saturation mutagenesis of twenty disease-associated regulatory elements at single base-pair resolution. Nat. Commun. 10, 3583 (2019).
Article ADS PubMed PubMed Central Google Scholar
Cui, R. et al. Improving fine-mapping by modeling infinitesimal effects. Nat. Genet. 56, 162–169 (2024).
Article CAS PubMed Google Scholar
Rhine, C. L. et al. Massively parallel reporter assays discover de novo exonic splicing mutants in paralogs of autism genes. PLoS Genet. 18, e1009884 (2022).
Article CAS PubMed PubMed Central Google Scholar
Gosai, S. J. et al. Machine-guided design of cell-type-targeting cis-regulatory elements. Nature 634, 1211–1220 (2024).
Spitz, F. & Furlong, E. E. M. Transcription factors: from enhancer binding to developmental control. Nat. Rev. Genet. 13, 613–626 (2012).
Ulirsch, J. C. Identification and Interpretation of Causal Genetic Variants Underlying Human Phenotypes. PhD thesis, Harvard Univ. (2022).
Zhou, W. et al. Efficiently controlling for case-control imbalance and sample relatedness in large-scale genetic association studies. Nat. Genet. 50, 1335–1341 (2018).
Article CAS PubMed PubMed Central Google Scholar
Loh, P.-R., Kichaev, G., Gazal, S., Schoech, A. P. & Price, A. L. Mixed-model association for biobank-scale datasets. Nat. Genet. 50, 906–908 (2018).
Article CAS PubMed PubMed Central Google Scholar
Loh, P.-R. et al. Efficient Bayesian mixed-model analysis increases association power in large cohorts. Nat. Genet. 47, 284–290 (2015).
Article CAS PubMed PubMed Central Google Scholar
Lonsdale, J. et al. The Genotype-Tissue Expression (GTEx) project. Nat. Genet. 45, 580–585 (2013).
Article CAS Google Scholar
Bycroft, C. et al. The UK Biobank resource with deep phenotyping and genomic data. Nature 562, 203–209 (2018).
Article ADS CAS PubMed PubMed Central Google Scholar
Nagai, A. et al. Overview of the BioBank Japan Project: study design and profile. J. Epidemiol. 27, S2–S8 (2017).
Article PubMed PubMed Central Google Scholar
Schoech, A. P. et al. Quantification of frequency-dependent genetic architectures in 25 UK Biobank traits reveals action of negative selection. Nat. Commun. 10, 790 (2019).
Article ADS CAS PubMed PubMed Central Google Scholar
Xue, J. R. et al. The functional and evolutionary impacts of human-specific deletions in conserved elements. Science 380, eabn2253 (2023).
Article CAS PubMed PubMed Central Google Scholar
Tippens, N. D. et al. Transcription imparts architecture, function and logic to enhancer units. Nat. Genet. 52, 1067–1075 (2020).
Article PubMed PubMed Central Google Scholar
Inoue, F., Kreimer, A., Ashuach, T., Ahituv, N. & Yosef, N. Identification and massively parallel characterization of regulatory elements driving neural induction. Cell Stem Cell 25, 713–727 (2019).
Article CAS PubMed PubMed Central Google Scholar
Magoč, T. & Salzberg, S. L. FLASH: fast length adjustment of short reads to improve genome assemblies. Bioinformatics 27, 2957–2963 (2011).
Article PubMed PubMed Central Google Scholar
Li, H. Minimap2: pairwise alignment for nucleotide sequences. Bioinformatics 34, 3094–3100 (2018).
Article CAS PubMed PubMed Central Google Scholar
Urbut, S. M., Wang, G., Carbonetto, P. & Stephens, M. Flexible statistical methods for estimating and testing effects in genomic studies with multiple conditions. Nat. Genet. 51, 187–195 (2019).
Article CAS PubMed Google Scholar
McLaren, W. et al. The Ensembl Variant Effect Predictor. Genome Biol. 17, 122 (2016).
Article PubMed PubMed Central Google Scholar
Roadmap Epigenomics Consortium. Integrative analysis of 111 reference human epigenomes. Nature 518, 317–330 (2015).
Article Google Scholar
Domcke, S. et al. A human cell atlas of fetal chromatin accessibility. Science 370, eaba7612 (2020).
Article CAS PubMed PubMed Central Google Scholar
Corces, M. R. et al. Lineage-specific and single-cell chromatin accessibility charts human hematopoiesis and leukemia evolution. Nat. Genet. 48, 1193–1203 (2016).
Article CAS PubMed PubMed Central Google Scholar
Corces, M. R. et al. Single-cell epigenomic analyses implicate candidate causal variants at inherited risk loci for Alzheimer’s and Parkinson’s diseases. Nat. Genet. 52, 1158–1168 (2020).
Article CAS PubMed PubMed Central Google Scholar
Calderon, D. et al. Landscape of stimulation-responsive chromatin across diverse human immune cells. Nat. Genet. 51, 1494–1505 (2019).
Article CAS PubMed PubMed Central Google Scholar
Abramov, S. et al. Landscape of allele-specific transcription factor binding in the human genome. Nat. Commun. 12, 2751 (2021).
Article ADS CAS PubMed PubMed Central Google Scholar
Grant, C. E., Bailey, T. L. & Noble, W. S. FIMO: scanning for occurrences of a given motif. Bioinformatics 27, 1017–1018 (2011).
Article CAS PubMed PubMed Central Google Scholar
DiMatteo, J. et al. Motif Liquidator Matrix. GitHub https://github.com/BradnerLab/pipeline/tree/motif_liquidator_matrix (2017).
Vorontsov, I. E. et al. HOCOMOCO in 2024: a rebuild of the curated collection of binding models for human and mouse transcription factors. Nucleic Acids Res. 52, D154–D163 (2024).
Article CAS PubMed PubMed Central Google Scholar
Castro-Mondragon, J. A. et al. JASPAR 2022: the 9th release of the open-access database of transcription factor binding profiles. Nucleic Acids Res. 50, D165–D173 (2022).
Article CAS PubMed PubMed Central Google Scholar
Razavi, R. et al. Extensive binding of uncharacterized human transcription factors to genomic dark matter. Preprint at bioRxiv https://doi.org/10.1101/2024.11.11.622123 (2024).
Jung, I. et al. A compendium of promoter-centered long-range chromatin interactions in the human genome. Nat. Genet. 51, 1442–1449 (2019).
Article CAS PubMed PubMed Central Google Scholar
Javierre, B. M. et al. Lineage-specific genome architecture links enhancers and non-coding disease variants to target gene promoters. Cell 167, 1369–1384 (2016).
Article CAS PubMed PubMed Central Google Scholar
Liu, Y. et al. Functional assessment of human enhancer activities using whole-genome STARR-sequencing. Genome Biol. 18, 219 (2017).
Article MathSciNet PubMed PubMed Central Google Scholar
Andersson, R. et al. An atlas of active enhancers across human cell types and tissues. Nature 507, 455–461 (2014).
Article ADS CAS PubMed PubMed Central Google Scholar
Gschwind, A. R. et al. An encyclopedia of enhancer-gene regulatory interactions in the human genome. Preprint at bioRxiv https://doi.org/10.1101/2023.11.09.563812 (2023).
Oliva, M. et al. The impact of sex on gene expression across human tissues. Science 369, eaba3066 (2020).
Article CAS PubMed PubMed Central Google Scholar
Komor, A. C. et al. Improved base excision repair inhibition and bacteriophage Mu Gam protein yields C:G-to-T:A base editors with higher efficiency and product purity. Sci. Adv. 3, eaao4774 (2017).
Article PubMed PubMed Central Google Scholar
Kleinstiver, B. P. et al. Engineered CRISPR-Cas9 nucleases with altered PAM specificities. Nature 523, 481–485 (2015).
Article ADS PubMed PubMed Central Google Scholar
Alpern, D. et al. BRB-seq: ultra-affordable high-throughput transcriptomics enabled by bulk RNA barcoding and sequencing. Genome Biol. 20, 71 (2019).
Article PubMed PubMed Central Google Scholar
Dobin, A. et al. STAR: ultrafast universal RNA-seq aligner. Bioinformatics 29, 15–21 (2013).
Article CAS PubMed Google Scholar
Love, M. I., Huber, W. & Anders, S. Moderated estimation of fold change and dispersion for RNA-seq data with DESeq2. Genome Biol. 15, 550 (2014).
Article PubMed PubMed Central Google Scholar
Siraj, L. et al. Functional dissection of complex and molecular trait variants at single nucleotide resolution. Zenodo https://doi.org/10.5281/zenodo.15297965 (2025).

Download references

Acknowledgements

We thank the Jackson Laboratory Genome Technologies Core for experimental support; and M. Stitzel, J. Ray, J. Fuxman Bass and S. Rong for suggestions and conversations about the manuscript. This work was directly supported by the Howard Hughes Medical Institute and by US National Institutes of Health grants UM1HG009435, R00HG010669, R01HG012872 and R35HG011329. Individual support was provided by the Novo Nordisk Foundation (NNF21SA0072102) to R.C. and T.R.J., and US National Institutes of Health grants T32GM007753 and T32GM144273 to L.S. The content is solely the responsibility of the authors and does not necessarily represent the official views of the National Institutes of Health.

Author information

Christopher M. Vockley
Present address: Eli Lilly and Company, Boston, MA, USA
These authors contributed equally: Jacob C. Ulirsch, Ryan Tewhey
These authors jointly supervised this work: Eric S. Lander, Pardis C. Sabeti, Hilary K. Finucane, Steven K. Reilly, Jacob C. Ulirsch, Ryan Tewhey

Authors and Affiliations

Broad Institute of Harvard and MIT, Cambridge, MA, USA
Layla Siraj, Sager J. Gosai, François Aguet, Christopher M. Vockley, Thouis R. Jones, Eric S. Lander & Pardis C. Sabeti
Program in Biophysics, Harvard Graduate School of Arts and Sciences, Boston, MA, USA
Layla Siraj
Harvard-Massachusetts Institute of Technology MD/PhD Program, Harvard Medical School, Boston, MA, USA
Layla Siraj
Analytic and Translational Genetics Unit, Massachusetts General Hospital, Boston, MA, USA
Layla Siraj, Masahiro Kanai, Qingbo S. Wang, Kristin Tsuo, Ran Cui, Hilary K. Finucane & Jacob C. Ulirsch
The Jackson Laboratory, Bar Harbor, ME, USA
Rodrigo I. Castro, Hannah B. Dewey, Susan Kales, John C. Butts, Daniel Berenzy, Kousuke Mouri & Ryan Tewhey
Graduate School of Biomedical Sciences, Tufts University School of Medicine, Boston, MA, USA
Hannah B. Dewey & Ryan Tewhey
Graduate School of Biomedical Sciences and Engineering, University of Maine, Orono, ME, USA
John C. Butts & Ryan Tewhey
Department of Genetics, Yale School of Medicine, New Haven, CT, USA
Thanh Thanh L. Nguyen & Steven K. Reilly
Program in Medical and Population Genetics, Broad Institute of Harvard and MIT, Cambridge, MA, USA
Masahiro Kanai, Qingbo S. Wang, Kristin Tsuo, Ran Cui, Hilary K. Finucane & Jacob C. Ulirsch
Stanley Center for Psychiatric Research, Broad Institute of Harvard and MIT, Cambridge, MA, USA
Masahiro Kanai, Qingbo S. Wang, Kristin Tsuo, Ran Cui, Hilary K. Finucane & Jacob C. Ulirsch
Center for Computational and Integrative Biology, Massachusetts General Hospital, Boston, MA, USA
Masahiro Kanai
Department of Statistical Genetics, Osaka University Graduate School of Medicine, Suita, Japan
Qingbo S. Wang & Yukinori Okada
Department of Genome Informatics, Graduate School of Medicine, University of Tokyo, Tokyo, Japan
Qingbo S. Wang & Yukinori Okada
Illumina Artificial Intelligence Laboratory, Illumina, San Diego, CA, USA
Petko P. Fiziev, François Aguet, Irfahan Kassam, Jeremy McRae & Jacob C. Ulirsch
Insitro, South San Francisco, CA, USA
Zachary R. McCaw
Program in Biological and Biomedical Sciences, Harvard Medical School, Boston, MA, USA
Sager J. Gosai & Jacob C. Ulirsch
Department of Organismic and Evolutionary Biology, Harvard University, Cambridge, MA, USA
Sager J. Gosai & Pardis C. Sabeti
Howard Hughes Medical Institute, Chevy Chase, MD, USA
Sager J. Gosai & Pardis C. Sabeti
The Novo Nordisk Foundation Center for Genomic Mechanisms of Disease, Broad Institute of MIT and Harvard, Cambridge, MA, USA
Ran Cui & Thouis R. Jones
Program in Computational and Systems Biology, Memorial Sloan Kettering Cancer Center, New York, NY, USA
Caleb A. Lareau
Altius Institute for Biomedical Sciences, Seattle, WA, USA
Sergey Abramov, Alexandr Boystov & Jeff Vierstra
Laboratory for Systems Genetics, RIKEN Center for Integrative Medical Sciences, Yokohama, Japan
Yukinori Okada
Harvard Medical School and Dana-Farber Cancer Institute, Boston, MA, USA
Alexander Gusev
Department of Biology, MIT, Cambridge, MA, USA
Eric S. Lander
Department of Systems Biology, Harvard Medical School, Boston, MA, USA
Eric S. Lander
Wu Tsai Institute, Yale University, New Haven, CT, USA
Steven K. Reilly

Authors

Layla Siraj
View author publications
Search author on:PubMed Google Scholar
Rodrigo I. Castro
View author publications
Search author on:PubMed Google Scholar
Hannah B. Dewey
View author publications
Search author on:PubMed Google Scholar
Susan Kales
View author publications
Search author on:PubMed Google Scholar
John C. Butts
View author publications
Search author on:PubMed Google Scholar
Thanh Thanh L. Nguyen
View author publications
Search author on:PubMed Google Scholar
Masahiro Kanai
View author publications
Search author on:PubMed Google Scholar
Daniel Berenzy
View author publications
Search author on:PubMed Google Scholar
Kousuke Mouri
View author publications
Search author on:PubMed Google Scholar
Qingbo S. Wang
View author publications
Search author on:PubMed Google Scholar
Petko P. Fiziev
View author publications
Search author on:PubMed Google Scholar
Kristin Tsuo
View author publications
Search author on:PubMed Google Scholar
Zachary R. McCaw
View author publications
Search author on:PubMed Google Scholar
Sager J. Gosai
View author publications
Search author on:PubMed Google Scholar
François Aguet
View author publications
Search author on:PubMed Google Scholar
Ran Cui
View author publications
Search author on:PubMed Google Scholar
Irfahan Kassam
View author publications
Search author on:PubMed Google Scholar
Jeremy McRae
View author publications
Search author on:PubMed Google Scholar
Christopher M. Vockley
View author publications
Search author on:PubMed Google Scholar
Caleb A. Lareau
View author publications
Search author on:PubMed Google Scholar
Sergey Abramov
View author publications
Search author on:PubMed Google Scholar
Alexandr Boystov
View author publications
Search author on:PubMed Google Scholar
Jeff Vierstra
View author publications
Search author on:PubMed Google Scholar
Yukinori Okada
View author publications
Search author on:PubMed Google Scholar
Alexander Gusev
View author publications
Search author on:PubMed Google Scholar
Thouis R. Jones
View author publications
Search author on:PubMed Google Scholar
Eric S. Lander
View author publications
Search author on:PubMed Google Scholar
Pardis C. Sabeti
View author publications
Search author on:PubMed Google Scholar
Hilary K. Finucane
View author publications
Search author on:PubMed Google Scholar
Steven K. Reilly
View author publications
Search author on:PubMed Google Scholar
Jacob C. Ulirsch
View author publications
Search author on:PubMed Google Scholar
Ryan Tewhey
View author publications
Search author on:PubMed Google Scholar

Contributions

J.C.U. and R.T. conceived the study. L.S., R.I.C., H.K.F., S.K.R., J.C.U. and R.T. designed experiments. S.K., D.B. and K.M. performed MPRA experiments. J.C.B. performed base-editing experiments. S.A., A.B. and J.V. provided allelic-imbalance datasets. L.S., R.I.C., H.B.D., M.K., Q.S.W., P.P.F., K.T., F.A., J.C.U. and R.T. analysed data. L.S., R.I.C., H.B.D., T.T.L.N., M.K., Q.S.W., Z.R.M., S.J.G., F.A., R.C., I.K., J.M., C.M.V., C.A.L., Y.O., A.G., E.S.L., H.K.F., S.K.R., J.C.U. and R.T. interpreted results. L.S., R.I.C., E.S.L., H.K.F., S.K.R., J.C.U. and R.T. drafted the manuscript. T.R.J., P.C.S., S.K.R. and R.T. secured funding. E.S.L., P.C.S., H.K.F., S.K.R., J.C.U. and R.T. supervised the study. All authors revised the manuscript and approved the final version.

Corresponding authors

Correspondence to Hilary K. Finucane, Steven K. Reilly, Jacob C. Ulirsch or Ryan Tewhey.

Ethics declarations

Competing interests

P.C.S. is a co-founder of, and former consultant to, Sherlock Biosciences and a former board member of the Danaher Corporation, and is a co-founder of, and equity holder of, Lyra Labs. P.C.S. and R.T. hold patents related to the application of MPRA. J.C.U., P.P.F., I.K. and J.M. are employees of Illumina. F.A. was an employee of Illumina when this work was performed and is currently an employee and equity holder of Predicta Biosciences. Q.S.W. is an employee of Calico Life Sciences LLC. C.M.V. is an employee of Eli Lilly and Company. Z.R.M. was an employee of insitro when this work was done. The remaining authors declare no competing interests.

Peer review

Peer review information

Nature thanks Arjun Bhattacharya and the other, anonymous, reviewer(s) for their contribution to the peer review of this work. Peer reviewer reports are available.

Additional information

Publisher’s note Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations.

Extended data figures and tables

Extended Data Fig. 1 Additional diagnostics of MPRA data quality.

a. Example of annotation-matched control (grey) and test (PIP > 0.1, burgundy) variant proportions in the MPRA library for LDL cholesterol, an exemplary trait. b. Distribution of barcodes per oligo across tested oligo elements. c. Correlations between normalized DNA and RNA counts (log₂ counts per million) across plasmids and cell-types for a representative library. Replicates 1, 2, and 3 (out of 5 total replicates) are shown due to space limitations. d. Pairwise Pearson r correlations of MPRA activity measurements [log₂(fold change)] of 91 positive technical controls shared across all libraries and cell-types. e. MPRA activity [log₂(fold change)] compared to normalized plasmid (DNA) counts. Significantly active elements are shown in blue (two-sided Wald test, Bonferroni-adjusted P < 0.01) after correcting for multiple hypothesis testing within each cell-type and library.

Extended Data Fig. 2 Additional characterization of inactive and active elements across annotations.

a. Correlation (Pearson r) of element activity at promoters and distal CREs between 4 tested cell-types. b. TF occupancy at active elements is compared to TF occupancy at inactive elements for 1,513 TFs. Enrichment is shown as an odds ratio (OR) and significant TFs are highlighted in burgundy (Bonferroni-adjusted P < 0.05). Error bars represent 95% CIs. c. Proportion of variants overlapping disjoint (with the exception of “Any CRE”) genomic annotations, stratified by whether the variant is active or an emVar by MPRA. Error bars represent 95% CIs. d. Proportion of 95% CSs containing at least one variant in the same set of genomic annotations. For comparison, MPRA-derived annotations are included. Error bars represent 95% CIs. e. Enrichment of variants in inactive vs active elements for overlap with ChIP-seq peaks from “Dark TFs”. TFs coloured in burgundy are significant at a Bonferroni-corrected P < 0.05. Error bars represent 95% CIs. f. Overlap of non-coding, non-promoter elements tested by MPRA with regions of the genome annotated as polycomb and heterochromatin. Elements are stratified by activity level and whether they contain eQTL or complex trait variants. Error bars represent 95% CIs. g. Baseline activity of MPRA across cell-types (normalized to 1 for the reference cell-type) for sequences obtained from annotated cell-type specific DHSs. Activity is highest when the MPRA cell-type matches the DHS cell-type. Error bars represent 95% CIs.

Extended Data Fig. 3 Additional characterization of emVars across annotations.

a. Proportion of variants that are emVars across cell-types and libraries. Circle size corresponds to the proportion of emVars and colour indicates average RNA counts. b. Total number of emVars with a PIP > 0.1 (size of circle) and proportion of tested variants that are emVars (colour scale) across complex traits. c. Proportion of eQTLs that are emVars in each cell-type. eQTLs are included only if they were fine-mapped with PIP > 0.5 in each indicated tissue/system. d. For eQTLs with PIP > 0.5 that are TARVs in a single cell-type, the difference in proportion of eQTLs fine-mapped for each tissue (y-axis) that are emVars in the cell-type is shown on the x-axis vs those that are cell-type specific emVars in the other 3 cell-types. Size is proportional to -log₁₀(P) of this association. e. Proportion of variants that are emVars for a given PIP. Solid lines represent locally estimated scatterplot smoothing across PIP centiles. Dashed lines represent linear regression. Colours represent whether variants were fine-mapped for complex traits or eQTLs. f. Enrichment of emVars in the highest vs lowest PIP bin for complex trait (purple) and eQTL (blue) variants for all variants or for variants in CREs. g. TF occupancy allelic effects are correlated (Pearson r and linear model) with MPRA allelic effects. emVars are shown in orange and non-emVars in purple.

Extended Data Fig. 4 Additional precision and recall analysis.

a. Precision-recall plots evaluating different methods for discriminating between high-confidence positive (high-PIP) and negative (low-PIP) eQTLs and complex trait variants. Error bars represent 95% CIs. b. Sensitivity-specificity plots evaluating different methods for discriminating between high-confidence positive (high-PIP) and negative (low-PIP) eQTLs and complex trait variants. Error bars represent 95% CIs. c,d. Precision-recall plots evaluating different methods (emVar or TARV) for discriminating between positives (high-PIP) and negative (low-PIP) complex trait (c) and eQTL (d) variants, varying the number and composition of cell-types included. Meta-analysis points across each number of included cell-types are shown in black. Error bars represent 95% CIs. e. Precision of each annotation across CS size. The background rate (1 / [CS size - 1]) is shown in grey and compared to emVar (green) or TARV (blue) annotations. f,g. Precision-recall plots evaluating the utility of cell-type matched CRE and E2G annotations on their own vs cell-type matched TARVs or cell-type matched E2G variants in conjunction with emVar status at discriminating between high-PIP and control (low-PIP) complex trait (f) and eQTL (g) variants. Meta-analysis points are indicated in black. Error bars represent 95% CIs.

Extended Data Fig. 5 Additional evidence for regulatory allelic heterogeneity and multiple causal variants.

a-c. Scatter plots of activity from all 4 diplotypes across 3 windows. Each point represents an activity measurement for one diplotype in one cell-type. Pearson correlations are shown for all points indicated in the plot. Lines represent linear fits of activity between windows stratified by cell-type. d-f. Same as a-c, except for allelic effects (i.e. each alternative diplotype vs the homozygous reference). g. Empirical correlations between overlapping windows for activity or allelic effect estimates based on the percentage of overlapping sequence shared between the windows. Correlations are smoothed by a sliding window. Points are scaled by the square root of the number of variant pairs. h,i. Correlations for estimates of activity (log₂(fold-change), top) and allelic effects (Δlog₂(fold-change), bottom) in each cell-type (h) and across cell-types (i) for variant pairs. j,k. Proportion of TARVs (j) or non-additive TARVs (k) across any diplotype stratified by independent vs same CS and complex trait vs eQTL variants. Error bars represent 95% CIs. l. Distribution of amplifying vs dampening interacting pairs across eQTLs and complex traits stratified by independent vs same CS and complex trait vs eQTL variants. m. Example of an amplifying non-additive variant pair, rs9294987 and rs9294988 (shaded purple), which are associated with systolic blood pressure. The closest gene is THBS2 and the variant pair jointly improve a Jun binding motif. Additive prediction is the sum of the re-coded allelic effects from variant 1 (TG vs CG) and variant 2 (TG vs TC). Double variant is the observed difference between TG and CC (FDR = 2.0 × 10⁻⁴⁰).

Extended Data Fig. 6 Additional saturation mutagenesis examples.

a,c,e. MPRA results of transcriptional activity for the reference (darker shade) and alternative (lighter shade) allele(s) in either K562 (blues) or HepG2 (purples) for TARVs rs11571842 (a), rs536864739 (c), and rs7953706 (e). Error bars indicate SEs. b,d,f. Nucleotide contribution scores across the elements containing TARVs from a,c, and e are highlighted by a dark yellow or green bar. Activity measurements for all positions when tested on the reference (top) or alternative (bottom) allele are depicted as lollipops indicating the change from baseline activity (Δ log₂ activity). Activity blocks (ABs) are labeled with a gray bar and matching TF motifs are highlighted with a black bar. Shaded boxes overlap allele(s) of interest, with a callout of the saturation mutagenesis constructed motif (MPRA) and catalogued TF motif PWM. g,h. Pearson correlation between the average allelic effect and mammalian constraint (PhyloP) for two separate loci. Positive values indicate activating sequences whereas negative values indicate repressive sequences.

Extended Data Fig. 7 Transcriptome analysis of rs191148279 base-edited K562 cells.

a. Volcano plot showing differential expression between base-edited (n = 9) and reference (n = 11) clones. Points coloured based on FDR-adjusted p-value (Wald test with Benjamini-Hochberg correction). Red indicates significant differential expression (FDR ≤ 0.01). C2orf88 is located in cis on chromosome 2 with the edited allele, while SPRED1 is located in trans on chromosome 15. b. Bar plot of log₂(fold-change) differences in gene expression between base-edited and reference clones for genes within +/− 1 Mb of rs191148279. Bars coloured based on FDR-adjusted p-value (Wald test with Benjamini-Hochberg correction). Red indicates significant differential expression (FDR ≤ 0.01). Error bars represent SE. Of note, the iron-transporter SLC40A1 (Ferroportin-1), located 640 kb from rs191148279, has an uncorrected p-value of 0.004 but is not significant after FDR correction (FDR = 0.33). c. Gene expression levels, measured by RT-qPCR (n = 2 for each clone from a), of C2orf88 and HIBCH relative to a control gene, TBP, in wild-type (blue) and base-edited (red) K562 cells. Base-edited cells carry the rs191148279 A allele along with a silent G > A substitution outside the GATA motif. P-values were calculated using a two-sided Mann-Whitney U test.

Extended Data Fig. 8 Saturation mutagenesis recapitulates allelic interactions.

a. Scatter plot between allelic pairs’ interaction effects measured in the original experiment and in saturation mutagenesis. Correlation is given using Spearman’s ρ. Size of points corresponds to the inverse variance of the interaction effect. b. MPRA results of transcriptional activity for the reference (darker shade) and alternative (lighter shade) allele(s) in K562 for TARVs rs1936950/rs1936951. Error bars indicate SEs. c. Nucleotide contribution scores across the 200 bp elements containing TARVs from b are highlighted by a dark yellow or green bar. Activity measurements for all positions when tested on the haplotype combinations in order of activity levels are depicted as lollipops indicating the change from baseline activity (Δ log2 activity). Activity blocks (ABs) are labeled with a gray bar and matching TF motifs are highlighted with a black bar. Shaded boxes overlap allele(s) of interest, with a callout of the saturation mutagenesis constructed motif (MPRA) and catalogued TF motif PWM. d,e. MPRA and saturation mutagenesis results of transcriptional activity for two interacting TARVs rs35081008 and rs34003091, similar to b and c except for results in both K562 (blues) and HepG2 (purples). f. Scatter plot of variant pair interactions between the centered TARV and all other substitutions made by saturation mutagenesis. Plotted is the proportion of non-additive substitutions (y-axis) that are amplifying (positive) or dampening (negative) with respect to the distance from the TARV (x-axis). Interactions are separated by whether the adjacent substitution resides in an AB/saturation mutagenesis matched motif or not.

Supplementary information

Supplementary Information

This file contains Supplementary Figs. 1–9, Supplementary Notes 1 and 2, and legends for Supplementary Tables 1–32.

Reporting Summary

Supplementary Table 1

Summary of fine-mapped and control variants.

Supplementary Table 2

Variant oligo sequences.

Supplementary Tables

This file contains Supplementary Tables 3, 5, 6, 8–15, 17, 19–23 and 25–31.

Supplementary Table 4

MPRA result summary.

Supplementary Table 7

Cell-type-specific activity and allelic effect with empirical Bayes shrinkage.

Supplementary Table 16

CRE-gene links for autosomal TARVs.

Supplementary Table 18

Oligo sequences for variant pairs

Supplementary Table 24

Summary of saturation-mutagenesis results

Supplementary Table 32

External data sources used in this article.

Peer Review File

Rights and permissions

Open Access This article is licensed under a Creative Commons Attribution 4.0 International License, which permits use, sharing, adaptation, distribution and reproduction in any medium or format, as long as you give appropriate credit to the original author(s) and the source, provide a link to the Creative Commons licence, and indicate if changes were made. The images or other third party material in this article are included in the article’s Creative Commons licence, unless indicated otherwise in a credit line to the material. If material is not included in the article’s Creative Commons licence and your intended use is not permitted by statutory regulation or exceeds the permitted use, you will need to obtain permission directly from the copyright holder. To view a copy of this licence, visit http://creativecommons.org/licenses/by/4.0/.

Reprints and permissions

About this article

Cite this article

Siraj, L., Castro, R.I., Dewey, H.B. et al. Functional dissection of complex trait variants at single-nucleotide resolution. Nature (2026). https://doi.org/10.1038/s41586-026-10121-6

Download citation

Received: 31 July 2024
Accepted: 09 January 2026
Published: 25 February 2026
Version of record: 25 February 2026
DOI: https://doi.org/10.1038/s41586-026-10121-6