Mitigating inconsistencies in GWAS follow-up analyses with LocusCompare2

Abstract

Gene implication methods (GIMs) are essential tools for identifying potential causal genes from GWAS signals, but considerable inter-method variability complicates follow-up studies. We present the LocusCompare2 platform to incorporate six popular GIMs and hundreds of QTL datasets, enabling validation across multiple GIMs and window settings to improve accuracy and reproducibility.

Genome-wide association studies (GWAS) have identified tens of thousands of trait- and disease-associated loci, though the functional significance of most remains elusive because approximately 90% are located within non-coding regions. To translate GWAS risk loci to target genes, large-scale efforts, such as GTEx¹ and eQTL Catalogue², have substantially expanded variants associated with molecular traits and elucidated the regulatory effects of GWAS variants across multiple bio-contexts. To jointly analyze GWAS and molecular quantitative trait loci (QTL), gene implication methods (GIMs) provide statistically principled ways to link GWAS hits to putative causal genes³. Various types of GIMs have been developed since 2010, broadly grouped into three categories, variant-level colocalization, transcriptome-wide association studies (TWAS), and Mendelian Randomization (MR; Box 1). Our survey of 1,765 Nature Genetics articles published between 2014 and 2023 identified 19 GIMs cited in 191 studies (Supplementary Fig. 1a; Table S1). Although the popularity of GIMs increased steadily (Supplementary Fig. 1b), the majority of studies (110 out of 191; 57.59%) used only a single GIM. Studies have shown TWAS, MR, and colocalization produce inconsistent results^4,5; the underlying reasons are multifactorial, including genetic heterogeneity when a GWAS locus harbors multiple independent signals. When genetic variants within a locus exert opposing effects on molecular traits, TWAS and MR signals might be weak despite potentially strong colocalization signals⁵. Additionally, linkage disequilibrium (LD) hitchhiking, where causal variants for molecular and disease traits are distinct but correlated through LD, violates the independence assumption inherent to MR and often leads to strong MR. LD between regulatory variants of nearby genes can also cause false-positive TWAS signals despite potentially weak colocalization signals⁵. Mismatched LD structure and underpowered QTL datasets can further weaken colocalization signals⁶. In a real-world study, 2 commonly used GIMs, eCAVIAR⁶ and SMR⁷, identified largely distinct causal genes⁸, each missing one of the 2 causal genes later validated by mouse models^9,10. Here, using real and simulated data, we demonstrate widespread inconsistencies across frequently used GIMs and illustrate their underlying causes. We further propose practical strategies to improve the precision and recall of causal gene prioritization and present LocusCompare2 as a web platform to support high-confidence causal gene identification.

Box 1. Characteristics of the 3 main categories of GIMs.

Variant-level colocalization methods

Most often, these methods assume that if a set of causal variants are shared between a GWAS and a molecular QTL signal, the GWAS effects are likely mediated through the QTL. Posterior probabilities are computed to provide evidence that shared variants affect both molecular and complex traits, with or without user-defined prior probabilities. For example, in COLOC, p₁ and p₂ are the prior probabilities that a SNP is associated with either trait 1 or 2, and p₁₂ stands for the prior probability that a variant is associated with both traits. Based on this, COLOC calculates posterior probabilities for five possible configurations: H₀, which indicates no association with either trait; H₁ or H₂, which indicate association with only one of the two traits; H₃, which suggests association with both traits but with two independent SNPs; and H₄, which suggests association with both traits with a shared causal variant.

Transcriptome-wide association studies (TWAS)

These methods construct hypothesis tests at the gene level, by examining the association between the complex trait and genetically predicted gene expression. TWAS methods first train a regression model to predict gene expression based on a reference molecular QTL dataset, accounting for linkage disequilibrium (LD) among SNPs and capturing the heterogeneous genetic effect sizes and directions within a gene. TWAS methods then perform association testing between predicted expression values derived from GWAS variants and complex traits using either individual-level or summary-level data.

Mendelian randomization (MR)

MR methods use molecular QTL SNPs, often the most statistically significant one, as instruments to assess whether the exposure (e.g., gene expression) has a causal or pleiotropic effect on the outcome (e.g., complex trait phenotypes). The findings derived from MR analysis may indicate either causality or pleiotropy. Early MR methods, such as Summary data-based Mendelian Randomization (SMR), use a single most significant SNP as the instrument variable. Later developments, including generalized SMR, use multiple independent SNPs as instrument variables to improve robustness. It is worth noting that SMR incorporates the heterogeneity in dependent instruments (HEIDI) test to distinguish pleiotropy from linkage.

Box 1 Fig.

Characteristics of the 3 main GIM categories.

Factors underlying inconsistencies across GIMs

To demonstrate the widespread inconsistency across GIMs, we selected 5 methods highly cited in Nature Genetics (Table 1, Supplementary Fig. 1c), as well as fastENLOC¹¹ because of its adoption by GTEx¹. We calculated pairwise Spearman correlation coefficients (ρ) across GIM results using GWAS on 36 human blood cell traits¹² (Table S2) and GTEx whole blood eQTL¹ (Fig. 1a). The analysis revealed moderate correlations among the colocalization methods (COLOC¹³, fastENLOC¹¹, and eCAVIAR⁶, ρ = 0.51-0.75), while TWAS and MR methods were strongly correlated (FUSION¹⁴, PrediXcan¹⁵, and SMR⁷, ρ = 0.80-0.82). In contrast, correlations between the remaining pairs were weak (ρ = 0.35– 0.49).

Table 1.

Brief overview of the 6 GIMs

Category	GIM	Description
Colocalization	COLOC13	Calculating colocalization posterior probability using Bayesian framework
TWAS	PrediXcan15	Identify trait-associated genes by leveraging genetically predicted expression and testing its association with the trait
MR	SMR7	Testing the causative or pleiotropic effect of expression on complex trait using a genetic variant
TWAS	FUSION14	Identifying genes whose cis-regulated expression is associated to complex traits accounting for LD
Colocalization	eCAVIAR6	Colocalization posterior probability is treated as the product of the posterior probabilities that a variant is causal in both the GWAS and the QTL study
Colocalization	fastENLOC11	Calculating colocalization probabilities enables the estimation of colocalization priors directly from the provided data

Fig. 1. Inconsistency across GIM results between 36 human blood cell traits GWAS and GTEx whole-blood eQTL.

(a) Spearman correlation coefficient (ρ) between each pair of GIMs. H₄, CLPP (colocalization posterior probability, the metric of eCAVIAR), and GRCP (gene-level variant colocalization probably, the metric for fastENLOC) are used to represent the results of COLOC, eCAVIAR, fastENLOC, and −log₁₀(P-value) is used to represent the results of FUSION, PrediXcan, and SMR. (b) Disjoint causal variant sets can weaken the colocalization signals. The bottom panel of the LocusZoom plot represents one cis-eQTL gene UBASH3B. The middle panel shows LocusZoom plot of this gene for GWAS trait Monocyte count; here, the colocalization signal in this trait is lost, whereas a significant colocalization signal is detected in the top panel, showing the LocusZoom plot of the same gene for GWAS trait Lymphocyte count. H₄, posterior probabilities of hypothesis 4 (suggesting association with both traits with a shared causal variant); H₃, hypothesis 3 (suggesting association with both traits but with two independent SNPs) for COLOC (see also Box 1). (c) A higher correlation of effect sizes between GWAS and eQTL can contribute to a significant TWAS signal. The bottom panel of the LocusZoom plot represents a cis-eQTL for gene LINCO1184. The middle panel shows LocusZoom plot of this gene in GWAS traits Red cell distribution width, with a correlation of effect sizes of 0.483. The top panel is a LocusZoom plot of the corresponding gene in GWAS trait Lymphocyte count, with a correlation of effect sizes of 0.087, and has a weaker TWAS signal compared to the middle panel. The purple diamond indicates the lead SNP of eQTLs, and r² indicates the LD between SNPs and the lead SNP.

Statistical assumptions

The first factor underlying inconsistencies across GIMs we address is different statistical assumptions. Colocalization methods search for causal variant sharing (Box 1), disjoint causal variant sets, i.e. QTL and GWAS signals that incorporate different causal variants and exhibit low LD, will weaken the colocalization signal. For instance, at the UBASH3B locus, eQTL strongly colocalized with Lymphocyte Count (LC) GWAS (H₄ = 0.98) (Fig. 1b, Table S3), but did not colocalize with Monocyte Count (MC) GWAS (H₄ = 0.0, H₃ = 1.00), where the inferred GWAS and eQTL causal variants were disjoint and showed low LD. Broadly, across all gene-trait pairs, lead variants from GWAS and eQTL showed significantly stronger LD in cases with strong colocalization signals, compared to those without evidence of colocalization (Supplementary Methods, Supplementary Fig. 2a).

TWAS methods search for sharing of genetic effect, and therefore both LD structure across variants and the effect size correlation between GWAS and QTL influence TWAS signals⁵. MR infers causal or pleiotropic effects of gene expression on complex traits using genetic variants (Box 1). Given the high concordance between TWAS and MR results (Fig. 1a), we analyzed them jointly to explore factors contributing to this consistency. LD contamination or strong co-regulation among neighboring genes can lead to false-positive TWAS and MR signals, even when GWAS and eQTL do not share causal variants. For example, within the region chr2: 218,350,000-218,830,000, joint analysis of Mean Corpuscular Hemoglobin Concentration (MCHC) GWAS and GTEx Whole Blood eQTL identified gene VIL1, located at the start of the central LD block, as the only gene with strong colocalization, TWAS, and MR signals (Supplementary Fig. 3a). Most downstream genes within the same LD block, including CNOT9, PLCD4, BCS1L, TTLL4, and CYP27A1, displayed significant TWAS and MR signals, despite lacking colocalization signals (Table S3). Manual inspection of their LocusZoom and LocusCompare plots revealed that these genes were unlikely to be causal genes, suggesting their TWAS signals are likely driven by LD hitchhiking (Supplementary Fig. 3b–h). Additionally, higher effect size correlation between shared GWAS and eQTL variants generally results in stronger TWAS and MR signals. Across all analyzed gene-trait pairs, those with TWAS and MR signals exhibited a significantly higher effect size correlation (Supplementary Fig. 2b). For example, at the LINC01184 locus, both LC and Red Cell Distribution Width (RCDW) GWAS traits showed strong colocalization signals with GTEx whole blood eQTLs. However, the correlation between effect sizes was lower for LC (Pearson’s r² = 0.087) than RCDW (r² = 0.483), leading to weaker TWAS and MR signals for LC (Fig. 1c, and Table S4). Taken together, variations in causal variant sharing, effect size correlations, and LD structures contribute to GIM inconsistencies and conflicting biological interpretations.

Input data and parameter settings

Differences in input data and hyperparameter settings also affect GIM results. In particular, variant-level colocalization results are sensitive to genomic window definitions. To illustrate this, we compared two commonly adopted colocalization window settings: the QTL-fixed-window, including variants within ±1Mb of gene transcription start site (TSS), consistent with the typical cis-eQTL definition, and the GWAS-fixed-window, including variants within ±500kb of the GWAS lead variant (Fig. 2a, Supplementary Methods). Given some cis-eQTL genes may overlap with multiple independent GWAS loci, we assigned each gene the highest colocalization posterior probability (See Box 1) within the QTL-fixed-window. The comparison revealed that many genes exhibited high colocalization probabilities exclusively under GWAS-fixed-window setting, while some genes did so only under QTL-fixed-window setting (Fig. 2b–d), suggesting divergent results under different window size settings.

Fig. 2. Illustration of the effect of fixed-window size settings for variant-level colocalization GIMs.

(a) Schematic diagram of different genomic window size settings as input for colocalization GIMs. GWAS-fixed-window loci mean the test region including all variants within ±500kb of the GWAS lead variant, while QTL-fixed-window loci mean all variants within ±1Mb of the transcription start site (TSS) of each gene. GWAS LD-based locus encompasses the lead SNP, variants with LD > 0.1 with the lead SNP, and an additional 50kb variant on either side. The combined LD-based locus extends this region by incorporating the union of GWAS and QTL LD-based loci, given that matched variants exist between GWAS and QTL loci. (b-d) Comparisons of the result of COLOC (b), fastENLOC (c), and eCAVIAR (d). The GTEx whole-blood eQTL and Mean platelet volume fixed-window GWAS levels reveal moderate Spearman’s correlation coefficients (ρ) for each of these methods, with data points shown in blue highlighting discrepancies between COLOC, fastENLOC, and eCAVIAR findings, respectively. The yellow dashed line indicates the threshold for the GIM on the x-axis and the dotted red line indicates the threshold for the GIM on the y-axis.

Apart from fixed-window settings, we evaluated two LD-based window settings: a GWAS LD-based window, defined by variants with LD r² > 0.1 to the lead SNP and ±50kb flanking regions, and a combined LD-based window, which further extends GWAS LD-based locus to incorporate a QTL LD-based window with at least 5 overlapping variants (Fig. 2a, Supplementary Methods). GWAS loci lacking a matching QTL LD-based window will be excluded from analyses. Across 36 cellular GWAS traits and GTEx whole-blood eQTL, both LD-based strategies improved concordance between colocalization results and those derived from TWAS and MR (ρ = 0.35– 0.54), primarily due to the narrower LD-informed loci that better reflect shared LD structure and filter out GWAS loci that lack overlapping QTLs.

Despite improving the consistency across GIMs, GWAS LD-based window settings tend to increase false-positives due to their localized testing window. For example, at gene CD300LD locus, the GWAS LD-based approach captured a local eQTL signal with COLOC H₄ = 0.784. However, LocusZoom plot revealed that this region is located outside the main eQTL signal (Fig. 3a). In contrast, the combined LD-based window included the adjacent eQTL loci and correctly identified two distinct causal variants for two traits (H₄ = 5 x 10⁻⁶, H₃ = 0.998). Nonetheless, even combined LD-based windows may still yield false-positives, particularly when secondary QTL signals are distant from the primary signal (Fig. 3b), necessitating wider windows for accurate signal resolution.

Fig. 3. Impact of window size selection on GIM inconsistency.

(a) The LocusZoom plot of gene CD300LD (GWAS: Plateletcrit) and GTEx Whole blood eQTL. The blue dashed lines indicate the colocalization test window under the GWAS LD-based approach (H₄ = 0.784), whereas the light brown dashed lines indicate the colocalization test window under the combined LD-based approach (H₄ = 5 x 10⁻⁶). (b) The LocusZoom plot of gene DVL3 (GWAS: Platelet count) and GTEx whole-blood eQTL. The combined LD-based strategy does not eliminate all false-positives when mapping the GWAS locus to a marginal eQTL locus that is distant from the main eQTL locus. Both LD-based approaches yield H₄ > 0.75. (c) GWAS fixed-window loci can introduce false-negatives because a wider window size may include multiple GWAS signals. The blue dashed lines indicate the colocalization test window under GWAS LD-based approach (H₄ = 1.000), while the green dashed lines represent the colocalization test window under the fixed-window approach (H₄ = 0.004). (d) When the confounding GWAS signal obtained under the GWAS-fixed-window approach, which is not aligned with the eQTL signal, is weaker than the main GWAS signal, colocalization GIMs can correctly identify the shared signal (H₄ = 1.000). The purple diamond indicates the lead SNP of eQTLs, and r² indicates the LD between SNPs and lead SNP.

However, wider window sizes are not always preferable. A GWAS-fixed-window approach generally uses larger windows than LD-based approaches, potentially encompassing multiple GWAS signals and increasing the risk of false-negatives. For example, at gene SLC12A7 locus, two independent Mean Corpuscular Volume (MCV) GWAS signals exhibiting low LD were included in the GWAS-fixed-window, with only one aligning with an eQTL signal. This led COLOC to report H₃ = 0.996, missing the true colocalization event (Fig. 3c). In contrast, the GWAS LD-based window correctly captured the colocalization event aligned with the eQTL signal (H₄ = 1.000). When the mismatched GWAS signal was weaker than the matched signal, as in the case of Platelet Distribution Width (PDW) GWAS, the GWAS-fixed-window still successfully identified the colocalization event (Fig. 3d). These results suggest that narrower window sizes may introduce false-positives, while wider fixed-window bias towards false-negatives. We recommend performing a first-pass scan using LD-based windows and confirming positive findings with fixed windows to increase confidence in true colocalization events.

Finally, differences in prior probability settings also impact variant-level colocalization results. For example, using the same prior probabilities settings (p₁=1x10⁻⁴, p₂=1x10⁻⁴, and p₁₂=1x10⁻⁵), COLOC and fastENLOC produced highly correlated results (ρ = 0.765). However, when fastENLOC inferred priors automatically while COLOC retained fixed priors, the correlation between fastENLOC and COLOC decreased to ρ = 0.61 and the number of putative causal genes for fastENLOC reduced from 65 to 37 (Supplementary Fig. 4).

Tissue and cell-type context

Inappropriate tissue and cell-type context can further complicate GIM result interpretation. Using Whole-blood eQTLs, 436 likely causal gene-trait associations, defined as genes consistently predicted by 6 GIMs, were identified across 36 cellular GWAS traits. In contrast, applying analyses to GTEx Lung eQTL, a less relevant tissue, the number of likely causal genes reduced to 259, with reduced concordance between TWAS and MR GIMs and colocalization GIMs (ρ = 0.31–0.42). Only 106 genes overlapped between whole blood and lung, underscoring context-specific differences in gene prioritization. Notably, incorporating molQTLs from multiple tissues, rather than only the target tissue, can improve power to detect regulatory effects, especially when the sample size is limited.

The LocusCompare2 platform

To facilitate integration of multiple GIMs and to standardize GIM parameter reporting, we developed the LocusCompare2 online platform (https://www.locuscompare2.com), a significantly expanded version of the original LocusCompare³ web server. LocusCompare2 enables cloud-based execution of the 6 GIMs, accepting GWAS summary statistics and allowing users to select from 280 pre-loaded eQTL datasets across diverse bio-contexts and customize GIM parameters (Supplementary Methods, Table S5). The platform streamlines data processing and accommodates diverse input requirements and formats with minimal user intervention (Fig. 4). Upon task completion, LocusCompare2 allows users to interactively explore and download data to their local environment.

Fig. 4. Workflow of the LocusCompare2 webserver.

LocusCompare2 pre-loads 280 eQTL datasets from GTEx Consortium and eQTL catalogue, as well as genes and variant information in a MySQL database. LocusCompare2 (https://www.locuscompare2.com) accepts GWAS files as input and users can customize parameters (e.g., GIM tool settings, GWAS/eQTL significance thresholds) on the website. LocusCompare2 automates all processes, including format checks, preprocessing, filtering, and GIM analyses, with minimal user intervention. Once task completion, LocusCompare2 allows users to explore results across multiple levels, including GWAS, eQTL, GIMs, and genes. Additionally, it provides interactive visualization such as LocusZoom, LocusCompare, and effect size correlation plots to help confirm findings.

To optimize multiple GIMs usage with LocusCompare2, researchers could use one GIM for discovery and another for validation. Alternatively, intersecting results from multiple GIMs, particularly for highly polygenic traits, will enhance robustness by increasing precision and reduce the validation space. Our simulations show that intersecting increasing number of GIMs boosted precision from 0.924 to 1, with the largest precision gain observed from 1 to 2 GIMs (0.924 to 0.989), while recall declined from 0.380 to 0.226 (Supplementary Fig. 5a). For cases with small eQTL sample sizes, such as rare or specialized cell types, combining multiple GIMs could better capture putative genes for downstream validation. In our simulations, using the union of 6 GIMs improved the recall from 0.380 to 0.503 compared to a single GIM, but reduced the precision from 0.924 to 0.737 (Supplementary Fig. 5b). We note that simulations do not fully reflect real-world complexity, which highlights the need for experimentally validated gold standards.

Overall, each dataset needs a customized setting for gene-implication analysis and there is currently no one-size-fits-all solution. Researchers should report GIM settings clearly to ensure reproducibility, yet reporting practices vary widely across studies. To that end, we provide an Excel template to help standardize the GIM parameter reporting, which can be included in related manuscripts as a supplementary table to ensure reproducibility (accessible through LocusCompare2 and Table S6).

Concluding remarks

GIMs employ distinct statistical frameworks and assumptions, leading to variations in GWAS causal gene predictions. Different window size selections and parameter settings further contribute to discrepancies across GIMs. To increase the robustness of detection of causal genes, we recommend using a combination of TWAS/MR and colocalization GIMs, and two window sizes to validate the gene implication results. To facilitate this process, we developed LocusCompare2, an integrated platform that streamlines the application of 6 widely used GIMs, allowing users to leverage multiple GIMs under diverse analytical settings. We hope LocusCompare2 will enhance the robustness of causal gene identification and enable precise functional dissection of the genetic architecture underlying complex traits and diseases.

Data availability

GTEx eQTL data can be accessed via https://gtexportal.org/home/datasets. eQTL Catalogue data can be accessed via https://www.ebi.ac.uk/eqtl/Studies/. Human blood cell GWAS summary statistics are available in the GWAS Catalog https://www.ebi.ac.uk/gwas/downloads/summary-statistics. Pre-computed predictive models for GTEx 49 tissues for FUSION are from http://gusevlab.org/projects/fusion/. Pre-computed transcriptome weights for GTEx 49 tissues for PrediXcan are downloaded from https://predictdb.org/. The simulated data and the LocusCompare2 results of 36 human blood cell traits GWAS summary statistics can be accessed via https://zenodo.org/records/15192686.

Code availability

The LocusCompare2 website is on https://www.locuscompare2.com. An open-source software at locuscompare2-standalone. Processing scripts for LocusCompare2 are at LocusCompare2_visualization.