(C) PLOS One
This story was originally published by PLOS One and is unaltered.
 . . . . . . . . . .



Variant-specific priors clarify colocalisation analysis [1]

['Jeffrey M. Pullin', 'Mrc Biostatistics Unit', 'University Of Cambridge', 'Cambridge', 'United Kingdom', 'Chris Wallace', 'Cambridge Institute Of Therapeutic Immunology', 'Infectious Disease']

Date: 2025-05 

Linking GWAS variants to their causal gene and context remains an ongoing challenge. A widely used method for performing this analysis is the coloc package for statistical colocalisation analysis, which can be used to link GWAS and eQTL associations. Currently, coloc assumes that all variants in a region are equally likely to be causal, despite the success of fine-mapping methods that use additional information to adjust their prior probabilities. In this paper we propose and implement an approach for specifying variant-specific prior probabilities in the coloc method. We describe and compare six source of information for specifying prior probabilities: non-coding constraint, enhancer-gene link scores, the output of the PolyFun method and three estimates of eQTL–TSS distance densities. Using simulations and analysis of ground-truth pQTL–eQTL colocalisations we show that variant-specific priors, particularly the eQTL–TSS distance density priors, can improve colocalisation performance. Furthermore, across GWAS–eQTL colocalisations variant-specific priors changed colocalisation significance in up to 14.1% of colocalisations, at some loci revealing the likely causal gene.

Evaluating whether two traits, such as disease risk and gene expression, are affected by the same genetic variants is crucial for understanding the molecular mechanisms through which genetic variants act. A widely used method for determining whether there is a shared cause—termed ‘colocalisation’—is ‘coloc’, which updates prior knowledge about the chance of two traits sharing a causal variant with observed genetic association data in a Bayesian statistical framework. However, currently coloc assumes that all variants are equally likely a priori—before looking at the data—to affect each trait. This assumption is made despite the existence of considerable information about the different types of variants and large amounts of data linking variants and traits. Here, we describe an approach for incorporating this variant-specific prior information into coloc. We describe several sources of prior information about variants and assess what impact they have on the performance of coloc. We find that some sources of information improve the performance of coloc and, when analysing disease risk and gene expression, can clarify which genes the variants that alter disease risk act through.

Data Availability: The eQTLGen data was downloaded from https://molgenis26.gcc .rug.nl/downloads/eqtlgen/cis-eqtl/2019-12-11-cis-eQTLsFDR0.05-ProbeLevel-CohortInfoRemoved-BonferroniAdded.txt.gz. The OneK1K data was downloaded from https://onek1k.s3.ap-southeast-2.amazonaws.com/esnp/esnp_table.tsv.gz . The eQTL catalogue v6 data (including GTEx v8) was downloaded from the eQTL Catalogue FTP site using the links specified in https://raw.githubusercontent.com/eQTL-Catalogue/eQTL-Catalogue-resources/master/tabix/tabix_ftp_paths.tsv . The FinnGen GWAS summary statistics were downloaded following the FinnGen consortium’s instructions https://www.finngen.fi/en/access_results . The UK Biobank summary statistics were downloaded from https://pheweb.org/ UKB-SAIGE/: https://pheweb.org/UKB-SAIGE/ download/250.2, https://pheweb.org/ UKB-SAIGE/download/401, https://pheweb .org/UKB-SAIGE/download/244. Baseline annotations used in the calculation of trait-specific PolyFun prior weights were downloaded from https://broad-alkesgroup- ukbb-ld.s3.amazonaws.com/UKBB_LD/baselineLF_v2.2.UKB.polyfun.tar.gz. The Open Targets Genetics data was downloaded from the Open Targets Genetics FTP site https://ftp.ebi.ac.uk/pub/databases/opentargets/genetics/ . The 1000 Genomes phase 3 haplotype data was downloaded from https://mathgen.stats.ox.ac.uk/impute/1000GP_Phase3.html .

Currently, coloc assumes that the prior probabilities and p 12 are uniform across variants in the region of interest. In this paper, inspired by the success of the use of additional variant-specific information in fine-mapping, and with the ultimate goal of narrowing the colocalisation gap, we propose an implementation of variant-specific prior probabilities in coloc. We describe and compare the effectiveness of various sources of information for specifying these prior probabilities using simulated data and pQTL–eQTL colocalisations. Finally, we demonstrate the effect of using variant-specific prior probabilities on eQTL–GWAS colocalisation analyses.

Different statistical methods have been developed to assess colocalisation, including SharePro, PWCoCo and coloc [ 7 , 17 , 18 ]. (See [ 19 ] for a recent benchmark of Bayesian colocalisation methods). Here, we focus on the widely used coloc method, implemented in the coloc R package, which uses a Bayesian approach to calculate the probability of hypotheses corresponding to different scenarios of shared and distinct causation of associations [ 7 ]. Specifically, coloc computes the posterior probability of five hypotheses:

Computational methods across statistical genetics have been designed to incorporate this additional variant-specific information to improve performance. For example, in statistical fine-mapping, methods such as PAINTOR [ 12 ] incorporate functional annotations in their estimation procedures, while the PolyFun method estimates the prior probability of causality for variants, for use in fine-mapping, based on functional data [ 13 ]. A key motivation for using additional data in fine-mapping is that it can ‘break ties’ between variants in strong linkage disequilibrium, which are otherwise impossible to distinguish. Incorporating functional annotation can greatly concentrates the posterior probabilities on a smaller number of variants. In an analysis of 49 UK Biobank traits, use of PolyFun prior probabilities with the SuSiE fine-mapping method [ 14 ], lead to 32% more variant–trait pairs with posterior causal probability >0.95 [ 13 ]. Beyond fine-mapping, locus-to-gene models, such as OpenTarget’s L2G model, [ 15 ] incorporate a range of genetic information, including distance and predicted variant effects. Incorporating distance information was also recently shown to greatly improve peak-to-gene link prediction [ 16 ]. However, despite these successes, currently no methods exist for incorporating additional variant-specific information into statistical colocalisation analyses.

Today, large amounts of functional annotations and other additional information about genetic variants are available, with initiatives such as The Encyclopedia of DNA Elements (ENCODE) [ 9 ] providing genome-wide catalogues of functional annotations. It is well established that GWAS associations are enriched among annotations such as DNase I Hypersensitive sites, implying that functional annotations can provide information about whether variants are causal [ 4 ]. More recently, studies have created genome-wide enhancer-gene maps using the activity-by-contact (ABC) score with the intention of prioritising variants that lie in enhancers [ 10 ]. Estimates of genomic constraint in non-coding regions have also been generated from large whole genome sequencing datasets, with the idea that highly constrained regions are more likely to have functional consequences [ 11 ]. Finally, there are large amounts of publicly available eQTL data that can be used to estimate the most likely positions of eQTLs relative to genes.

Genome-wide association studies (GWAS) have uncovered hundreds of thousands of disease associated variants [ 1 , 2 ]. However, more than 90% of these variants lie in non-coding regions of the genome, making elucidating their functional consequence challenging [ 3 , 4 ]. Instead, GWAS associations generally lie in open chromatin regions and are enriched in gene regulatory regions, suggesting they act by modulating gene expression [ 4 – 6 ]. Therefore, a widely used approach to interrogate variant function is assessing whether GWAS and variants associated with gene expression—expression quantitative trait loci (eQTLs)—are caused by the same variants using statistical colocalisation methods [ 7 ]. However, most GWAS hits do not have evidence of colocalisation, representing a large ‘colocalisation gap’ [ 8 ].

Results

Incorporating variant-specific prior probabilities in coloc We implemented the ability to specify variant-specific priors in the widely used coloc package for statistical colocalisation analysis (Fig 1B). The coloc.abf(), coloc.susie() and coloc.bf_bf() functions in coloc have been updated to include two new arguments: prior_weights1 and prior_weights2, which specify non-negative weights for the probability of a variant being causal for trait 1 and trait 2 respectively. In addition, for consistency we added the ability to use priors weights in the finemap.abf() function. Users of coloc can supply any weights they think are informative for the causality of variants a priori, including the methods of specifying weights we assess in this manuscript, information about variant level properties in trait relevant contexts, such as single-cell enhancer-gene maps [22], and priors that have been empirically estimated for their trait of interest, for example using PolyFun. As coloc does not estimate variant-specific priors, using variant-specific priors is not meaningfully slower than standard coloc (S1 Fig). For mathematical details of the implementation of variant-specific priors see Methods. PPT PowerPoint slide

PNG larger image

TIFF original image Download: Fig 1. Example prior probabilities of causality and conceptual figure. (a) Prior probabilities from the six sources of prior probabilities: eQTLGen estimated eQTL-TSS distance density, OneK1K round 1 estimated eQTL-TSS distance density, OneK1K round 2+ estimated eQTL-TSS distance density, genome-wide enhancer-gene link scores calculated using the ABC score, Gnocchi non-coding constraint score and previously calculated PolyFun prior probabilities. The probabilities are displayed in a +/-500kB region around the canonical TSS, shown by the dashed black line, of the VAV3 genes. (b) A conceptual overview of coloc with variant-specific priors. The coloc method takes summary statistic for two traits and test colocalisation, now with the ability to all specify variant specific priors. The formulas shown are proportional to using either variant-specific or uniform priors. https://doi.org/10.1371/journal.pgen.1011697.g001

Specification of variant-specific prior probabilities Various methods are available for specifying variant-specific weights. In this paper we assessed four types of sources for specifying variant-specific weights. Three sources were drawn from public datasets: the output of the PolyFun method [13], ‘Gnocchi’ score, a measure of non-coding genomic constraint [11] and genome-wide activity-by-contact enhancer-gene link scores [10, 23]. For these sources we accessed publicly available data and performed minor processing, mainly to account for SNPs which did not have specific prior values (for further details see Methods). We sought to use empirical eQTL-TSS distance densities as an additional source of prior information. We estimated densities from all significant eQTLs in eQTLGen [24] and both significant round 1 eQTLs and significant round 2-5 conditional eQTLSs OneK1K dataset [25] after finding that the estimated densities were not highly tissues specific (S2 Fig) and that these datasets had more distal eQTLs relative to the TSS compared to other datasets (See Methods for details). Overall, we used 6 sources of variant-specific prior probabilities (Table 1), shown in Fig 1A for a Kb window around the TSS of VAV3. The eQTL-TSS density priors place most mass in a small region around the TSS, with the OneK1K round 2+ estimated prior being less concentrated around the TSS. The ABC score-prior is also centred around the TSS, but is less smooth. PPT PowerPoint slide

PNG larger image

TIFF original image Download: Table 1. Methods for specifying variant-specific prior probabilities. https://doi.org/10.1371/journal.pgen.1011697.t001

Simulation analysis First, as a proof of concept, we performed simulations to validate our implementation and highlight the potential benefits of incorporating variant-specific prior information into colocalisation analysis. We simulated association summary statistics for Kb regions around the TSS of three genes with patterns of high LD. High LD regions were chosen as high LD makes determining colocalisation more challenging. Summary statistics were simulated using simGWAS [26] under both H 3 and H 4 , using the eQTLGen-calculated distance density as the true density of causal variants (See Methods for details). Notably, the simulations recapitulate the tendency of to be close to concentrated close to 0 or 1, for example as in the colocalisation calculated in the OpenTargets Genetics platform. (S4b Fig, see Methods for details). The original version of coloc, which assumes a single causal variant, was run on the simulated data, with and without using the simulation ground-truth distance density as the variant-specific prior. Overall, using variant-specific priors improved the accuracy of colocalisation. Across all three loci, using variant-specific priors generally increased the computed when the data was simulated under H 4 and decreased it when the data was the simulated under H 3 (Fig 2A). This finding demonstrates the ability of variant-specific priors to improve colocalisation accuracy, at least when the specified prior matches the true distribution of causal variants. In addition, we assessed how the effect of the variant-specific priors was effected by both the distance of the causal variant to the TSS and the value of when uniform priors were used. We found that, as intended, variants, close the TSS had the largest increase in while distal variants had smaller increases (Fig 2B). We also observed that simulations with moderate values of using uniform priors had the largest change in when variant-specific priors were used (Fig 2C). This observation is partly due to a ‘ceiling effect’: as , the many values close to 1 cannot be increased by a large amount by the variant-specific priors. The observation also highlights that variant-specific priors may only be able to change whether colocalisation is called at the widely used 0.8 threshold in the minority of cases where is close to 0.8 using uniform priors. PPT PowerPoint slide

PNG larger image

TIFF original image Download: Fig 2. Simulation analysis. (a) values calculated using uniform priors or the simulation ground-truth distribution of causal variants as the variant-specific prior. Simulations are conducted under either coloc hypotheses H 3 or H 4 in 1MB (+/-500kb) regions around the TSS of the listed genes. (b) For H 3 and H 4 simulations, the calculated using variant-specific priors - calculated using uniform priors against the position of the simulated causal variant. c) For H 3 and H 4 simulations, scatter plot of the calculated using variant-specific priors against calculated using uniform priors. https://doi.org/10.1371/journal.pgen.1011697.g002

[END]
---
[1] Url: https://journals.plos.org/plosgenetics/article?id=10.1371/journal.pgen.1011697

Published and (C) by PLOS One
Content appears here under this condition or license: Creative Commons - Attribution BY 4.0.

via Magical.Fish Gopher News Feeds:
gopher://magical.fish/1/feeds/news/plosone/