MungeSumstats: a Bioconductor package for the standardization and quality control of many GWAS summary statistics

1Genome‐wide association studies (GWAS) summary statistics are used to distribute the most important outputs of GWASs in a manner which does not require the transfer of individual-level personally identifiable information from participants. Summary statistics from past studies tend to become more valuable over time as it becomes possible to meta-analyze and integrate them with new annotation information through approaches such as Linkage Disequilibrium Score Regression (LDSC), Generalized Gene-Set Analysis of GWAS Data, MAGMA and multi-phenotype investigations. Summary statistics are also commonly integrated for use in the meta-analysis of GWAS. However, these tools and this integration require a standardized data format which was historically lacking from the field. The diversity of data formats in summary statistics has been a result of the phenotypes in question, for example disease-control or quantitative trait, the software used to perform the analysis, such as PLINK and GCTA or just the preference of the consortium in question.

2There have been movements to standardize the summary statistic file format such as the NHGRI-EBI GWAS Catalogue standardized format and the SMR Tool binary format. More recently, the variant call format to store GWAS summary statistics (GWAS-VCF) has been developed which has manually converted over 10 000 GWAS to this format. While GWAS-VCF offers a standardized format that future GWAS consortium may adopt, there are still a multitude of past, publicly available GWAS which have not been standardized. For instance, although their summary statistics are publicly available, the GWAS for Cerebral small vessel disease is not yet available in VCF format via IEU GWAS. Furthermore, as VCF is not yet the standard for sharing files between geneticists, unpublished GWAS shared internally within genetics consortia or provided by personal genetics companies are still found in a variety of summary statistic formats. As such, there is a need for tools to move between the various formats in which summary statistics are stored.

3The standardization of GWAS summary statistics also requires quality control to ensure cohesive integration. For example, checking if the non-effect allele from the summary statistics matches the reference sequence from a reference genome to ensure consistent directionality of allelic effects across GWAS. In addition, downstream analysis tools often require a degree of quality control which, in the case of meta-analysis, must be applied across all GWAS. One such example is the removal of all non-biallelic SNPs is a common requirement of all downstream analysis.

4To address these issues, we introduce MungeSumstats a Bioconductor R package for the rapid standardization and quality control of many GWAS summary statistics. MungeSumstats can handle the most common summary statistic formats as well as GWAS-VCFs to enable the integrative meta-analysis of diverse GWAS. MungeSumstats also offers a comprehensive and tuneable quality control protocol with defaults for common, best-practice approaches. MungeSumstats capitalizes on R’s familiar interface, is readily accessible through Bioconductor and utilizes an intuitive approach, running with a single line of input code.

5To demonstrate the diversity in summary statistics across GWAS, we analyzed a public repository of over 200 publicly available GWAS. From this, the most common summary statistics were derived (see Fig. 1 for the 12 most common file header formats).

6A total of 327 summary statistic files were derived from the analysis which corresponded to 127 unique formats. Thus, on average, every 2.5 summary statistic files had a unique format, showing the clear disparity across GWAS. The 12 most common formats, shown in Figure 1, accounted for approximately 47% all summary statistics. MungeSumstats has been tested on these 12 most common formats and is able to standardize their summary statistics.

7MungeSumstats was implemented using the R programming language (v 4.0) and Bioconductor S4 data infrastructure (v 3.13) enabling the full analysis of summary statistics within the R environment. The package removes the need for external software to perform the standardization and quality control steps.

8MungeSumstats’ implementation ensures both memory and speed efficiency through the use of R data.table (v.1.14.0), which can take advantage of multi-core parallelization. Moreover, MungeSumstats benefits from Bioconductor’s infrastructure for efficient representation of full genomes and their SNPs, using BSgenome (v 1.59.2) SNP reference genomes. Either Ensembl’s GRCh37 or GRCh38 are queried dependent on the build for the particular GWAS. Numerous of MungeSumstats’ quality control steps for summary statistics require the use of a reference genome. For example, an allele flipping test is run (see Table 1) to ensure consistent directionality of allelic effect and frequency variables. The effect or alternative allele is always assumed to be the second allele (A2), in line with the approach for GWAS-VCF. Moreover, MungeSumstats can impute any missing, essential information like SNP ID, base-pair position and effect/non-effect allele.

9Using these two infrastructures, MungeSumstats conducts more than 30 checks on the inputted summary statistics file (see Table 1 for a description of their use). MungeSumstats is also written to ensure the ease of addition of further checks so if users have summary statistics which can’t currently be handled in MungeSumstats, these can be incorporated easily in future releases. Finally, MungeSumstats returns a reformatted, tabular summary statistics file, a VCF or an R native data object (data.table, VRanges or GRanges) with standardized columns for the information necessary for downstream analysis.

10The quality control and standardization checks conducted. Most checks are optional and can be set by the user. Here, CHR is chromosome, BP is Base-pair position, A1 is the non-effect allele, A2 is the effect allele, N is the sample size, INFO is imputation information score, FRQ is the minor allele frequency (MAF) of the SNP, SNP ID is the single nucleotide polymorphism reference ID, P is the unadjusted P-value, Z is z-score, OR is odds ratio, LOG_ODDS is the log odds ratio, BETA is the effect size estimate relative to the alternative allele and SIGNED_SUMSTAT is the directional effect size estimate for the summary statistics.

11Once MungeSumstats is installed, usage involves a single line of code or one function call (format_sumstats) with the path to the summary statistics file of interest. Then, the path to the reformatted, standardized summary statistic file is returned. MungeSumstats also offers adjustable parameters to manage the quality control steps. These include options to adjust the imputation information score (INFO) cut-off threshold, the number of samples (N) outliers cut-off threshold and whether to remove mitochondrial SNPs or SNPs on the X or Y chromosome (see Table 1). Quality control steps which use a reference genome can also be adjusted such as whether to filter SNPs based on their RS ID’s presence on the reference genome, whether to check for allele flipping and whether to remove multi-allelic or strand-ambiguous SNPs. These parameters ensure MungeSumstats can be adjusted to the user’s analysis pipelines.

12Here, we presented MungeSumstats, a Bioconductor package for the standardization and quality control of GWAS summary statistics. This package enables integration of summary statistics of vastly different formats, simplifying meta-analysis and summary statistics use in other secondary research applications. The package provides an efficient, user-friendly R-native approach, returning a standardized, tabular format file, VCF or R native data object. This ensures that the summary statistics are accessible to the average user. Moreover, MungeSumstats is written to permit future development of additional standardization steps if users encounter issues with their specific GWAS.