To download FinnGen summary statistics you will need to fill the online form at this link. You will then receive an email containing the detailed instructions for downloading the data.

Release 13 contains

• GWAS summary association statistics

• Fine-mapping results

When using these results in publications, please remember to:

1) Acknowledge the FinnGen study. You can use the following text:

“We want to acknowledge the participants and investigators of the FinnGen study”

2) Cite our latest publication:

Kurki M.I., et al. . Nature 2023 Jan;613(7944):508-518. doi: 10.1038/s41586-022-05473-8. Epub 2023 Jan 18.

Furthermore, if possible, include "FinnGen" as a keyword for your publication.

If you want to cite this website, use the following citation:

The manifest file with the link to all the downloadable summary stats is available at:

Please use the following description when referring to our project:

The FinnGen study is a large-scale genomics initiative that has analyzed over 500,000 Finnish biobank samples and correlated genetic variation with health data to understand disease mechanisms and predispositions. The project is a collaboration between research organisations and biobanks within Finland and international industry partners.

When using these results in publications, please remember to:

1. Acknowledge the FinnGen study. You can use the following text:

“We want to acknowledge the participants and investigators of the FinnGen study”

1. Cite our latest publication:

Kurki M.I., et al. . Nature 2023 Jan;613(7944):508-518. doi: 10.1038/s41586-022-05473-8. Epub 2023 Jan 18.

Furthermore, if possible, include "FinnGen" as a keyword for your publication.

If you want to cite this website, use the following citation:

FinnGen individuals were genotyped with Illumina and Affymetrix chip arrays (Illumina Inc., San Diego, and Thermo Fisher Scientific, Santa Clara, CA, USA).

Chip genotype data were imputed using the population-specific SISu v4.2 imputation reference panel of 8,554 whole genomes.

Merged imputed genotype data is composed of 141 data sets that include samples from multiple cohorts.

• Total number of individuals: 519,972

• Total number of variants (merged set): 21,331,644

• Reference assembly: GRCh38/hg38

Timeline for releases:

[1] samples used for PheWAS.

The PheWeb portal can be used to browse results from FinnGen's predetermined endpoints (or 'phenotypes') a.k.a. core analysis results. FinnGen PheWeb tutorial is available .

These were analysed for genetic associations, which allows for disproportionate case-control numbers and corrects for relatedness between samples with a sparse genetic relatedness matrix.

The results from each association run are uploaded onto the PheWeb portal, which can be accessed by clicking this link:

The figure below shows the a table of the first few endpoints ('phenotypes') in FinnGen with the highest numbers of GWAS significant loci, along with the summary of case-control analyses and the number of hits.

You can reorder the table by clicking on the appropriate header value (in the figure above, we clicked on GWAS significant loci to order the table based on the number of GWAS loci).

FinnGen is a research project in genomics and personalized medicine. It is large public-private partnership that has collected and analysed genome and health data from 500,000 Finnish biobank donors to understand the genetic basis of diseases. FinnGen is now expanding into understanding the progression and biological mechanisms of diseases. FinnGen provides a world-class resource for further breakthroughs in disease prevention, diagnosis, and treatment and a outlook into our genetic make-up.

FinnGen results are subjected to one year embargo and, after that, available to the larger scientific community via the Pheweb browser or through data download.

Summary association statistics

GWAS summary statistics (tab-delimited, bgzipped, genome build 38, tabix index files included) are named as {endpoint}.gz. For example, endpoint I9_CHD has I9_CHD.gz and I9_CHD.gz.tbi.

To learn more about the methods used, see section GWAS.

The {endpoint}.gz have the following structure:

Two fine-mapping methods were used:

Fine-mapping results are tab-delimited and bgzipped.

SuSiE results have the following filename pattern:

• {endpoint}.SUSIE.cred.bgz

• {endpoint}.SUSIE.cred_99.bgz

• {endpoint}.SUSIE.snp.bgz

FINEMAP results have the following filename pattern:

• {endpoint}.FINEMAP.config.bgz

• {endpoint}.FINEMAP.region.bgz

• {endpoint}.FINEMAP.snp.bgz

To learn more about the methods used, see section .

{endpoint}.SUSIE.cred.bgz contain credible set summaries from SuSiE fine-mapping for all genome-wide significant regions. {endpoint}.SUSIE.cred_99.bgz contain the 99% credible set summaries while the default is 95%. They have the following structure:

{endpoint}.SUSIE.snp.bgz contain variant summaries with credible set information and have the following structure:

{endpoint}.FINEMAP.config.bgz contain summary fine-mapping variant configurations from FINEMAP method and have the following structure:

{endpoint}.FINEMAP.region.bgz contain summary statistics on number of independent signals in each region and have the following structure:

{endpoint}.FINEMAP.snp.bgz has summary statistics of variants and into what credible set they may belong to. Columns:

Linkage disequilibrium (LD) was estimated from for each chromosome. Use the tool for further usage of the bcor files.

ldstore --bcor FG_LD_chr1.bcor --incl-range 20000000-50000000 --table output_file_name.table

To learn more about the methods used, see section .

The variant annotation has measures (HWE, INFO, ...) listed per batch.

• Hail v0.2

• Cromwell-42

• Wdltool-0.14

• Plink 1.9 and 2.0

• BCFtools 1.7 and 1.9

• Eagle 2.3.5

• Beagle 4.1 (version 27Jan18.7e1)

• R 3.4.1 (packages: data.table 1.10.4, sm 2.2-5.4)

For matters related to this documentation, send us an email to finngen-info@helsinki.fi.

for the latest updates on the project as well as additional background information please consider visiting the study website https://www.finngen.fi/en or follow FinnGen on twitter @FinnGen_FI.

If you want to host FinnGen summary statistics on your website, please get in contact with us at: finngen-servicedesk@helsinki.fi.

The BCOR files were created using LDstore from the Finnish SISu panel v4.2.

The panel has been divided per chromosome. For example, to use the LD information in the first chromosome, FG_LD_chr1.bcor would be the file to use.

Settings used

• number of samples: 3775

• window size: 1500 kb

• accuracy: low

• number of threads: 96

• LD threshold to include correlations: 0.05

can be downloaded via:

And an example to extract variant range 20 Mb - 50 Mb from chromosome 7 is as follows:

It is not preferred to use these LD estimate files for e.g. fine-mapping, since many of the fine-mapping methods (e.g. SuSiE) require in-sample LD information for good results!

v4.2 consists of 8,554 WGS of Finnish individuals from 5 research cohorts from:

1. METSIM (PIs Markku Laakso and Mike Boehnke)

2. FINRISK (PI Pekka Jousilahti)

3. Corogene (PI Juha Sinisalo)

Chip genotype data processing and QC Samples were genotyped with Illumina (Illumina Inc., San Diego, CA, USA) and Affymetrix arrays (Thermo Fisher Scientific, Santa Clara, CA, USA).

Genotype calls were made with GenCall and zCall algorithms for Illumina and AxiomGT1 algorithm for Affymetrix data.

Chip genotyping data produced with previous chip platforms and reference genome builds were lifted over to build version 38 (GRCh38/hg38) following the protocol described here:

In sample-wise quality control steps, individuals with ambiguous gender, high genotype missingness (>5%), excess heterozygosity (+-4SD) and non-Finnish ancestry were excluded. In variant-wise quality control steps, variants with high missingness (>2%), low HWE P-value (<1e-6) and low minor allele count (MAC<3) were excluded.

Before imputation, chip-genotyped samples were pre-phased with using the default parameters, except the number of conditioning haplotypes, which was set to 20,000.

Genotype imputation was done with the population-specific .

The reference panel variant call set was produced with the GATK HaplotypeCaller algorithm by following GATK best practices for variant calling.

Genotype-, sample- and variant-wise QC was carried out iteratively by using the and the resulting high-quality WGS data for 8,554 individuals were phased with as described in the previous section.

Genotype imputation was carried out by using the population-specific SISu v4.2 imputation reference panel with (version 27Jan18.7e1) as described in the following protocol: .

Post-imputation quality control involved checking the expected conformity of the imputation INFO-value distribution, MAF differences between the target dataset and the imputation reference panel and checking chromosomal continuity of the imputed genotype calls.

This is a description of the quality control procedures applied before running the GWAS.

PCA

The PCA for population structure has been run in the following way:

Variant filtering and LD pruning

The sisu version 4.2 imputation panel is pruned iteratively, until a target number of SNPs is reached:

9,641,808 starting variants: only variants with a minimum info score of 0.9 in all batches are kept.

The script starts with [500.0, 50.0, 0.9] params in plink (window,step,r2). It then decreases 0.05 in r2 iteratively pruning the imputation panel until the threshold of 200,000 snps is reached. Once the SNP count falls under 200,000 the closest pruning is returned.

If the higher r2 is closer, 200,000 snps are randomly selected, else the last pruned snps are returned.

Plink flags used: --snps-only --chr 1-22 --max-alleles 2 --maf 0.01 .

For this run 180,042 snps are returned.

PCA outlier detection

Then, FinnGen data was merged with the 1k genome project (1kgp) data, using the variants mentioned above. A round of PCA was performed and a bayesian algorithm was used to spot outliers. This process got rid of 15,898 FinnGen samples. The figure below shows the scatter plots for the first 3 PCs. Outliers, in green, are separated from the FinnGen red cluster.

While the method automatically detected as being outliers the 1kg samples with non European and southern European ancestries, it did not manage to exclude some samples with Western European origins. Since the signal from these samples would have been too small to allow a second round to be performed without detecting substructures of the Finnish population, another approach was used. The FinnGen samples that survived the first round were used to compute another PCA. The EUR and FIN 1kg samples were then projected onto the space generated by the first 3 PCs. Then, the centroid of each cluster was calculated and used to calculate the squared mahalanobis distance of each FinnGen sample to each of the centroids. Being the squared distance a sum of squared variables (with unitary variance, due to the mahalanobis distance), we could see it as a sum of 3 independent squared variables. This allowed us to map the squared distance into a probability (chi squared with 3 degrees of freedom). Therefore, for each cluster, a probability of being part of it was computed. Then, a threshold of 0.95 was used to exclude FinnGen samples whose relative chance of being part of the Finnish cluster was below the level. This method produced another 878 outliers. The figure below shows the first three principal components.

FIN 1kg samples are in purple, while EUR 1kgp samples are in Blue. Samples in green are FinnGen samples who are flagged as being non Finnish, while red ones are considered Finnish.

Then all pairs of FinnGen samples up to second degree were returned. The figure below shows the distribution of kinship values.

Then, the previously defined “non Finnish” samples were excluded and 2 algorithms were used to return a unique subset of unrelated samples:

• one called greedy would continuously remove the highest degree node from the network of relations, until no more links are left in the network.

• one called native, based on a native implementation of python’s networkx package, performed on each subgraph of the network.

The largest independent set of either algorithm would be used to keep those sample, while flagging the others as “outliers” for the final PCA.

Then, the subset of outliers who also belong to the set of duplicates/twins was identified.

To compute the final step the Finngen samples were ultimately separated in three groups:

• 281,515 inliers: unrelated samples with Finnish ancestry.

• 218,975 outliers: non duplicate samples with Finnish ancestries, but who are also related to the inliers.

• 19,482 rejected samples: either of non Finnish ancestry or are twins/duplicates with relations to other samples.

Finally, the PCA for the inliers was calculated, and then outliers were projected on the same PC space, allowing to calculate covariates for a total of 500,490 samples.

Of the 500,490 non-duplicate population inlier samples from PCA, we excluded 345 samples from analysis because of missing minimum phenotype data, and 31 samples because of failing sex check with F thresholds of 0.4 and 0.7. A total of 500,186 samples were used for core analysis. There are 281,909 females and 218,277 males among these samples.

Documentation from the original developers of the algorithm can be found here: .

We included 2,755 endpoints in the analysis, which consisted of 2,499 binary endpoints and 3 quantitative endpoints (HEIGHT_IRN, WEIGHT_IRN, BMI_IRN). Endpoints with less than 50 cases among the 500,348 samples were excluded, as well as endpoints labeled with an OMIT tag in the endpoint definition file.

The quantitative endpoints HEIGHT and WEIGHT were acquired from minimum phenotype data. After that, phenotype BMI was formed from them, and all of them were inverse normal transformed.

For regenie step 1 LOCO prediction computation for each endpoint, we used age, sex, 10 PCs, Finngen 1 or 2 chip or legacy genotyping batch as covariates. For sex-specific phenotypes, sample sex was left out from the covariates. We excluded covariates that had less than 10 cases.

For calculating genetic relatedness in regenie step 1, we included variants 1) imputed with an INFO score > 0.95 in all batches and 2) > 97 % non-missing genotypes and 3) MAF > 1 %. The remaining variants were LD pruned with a 1.5Mb window and r2 threshold of 0.2. This resulted in a set of 188,153 well-imputed not rare variants for relatedness calculation.

We used a genotype block size of 1,000 in regenie step 1.

We ran association tests with regenie for each of the 2,755 endpoints for each variant with a total minimum allele count of 5. We used the approximate Firth test for variants with an initial p-value of less than 0.01 and computed the standard error based on effect size and likelihood ratio test p-value (regenie options --firth --approx --pThresh 0.01 --firth-se).

HLA imputation

The HLA data was imputed from R13 genotype data, using HIBAG models created by Jarmo Ritari from the Finnish Blood Bank. More information can be found in the repository:

https://github.com/FRCBS/HLA-imputation

as well as in the publication:

Ritari J, Hyvä rinen K, Clancy J, FinnGen, Partanen J, Koskela S. Increasing accuracy of HLA imputation by a population-specific reference panel in a Finngen biobank cohort. NAR Genomics and Bioinformatics, Volume 2, Issue 2, June 2020, lqaa030, https://doi.org/10.1093/nargab/lqaa030

Variant summary

A snp-stats report was generated with qctool

Association testing

Association testing was performed using Regenie 2.2.4, or for some endpoints Regenie 3.3. Same settings were used as in the core GWAS analysis. See the Association tests page for more information.

Association summary

A summary was created from the regenie summary statistic outputs. This summary contains the most significant variant (by p-value) for each phenotype. Pheweb links to phenotype and gene pages have been added as additional columns.

Registries

The disease endpoints were defined using nationwide registries:

• Drug purchase and Drug Reimbursement

• Digital and Population Data Services Agency

• Statistics Finland

We harmonized over the International Classification of Diseases (ICD) revisions 8, 9 and 10, cancer-specific ICD-O-3, (NOMESCO) procedure codes, Finnish-specific Social Insurance Institute (KELA) drug reimbursement codes and ATC-codes.

These registries spanning decades were electronically linked to the cohort baseline data using the unique national personal identification numbers assigned to all Finnish citizens and residents.

A full list of FinnGen endpoints is .

The endpoints with fewer than 50 cases, and developmental “helper” endpoints were excluded from the final PheWas (“OMIT” tag in the endpoint definition file).

(Risteys = intersection in Finnish) allows browsing of the FinnGen data at the phenotype level, including endpoint definitions, statistics about number of individuals, gender distribution, and longitudinal relationships. Please also note the R13 specific page

We used regenie for the FinnGen R13 release. Regenie's main advantages are fast leave-one-chromosome-out relatedness calculation which avoids proximal contamination, and use of an approximate Firth test which gives more reliable effect size estimates for rare variants.

Regenie version 3.3 was used for all the endpoints.

Links:

• regenie preprint

• regenie GitHub repository

We analyzed:

• ​2,755 endpoints

  • 2,752 binary endpoints

  • 3 quantitative endpoints (HEIGHT_IRN, WEIGHT_IRN, BMI_IRN)

• 500,186 samples

We included the following covariates in the model: sex, age, 10 PCs, Finngen chip version 1 or 2 , and legacy genotyping batch.

Colocalizations in FinnGen

Current primary co-localization method in FinnGen is Coloc method based on Susie finemapped results (Wallace et al. 2021). We also report previous colocalization approach metrics (CLPP and CLPA, see below). CLPP metric uses the probabilistic model for integrating GWAS and eQTL data presented in eCAVIAR (Hormozdiari et al. 2016). Compared to eCAVIAR, we are using SuSiE (Wang et al. 2019) to fine-map our inputs and provide an additional colocalization metric (CLPA). We use coloc version 5 to finemap the whole finemapped regions returned by SuSie.

Our goal is to extract a list of genomic regions that show colocalization between two phenotypes p1 and p2. Further, we assume that the summary statistics of p1 and p2 have been finemapped. The finemapping output for each phenotype contains three columns: the variant identifier (VAR), posterior inclusion probability (PIP), and the credible set (CS) identifier.

Methods

SuSie-Coloc

In the coloc framework, any pair of finemapped regions between two traits can be given one of the following five hypotheses:

H0: No association between either traits and the genomic regions

H1: Association between trait 1 and genomic region

H2: Association between trait 2 and genomic region

H3: Associations to traits 1 and 2, but different causal variants

H4: Associations to traits 1 and 2 for the same causal variant, i.e. colocalization.

One part of the colocalization results is posterior probabilities output by coloc for each of these hypotheses.

Since the move to implement colocalization using the Coloc package , the colocalization pipeline uses the per-variant log-bayes factors of the finemapped regions, instead of using only the credible set variants to identify the colocalization. This makes it possible to provide the posterior probability estimates for colocalization hypotheses that coloc creates. Not relying on credible set variants is more sensitive, for example if a FinnGen phenotype association is finemapped to a single credible set variant and that variant happens to be missing in other resources (e.g. UKBB pQTL data)

Due to the move to coloc, and no longer requiring the signals credible set variants to overlap, we do some quality control filtering to the colocalization results:

• We require the signals to have at least 90% of their PIP probability mass in the shared region of the colocalization. This measure lets us discard the colocalizations where the signals are not actually in the same region.

• We require the posterior probability of colocalization, i.e. probability of hypothesis 4 to be at least 0.8.

• We require the log bayes factors of the credible sets to be at least 0.9.

The Causal Posterior Probability (CLPP) is computed between two credible sets cs1 and cs2, with cs1 coming from a given phenotype p1 and cs2 coming from phenotype p2. CLPP is defined as follows: For vectors x and y, containing the PIP for variants in cs1 and cs2, respectively, CLPP is calculated by

This CLPP calculation is similar to equation 8 in Hormozdiari et al. 2016.

CLPP is dependent on the credible set size. By definition, any credible set size > 1 will yield a CLPP < 1.

We derived another colocalization metric called causal posterior agreement (CLPA) that is independent of credible set size.

The picture below shows how colocalizations are defined.

This rough example shows why we mostly use CLPA since it is independent of sample size.

The colocalization is performed between FinnGen endpoints as well as between FinnGen endpoints and various QTL resources, as shown in the image below.

These resources are listed below:

We use the following FinnGen data sources:

• The SuSiE finemapping results for the release GWAS were used as the FinnGen data. This resource was used as the FinnGen endpoints to colocalize against every other resource.

• FinnGen Somascan proteomics: EA5 proteomics, pQTLs using Somascan V4.1 assay, 865 unrelated samples from R12 imputed genotypes.

• FinnGen Olink proteomics: EA5 proteomics, Olink proteomics pQTLs, released in 11th October 2023. Olink Explore 3072 library, 1732 unrelated samples from R12 imputed genotypes.

• FinnGen Metabolomics data: mQTL data from NMR measurements for 34,218 samples.

Expression QTL datasets

• GTEx v8: SuSiE fine-mapping, 49 tissues, donors of mixed ancestry, Aguet et al. (2019, BioRxiv) (49 tissues only involve tissues with a sample size of n >= 50). Fine-mapping performed by Hilary Finucane, Jacob Ulirsch, Masahiro Kanai from the . Effect size interpretation: change in normalised gene expression (sd units) per alternate allele. Normalization = inverse normal transformation.

• EMBL-EBI (European Bioinformatics Institute) . eQTL data from 24 tissues/cell types, 16 RNAseq sources, 6 Microarray, SuSiE fine-mapping, donors of 88% European ancestry, Kerimov et al. (2021,Nature Genetics, doi: 10.1038/s41588-021-00924-w). For RNAseq data, four quantification methods (gene expression, exon expression, transcript usage, txrevise event usage). Fine-mapping was performed by . Effect size interpretation: change in normalised gene expression (sd units) per alternate allele. Normalization = inverse normal transformation.

• (RNAseq), muscle and adipose tissue.

Metabolom QTL datasets

GeneRISK: 186 lipid species QTLs, SuSiE fine-mapping of Widen et al. (2020), 7632 Finnish samples. Effect size interpretation: change in standard deviation of the lipid species per alternate allele.

Biomarkers

UK Biobank: 36 continuous endpoints, 57 biomarkers from UKBB prepared by , SuSiE fine-mapping. Effect size interpretation for quantitative traits: change in standard deviation of the normalized outcome per alternate allele. Effect size interpretation for binary traits increase in log(odds ratios) per alternate allele.

Release outputs

The following resources are released each release (with release number changing between releases):

In addition to the datafiles, the following documentation is also included in each release:

Acknowledgements

We thank the following people for helping us assembling the QTL resources:

• Kaur Alasoo and Nurlan Kerimov provided us the fine-mapped EMBL-EBI eQTL catalogue datasets.

• Hilary Finucane, Jacob Ulirsch, Masahiro Kanai gave us access to their fine-mapped GTEx data.

We use two state-of-the-art methods, FINEMAP (Benner, C. et al., 2016; Benner, C. et al., 2018) and SuSiE (Wang, G. et al., 2020) to fine-map genome-wide significant loci in FinnGen endpoints.

Briefly, there are three main steps:

1. Preprocessing

For each genome-wide significant locus (default configuration: P < 5e-8), we define a fine-mapping region by taking a 3 Mb window around a lead variant (and merge regions if they overlap). If a merged window exceeds 10MB, we iteratively shrink the window by 10%, until the merged window fits into 10MB or is split into merged windows that each fit into 10MB. We preprocess an input GWAS summary statistics into separate files per region for the following steps.

Note: Due to the complexity of the HLA region and the difficulty in finemapping this region, by default we exclude the region 25Mb to 34Mb (hg38; inclusive) of chromosome 6 during the preprocessing step. Therefore, the HLA region is not finemapped as part of the standard FinnGen finemapping pipelines.

2. LD computation

We compute in-sample dosage LD using LDstore2 for each fine-mapping region.

3. Fine-mapping

With the inputs of summary statistics and in-sample LD from the steps 1-2, we conduct fine-mapping using FINEMAP and with the maximum number of causal variants in a locus L = 10.

The "Credible Sets"-table on a phenotype page in the shows the SuSiE-fine-mapped credible sets of that phenotype. The variant shown per credible set is the maximum PIP (posterior inclusion probability) variant of that credible set. In addition to the causal variants, variants that were in sufficient LD (Pearson r^2 > 0.05), had a small enough p-value (pval < 0.01), and were close enough to the lead variant (distance to lead variant < 1.5 megabases) were clumped together with the credible set. Variants have been compared against GWAS Catalog and annotated. The LD grouping, annotation and GWAS Catalog comparison were done using the autoreporting pipeline.

The columns of the table are explained below:

Variant Selection

Loss of function (LoF) variants were generated from vcf files with VEP (https://github.com/Ensembl/ensembl-vep). LoF variants are defined as having one of the following consequences:

• frameshift_variant

• splice_donor_variant

• stop_gained

• splice_acceptor_variant

Also, a max_maf (0.01) and minimum info score (0.8) filters are applied. This leaves 4,793 genes on autosomal chromosomes that can be used for the association tests.

We used all 2,752 passing core binary phenotypes in the analyses with the exception of PD_DEMENTIA_EXMORE which failed to converge during the association test.

The null models from core additive GWAS runs were used. Tests are performed with regenie --step2 in burden mode using a max mask (i.e. using the maximum number of ALT alleles across sites). Only autosomal chromosomes (1-22) were used in the analysis.