Welcome to the Robust and Accurate Imputation from Summary Statistics (RAISS) documentation!

What is RAISS ?

RAISS is a python package to impute missing SNP summary statistics from neighboring SNPs in linkage desiquilibrium.

The statistical model used to make the imputation is described in [PZS+14], [LBR+13] and [JSPA19]. The implementation and performances of RAISS are described in [JSPA19].

The imputation execution time is optimized by precomputing Linkage desiquilibrium between SNPs.

Dependencies

If you need to compute LD matrices from your reference panel, RAISS requires plink version 1.9 for the precomputation of LD: https://www.cog-genomics.org/plink2

Installation

pip3 install git+https://gitlab.pasteur.fr/statistical-genetics/raiss.git
pip3 install -r https://gitlab.pasteur.fr/statistical-genetics/raiss/-/raw/master/requirements.txt

Alternatively, RAISS is available as a docker container:

https://quay.io/repository/biocontainers/raiss?tab=tags

Precomputation of LD-correlation

The imputation is based the Linkage desiquilibrium between SNPs.

To save computation time, the LD is computed before imputation and saved as tabular format. To limit the number of SNP pairs, the LD is computed between pairs of SNPs in a approximately LD-independent regions. Regions defined by [BP15] or computed using bigsnpr ([Pri21]) are provided in the package data folder for all several genome build and ancestries.

To compute the LD you need to specify a reference panel splitted by chromosomes (bed, fam and bim formats of plink, see PLINK formats )

# path to the Region file
region_berisa = "/mnt/atlas/PCMA/WKD_Hanna/cleaned_jass_input/Region_LD.csv"
# Path to the reference panel
ref_folder="/mnt/atlas/PCMA/1._DATA/ImpG_refpanel"
# path to the folder to store the results
ld_folder_out = "/mnt/atlas/PCMA/WKD_Hanna/impute_for_jass/berisa_ld_block"
raiss.LD.generate_genome_matrices(, ...)

Since 2021, LD can also be specified as scipy sparse matrix (.npz), the index must be provided in a separate file (one id by line)

Input format:

GWAS results files must be provided in the tabular format by chromosome (tab separated) all in the same folder with the following columns with the same header:

rsID

pos

A0

A1

Z

rs6548219

30762

A

G

-1.133

The files names must follow this pattern “z_{GWAS_TAG}_chr{1..22}.txt” This format can be obtained with the JASS PreProcessing package.

Launching imputation on one chromosome

RAISS has an interface with the command line.

Here is an example for imputing the file z_FINNS_IL-6_chr22.txt

raiss --chrom chr22 --gwas FINNS_IL-6 --ld-folder ${path-to-ld-matrices} --ref-folder ${path-to-the-reference-panel}  --zscore-folder ${path-to-input data} --output-folder ${folder to store imputed files} --eigen-threshold 0.000001

see details and all parameters in Command Line Usage bellow and note that the eigen_threshold parameter should be adapted to your reference panel (see the Optimizing RAISS parameters for your data section)

If you have access to a cluster, an efficient way to use RAISS is to launch the imputation of each chromosome on a separate cluster node. The script launch_imputation_all_gwas.sh contain an example of raiss usage with a SLURM scheduler.

Output

The raiss package outputs imputed GWAS files in the tabular format:

rsID

pos

A0

A1

Z

Var

ld_score

imputation_R2

rs3802985

198510

T

C

0.334

-1.0

inf

2.0

rs111876722

201922

C

T

0.297

0.09

5.412

0.91578

Variance is set to -1 for variants present in the input dataset

Optimizing RAISS parameters for your data

Raiss package contains a subcommand raiss performance-grid-search to assess its performance on your data and fine tune RAISS parameter.

Test procedure : 1. Mask N SNPs on a chromosome 2. Impute masked files for different values of the –eigen-threshold and the –minimum-ld parameters 3. Compute correlation (and other statistics) between genotype Z-values to imputed Z-values

To perform this test follow this procedure :

  1. Create a folder to store masked z-score files

  2. Create a folder to store z-score files imputed with different parameter

  3. Adapt the following command snippet to apply the command to your data:

Here the command example run for the harmonized GWAS files z_FINNS-IL-6.txt for eigen_threshold values in [0.000001, 0.001,0.1] and ld-threshold in [0,10,20]. Note that if you allocate more than 1 cpu to the job, different parameter combination will be tested in parallel.

raiss --ld-folder ${path-to-ld-matrices} --ref-folder ${path-to-the-reference-panel} --gwas FINNS_IL-6 --chrom chr22 --ld-type ${'scipy' or 'plink'} performance-grid-search --harmonized-folder ${path-to-harmonized data} --masked-folder ${folder to store partially masked input data} --imputed-folder $ --output-path ${path-to-save the performance report} --eigen-ratio-grid '[0.000001, 0.001,0.1]' --ld-threshold-grid '[0,10,20]' --n-cpu ${number of cpu to be allocated to the job}

The file Perf_GWAS_TAG (Perf_FINNS_IL-6 here) ressembles the following output:

Performance Report

eigen_ratio

min_ld

N_SNP

fraction_imputed

cor

mean_absolute_error

median_absolute_error

min_absolute_error

max_absolute_error

SNP_max_error

0.1

0

2970

0.594

0.978

0.277

0.171

1.47e-05

6.92

rs5756504

0.1

2

2970

0.594

0.978

0.277

0.171

1.47e-05

6.92

rs5756504

0.1

5

2840

0.568

0.978

0.277

0.169

1.47e-05

6.92

rs5756504

0.1

7

2550

0.51

0.978

0.275

0.164

0.000285

6.92

rs5756504

0.15

0

2470

0.494

0.976

0.282

0.172

2.43e-05

4.22

rs59411032

0.15

2

2470

0.494

0.976

0.282

0.172

2.43e-05

4.22

rs59411032

0.15

5

2450

0.49

0.976

0.281

0.172

2.43e-05

4.22

rs59411032

0.15

7

2320

0.465

0.976

0.282

0.172

0.00044

4.22

rs59411032

0.2

0

2040

0.409

0.973

0.291

0.168

6.61e-05

4.37

rs5752798

0.2

2

2040

0.409

0.973

0.291

0.168

6.61e-05

4.37

rs5752798

0.2

5

2040

0.408

0.973

0.291

0.168

6.61e-05

4.37

rs5752798

0.2

7

2020

0.403

0.973

0.291

0.168

6.61e-05

4.37

rs5752798

  • eigen_ratio : eigen ratio parameter that was tested.

  • min_ld : eigen ratio parameter that was tested.

  • N_SNP : number of the masked SNPs that were successfully imputed (i.e. not filtered out by the R2 criteria and/or min_ld criteria)

  • fraction_imputed : fraction of the masked SNPs that were successfully imputed (N_SNP / total_number_of_masked_SNP)

  • cor : the correlation between imputed and genotyped Z-scores.

  • mean_absolute_error : \(\mathbb{E}|Z_{imputed} - Z_{masked}|\)

  • median_absolute_error : \(median|Z_{imputed} - Z_{masked}|\)

  • min_absolute_error : \(min|Z_{imputed} - Z_{masked}|\)

  • max_absolute_error : \(max|Z_{imputed} - Z_{masked}|\)

  • SNP_max_error : \(argmax|Z_{imputed} - Z_{masked}|\)

To pick the best parameters, we recommend to search for a compromise between low imputation error and an high fraction of masked SNPs imputed (a trade-off between fraction_imputed and mean_absolute_error).

The optimal eigen_ratio and min_ld can vary depending on the density of your reference panel and input data. Hence, we recommend to run a grid search to pick the best parameter for your data. However, so far, we never observed a difference of performance from one chromosome to another. We suggest testing on the chr22 for computational efficiency. Note that this performance report evaluate RAISS on limited number of SNPs. Hence, we recommend to complement this report with our sanity check report_snps that check if the number of new genome wide significant hits is coherent with the number of imputes SNPs.

Generating report and sanity ckecks

Once you have imputed your dataset, you might want to know how many SNPs have been imputed and with which imputation quality. More importantly, you might want to make sure that the distribution of imputed signal is coherent with what is expected.

Indeed, if the population of your reference panel is too different from the one of your study, the Linkage desequilibrium matrices derived from your panel might not be adapted and might generate false positive

To run a sanity check on your data use the raiss sanity-check command

raiss sanity-check --trait z_{trait} --harmonized-folder {folder_input_data_files} --imputed-folder {folder_imputed_files}

Here is a sanity check log example :

Number of SNPs
in harmonized file: 7245111
in imputed file: 10665361
Proportion imputed : 0.32068769167775946


Number of SNPs by level of significance
                 harmonized  harmonized_by_1Mbregion  imputed  imputed_by_1MBregion  imputed_proportion  imputed_hit_OR
5.00e-02-1.00e+00     6506492                       16  9628715                    14            0.324262        0.932818
5.00e-06-5.00e-02      738573                     2682  1036583                  2684            0.287493        0.999821
5.00e-08-5.00e-06          43                       17       60                    20            0.283333        1.080485
0.00e+00-5.00e-08           3                        1        3                     1            0.000000        0.999448

Number of imputed SNPs by level of imputation quality
(0.8, 0.9]     2086501
(0.6, 0.8]     1333749
(0.0, 0.6]           0
(0.9, 0.95]          0
(0.95, 1.0]          0

The first block of the log gives you information about total number of SNPs contained in the harmonized file and imputed file.

The second block reports the number of SNPs by significance strata in the harmonized file and imputed file. the _by_1MBregion columns reports the number of 1Mb region reaching significance (minimal p-value of snps contained in the region). This provide you with a rough estimate of the number of LD independent hits before and after imputation. the imputed_hit_OR compares the imputed_by_1MBregion and harmonized_by_1Mbregion and reports if the number of new hit in imputed data is reasonable knowing the increased coverage.

The third block provides the number of imputed SNPs by imputation quality strata.

Command Line Usage

raiss command launch imputation for one trait on one chromosome

usage: raiss [-h] [--chrom CHROM] [--gwas GWAS] [--ref-folder REF_FOLDER]
             [--ld-folder LD_FOLDER] [--zscore-folder ZSCORE_FOLDER]
             [--output-folder OUTPUT_FOLDER] [--window-size WINDOW_SIZE]
             [--buffer-size BUFFER_SIZE]
             [--l2-regularization L2_REGULARIZATION]
             [--eigen-threshold EIGEN_THRESHOLD] [--R2-threshold R2_THRESHOLD]
             [--ld-type LD_TYPE] [--ref-panel-prefix REF_PANEL_PREFIX]
             [--ref-panel-suffix REF_PANEL_SUFFIX] [--minimum-ld MINIMUM_LD]
             {sanity-check,performance-grid-search} ...

Named Arguments

--chrom

chromosome to impute to the chrd+ format

--gwas

GWAS to impute to the consortia_trait format

--ref-folder

reference panel location (used to determine which snp to impute)

--ld-folder

Location LD correlation matrices

--zscore-folder

Location of the zscore files of the gwases to impute

--output-folder

Location of the impute zscore files

--window-size

Size of the non overlapping window

Default: 500000

--buffer-size

Size of the buffer around the imputation window

Default: 125000

--l2-regularization

Size of the small value added to the diagonal of the covariance matrix before inversion

Default: 0.1

--eigen-threshold

threshold under which eigen vectors are removed for the computation of the pseudo inverse

Default: 0.1

--R2-threshold

R square (imputation quality) threshold bellow which SNPs are filtered from the output

Default: 0.6

--ld-type

‘Two options are available : ‘plink’ or ‘scipy’. Ld can be supplied as plink command –ld-snp-list output files (see raiss.ld_matrix.launch_plink_ld to compute these data using plink) or as a couple of a scipy sparse matrix (.npz )and an .csv containing SNPs index

Default: “plink”

--ref-panel-prefix

prefix for the reference panel files

Default: “”

--ref-panel-suffix

suffix for the reference panel files

Default: “.bim”

--minimum-ld

this parameter ensure that their is enough typed SNPs around the imputed loci to perform a high accuracy imputation

Default: 4

Sub-commands

sanity-check

count imputed SNPs and significant hit before and after imputation

raiss sanity-check [-h] --trait TRAIT --harmonized-folder HARMONIZED_FOLDER
                   --imputed-folder IMPUTED_FOLDER [--output-path OUTPUT_PATH]
                   [--chr-list CHR_LIST]

Named Arguments

--trait

trait on which to run the sanity check

--harmonized-folder

folder containing data before imputation

--imputed-folder

folder containing imputed files

--output-path

path+prefix of the report file, the file will be suffixed by the trait name and .txt

Default: “./raiss_report”

--chr-list

Default: “range(1,23)”

Indices and tables

imputation_launcher

Imputation launcher

ld_matrix

Function set to compute LD correlation from a reference panel in predefined Region

stat_models

This module contain the statistical library for imputation.

windows

implement the imputation window is sliding along the genome:

filter_format_output

Module to filter SNPs on imputation quality and format output for JASS

imputation_R2

Module for imputation performance evaluation
[BP15]

Tomaz Berisa and Joseph K. Pickrell. Approximately independent linkage disequilibrium blocks in human populations. Bioinformatics, 32(2):283–285, 2015. doi:10.1093/bioinformatics/btv546.

[JSPA19] (1,2)

Hanna Julienne, Huwenbo Shi, Bogdan Pasaniuc, and Hugues Aschard. RAISS: robust and accurate imputation from summary statistics. Bioinformatics, 35(22):4837–4839, 06 2019. URL: https://doi.org/10.1093/bioinformatics/btz466, arXiv:https://academic.oup.com/bioinformatics/article-pdf/35/22/4837/30706731/btz466.pdf, doi:10.1093/bioinformatics/btz466.

[LBR+13]

Donghyung Lee, T Bernard Bigdeli, Brien P Riley, Ayman H Fanous, and Silviu-Alin Bacanu. Dist: direct imputation of summary statistics for unmeasured snps. Bioinformatics, 29(22):2925–2927, 2013.

[PZS+14]

Bogdan Pasaniuc, Noah Zaitlen, Huwenbo Shi, Gaurav Bhatia, Alexander Gusev, Joseph Pickrell, Joel Hirschhorn, David P. Strachan, Nick Patterson, and Alkes L. Price. Fast and accurate imputation of summary statistics enhances evidence of functional enrichment. Bioinformatics (Oxford, England), 30(20):2906–2914, 2014. arXiv:arXiv:1309.3258v1, doi:10.1093/bioinformatics/btu416.

[Pri21]

Florian Privé. Optimal linkage disequilibrium splitting. Bioinformatics, 38(1):255–256, 07 2021. URL: https://doi.org/10.1093/bioinformatics/btab519, arXiv:https://academic.oup.com/bioinformatics/article-pdf/38/1/255/41891000/btab519.pdf, doi:10.1093/bioinformatics/btab519.