PRSice User Manual

PRSice: Polygenic Risk Score software
v1.2
Jack Euesden
[email protected]
Cathryn M. Lewis
Paul F. O’Reilly
January 30, 2015
Contents
1
Overview
3
2
R packages required
3
3
Quickstart
3.1 Input Data . . . . . . . . . . . . . . . . . . .
3.1.1 Base data set . . . . . . . . . . . . . .
3.1.2 Target Dataset . . . . . . . . . . . . .
3.1.3 PLINK2 Executable FIle . . . . . . .
3.2 Outputs . . . . . . . . . . . . . . . . . . . . .
3.2.1 Figures . . . . . . . . . . . . . . . . .
3.2.2 PRS model-fit . . . . . . . . . . . . .
3.2.3 Scores for each Individual . . . . . . .
3.3 Summary-Summary Statistic Based Analysis
4
User Options
5
Detail on command line options
5.1 Target Data Set Format . . . . . . . . . . . .
5.2 Target Data Set phenotype . . . . . . . . . .
5.3 Covariates . . . . . . . . . . . . . . . . . . . .
5.4 Graphical Parameters . . . . . . . . . . . . .
5.5 Linkage Disequilibrium . . . . . . . . . . . . .
5.5.1 Clumping . . . . . . . . . . . . . . . .
5.5.2 Pruning . . . . . . . . . . . . . . . . .
5.6 Dosage Data . . . . . . . . . . . . . . . . . .
5.7 High-Resolution Scoring . . . . . . . . . . . .
5.8 Multiple Phenotypes . . . . . . . . . . . . . .
5.9 Summary-Summary Statistic Based Analysis
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
3
5
5
5
5
6
6
6
6
6
8
1
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
15
15
15
15
16
16
16
16
17
17
17
18
5.10 Miscellaneous . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 20
6
Outputs
21
7
Acknowledgements
23
References
24
2
1 Overview
PRSice (pronounced ‘precise’) is a software package for calculating, applying, evaluating
and plotting the results of polygenic risk scores. PRSice can run at high-resolution to
provide the best-fit PRS as well as provide results calculated at broad P -value thresholds, illustrating results corresponding to either, can thin SNPs according to linkage
disequilibrium and P -value (“clumping”), handles genotyped and imputed data, can
calculate and incorporate ancestry-informative variables, and facilitates the systematic
application of PRS across multiple traits.
PRSice is a software package written in R, including wrappers for bash data management
scripts and PLINK2 ([1]) to minimise computational time; thus much of its functionality relies entirely on computations written originally by Shaun Purcell in PLINK 1 and
Christopher Chang in PLINK 2. PRSice runs as a command-line program, compatible
for Unix/Linux/Mac OS, with a variety of user-options and is freely available for download from: http://PRSice.info.
2 R packages required
• ggplot2
• plyr
• batch
• fmsb
• gtx
NB: Users will also need to install dependancies for these
3 Quickstart
As standard, PRSice has four input files and three main outputs:
Inputs
• PRSice.R file: includes R scripts, bash and PLINK2 wrappers
• PLINK2 executable files (Linux and Mac)
• Base data set: GWAS summary results, which the PRS is based on
• Target data set: Raw genotype data of ‘target phenotype’
3
Outputs
• Figures illustrating the model fit of the PRS analyses
• Data on the model-fit of the PRS analyses
• PRS for each individual in the target data (for best-fit PRS as default)
Simple command for running PRSice on toy data provided:
R --file=PRSice_v1.2.R -q --args \
plink ./plink_1.9_mac_160914 \
base TOY_BASE_GWAS.assoc \
target TOY_TARGET_DATA \
slower 0 \
sinc 0.01 \
supper 0.5
This runs PRSice with PRS calculated at P -value thresholds between PT = 0 and
PT = 0.5 at increments of 0.01, where the target phenotype is a binary trait. Ancestryinformative covariates are calculated in the target data set and the first two used to
adjust for population structure by default. Clumping is applied by default to thin SNPs
according to linkage disequilibrium and P -value (the SNP with smallest P -value in each
250 kb window is retained and all those in LD, r2 > 0.1, with this SNP are removed).
Below is a different run with more non-default options added (see Section 5 for the full
list of user-options):
R --file=PRSice_v1.2.R -q --args \
plink ./plink_1.9_mac_160914 \
base TOY_BASE_GWAS.assoc \
target TOY_TARGET_DATA \
slower 0 \
sinc 0.02 \
supper 0.3 \
covary F \
best.thresh.on.bar F \
report.individual.scores T
This runs PRSice with PRS calculated between PT = 0 and PT = 0.3 at increments of
0.02, where the target phenotype is a binary trait. No covariates are included. Polygenic
risk scores for each individual at the best-fit PRS are printed to a file
PRSice SCORES AT BEST-FIT-PRS.txt. The best-fit PRS bar will not be included in the
bar plot.
4
3.1 Input Data
3.1.1
Base data set
Base data must be provided as a whitespace delimited file containing association analysis
results for SNPs on the base phenotype. Columns with the following header names must
be present: SNP, CHR, BP, A1, A2, OR or BETA, P - containing SNP name (eg. rs number),
chromosome number, base pair position, effect allele (A1), non-effect allele (A2), effect
size estimate as an odds ratio (binary phenotype) or continuous effect beta (continuous
phenotype) and P -value for association.
Thus, the first few lines of the base data set may look like the following for a binary trait:
SNP
rs3094315
rs3131972
rs3131971
CHR
1
1
1
BP
752566
752721
752894
A1
A
A
T
A2
G
G
C
OR
0.9912
1.007
1.003
SE
0.0229
0.0228
0.0232
A1
A
A
T
A2
G
G
C
BETA
-0.008
0.0069
0.002
SE
0.0229
0.0228
0.0232
P
0.7009
0.769
0.8962
or for continuous traits:
SNP
rs3094315
rs3131972
rs3131971
CHR
1
1
1
BP
752566
752721
752894
P
0.7009
0.769
0.8962
NB: PRSice currently only supports autosomal SNPs.
Strand flips are automatically detected and accounted for.
3.1.2
Target Dataset
The target data set must be supplied in PLINK binary format, with the extensions .bed
.bim .fam - where bed contains compressed genotype data. Missing phenotype data
can be coded as NA, or -9 for binary traits.
3.1.3
PLINK2 Executable FIle
If the PLINK2 executable file is not in the working directory then the path to it must
be given.
5
3.2 Outputs
3.2.1
Figures
• A bar plot named PRSice BARPLOT [date].png , where [date] is today’s date, is
generated, displaying the model fit of the PRS at different broad P -value thresholds
(fig 1a).
• A high-resolution plot named PRSice HIGH-RES PLOT [date].png , where [date]
is today’s date, displaying the model fit of PRS calculated at a large number P value thresholds. A green line connects points showing the model fit at the broad
P -value thresholds used in the corresponding bar plot (fig 1b).
3.2.2
PRS model-fit
A file containing the PRS model fit across thresholds is named PRSice RAW RESULTS DATA.txt;
this is stored as threshold, P -value, variance in target phenotype explained, r2 , and number of SNPs at this threshold.
3.2.3
Scores for each Individual
A file containing PRS for each individual at the best-fit PRS named
PRSice SCORES AT BEST-FIT-PRS.txt (by default) or
PRSice SCORES AT ALL THRESHOLDS.txt (as required).
NB: PRSice also supports multiple base data sets and multiple phenotypes
for target data. For more details on the options used to implement this, see
below.
3.3 Summary-Summary Statistic Based Analysis
While the main function of PRSice is to apply polygenic risk scores that use effect size
estimates from summary statistic data in a target data set containing raw genotype data,
there is also the option to perform a summary-summary statistic analysis that exploits
the gtx R package by Toby Johnson (gtx). An example command is shown below:
R --file=PRSice_v1.2.R -q --args \
plink ./plink_1.9_mac_160914 \
base TOY_BASE_GWAS.assoc \
target TOY_TARGET_GWAS.assoc \
sumsum T \
size.targ 2000 \
slower 0 \
sinc 0.02 \
supper 0.3 \
6
clump.snps F
This command includes the option to use genotype data for clumping by LD. See below
for more details on the summary-summary analysis option.
7
4 User Options
Essential Arguments
• target: Location of the PLINK data used for the target data set. This must be
supplied without the file’s extension (i.e. bed bim fam)
• base: Location and full name of the GWAS results for the base data set - see the
note above on column naming.
PLINK
• plink: The location and name of the binary file for executing plink. NB: this is
to be the more recent version of PLINK, ‘1.9’ if the target data set is in genotype
format, and the older version, ‘1.07’ if dosage data is used. Default value is NA.
Target Data set Phenotype
• pheno.file: Location of the file containing external phenotype data, if it is not
coded in the genotype file. This file must have two columns, individual ID and
phenotype, with no header line. Missing data is coded NA, and NA and -9 for
binary traits. If NA, phenotypes will be extracted from the genotype data. Default
value is NA
• binary.target: If T, phenotype in target data is assumed to be binary. If F,
phenotype in target data is assumed to be continuous. Default value is T.
Target Data set Format
• geno.is.ped: If geno.is.ped T, target data set will be converted from pedigree
format (.ped, .map) to binary format before analysis begins. Default value is F
• geno.as.list: If geno.as.list T, the argument passed to target is treated as
a general path to multiple genotype datasets, one per chromosome. In this case,
every instance of the string CHRNUM will be substituted for values between 1-22 in
order to analyse all chromosomes. Default value is F.
Covariates
• covary: If T, covariates are used when testing model fit of polygenic score on
phenotype. If F, covariates are not used. Default value is T
• user.covariate.file: Location of file containing custom covariates. This must
have column names individual ID (headed IID) and then covariates. If covary
T but no user covariate file is supplied, ancestry-informative dimensions will be
generated automatically from the data. Default value is NA
8
• covariates: Names of covariates to adjust for. If covary T, but user.covariate.file
NA, these are the number of ancestry informative covariates to calculate and use,
denoted C1,C2,C3 etc. Default value is C1,C2
• ancestry.dim: If covary T, but user.covariate.file NA, ancestry informative
dimensions are generated automatically. This specifies the method that should be
used. Users can select PCA or MDS. Default value is MDS.
Figures
• ggfig: If T, ggplot2 will be used to generate figures. If F, base graphics will be
used to generate figures. Default value is T
• barchart.levels: Thresholds which should be plotted on the bar chart. Default
value is 0.001,0.05,0.1,0.2,0.3,0.4,0.5
• barpalatte: Colour palette used to colour bars on barplot, if ggfig T and
bar.col.is.pval F. Default value is YlOrRd
• best.thresh.on.bar: If fastscore F and best.thresh.on.bar T, the most predictive threshold from high-resolution scoring will be identified automatically and
added to the barchart. If best.thresh.on.bar F, only thresholds listed in barchart.levels will be plotted. Default value is F if fastscore T. If fastscore F,
which is default, default value is T
• scatter.R2: If scatter.R2 T, variance explained will be used as y axis on the
high-resolution plot. If scatter.R2 F, − log10 (P ) will be used as y axis on highresolution plot. If fastscore T, a high-resolution plot won’t be generated. Default
value is F
• figname: An optional prefix for figures and text-files. Default value is PRSice
• bar.col.is.pval: If ggfig T, and bar.col.is.pval T, bars are coloured by
association with phenotype. If bar.col.is.pval F, bars are coloured by P -value
threshold. Default value is F
• bar.col.is.pval.lowcol: If bar.col.is.pval T, bar.col.is.pval.lowcol will
be used to colour the poorest predicting thresholds. Default value is dodgerblue
• bar.col.is.pval.highcol: If bar.col.is.pval T, bar.col.is.pval.highcol
will be used to colour the best predicting thresholds. Default value is firebrick.
Clumping
• clump.snps: If clump.snps T, base SNPs will be clumped to remove linkage disequilibrium. If clump.snps F, base SNPs will not be clumped. This is currently
unsupported with dosage data. Default value is T
• clump.p1: The clumping threshold, P -value, for index SNPs. Default value is 1
9
• clump.p2: The clumping threshold, P -value, for clumped SNPs. Default value is
1
• clump.r2: The LD threshold, r2 , for clumping. Default value is 0.1
• clump.kb: The distance threshold for clumping, in Kb. Default value is 250.
Pruning
• prune.snps: If prune.snps T, LD will be stripped using pruning, which is agnostic to base P -value. If prune.snps F, pruning will not be used. NB: pruning and
clumping can not both be used. Unless clump.snps is set to F, this will override
prune.snps T. Default value is F
• prune.kb.wind: Window size for pruning, in Kb. Default value is 50
• prune.kb.step: Increment for sliding window, for pruning, in Kb. Default value
is 2
• prune.kb.r2: Pairwise LD for pruning, as r2 . Default value is 0.8.
High-Resolution Scoring
• slower: The lower bound at which polygenic risk score is calculated at, as PT .
Default value is 0.0001
• supper: The upper bound at which polygenic risk score is calculated at, as PT .
Default value is 0.5
• sinc: The increment size between slower and supper. Polygenic risk scores will
be calculated at each increment between the two bounds. Default value is 0.00005
• fastscore: If fastscore T, scores will only be calculated at the thresholds specified by barchart.levels. If fastscore F, scores will be calculated at incremements of sinc between slower and supper. This is high-resolution scoring.
Default value is F
• mend.score: If mend.score T, the first n SNPs will be added from the base data
set, sorted by P -value, one by one in order to verify that there are no individual
loci of large effect influencing target phenotype. If mend.score F, these will not
be added. Default value is F
• mend.score.len: If mend.score T, a number of SNPs will be added one by one to
calculate PRS. mend.score.len sets the number of SNPs to add before thresholds
are used. Default value is 100
• score.at.1: If score.at.1 T, an extra threshold will be added at PT = 1. This
allows users to compare a polygenic risk score calculated with all SNPs to the
next-best score. If score.at.1 F, this will not be calculated. Default value is F.
10
Dosage Data
• dosage: If dosage T, PRSice will expect to read in dosage data. If dosage F,
PRSice will expect to read in genotype data. Default value is dosage F
• dosage.format: The format of dosage files can be specified to use default values
for some of the options below. Currently only ‘gen’ is supported by this option.
Default value is gen
• dos.skip0: The number of columns in dosage file before column containing SNP
names. Default value is 1
• dos.skip1: The number of columns in dosage file between SNP names and A1.
Default value is 1
• dos.coding: The value that dosage probabilities sum to per variant per individual.
Default value is 1
• dos.format: Number of columns dosage data occumpies per variant per individual.
Default value is 3
• dos.sep.fam: The path to an external fam file containing individual-level data for
the dosage file. Default value is NA
• dos.fam.is.samp: If dos.fam.is.samp T, the fam file supplied will be automatically converted from gtool’s ‘sample’ format before being read in. If dos.fam.is.samp
F, the fam file will be read in like a normal 6-column PLINK fam file. Default value
is F
• dos.impute2: If dos.impute2 T, dosage data will be read in using impute2 format
defaults. If dos.impute2 F, these must be set manually. Default value is F
• dos.path.to.lists: If there are dosage files for each chromosome, with a regular pattern in which CHRNUM is a value between 1 and 22 for each dosage file,
dos.path.to.lists can be set to a single string, e.g.
/path/to/chr CHRNUM/chr CHRNUM.dat. Default value is NA
• dos.list.file: If dosage file names are listed in a separate file, these will be read
in automatically from the external file. These file names must have one line per
file. Default value is NA.
Multiple Phenotypes
• multiple.target.phenotypes: If multiple.target.phenotypes T, multiple columns
will be read from the external phenotype file supplied using pheno.file. If this
is used, the file pheno.file must have column names. The first column must be ID.
The other columns must be names of phenotypes. If multiple.target.phenotypes
F, only the first two columns from the phenotype file will be used. Default value
is F
11
• target.phenotypes: A list of names of target phenotypes to be read from a
phenotype file, if multiple.target.phenotypes T. These must be separated by
commas. Polygenic risk score for base phenotype will be regressed on each of these
in turn. Default value is NA
• target.phenotypes.binary: A vector of logical T and F’s separated by commas.
This determines whether the phenotypes listed by target.phenotypes are binary
(i.e. case-control) or quantitative (i.e. continuous) traits. This must have the same
number of items as target.phenotypes. NA, -9 and empty values will be treated
as missing values for binary phenotypes, binary phenotypes must be coded 0,1 NA
and empty values will be treated as a missing value for quantitative traits. Default
value is NA
• multiple.base.phenotypes: If multiple.base.phenotypes T, the argument passed
to base will be interpreted as a general file path, in which the string PHEN.NAME
will be evaluated as the names of different base phenotypes. Default value is F
• base.phenotypes.names: A vector of base phenotype names. These will be substituted for PHEN.NAME in the argument passed to base, and used to read multiple
GWAS results in series.
Summary-Summary Statistic Based Analysis
• sumsum: If sumsum T, PRSice will expect both target and base datasets to be
GWAS results files, and will use the method of Toby Johnson ([2] as implemented
in the R package gtx) to evaluate evidence for shared genetic aetiology between
base and target phenotypes. If using sumsum T, a number of changes need to be
made to input formats - see below. If sumsum F, polygenic risk scores will be
calculated, using genotype data, as decribed above. Default value is F
• clump.ref: If sumsum T, input data can be clumped to ensure linkage equilibrium. This is performed on the base data set, and is elected using clump.snps T.
clump.ref is used to select genotype data to be used to clump the base data set.
Default value is NA
• size.targ: The sample size of target data set. Default value is NA.
Quantile Plots
• quantiles: If quantiles T, a plot will be produced displaying association between target phenotype and PRS at the most predictive threshold, divided into
a number of quantiles, as shown in figure 1c. If quantiles F, this plot will not
be produced. (NB. these will be adjusted for covariates unless covary F). Default
value is F
• num.quantiles: Number of quantiles to split PRS into when producing quantile
plot. Default value is 5
12
• quant.ref: Reference quantile to use when produing quantile plot. Default value
is 3.
Miscellaneous
• wd: The folder which is to be used to store output data, temporary files and figures.
Default value is ./ , i.e. current directory
• print.time: If print.time T, running time will be printed at end of output. If
print.time F, running time will not be printed. Default value is T
• cleanup: If cleanup T, all temprorary files will be removed at the end of the
analysis. If cleanup F, these will be left in the working directory. Default value
is T
• remove.mhc: If remove.mhc T. the MHC region between 26 and 33Mb on chromosome 6 will be removed when calculating polygenic risk scores. If remove.mhc
F, this region will not be removed. Default value is F. NB, not valid for dosage
data
• for.meta: If for.meta T, coefficients and standard errors are reported for each
regression model, into the file PRSICE RAW RESULTS DATA.txt. This allows metaanalysis of scores across different target data sets. If for.meta F, these extra
columns are not reported. Default value is F
• report.individual.scores: If report.individual.scores T, a file called
PRSice SCORES AT ALL THRESHOLDS.txt or
PRSice SCORES AT BEST-FIT-PRS.txt (see below) will be produced containing every individual’s polygenic risk score at every threshold (or just the most predictive
threshold) and written to the working directory. NB, this file may be very large
depending on the number of thresholds used. If report.individual.scores F,
this file will not be written. Default value is T
• report.best.score.only: If report.best.score.only F, and
report.individual.scores T, polygenic risk scores for all individuals at all thresholds will be written to a file called PRSice SCORES AT ALL THRESHOLDS.txt. If
report.best.score.only F, scores for every individual at the most predictive
threshold will be written to a file called PRSice SCORES AT BEST-FIT-PRS.txt.
Default value is T
• plink.silent: if plink.silent T, PLINK will not output to the terminal whilst
PRSice is running. If plink.silent F, PLINK’s outputs will print to the terminal
as PRSice runs. Default value is T
• no.regression: if no.regression T, phenotype will not be regressed on polygenic risk scores, but they will be calculated and printed to a file, if
report.individual.scores T. If no.regression F, phenotype will be regressed
on polygenic risk scores. Default value is F
13
• allow.no.sex: If allow.no.sex F, individuals with missing sex data in the genotype file will be given missing phenotype data and excluded from regression models.
If allow.no.sex T, individuals with missing sex data but non-missing genotype
data will be included in regression models. Default value is F
• debug.mode: If debug.mode T, more text including warning and error messages
from bash, R and PLINK will be reported to the terminal. If debug.mode F, these
will be suppressed. Default value is F.
14
5 Detail on command line options
5.1 Target Data Set Format
If genotype data is stored chromosome by chromosome, this data can be analysed without
merging, by using the option geno.as.list T. If this is the case, the string supplied
to target must contain the characters CHRNUM - this will be iteratively replaced with
chromosome numbers from 1 to 22 and scores calculated across the genome.
5.2 Target Data Set phenotype
By default, the software identifies binary phenotype, coded either 1,2 or 0,1 from a column in the ped or fam file. If the phenotype data stored in this file is continous, the
option binary.target F must be used. If binary phenotype data is stored 1,2 then -9
and 0 will be coded as missing. If it is coded 0,1 then -9 will be coded as missing. NA
will also be coded as missing.
If the phenotype data is stored in an external file, this file must have two columns,
individual ID and phenotype, with no header line. This is specified with the option
pheno.file "/path/to/pheno.file".
5.3 Covariates
By default, the software calculates two ancestry informative dimensions using Multidimensional Scaling (MDS) and uses these as covariates when predicting target data
set phenotype using polygenic score. Using covariates altogether can be disabled using
covary F.
A different number of ancestry informative dimensions can be used by specifying
covariates C1,C2,C3, for example to use three covariates. PCA or MDS can be
used for calculating ancestry informative dimensions, using ancestry.dim "PCA” or
ancesrtry.dim "MDS" respectively.
NB: if covariates are automatically generated, output files should be inspected to check
for outliers etc.
External covariates can also be used. These are added using the option
user.covariate.file "path/to/file.covary". This file must have a first column
headed IID containing individual ID. The remaining columns must be covariates with a
header line. The covariates to use from this file must be specified based on the column
name, using the option covariates name1,name2 etc.
15
5.4 Graphical Parameters
The default output is a barchart and a high-resolution plot (fig 1a) using ggplot2.
Base graphics can be used instead using ggfig F - NB, plotting with base graphics allows fewer options. The levels of PT to use in the barchart are specified using
barchart.levels 0.1,0.2,0.3 , for example. If fastscore F, the most predictive
polygenic score can be identified from the high-resolution plot and added to the barchart using best.thresh.on.bar T. If ggfig T, the user can specify the colour scheme
used for the barchart using barpalatte. If high-resolution scoring is used, the y axis
of the high-resolution plot is − log10 (P ) by default - this can be switched to R2 (or
Nagelkerke’s pseudo R2 for binary target phenotypes) using scatter.R2 T
The extension name for all output files, which is PRSice by default, can be altered
using figname.
If ggfig T, barplots are coloured by the predictive ability of that score on phenotype, − log10 (P ), as a default. The colours of the barplot can be set by selecting
the low and high colours to use for gradients using bar.col.is.pval.lowcol and
bar.col.is.pval.highcol. Bars can be coloured by P -value threshold instead, using bar.col.is.pval F. In this case, barpalatte is used.
5.5 Linkage Disequilibrium
By default, base SNPs are clumped using base P -values and LD data from the target
dataset, in order to obtain SNPs in linkage equilibrium. This option is disabled when
using dosage data.
5.5.1
Clumping
Clumping may be disabled using clump.snps F.
Predefined clumping parameters can be changed by updating the clump.p1 clump.p2
clump.r2 and clump.kb arguments.
5.5.2
Pruning
An alternative method for obtaining SNPs in linkage equilibrium is LD-informed pruning. The LD structure of the target data set is used to obtain SNPs in linkage equilibrium in the base dataset, and is agnostic of base P -value. This option is enabled using
prune.snps T. NB: if this is used, clump.snps must be set to F, as the two methods
are mutually exclusive. If both are set to T, clumping will be used by default.
Pruning takes three arguments: the window size in Kb, the step size in Kb, and the r2
to prune to within each window. These arguments are prune.kb.wind, prune.kb.step,
and prune.kb.r2 respectively, and default to 50, 2 and 0.8.
16
5.6 Dosage Data
Dosage data is not used by default and requires a number of defaults to run correctly.
If using dosage format data, this must be specified using dosage T
Individual-level data must be provided using dos.sep.fam "/path/to/file.fam". By
defalt this must be in the format of a PLINK .fam file. Users may want to use the
.sample files which are generated by impute2 - if so, this must be specified with
dos.fam.is.samp T - then the input file can be reformatted correctly.
A target data set can be provided in three ways.
• A single file containing dosage data - this can be specified using
target "/path/to/dosages.dos"
• A single filepath, where a certain part varies for dosage files across the 22 chromosomes. This is specified using
dos.path.to.lists "/path/to/chr CHRNUM/chr CHRNUM.impute". CHRNUM will
be substituted for numbers 1-22 in order to read in autosome data.
• A file that contains the full paths and names of each dosage file, specified using
dos.list.file "/path/to/list of dos files".
Other dosage options, indicating the format of the input data, are specified using
dos.skip0 dos.skip1 dos.coding and dos.format. These options can be set to their
defaults for reading in output from impute2 using dos.impute2 T.
5.7 High-Resolution Scoring
If fastscore T, polygenic scores will only be calculated at the levels specified by
barchart.levels. If fastscore F, scores every few thresholds between a lower and
upper bound will be calculated. By default, these are increments of 0.00005 from 0.0001
to 0.5 for genotype data. We recommend changing these to increments of 0.001 from
0.001 to 0.5 for dosage data in order to offset the increased running time for dosage files.
The size of these increments, and the upper and lower bounds used, can be set using the sinc, slower and supper options respectively.
fastscore allows a second figure to be produced, a high-resolution plot showing the
result of high-resolution scoring, and allows users to add an extra bar to the barchart to
indicate the most predictive threshold, using best.thresh.on.bar T.
5.8 Multiple Phenotypes
PRSice supports the analysis of pairwise comparisons between multiple base and multiple
target phenotypes. Multiple target phenotypes must be stored in a single external phenotype file, specified using pheno.file. If multiple.target.phenotypes T, this file
17
must have column names ID and then phenotype names. The option to use multiple target phenotypes is selected using multiple.target.phenotypes T. Target phenotypes
are selected from this file using target.phenotypes, a vector of names. Whether these
are binary or quantitative traits must be specified using target.phenotypes.binary,
a vector of logical TRUE’s and FALSE’s. target.phenotypes.binary must have the
same number of items as target.phenotypes.
Multiple base phenotypes are specified using multiple.base.phenotypes T. If
multiple.base.phenotypes T, the names of phenotyoes supplied to
base.phenotypes.names will be substituted for PHEN.NAME in the argument passed to
base. For example, if base /path/to/PHEN.NAME.assoc and base.phenotype.names
DIS1,DIS2, then two files, /path/to/DIS1.assoc and /path/to/DIS2.assoc would be
read in as base data sets.
If both multiple target and multiple base phenotypes are supplied, PRSice will perform pairwise tests of every base phenotype predicting every target phenotype at every
threshold, and save the most predictive threshold from each comparison. For more detail
on the outputs generated when using multiple base and target phenotypes see below.
5.9 Summary-Summary Statistic Based Analysis
PRSice can use GWAS summary data in both the base and target data sets to evaluate
evidence for shared genetic aetiology, using the method of Johnson et al, as implemented
in gtx ([2]). This is enabled using sumsum T.
If sumsum T, target must specify the path to a file containing GWAS summary results. This must contain columns labelled SNP, SE, A1, A2, for marker ID, standard
error, effect allele, non-effect allele respectively. Users must also include a column for
effect size, either OR for binary phenotypes or BETA for quantitative traits. The sample
size of the target data set must also be set using size.targ.
The base data set must contain columns labelled SNP, P, A1, A2 for marker ID, P value, effect allele and non-effect allele respectively.
A sample of genotype data, e.g. HapMap (available here), can be used to clump the
base GWAS data before testing for evidence for shared genetic aetiology. This is selected using clump.ref, where the path provided is the name of three plink binary files
(i.e. .bed, .bim, .fam). NB: The authors of gtx recommend using more stringent clumping parameters; these must be set manually. The authors recommend clump.p1 0.5
clump.p2 0.5 clump.kb 300 clump.r2 0.05.
When using sumsum T, many other options are no longer valid. The following options
are still valid:
18
• plink.silent
• print.time
• cleanup
• debug.mode
• slower
• sinc
• supper
• ggfig
• target
• base
• clump.ref
• clump.p1
• clump.p2
• clump.r2
• clump.kb
• plink
• binary.target
• size.targ
• figname
• bar.col.is.pval.lowcol
• bar.col.is.pval.highcol
• barchart.levels
19
5.10 Miscellaneous
Running time is printed by default - this can be disabled using print.time F. Profile
lists and supporting files are also removed from the working directory by default. This
can be disabled using cleanup F.
By default all intermediate and output files are printed to the current directory. This
can be changed using wd "path/to/wd/".
The MHC region on chromosome 6 (26-33Mb) is frequently omitted from polygenic
risk scores, as the long-range linkage disequilibrium in this region makes linkage equilibrium difficult to obtain. This region can be removed using remove.mhc T
The option for.meta T reports coefficients and standard errors for each regression model
in the output file PRSice RAW RESULTS DATA.txt, allowing meta-analysis across target
data sets to be performed, at a given threshold.
20
6 Outputs
The script produces a large amount of supporting data. Two figures are generated a barplot and a high-resolution plot of polygenic score threshold, PT , plotted against
model-fit across thresholds. The high-resolution plot depicts model-fit as − log10 (P ).
The barplot depicts model-fit as improvement in model-fit, as Nagelkerke’s pseudo R2 ,
provided by adding polygenic score as a predictor to the model. The barplot appends
the Wald test P -value for polygenic score above each bar.
The high-resolution plot shows a green line that connects points corresponding to the
model-fit at the broad P -value thresholds considered in the bar plot. This demonstrates
the performance of high-resolution thresholds in comparison to the more traditional candidate thresholds.
The third output file contains polygenic scores for each individual at each threshold.
This file is called PRSice SCORES AT ALL THRESHOLDS.txt. The P -value thresholds used
for each column are specified in the column headings, and the first column is individuals’
IDs.
The fourth output file contains all the raw data used to generate the high-resolution
plot. This is named PRSice RAW RESULTS DATA.txt and is written to the working directory. It is in the following format:
thresh
1e-04
0.0001
2e-04
0.0003
p.out
0.9138
0.8689
0.8689
0.8689
r2.out
44.8535e-06
11.1294e-05
11.1294e-05
11.1294e-05
nsnps
46
47
51
67
If multiple base and target phenotypes are used, additional files are generated. The file
PRSice ALL BEST THRESHOLDS BASE AND TARGET.txt contains the P -value for the most
predictive model between each pair of phenotypes in a matrix. This is also displayed
visually in a heatmap called PRSice HEATMAP.png. NB: we reccomend using ggfig T
for this, as heatmaps produced using base graphics are less readable.
21
(a) Bar plot generated by PRSice as default, demonstrating PRS-model fit across a small number of thresholds,
PT .
(b) High-resolution plot generated by
PRSice as default, showing PRS model-fit
over a large number of thresholds, PT .
(c) Quantiles plot generated using quantiles T showing
the effect of increasing score on risk of disease, at the most
predictive threshold, PT .
Figure 1: : Figures generated by PRSice run at high-resolution. The best-fit bar, added
to the bar plot, is calculated from a high-resolution run, and so is not available if the
high resolution option is not used. The quantiles plot demonstrates that increasing PRS
is associated with increasing odds of phenotype.
22
7 Acknowledgements
We would like to thank all beta testers for their help and invaluable suggestions:
• Jonathan Coleman
• Simone de Jong
• Niamh Mullins
• Robert Power
23
References
[1] C. C. Chang, C. C. Chow, L. C. A. M. Tellier, S. Vattikuti, S. M. Purcell, and J. J.
Lee. Second-generation PLINK: rising to the challenge of larger and richer datasets.
ArXiv e-prints, October 2014.
[2] Toby Johnson. gtx: Genetics ToolboX, 2013. R package version 0.0.8.
24