RASQUAL (Robust Allele Specific QUAntification and quality controL) maps QTLs for sequenced based cellular traits by combining population and allele-specific signals.
09-08-2018 : Introduced assay specific QC filters in createASVCF.sh.
How to build & install
Please make sure CLAPACK and GSL (GNU Scientific Library) are installed in your environment (if you don't have them, then see below for installation tips). GSL is usually installed in /usr directory. Please check /usr/include/gsl and /usr/lib/libgsl.a are existing.
To build and install RASQUAL, firstly go to the source directory (src), then set environment variables appropriately to point to the CLAPACK library and GSL. Finally use "make" to build and install RASQUAL which will be installed in "$RASQUALDIR/bin".
RASQUALDIR=/path/to/rasqualdir/ cd $RASQUALDIR/src # Not run! Please export your environment. export CFLAGS="-I/path/to/your/CLAPACK-*.*.*.*/INCLUDE -I/path/to/your/CLAPACK-*.*.*.*/F2CLIBS -I/path/to/your/gsl-*.*" export LDFLAGS="-L/path/to/your/CLAPACK-*.*.*.* -L/path/to/your/CLAPACK-*.*.*.*/F2CLIBS -I/path/to/your/gsl-*.*/lib" make make install
QTL mapping with RASQUAL
This section describes how to prepare the input files for RASQUAL. If you prefer, Kaur Alasoo has now developed a set of tools in Python that hopefully simplify the process of creating RASQUAL input files here. RASQUAL needs two input data files:
- A fragment (read) count table, with sample specific offsets (such as library size)
- A VCF file with *phased* SNP genotypes and allele-specific counts.
An example of each of these files can be found in the data directory. In the usage example below, RASQUAL takes SNP information from a tabix-indexed VCF file as standard input, while the count table and sample specific offsets are binary files (Y.bin and K.bin, respectively). Tabix-indexing is not strictly necessary but allows for genotype and allelic count information to be accessed quickly from the command line. Using the example data files, you can use the following commands to map expression QTLs for two genes (C11orf21 and TSPAN32) using RASQUAL:
# make sure tabix is installed in your environment cd $RASQUALDIR tabix data/chr11.gz 11:2315000-2340000 | bin/rasqual -y data/Y.bin -k data/K.bin -n 24 -j 1 -l 378 -m 62 \ -s 2316875,2320655,2321750,2321914,2324112 -e 2319151,2320937,2321843,2323290,2324279 \ -t -f C11orf21 -z tabix data/chr11.gz 11:2315000-2340000 | bin/rasqual -y data/Y.bin -k data/K.bin -n 24 -j 2 -l 378 -m 60 \ -s 2323227,2323938,2324640,2325337,2328175,2329966,2330551,2331219,2334884,2335715,2338574,2339093 \ -e 2323452,2324188,2324711,2325434,2328220,2330040,2330740,2331248,2334985,2337897,2338755,2339430 \ -t -f TSPAN32 -z
Sample size (in this example, N=24) is given by -n option and the feature ID is given by -j option. Here only two genes exist in this example, thereby j=1, 2. In reality, you may have e.g. >10,000 features in your data, of which you may want to map QTL e.g. for the 12,345th feature, you must set -j 12345. You need to provide the number of testing SNPs and feature SNPs in the cis-window a priori (-l and -m, respectively). RASQUAL also requires the feature start and end positions (as comma separated values without space in ascending order for more than one positions, e.g. such as for a union of exons in this example) as inputs (-s and -e, respectively). By default, RASQUAL outputs QTL mapping results for all tested SNPs, but you can also specify only the lead QTL SNP (-t option). In the output, you can also specify the feature name by -f option. To take account of genotype uncertainty, imputation quality score (R square value) are used in this example (-z option; see the section below).
On output, RASQUAL provides the following values for each tested SNP:
- Feature ID*
- rs ID*
- SNP position*
- Ref allele
- Alt allele
- Allele frequency (not MAF!)*
- HWE Chi-square statistic
- Imputation quality score (IA)
- Log_10 Benjamini-Hochberg Q-value
- Chi square statistic (2 x log Likelihood ratio)*
- Effect size (Pi)
- Sequencing/mapping error rate (Delta)
- Reference allele mapping bias (Phi)
- SNP ID within the region
- No. of feature SNPs
- No. of tested SNPs*
- No. of iterations for null hypothesis
- No. of iterations for alternative hypothesis
- Random location of ties (tie lead SNP; only useful with -t option)
- Log likelihood of the null hypothesis
- Convergence status (0=success)
- Squared correlation between prior and posterior genotypes (fSNPs)
- Squared correlation between prior and posterior genotypes (rSNP)*
You may need columns with (*) for the downstream analysis.
You can find an example expression data for C11orf21 and TSPAN32 genes in the data directory. There are two files: a read/fragment count table (Y.txt) and sample specific offset data (K.txt), both have to be organised in feature by sample format (i.e., row: gene; col: sample). We have included an R script to allow you to convert the count and offset files (text) into binary format for us by RASQUAL. The script converts a table data into a vector of double precision values.
cd $RASQUALDIR RHOME=/software/R-3.0.0/bin/ $RHOME/R --vanilla --quiet --args data/Y.txt data/K.txt < R/txt2bin.R > log
Note here that, the row number of the fragment count table corresponds to the feature ID given by -j option previously explained. For example, if you set -j 12345 in the RASQUAL command, you would map QTL for the 12,345th feature (row) from the top of the fragment count table.
You will also need to prepare custom VCF files containing the allele specific counts of your target cellular trait at all SNPs. The files have to contain an additional subfield, "AS", located within the genotype field consisting of two integers, the reference and alternative allele counts, separated by a comma. For example, sample i is heterozygous at a SNP and has 1 and 10 reads overlapping the reference and alternative alleles respectively, the genotype field for the sample becomes
... FORMAT ... Sample_i ... ... GT:AS ... 0|1:1,10 ...
An example VCF file (chr11.gz) can be found in the data directory. Note that, phased genotypes are required for QTL mapping with RASQUAL. Currently, SNP genotypes are only used to map QTLs, but short INDELs and some form of structural variations will be able to use shortly.
Note on QC for genotype error correction: We have found that, in rare cases, RASQUAL may aggressively overcorrect genotyping errors to obtain a higher likelihood ratio. To detect this, we recommend that you always check the squared correlation between prior and posterior rSNP genotypes on the 25th column of the output. Cases where there is a very large change in genotypes between the prior and posterior should be treated with caution. Another approach that you can use to detect these instances, is to stop updating posterior genotypes (--no-posterior-update) or use the nominal genotype 0, 1 and 2 (--fix-genotype option) and compare the Chi-square statistics with and without genotype correction. Cases where weakly significant QTLs become highly significant with genotype error correction should be treated with caution.
There are 4 options to incorporate genotype uncertainty in RASQUAL:
Allelic probability (AP)
Allelic probability can be obtained from the standard 2-step imputation scheme, where genotypes are first phased then imputed on haplotype-by-haplotype basis. The custom AP field consists of the allelic probabilities (in Log10 scale) of two haplotypes from one individual, separated by a comma:
... FORMAT ... Sample_i ... ... AP:AS ... 0.0,-5.0:1,10 ...
Genotype likelihood (GL & GP)
You may also use genotype likelihood (in Log10 scale) from conventional genotype imputation in conjunction with phased genotype data:
... FORMAT ... Sample_i ... ... GT:GL:AS ... 0|1:-4.0,-0.1,-0.6:1,10 ...
Likewise, you can also provide the genotype likelihood in nominal scale [0-1] with GP FORMAT:
... FORMAT ... Sample_i ... ... GT:GP:AS ... 0|1:0.01,0.99,0.0:1,10 ...
Genotype dosage can also be utilised to take account of genotype uncertainty:
... FORMAT ... Sample_i ... ... GT:DS:AS ... 0|1:1.1:1,10 ...
Imputation quality score (R square value; RSQ)
Imputation methods often provide a quality score for each SNP locus that approximates squared correlation coefficient between true and observed genotypes (e.g., R^2 from MaCH or Beagle; I^2 from IMPUTE2 ). RASQUAL can convert the score into genotyping error rate to handle uncertainly:
... INFO FORMAT ... Sample_i ... ... ...;RSQ=0.9;... GT:AS ... 0|1:1,10 ...
If you want to prioritise RSQ, you need to specify -z option for RASQUAL (see above example).
Population allele frequency
If genotype information is not available, RASQUAL can run in SNP free mode (--population-allele-frequency option). In this case, you just need to provide the population allele frequency (not minor allele frequency!) for the SNP in VCF files. Note that, this mode is only feasible when the causal SNP is a feature SNP (e.g., ChIP-seq QTL).
... INFO FORMAT ... Sample_i ... ... ...;AF=0.4;... AS ... 1,10 ...
We strongly recommend using the AP field for QTL mapping. If there are multiple subfields of AP, GL and DS, AP is prioritised over both GL and DS, and GL is prioritised over DS. If neither of GL, DS nor AP is provided, GL is used as AP.
Sample specific offset terms (K.txt) can be calculated from the count table. See the script makeOffset.R in the R directory. The usage is:
# Not run! $RHOME/R --vanilla --quiet --args data/your.Y.txt < R/makeOffset.R > log # With GC content; not run! $RHOME/R --vanilla --quiet --args data/your.Y.txt data/gcc.txt < R/makeOffset.R > log
Note that you need to prepare a GC content file (gcc.txt in this example) to apply GC correction for the read count at each feature. The file is a vector of GC% values for all features as a text file (separated by either, a comma, a tab or a line break). In order to obtain the GC% for each feature, we normally extract the reference sequence overlapping with the feature annotation, count G/C bases and then divide the count by the total feature length. Note also that you may not convert the test data (Y.txt in the data directory) into K.txt because the offset was calculated and extracted from the complete expression data (>50K genes).
Real data is usually affected by hidden confounding factors, such as sequencing batch, sample preparation date etc, that can reduce power to detect QTLs. RASQUAL handles covariates as an input (-x option). The following is the same eQTL mapping example above, but with covariates:
tabix data/chr11.gz 11:2315000-2340000 | bin/rasqual -y data/Y.bin -k data/K.bin -n 24 -j 1 -l 378 -m 62 \ -s 2316875,2320655,2321750,2321914,2324112 -e 2319151,2320937,2321843,2323290,2324279 \ -z -t -f C11orf21 \ -x data/X.bin
The covariate file is based on a sample-by-variable table (see X.txt in the data directory). You can convert the file into binary format for RASQUAL by using the same R script:
$RHOME/R --vanilla --quiet --args data/Y.txt data/K.txt data/X.txt < R/txt2bin.R > log
Those confounding factors are not often observed but can be captured by principal component analysis (PCA). In our example, we applied PCA onto log FPKMs with and without permutation and picked the first N components whose explained variances are greater than those from permutation result as covariates for subsequent analyses. A sample code is also available in the R directory:
# Not run! $RHOME/R --vanilla --quiet --args data/your.Y.txt data/your.K.txt < R/makeCovariates.R > log
Note that, the result of PCA is always sensitive to few outliers (just one or two), which explain almost all variance in the data. Using those PCs as covariates hurts your QTL mapping result. We strongly recommend to spend some time to explore and clean up your data first.
Installation tips for CLAPACK and GSL
You first need to get the latest library from http://www.netlib.org/clapack/. Then, compile it like
tar zxvf clapack.tgz cd CLAPACK-3.2.1 mv make.inc.example make.inc make
When it has been done, you will find three archive files in the directory which need to be either linked or renamed, such as
ln -s lapack_LINUX.a liblapack.a ln -s tmglib_LINUX.a libtmglib.a ln -s blas_LINUX.a libblas.a
before compiling RASQUAL.
You may also need to get GSL (GNU Scientific Library) from http://www.gnu.org/software/gsl/ (if it is not installed). Then, compile it like
tar zxvf gsl-*.*.tar.gz cd gsl-*.* ./configure --prefix=$PWD make make install
Creating a VCF file with AS counts
We provide a useful script to create a VCF file with AS counts from a master VCF file and a set of BAM files. To use the script you need to compile some C codes called from the script:
export RASQUALDIR=/path/to/rasqualdir/ cd $RASQUALDIR/src/ASVCF make
The command to produce the VCF with AS counts is:
# Usage: bash ./createASVCF.sh # Usage: ./createASVCF.sh [paired_end|single_end] bam_list_file vcf_input_file vcf_output_file assay_type # For single-end reads from RNA-seq: bash ./createASVCF.sh single_end bam.list master.vcf.gz master.vcf.new.gz rna # For paired-end reads form ATAC-seq: bash ./createASVCF.sh paired_end bam.list master.vcf.gz master.vcf.new.gz atac
which creates a VCF output file. The VCF inpuy file must be tabix indexed and the BAM list file is a text file which contains absolute paths to your BAM files for which AS counts are produced:
# bam.list.txt /path/to/your/bam/sample1.bam /path/to/your/bam/sample2.bam /path/to/your/bam/sample3.bam ...
The order of the samples MUST be the same as that in the master VCF. The chromosome IDs in the master VCF MUST also be the same as in those BAM files. Before using the script, please make sure the latest tabix (http://www.htslib.org/doc/tabix.html) is installed in your environment.
Note that, assay_type specifies the minimum and maximum fragment length to filter out extremely short/long paired-end fragments and can be omitted in general. The latest createASVCF.sh can filter out AS read by means of QC criteria (mapping quality, max number of mismatches, etc.). See the usage of qcFilterBam in the ASVCF directory. You may also want to filter out reads a priori using other software, such as GATK ASEReadCounter (https://www.broadinstitute.org/gatk/gatkdocs/org_broadinstitute_gatk_tools_walkers_rnaseq_ASEReadCounter.php).
The permutation test is also implemented in RASQUAL. The -r/--random-permutation option generates a random permutation for each feature to break the correlation between genotype and total feature count as well as AS counts.
Tips to speed up QTL mapping
RASQUAL is now multithreaded in order to speed up execution times - this requires the pthread library to be installed. You need to specify the additional option --n-threads to use this function. Using the above example, you can use the following commands to map the eQTL with 10 threads:
tabix data/chr11.gz 11:2315000-2340000 | bin/rasqual -y data/Y.bin -k data/K.bin -n 24 -j 1 -l 409 -m 63 \ -s 2316875,2320655,2321750,2321914,2324112 -e 2319151,2320937,2321843,2323290,2324279 \ -t -f C11orf21 -z --n-threads 10
Filters for fSNPs
To maximize power to detect QTLs, RASQUAL uses all fSNPs with MAF>0.0, pHWE>0.0 and imputation quality score RSQ>0.0. However, RASQUAL can take a long time to map QTLs with a large number of fSNPs in a feature (for example, in very long genes). Therefore you may want to reduce the number of fSNPs with additional filters. We introduced the following new options --minor-allele-frequency-fsnp, --imputation-quality-fsnp and --hardy-weinberg-pvalue-fsnp to eliminate some of fSNPs which may not contribute much information.
Conditional analysis (UNDER DEVELOPMENT!)
There is currently an issue with conditional analysis using RASQUAL, and we recommend not using this option at the current time.
To map subsidiary QTLs conditional on the lead QTL variant(s) can be performed with the following option:
bin/rasqual ... -k2 rs0001:0.1,rs0002:0.2 ...
You may introduce any number of variants with their effect sizes (Pi values) as comma separated values where each variant ID and its Pi value is separaed by a colon (:).
To save memory, each variant ID in the VCF file must be shorter than 100 characters.