Introduction

Whole-genome sequencing (WGS) data is being generated at an unprecedented rate

• 1000 Genomes Project Phase 3 (1KG)
  • 81 million variants and 2,504 individuals
  • https://www.internationalgenome.org/
• Variant Call Format (VCF)
  • a generic and flexible text-based format
  • VCF files are large and data retrieval is relatively slow

Methods

CoreArray (C++ library)

• designed for large-scale data management of genome-wide variants
• data format (GDS) to store multiple array-oriented datasets in a single file

Two R packages

• gdsfmt – R interface to CoreArray Genomic Data Structure (GDS) files
• SeqArray – specifically designed for data management of genome-wide sequence variants from Variant Call Format (VCF) files

Methods – Advantages

• SeqArray provides the same capabilities as VCF

• Stores data in a binary and array-oriented manner

  • efficient access using the R language

• Genotypes are stored in a compressed manner

  • 2-bit array to store alleles (95% sites are bi-allelic)
  • rare variants: highly compressed
  • 1KG, 203.5 billion genotypes, saved in 4.3G (2.26% if a byte stores a genotype)

• Parallel access

  • multiple cluster nodes and/or cores

Methods – File Contents

File: SeqArray/extdata/CEU_Exon.gds (387.3K)
|--+ description   [  ] *
|--+ sample.id   { Str8 90 ZIP_ra(30.8%), 222B }
|--+ variant.id   { Int32 1348 ZIP_ra(35.7%), 1.9K }
|--+ position   { Int32 1348 ZIP_ra(86.4%), 4.6K }
|--+ chromosome   { Str8 1348 ZIP_ra(2.66%), 91B }
|--+ allele   { Str8 1348 ZIP_ra(17.2%), 928B }
|--+ genotype   [  ] *
|  \--+ data   { Bit2 2x90x1348 ZIP_ra(28.4%), 16.8K } *
|--+ phase   [  ]
|  \--+ data   { Bit1 90x1348 ZIP_ra(0.36%), 55B } *
|--+ annotation   [  ]
|  |--+ id   { Str8 1348 ZIP_ra(41.0%), 5.8K }
|  |--+ qual   { Float32 1348 ZIP_ra(0.91%), 49B }
|  |--+ filter   { Int32,factor 1348 ZIP_ra(0.89%), 48B } *
|  |--+ info   [  ]
|  |  |--+ AA   { Str8 1348 ZIP_ra(24.2%), 653B } *
|  |  \--+ HM2   { Bit1 1348 ZIP_ra(117.2%), 198B } *
|  \--+ format   [  ]
|     \--+ DP   [  ] *
|        \--+ data   { Int32 90x1348 ZIP_ra(33.8%), 160.3K }
\--+ sample.annotation   [  ]
   \--+ family   { Str8 90 ZIP_ra(34.7%), 135B }

Methods – Key Functions

Table 1: The key functions in the SeqArray package.

Benchmark

• Dataset
  • 1000 Genomes Project Phase 3, chromosome 1
  • 6,468,094 variants, 2,504 individuals
  • original VCF.gz file: 1.2G
  • reformat to a SeqArray file: 458M (zlib compression)
• Calculate the frequencies of reference alleles
  1. R code (sequential version)
  2. R code (parallel version)
  3. R and C++ integration via the Rcpp package

Benchmark – Test 1 (sequentially)

# load the R package
library(SeqArray)

# open the file
genofile <- seqOpen("1KG_chr1.gds")

# apply a user-defined function over variants
system.time(afreq <- seqApply(genofile, "genotype",
    FUN = function(x) { mean(x==0L, na.rm=TRUE) },
    as.is="double", margin="by.variant")
)

10.8 minutes on Linux with Intel Xeon CPU @2GHz and 128GB RAM function(x) { mean(x==0L, na.rm=TRUE) } is a user-defined function, where x is an integer matrix:

                           sample
  allele [,1] [,2] [,3] [,4] [,5]
    [1,]    0    1    0   NA    1
    [2,]    0    0    0    1    0

0 – reference allele, 1 – the first alternative allele

Benchmark – Test 2 (in parallel)

seqParallel() splits genotypes into 4 non-overlapping parts according to different cores.

# load the R package
library(parallel)

# create a computing cluster with 4 cores
seqParallelSetup(4)

# run in parallel
system.time(afreq <- seqParallel(gdsfile=genofile,
    FUN = function(f) {
        seqApply(f, "genotype", as.is="double", margin="by.variant",
            FUN = function(x) mean(x==0L, na.rm=TRUE))
    }, split = "by.variant")
)

3.1 minutes (vs. 10.8m in Test 1)

Benchmark – Test 3 (C++ Integration)

library(Rcpp)

# dynamically define an inline C/C++ function in R
cppFunction('double RefAlleleFreq(IntegerMatrix x) {
    int nrow = x.nrow(), ncol = x.ncol();
    int cnt=0, zero_cnt=0, g;
    for (int i = 0; i < nrow; i++) {
        for (int j = 0; j < ncol; j++) {
            if ((g = x(i, j)) != NA_INTEGER) {
                cnt ++;
                if (g == 0) zero_cnt ++;
            }
    }}
    return double(zero_cnt) / cnt;
}')

system.time(
    afreq <- seqApply(genofile, "genotype", RefAlleleFreq,
        as.is="double", margin="by.variant")
)

1.5 minutes (significantly faster! vs. 10.8m in Test 1)

Conclusion

SeqArray is of great interest to

• R users involved in data analyses of large-scale sequence variants
• particularly those with limited experience of parallel / high-performance computing

SeqVarTools (Bioconductor)

• variant analysis, such like allele frequencies, HWE, Mendelian errors, etc
• functions to display genotypes / annotations in a readable format

SNPRelate (Bioconductor)

• a parallel computing toolset for relatedness and principal component analysis

Resource

https://gds-stat.s3.amazonaws.com/download/1000g/index.html

1000 Genomes Project Phase 3:

• Autosomes (2.60GB, 2,504 individuals and 81,271,745 variants): 1KG_ALL.autosome.phase3_shapeit2_mvncall_integrated_v5a.20130502.genotypes.gds
• Chromosome X (94.1MB, 2,504 individuals and 3,468,093 variants): 1KG_ALL.chrX.phase3_shapeit2_mvncall_integrated_v1b.20130502.genotypes.gds
• Chromosome Y (2.70MB, 1,233 males and 62,042 variants): 1KG_ALL.chrY.phase3_integrated_v2a.20130502.genotypes.gds

Acknowledgements

Department of Biostatistics at University of Washington – Seattle

Genetic Analysis Center:

• Stephanie M. Gogarten
• David Levine
• Cathy Laurie