inst/scripts/make-data_MafDb.gnomAD.r2.1.hs37d5.R

In rcastelo/GenomicScores: Infrastructure to work with genomewide position-specific scores

## This README file explains the steps to download and freeze in this
## annotation package the gnomAD genome allele frequencies. If you use
## these data please cite the following publication:

## Karczewski et al. Variation across 141,456 human exomes and genomes
## reveals the spectrum of loss-of-function intolerance across human
## protein-coding genes. bioRxiv, 531210, 2019.
## doi: http://dx.doi.org/10.1101/531210

## See http://gnomad.broadinstitute.org/terms for further information
## on using this data for your own research

## The data were downloaded from as follows:
##
## allchr=`seq 1 22`" X"
## for chr in $allchr ; do {
##   gsutil -m cp -r gs://gnomad-public/release/2.1/vcf/genomes/gnomad.genomes.r2.1.sites.chr$chr.vcf.bgz.tbi .
##   gsutil -m cp -r gs://gnomad-public/release/2.1/vcf/genomes/gnomad.genomes.r2.1.sites.chr$chr.vcf.bgz .
## } done

## The following R script processes the downloaded data to
## transform the allele frequencies into minor allele frequencies
## and store them using only one significant digit for values < 0.1,
## and two significant digits for values > 0.1, to reduce the memory
## footprint through RleList objects

library(Rsamtools)
library(GenomicRanges)
library(GenomeInfoDb)
library(VariantAnnotation)
library(BSgenome.Hsapiens.1000genomes.hs37d5)
library(doParallel)

downloadURL <- "http://gnomad.broadinstitute.org/downloads"
citationdata <- bibentry(bibtype="Article",
                         author=c(person("Konrad J Karczewski"), person("et al.")),
                         title="Variation across 141,456 human exomes and genomes reveals the spectrum of loss-of-function intolerance across human protein-coding genes",
                         journal="bioRxiv",
                         pages="531210",
                         year="2019",
                         doi="10.1101/531210")

registerDoParallel(cores=4) ## up to 90 Gb per process

## quantizer function. it maps input real-valued [0, 1] allele frequencies
## to positive integers [1, 255] so that each of them can be later
## coerced into a single byte (raw type). allele frequencies between 0.1 and 1.0
## are quantized using two significant digits, while allele frequencies between
## 0 and 0.1 are quantized using one significant digit. quantized values are
## add up one unit to make them positive and keep the value 0 to encode for NAs
## since there is no NA value in the raw class (Sec. 3.3.4 NA handling, R Language Definition)
.quantizer <- function(x) {
  .ndec <- function(x) {
    spl <- strsplit(as.character(x+1), "\\.")
    spl <- sapply(spl, "[", 2)
    spl[is.na(spl)] <- ""
    nchar(spl)
  }

  maskNAs <- is.na(x)
  x[!maskNAs & x > 0.1] <- signif(x[!maskNAs & x > 0.1], digits=2)
  x[!maskNAs & x <= 0.1] <- signif(x[!maskNAs & x <= 0.1], digits=1)
  x[maskNAs] <- NA
  nd <- .ndec(x)
  x[nd < 3] <- x[nd < 3] * 100
  x[nd >= 3] <- x[nd >= 3] * 10^nd[nd >= 3] + 100 + (nd[nd >= 3] - 3) * 10
  x <- x + 1
  x[maskNAs] <- 0
  q <- as.integer(sprintf("%.0f", x)) ## coercing through character is necessary
                                      ## to deal with limitations of computer
                                      ## arithmetic such as with as.integer(0.29*100)
  if (any(q > 255))
    stop("current number of quantized values > 255 and cannot be stored into one byte")
  q
}
attr(.quantizer, "description") <- "quantize [0.1-1] with 2 significant digits and [0-0.1) with 1 significant digit"

## dequantizer function
.dequantizer <- function(q) {
  x <- q <- as.integer(q)
  x[x == 0L] <- NA
  x <- (x - 1L) / 100L
  maskNAs <- is.na(x)
  q <- q - 1L
  sel <- !maskNAs & q > 100L
  x[sel] <- ((q[sel] - 100L) %% 10L) / 10^(floor((q[sel] - 100L) / 10L) + 3L)
  x
}
attr(.dequantizer, "description") <- "dequantize [0-100] dividing by 100, [101-255] subtract 100, take modulus 10 and divide by the corresponding power in base 10"

vcfFilename <- "gnomad.genomes.r2.1.sites.chr21.vcf.bgz" ## just pick one file
genomeversion <- "hs37d5"
pkgname <- sprintf("MafDb.gnomAD.r2.1.%s", genomeversion)
dir.create(pkgname)

path2vcfs <- "/projects_fg/GenomicScores/gnomad/GRCh37/r2.1/genomes"

vcfHeader <- scanVcfHeader(file.path(path2vcfs, vcfFilename))

Hsapiens <- hs37d5
stopifnot(all(seqlengths(vcfHeader)[1:24] == seqlengths(Hsapiens)[1:24])) ## QC

## save the GenomeDescription object
refgenomeGD <- GenomeDescription(organism=organism(Hsapiens),
                                 common_name=commonName(Hsapiens),
                                 provider=provider(Hsapiens),
                                 provider_version=providerVersion(Hsapiens),
                                 release_date=releaseDate(Hsapiens),
                                 release_name=releaseName(Hsapiens),
                                 seqinfo=seqinfo(Hsapiens))
saveRDS(refgenomeGD, file=file.path(pkgname, "refgenomeGD.rds"))

## read INFO column data
infoCols <- rownames(info(vcfHeader))
AFcols <- infoCols[grep("^AF", infoCols)]
exclude <- grep("raw|male", AFcols)
if (length(exclude) > 0)
  AFcols <- AFcols[-exclude] ## remove columns such as those for maximum allele frequency among populations

message("Starting to process variants")

## restrict VCF INFO columns to AC and AN values
vcfPar <- ScanVcfParam(geno=NA,
                       fixed=c("ALT", "FILTER"),
                       info=AFcols)

foreach (chr=seqlevels(vcfHeader)[1:23]) %dopar% {

  message(sprintf("Reading VCF from chromosome %s", chr))

  ## the whole VCF for the chromosome into main memory (up to 55Gb of RAM for chromosome 1)
  vcf <- readVcf(file.path(path2vcfs, sprintf("gnomad.genomes.r2.1.sites.chr%s.vcf.bgz", chr)),
                 genome=genomeversion, param=vcfPar)

  ## discard variants not passing all FILTERS
  mask <- fixed(vcf)$FILTER == "PASS"
  vcf <- vcf[mask, ]
  gc()

  ## mask variants where all alternate alleles are SNVs
  evcf <- expand(vcf)
  maskSNVs <- sapply(relist(isSNV(evcf), alt(vcf)), all)
  rm(evcf)
  gc()

  ## treat snvs and nonSNVs separately
  vcfsnvs <- vcf[maskSNVs, ]
  vcfnonsnvs <- vcf[!maskSNVs, ]

  ##
  ## SNVs
  ##

  ## fetch SNVs coordinates
  rr <- rowRanges(vcfsnvs)

  ## clean up the ranges
  mcols(rr) <- NULL
  names(rr) <- NULL
  gc()

  ## fill up SeqInfo data
  si <- seqinfo(vcf)
  seqlengths(rr) <- seqlengths(seqinfo(Hsapiens))[match(seqnames(si), seqnames(seqinfo(Hsapiens)))]
  isCircular(rr) <- isCircular(seqinfo(Hsapiens))[match(seqnames(si), seqnames(seqinfo(Hsapiens)))]
  genome(rr) <- genome(seqinfo(Hsapiens))[match(seqnames(si), seqnames(seqinfo(Hsapiens)))]

  ## according to https://samtools.github.io/hts-specs/VCFv4.3.pdf
  ## "It is permitted to have multiple records with the same POS"
  ## and according to the release notes of gnomAD v2.1
  ## "all multi-allelic sites have been split. This means that
  ## multiple lines now have the same chromosome and position."
  ## in such a case we take the maximum MAF by looking at repeated positions
  rrbypos <- split(rr, start(rr))
  rr <- rr[!duplicated(rr)]
  ## put back the genomic order
  mt <- match(as.character(start(rr)), names(rrbypos))
  stopifnot(all(!is.na(mt))) ## QC
  rrbypos <- rrbypos[mt]
  rm(mt)
  gc()

  ## fetch allele frequency data
  afValues <- info(vcfsnvs)
  clsValues <- sapply(afValues, class)

  for (j in seq_along(AFcols)) {
    afCol <- AFcols[j]

    message(sprintf("Processing %s SNVs allele frequencies from chromosome %s", afCol, chr))
    mafValuesCol <- afValues[[afCol]]
    if (clsValues[afCol] == "numeric" || clsValues[afCol] == "Numeric") {
      mafValuesCol <- as.numeric(mafValuesCol)
    } else if (clsValues[afCol] == "character") {
      mafValuesCol <- as.numeric(sub(pattern="\\.", replacement="0", mafValuesCol))
    } else if (clsValues[afCol] == "CompressedNumericList") { ## in multiallelic variants take
      mafValuesCol <- sapply(mafValuesCol, max)               ## the maximum allele frequency
    } else if (clsValues[afCol] == "CompressedCharacterList") {
      mafValuesCol <- sapply(NumericList(lapply(mafValuesCol, sub, pattern="\\.", replacement="0")), max)
    } else {
      stop(sprintf("Uknown class for holding AF values (%s)", clsValues[afCol]))
    }

    ## allele frequencies from gnomAD are calculated from alternate alleles,
    ## so for some of them we need to turn them into minor allele frequencies (MAF)
    ## for biallelic variants, in those cases the MAF comes from the REF allele
    maskREF <- !is.na(mafValuesCol) & mafValuesCol > 0.5
    if (any(maskREF))
      mafValuesCol[maskREF] <- 1 - mafValuesCol[maskREF]

    mafValuesCol <- relist(mafValuesCol, rrbypos)
    maskREF <- relist(maskREF, rrbypos)
    mafValuesCol <- sapply(mafValuesCol, max) ## in multiallelic variants
                                              ## take the maximum allele frequency
    maskREF <- sapply(maskREF, any) ## in multiallelic variants, when any of the
                                    ## alternate alleles has AF > 0.5, then we
                                    ## set to TRUE maskREF as if the MAF is in REF

    q <- .quantizer(mafValuesCol)
    x <- .dequantizer(q)
    f <- cut(x, breaks=c(0, 10^c(seq(floor(min(log10(x[x!=0]), na.rm=TRUE)),
                                     ceiling(max(log10(x[x!=0]), na.rm=TRUE)), by=1))),
             include.lowest=TRUE)
    err <- abs(mafValuesCol-x)
    max.abs.error <- tapply(err, f, mean, na.rm=TRUE)

    ## build an integer-Rle object of MAF values using the 'coverage()' function
    obj <- coverage(rr, weight=q)[[chr]]
    ## build an integer-Rle object of maskREF using the 'coverage()' function
    maskREFobj <- coverage(rr, weight=maskREF+0L)[[chr]]

    ## build ECDF of MAF values
    if (length(unique(mafValuesCol)) <= 10000) {
      Fn <- ecdf(mafValuesCol)
    } else {
      Fn <- ecdf(sample(mafValuesCol, size=10000, replace=TRUE))
    }

    ## coerce to raw-Rle, add metadata and save
    if (any(runValue(obj) != 0)) {
      runValue(obj) <- as.raw(runValue(obj))
      runValue(maskREFobj) <- as.raw(runValue(maskREFobj))
      metadata(obj) <- list(seqname=chr,
                            provider="BroadInstitute",
                            provider_version="r2.1",
                            citation=citationdata,
                            download_url=downloadURL,
                            download_date=format(Sys.Date(), "%b %d, %Y"),
                            reference_genome=refgenomeGD,
                            data_pkgname=pkgname,
                            qfun=.quantizer,
                            dqfun=.dequantizer,
                            ecdf=Fn,
                            max_abs_error=max.abs.error,
                            maskREF=maskREFobj)
      saveRDS(obj, file=file.path(pkgname, sprintf("%s.%s.%s.rds", pkgname, afCol, chr)))
    } else {
      warning(sprintf("No MAF values for SNVs in chromosome %s", chr))
    }
  }

  rm(vcfsnvs)
  gc()

  ##
  ## nonSNVs
  ##

  ## fetch nonSNVs coordinates
  rr <- rowRanges(vcfnonsnvs)

  ## fill up SeqInfo data
  si <- seqinfo(vcf)
  seqlengths(rr) <- seqlengths(seqinfo(Hsapiens))[match(seqnames(si), seqnames(seqinfo(Hsapiens)))]
  isCircular(rr) <- isCircular(seqinfo(Hsapiens))[match(seqnames(si), seqnames(seqinfo(Hsapiens)))]
  genome(rr) <- genome(seqinfo(Hsapiens))[match(seqnames(si), seqnames(seqinfo(Hsapiens)))]

  ## clean up the GRanges
  mcols(rr) <- NULL
  names(rr) <- NULL
  gc()

  ## re-order by chromosomal coordinates to deal with wrongly-ordered nonSNVs over multiple VCF lines
  ord <- order(rr)
  rr <- rr[ord]

  ## fetch allele frequency data
  afValues <- info(vcfnonsnvs)
  clsValues <- sapply(afValues, class)
  rm(vcf)
  rm(vcfnonsnvs)
  gc()

  ## re-order by chromosomal coordinates
  afValues <- afValues[ord, ]

  ## according to https://samtools.github.io/hts-specs/VCFv4.3.pdf
  ## "It is permitted to have multiple records with the same POS"
  ## and according to the release notes of gnomAD v2.1
  ## "all multi-allelic sites have been split. This means that
  ## multiple lines now have the same chromosome and position."
  ## in such a case we take the maximum MAF by looking at repeated positions
  posids <- paste(start(rr), end(rr), sep="-")
  rrbypos <- split(rr, posids)
  rr <- rr[!duplicated(rr)]
  ## put back the genomic order
  posids <- paste(start(rr), end(rr), sep="-")
  mt <- match(posids, names(rrbypos))
  stopifnot(all(!is.na(mt))) ## QC
  rrbypos <- rrbypos[mt]
  saveRDS(rr, file=file.path(pkgname, sprintf("%s.GRnonsnv.%s.rds", pkgname, chr)))

  rm(ord)
  rm(posids)
  rm(mt)
  gc()

  for (j in seq_along(AFcols)) {
    afCol <- AFcols[j]

    message(sprintf("Processing %s nonSNVs allele frequencies from chromosome %s", afCol, chr))
    mafValuesCol <- afValues[[afCol]]
    if (clsValues[afCol] == "numeric" || clsValues[afCol] == "Numeric") {
      mafValuesCol <- as.numeric(mafValuesCol)
    } else if (clsValues[afCol] == "character") {
      mafValuesCol <- as.numeric(sub(pattern="\\.", replacement="0", mafValuesCol))
    } else if (clsValues[afCol] == "CompressedNumericList") { ## in multiallelic variants take
      mafValuesCol <- sapply(mafValuesCol, max)               ## the maximum allele frequency
    } else if (clsValues[afCol] == "CompressedCharacterList") {
      mafValuesCol <- sapply(NumericList(lapply(mafValuesCol, sub, pattern="\\.", replacement="0")), max)
    } else {
      stop(sprintf("Uknown class for holding AF values (%s)", clsValues[afCol]))
    }

    ## allele frequencies from gnomAD genomes are calculated from alternate alleles,
    ## so for some of them we need to turn them into minor allele frequencies (MAF)
    ## for biallelic variants, in those cases the MAF comes from the REF allele
    maskREF <- !is.na(mafValuesCol) & mafValuesCol > 0.5
    if (any(maskREF))
      mafValuesCol[maskREF] <- 1 - mafValuesCol[maskREF]

    mafValuesCol <- relist(mafValuesCol, rrbypos)
    maskREF <- relist(maskREF, rrbypos)
    mafValuesCol <- sapply(mafValuesCol, max) ## in multiallelic variants
                                              ## take the maximum allele frequency
    maskREF <- sapply(maskREF, any) ## in multiallelic variants, when any of the
                                    ## alternate alleles has AF > 0.5, then we
                                    ## set to TRUE maskREF as if the MAF is in REF

    q <- .quantizer(mafValuesCol)
    x <- .dequantizer(q)
    f <- cut(x, breaks=c(0, 10^c(seq(floor(min(log10(x[x!=0]), na.rm=TRUE)),
                                     ceiling(max(log10(x[x!=0]), na.rm=TRUE)), by=1))),
             include.lowest=TRUE)
    err <- abs(mafValuesCol-x)
    max.abs.error <- tapply(err, f, mean, na.rm=TRUE)

    ## build ECDF of MAF values
    if (length(unique(mafValuesCol)) <= 10000) {
      Fn <- ecdf(mafValuesCol)
    } else {
      Fn <- ecdf(sample(mafValuesCol, size=10000, replace=TRUE))
    }

    ## coerce the quantized value vector to an integer-Rle object
    obj <- Rle(q)
    ## coerce the maskREF vector to an integer-Rle object
    maskREFobj <- Rle(maskREF+0L)

    ## coerce to raw-Rle, add metadata and save
    if (any(runValue(obj) != 0)) {
      runValue(obj) <- as.raw(runValue(obj))
      runValue(maskREFobj) <- as.raw(runValue(maskREFobj))
      metadata(obj) <- list(seqname=chr,
                            provider="BroadInstitute",
                            provider_version="r2.1",
                            citation=citationdata,
                            download_url=downloadURL,
                            download_date=format(Sys.Date(), "%b %d, %Y"),
                            reference_genome=refgenomeGD,
                            data_pkgname=pkgname,
                            qfun=.quantizer,
                            dqfun=.dequantizer,
                            ecdf=Fn,
                            max_abs_error=max.abs.error,
                            maskREF=maskREFobj)
      saveRDS(obj, file=file.path(pkgname, sprintf("%s.RLEnonsnv.%s.%s.rds", pkgname, afCol, chr)))
    } else {
      warning(sprintf("No MAF values for nonSNVs in chromosome %s", chr))
    }
  }
}

## save rsIDs assignments from gnomAD
## iterating through every chromosome VCF file
vcfPar <- ScanVcfParam(geno=NA,
                       fixed=c("ALT", "FILTER"),
                       info=NA)

message("Starting to process variant identifiers")

rsIDs <- character(0)  ## to store rsIDs annotated by the gnomAD project
rsIDgp <- GPos()       ## to store positions of rsIDs
maskSNVs <- logical(0) ## to store a mask whether the variant is an SNV or not

nTotalVar <- 0
for (chr in seqlevels(vcfHeader)[1:23]) {
  vcf <- readVcf(file.path(path2vcfs, sprintf("gnomad.genomes.r2.1.sites.chr%s.vcf.bgz", chr)),
                 genome=genomeversion, param=vcfPar)

  ## discard variants not passing all FILTERS
  mask <- fixed(vcf)$FILTER == "PASS"
  vcf <- vcf[mask, ]
  gc()

  nVar <- nrow(vcf)
  nTotalVar <- nTotalVar + nVar
  rr <- rowRanges(vcf)
  mcols(rr) <- NULL
  whrsIDs <- grep("^rs", names(rr))
  evcf <- expand(vcf)
  maskSNVs <- c(maskSNVs, sapply(relist(isSNV(evcf), alt(vcf)), all)[whrsIDs])
  rm(evcf)
  gc()
  rrTmp <- rr[whrsIDs]
  mcols(rrTmp) <- NULL
  rrTmp <- resize(rrTmp, width=1, fix="start", ignore.strand=TRUE)
  idTmp <- names(rrTmp)
  names(rrTmp) <- NULL
  gpTmp <- as(rrTmp, "GPos")

  ## just in case there are multiple rsID assignments separated by
  ## semicolons like it happens with ExAC, just assign the first one
  idTmp <- strsplit(idTmp, ";")
  idTmp <- sapply(idTmp, "[", 1)

  rsIDs <- c(rsIDs, idTmp)
  rsIDgp <- c(rsIDgp, gpTmp)
  rm(rrTmp)
  rm(gpTmp)
  rm(idTmp)
  gc()

  message(sprintf("%d variant identifiers processed from chromosome %s", nVar, chr))
}
message(sprintf("%d variants processed in total", nTotalVar))

## save the total number of variants
saveRDS(nTotalVar, file=file.path(pkgname, "nsites.rds"))

## store mask flagging SNVs
rsIDgp$isSNV <- Rle(maskSNVs)

## double check that all identifiers start with 'rs'
stopifnot(identical(grep("^rs", rsIDs), 1:length(rsIDs))) ## QC

## chop the 'rs' prefix and convert the character ids into integer values
rsIDs <- as.integer(sub(pattern="^rs", replacement="", x=rsIDs))
gc()

## calculate the indices that lead to an increasing values of the (integer) ids
rsIDidx <- order(rsIDs)

## order increasingly the integer values of rs IDs
rsIDs <- rsIDs[rsIDidx]

## save the objects that enable the search for rs ID
saveRDS(rsIDs, file=file.path(pkgname, "rsIDs.rds"))
saveRDS(rsIDidx, file=file.path(pkgname, "rsIDidx.rds"))
saveRDS(rsIDgp, file=file.path(pkgname, "rsIDgp.rds"))