[Seqinr-commits] r1651 - pkg/man

noreply at r-forge.r-project.org noreply at r-forge.r-project.org
Tue Sep 22 15:10:31 CEST 2009


Author: lobry
Date: 2009-09-22 15:10:31 +0200 (Tue, 22 Sep 2009)
New Revision: 1651

Added:
   pkg/man/fastacc.Rd
Log:
 

Added: pkg/man/fastacc.Rd
===================================================================
--- pkg/man/fastacc.Rd	                        (rev 0)
+++ pkg/man/fastacc.Rd	2009-09-22 13:10:31 UTC (rev 1651)
@@ -0,0 +1,198 @@
+\name{fastacc}
+\alias{fastacc}
+\title{Fast Allele in Common Count}
+\description{
+The purpose of this function is to compute as fast as possible the number
+of allele in common between a target (typically the genetic profile observed at a
+crime scene, possibly a mixture with dropouts) and a database reference (typically
+genetic profile of individuals). Both are assumed to be pre-encoded at the bit
+level in a consistent way.
+}
+\usage{
+fastacc(target, database)
+}
+\arguments{
+  \item{target}{the \code{\link{raw}} encoding of the target, typically 40 octets for a core-CODIS profile in 2009}
+  \item{database}{the \code{\link{raw}} encoding of the database. If there are n entries
+in the database, then the database must n times longer than the target.}
+}
+\details{
+This function is an RFC state. Comments are welcome.
+
+Genetic profiles are encoded at the bit level. One bit represents one allele.
+Count is based on a logical AND at bit level. Bit count is encoded at C level
+using the precomputed approach: one indirection with an auxiliary table
+of size 256 called \code{bits_in_char} which is pre-computed at R level and
+passed at C level.
+}
+\value{
+A vector of \code{\link{integer}} giving for each entry in the database how many
+alleles are in common between the entry and the target.
+}
+\references{
+\code{citation("seqinr")}
+}
+\author{
+J.R. Lobry
+}
+
+\section{Warning }{Experimental, first release schedulded for seqinr  2.0-6 by the end of 2009}
+
+\seealso{
+FIXME
+}
+\examples{
+#
+# NOTE:
+#
+# This example section is a proof-of-concept stuff. Most code should be
+# enbeded in documented functions to avoid verbosity. But at the RFC stage
+# this is perhaps not a too bad idea to show how powerfull R is.
+#
+
+#
+# Let's start from the 16 loci available in the AmpFLSTR kit:
+#
+
+path <- system.file("abif/AmpFLSTR_Bins_v1.txt", package = "seqinr")
+resbin <- readBins(path)
+codis <- resbin[["Identifiler_CODIS_v1"]]
+names(codis)
+
+#
+# We count how many different alleles are present per locus:
+#
+
+na <- unlist(lapply(codis, function(x) length(x[[1]])))
+na
+
+#
+# The number of octets required to encode a genetic for each locus is then:
+#
+
+ceiling(na/8)
+
+#
+# We need then a total of 40 octets to code these profiles:
+#
+
+sum(ceiling(na/8))
+
+#
+# Let's definene a function to encode a profile at a given locus, and vice versa :
+#
+
+prof2raw <- function(profile, alleles) {
+  if (!is.ordered(alleles)) stop("ordered factor expected for alleles")
+  if (!is.character(profile)) stop("vector of character expected for profile")
+  noctets <- ceiling(length(alleles)/8)
+  res.b <- rawToBits(raw(noctets))
+  for (i in 1:length(profile)) {
+    res.b[which(profile[i] == alleles)] <- as.raw(1)
+  }
+  return(packBits(res.b, type = "raw"))
+}
+
+raw2prof <- function(rawdata, alleles) {
+  if (!is.ordered(alleles)) stop("ordered factor expected for alleles")
+  if (!is.raw(rawdata)) stop("vector of raw expected for rawdata")
+  res <- as.character(alleles)[as.logical(rawToBits(rawdata))]
+  return(paste(res, collapse = ", "))
+}
+
+#
+# Let now code all alleles present in codis as ordered factors:
+#
+
+allalleles <- lapply(codis, function(x) factor(x[, 1], levels = x[, 1], ordered = TRUE))
+
+#
+# Let's play with our encoding/decoding utilities with first locus:
+#
+
+allalleles[[1]] #  <8 8 9 10 11 12 13 14 15 16 17 18 19 >19
+res <- prof2raw(c("8", "9", "13", "14", ">19"), allalleles[[1]])
+res # c6 20
+rawToBits(res) # 00 01 01 00 00 00 01 01 00 00 00 00 00 01 00 00
+raw2prof(res, allalleles[[1]]) #  "8, 9, 13, 14, >19"
+
+#
+# Let define a profile with all possible alleles:
+#
+
+ladder <- unlist(lapply(allalleles, function(x) prof2raw(as.character(x),x)))
+names(ladder) <- NULL
+stopifnot(identical(as.integer(ladder), 
+ c(255L, 63L, 255L, 255L, 255L, 63L, 255L, 63L, 255L, 31L, 255L, 
+ 63L, 255L, 255L, 7L, 255L, 3L, 255L, 63L, 255L, 255L, 255L, 255L, 
+ 15L, 255L, 127L, 255L, 3L, 255L, 255L, 255L, 255L, 3L, 3L, 255L, 
+ 15L, 255L, 255L, 255L, 7L))) # simple sanity check
+
+#
+# Let's make a simulated database. Here we use a random sampling
+# with a uniform distribution between all possible profile possible
+# at a given locus. A more realist sampling for an individual database
+# would be to sample only two alleles at each locus according to
+# observed frequencies in populations. 
+#
+
+n <- 10^5 # the number of records in the database
+DB <- sapply(ladder, function(x) as.raw(sample(0:as.integer(x), size = n, replace = TRUE)))
+
+#
+# Now we make sure that the target is in the database:
+#
+
+target <- DB[666, ]
+DB <- as.vector(t(DB)) # put DB as a flat database (is it usefull?) 
+
+#
+# Now we compute the number of alleles in common between the
+# target and all the entries in the DB:
+#
+
+system.time(res <- fastacc(target,DB)) # Fast, isn't it ?
+stopifnot(which.max(res) == 666) # sanity check
+
+#
+# Don't run : too tedious for routine check. We check here that complexity is
+# linear in time up to a 10 10^6 database size (roughly the size of individual
+# profiles at the EU level)
+#
+
+\dontrun{
+maxn <- 10^7
+DB <- sapply(ladder, function(x) as.raw(sample(0:as.integer(x),
+  size = maxn, replace = T)))
+target <- DB[666, ]
+DB <- as.vector(t(DB))
+
+np <- 10
+nseq <- seq(from = 10^5, to = maxn, length = np)
+res <- numeric(np)
+i <- 1
+for (n in nseq) {
+  print(i)
+  res[i] <- system.time(tmp <- fastacc(target, DB[1:n]))[1]
+  stopifnot(which.max(tmp) == 666)
+  i <- i + 1
+}
+dbse <- data.frame(list(nseq = nseq, res = res))
+
+x <- dbse$nseq
+y <- dbse$res
+plot(x, y, type = "b", xlab = "Number of entries in DB", ylab = "One query time [s]",
+las = 1, xlim = c(0, maxn), ylim = c(0, max(y)), main = "Data base size effect on query time")
+lm1 <- lm(y ~ x - 1)
+abline(lm1, col = "red")
+legend("topleft", inset = 0.01, legend = paste("y =", formatC(lm1$coef[1],
+digits = 3), "x"), col = "red", lty = 1)
+
+#
+# On my laptop the slope is 2.51e-08, that is a 1/4 of second to scan a database
+# with 10 10^6 entries.
+#
+}
+
+## end
+}



More information about the Seqinr-commits mailing list