matchPDict-exact: Matching a dictionary of patterns against a reference

In anandhupresannan/biostrings: Efficient manipulation of biological strings

Description

A set of functions for finding all the occurrences (aka "matches" or "hits") of a set of patterns (aka the dictionary) in a reference sequence or set of reference sequences (aka the subject)

The following functions differ in what they return: matchPDict returns the "where" information i.e. the positions in the subject of all the occurrences of every pattern; countPDict returns the "how many times" information i.e. the number of occurrences for each pattern; and whichPDict returns the "who" information i.e. which patterns in the input dictionary have at least one match.

vcountPDict and vwhichPDict are vectorized versions of countPDict and whichPDict, respectively, that is, they work on a set of reference sequences in a vectorized fashion.

This man page shows how to use these functions (aka the *PDict functions) for exact matching of a constant width dictionary i.e. a dictionary where all the patterns have the same length (same number of nucleotides).

See ?`matchPDict-inexact` for how to use these functions for inexact matching or when the original dictionary has a variable width.

Usage

matchPDict(pdict, subject,
           max.mismatch=0, min.mismatch=0, with.indels=FALSE, fixed=TRUE,
           algorithm="auto", verbose=FALSE)
countPDict(pdict, subject,
           max.mismatch=0, min.mismatch=0, with.indels=FALSE, fixed=TRUE,
           algorithm="auto", verbose=FALSE)
whichPDict(pdict, subject,
           max.mismatch=0, min.mismatch=0, with.indels=FALSE, fixed=TRUE,
           algorithm="auto", verbose=FALSE)

vcountPDict(pdict, subject,
            max.mismatch=0, min.mismatch=0, with.indels=FALSE, fixed=TRUE,
            algorithm="auto", collapse=FALSE, weight=1L,
            verbose=FALSE, ...)
vwhichPDict(pdict, subject,
            max.mismatch=0, min.mismatch=0, with.indels=FALSE, fixed=TRUE,
            algorithm="auto", verbose=FALSE)

Arguments

pdict	A PDict object containing the preprocessed dictionary. All these functions also work with a dictionary that has not been preprocessed (in other words, the pdict argument can receive an XStringSet object). Of course, it won't be as fast as with a preprocessed dictionary, but it will generally be slightly faster than using matchPattern/countPattern or vmatchPattern/vcountPattern in a "lapply/sapply loop", because, here, looping is done at the C-level. However, by using a non-preprocessed dictionary, many of the restrictions that apply to preprocessed dictionaries don't apply anymore. For example, the dictionary doesn't need to be rectangular or to be a DNAStringSet object: it can be any type of XStringSet object and have a variable width.
subject	An XString or MaskedXString object containing the subject sequence for matchPDict, countPDict and whichPDict. An XStringSet object containing the subject sequences for vcountPDict and vwhichPDict. If pdict is a PDict object (i.e. a preprocessed dictionary), then subject must be of base class DNAString. Otherwise, subject must be of the same base class as pdict.
max.mismatch, min.mismatch	The maximum and minimum number of mismatching letters allowed (see ?isMatchingAt for the details). This man page focuses on exact matching of a constant width dictionary so max.mismatch=0 in the examples below. See ?`matchPDict-inexact` for inexact matching.
with.indels	Only supported by countPDict, whichPDict, vcountPDict and vwhichPDict at the moment, and only when the input dictionary is non-preprocessed (i.e. XStringSet). If TRUE then indels are allowed. In that case, min.mismatch must be 0 and max.mismatch is interpreted as the maximum "edit distance" allowed between any pattern and any of its matches. See ?`matchPattern` for more information.
fixed	Whether IUPAC ambiguity codes should be interpreted literally or not (see ?isMatchingAt for more information). This man page focuses on exact matching of a constant width dictionary so fixed=TRUE in the examples below. See ?`matchPDict-inexact` for inexact matching.
algorithm	Ignored if pdict is a preprocessed dictionary (i.e. a PDict object). Otherwise, can be one of the following: "auto", "naive-exact", "naive-inexact", "boyer-moore" or "shift-or". See ?matchPattern for more information. Note that "indels" is not supported for now.
verbose	TRUE or FALSE.
collapse, weight	collapse must be FALSE, 1, or 2. If collapse=FALSE (the default), then weight is ignored and vcountPDict returns the full matrix of counts (M0). If collapse=1, then M0 is collapsed "horizontally" i.e. it is turned into a vector with length equal to length(pdict). If weight=1L (the default), then this vector is defined by rowSums(M0). If collapse=2, then M0 is collapsed "vertically" i.e. it is turned into a vector with length equal to length(subject). If weight=1L (the default), then this vector is defined by colSums(M0). If collapse=1 or collapse=2, then the elements in subject (collapse=1) or in pdict (collapse=2) can be weighted thru the weight argument. In that case, the returned vector is defined by M0 %*% rep(weight, length.out=length(subject)) and rep(weight, length.out=length(pdict)) %*% M0, respectively.
...	Additional arguments for methods.

Details

In this man page, we assume that you know how to preprocess a dictionary of DNA patterns that can then be used with any of the *PDict functions described here. Please see ?PDict if you don't.

When using the *PDict functions for exact matching of a constant width dictionary, the standard way to preprocess the original dictionary is by calling the PDict constructor on it with no extra arguments. This returns the preprocessed dictionary in a PDict object that can be used with any of the *PDict functions.

Value

If M denotes the number of patterns in the pdict argument (M <- length(pdict)), then matchPDict returns an MIndex object of length M, and countPDict an integer vector of length M.

whichPDict returns an integer vector made of the indices of the patterns in the pdict argument that have at least one match.

If N denotes the number of sequences in the subject argument (N <- length(subject)), then vcountPDict returns an integer matrix with M rows and N columns, unless the collapse argument is used. In that case, depending on the type of weight, an integer or numeric vector is returned (see above for the details).

vwhichPDict returns a list of N integer vectors.

Author(s)

H. Pagès

References

Aho, Alfred V.; Margaret J. Corasick (June 1975). "Efficient string matching: An aid to bibliographic search". Communications of the ACM 18 (6): 333-340.

See Also

PDict-class, MIndex-class, matchPDict-inexact, isMatchingAt, coverage,MIndex-method, matchPattern, alphabetFrequency, DNAStringSet-class, XStringViews-class, MaskedDNAString-class

Examples

## ---------------------------------------------------------------------
## A. A SIMPLE EXAMPLE OF EXACT MATCHING
## ---------------------------------------------------------------------

## Creating the pattern dictionary:
library(drosophila2probe)
dict0 <- DNAStringSet(drosophila2probe)
dict0                                # The original dictionary.
length(dict0)                        # Hundreds of thousands of patterns.
pdict0 <- PDict(dict0)               # Store the original dictionary in
                                     # a PDict object (preprocessing).

## Using the pattern dictionary on chromosome 3R:
library(BSgenome.Dmelanogaster.UCSC.dm3)
chr3R <- Dmelanogaster$chr3R         # Load chromosome 3R
chr3R
mi0 <- matchPDict(pdict0, chr3R)     # Search...

## Looking at the matches:
start_index <- startIndex(mi0)       # Get the start index.
length(start_index)                  # Same as the original dictionary.
start_index[[8220]]                  # Starts of the 8220th pattern.
end_index <- endIndex(mi0)           # Get the end index.
end_index[[8220]]                    # Ends of the 8220th pattern.
nmatch_per_pat <- elementNROWS(mi0)  # Get the number of matches per pattern.
nmatch_per_pat[[8220]]
mi0[[8220]]                          # Get the matches for the 8220th pattern.
start(mi0[[8220]])                   # Equivalent to startIndex(mi0)[[8220]].
sum(nmatch_per_pat)                  # Total number of matches.
table(nmatch_per_pat)
i0 <- which(nmatch_per_pat == max(nmatch_per_pat))
pdict0[[i0]]                         # The pattern with most occurrences.
mi0[[i0]]                            # Its matches as an IRanges object.
Views(chr3R, mi0[[i0]])              # And as an XStringViews object.

## Get the coverage of the original subject:
cov3R <- as.integer(coverage(mi0, width=length(chr3R)))
max(cov3R)
mean(cov3R)
sum(cov3R != 0) / length(cov3R)      # Only 2.44% of chr3R is covered.
if (interactive()) {
  plotCoverage <- function(cx, start, end)
  {
    plot.new()
    plot.window(c(start, end), c(0, 20))
    axis(1)
    axis(2)
    axis(4)
    lines(start:end, cx[start:end], type="l")
  }
  plotCoverage(cov3R, 27600000, 27900000)
}

## ---------------------------------------------------------------------
## B. NAMING THE PATTERNS
## ---------------------------------------------------------------------

## The names of the original patterns, if any, are propagated to the
## PDict and MIndex objects:
names(dict0) <- mkAllStrings(letters, 4)[seq_len(length(dict0))]
dict0
dict0[["abcd"]]
pdict0n <- PDict(dict0)
names(pdict0n)[1:30]
pdict0n[["abcd"]]
mi0n <- matchPDict(pdict0n, chr3R)
names(mi0n)[1:30]
mi0n[["abcd"]]

## This is particularly useful when unlisting an MIndex object:
unlist(mi0)[1:10]
unlist(mi0n)[1:10]  # keep track of where the matches are coming from

## ---------------------------------------------------------------------
## C. PERFORMANCE
## ---------------------------------------------------------------------

## If getting the number of matches is what matters only (without
## regarding their positions), then countPDict() will be faster,
## especially when there is a high number of matches:

nmatch_per_pat0 <- countPDict(pdict0, chr3R)
stopifnot(identical(nmatch_per_pat0, nmatch_per_pat))

if (interactive()) {
  ## What's the impact of the dictionary width on performance?
  ## Below is some code that can be used to figure out (will take a long
  ## time to run). For different widths of the original dictionary, we
  ## look at:
  ##   o pptime: preprocessing time (in sec.) i.e. time needed for
  ##             building the PDict object from the truncated input
  ##             sequences;
  ##   o nnodes: nb of nodes in the resulting Aho-Corasick tree;
  ##   o nupatt: nb of unique truncated input sequences;
  ##   o matchtime: time (in sec.) needed to find all the matches;
  ##   o totalcount: total number of matches.
  getPDictStats <- function(dict, subject)
  {
    ans_width <- width(dict[1])
    ans_pptime <- system.time(pdict <- PDict(dict))[["elapsed"]]
    pptb <- pdict@threeparts@pptb
    ans_nnodes <- nnodes(pptb)
    ans_nupatt <- sum(!duplicated(pdict))
    ans_matchtime <- system.time(
                       mi0 <- matchPDict(pdict, subject)
                     )[["elapsed"]]
    ans_totalcount <- sum(elementNROWS(mi0))
    list(
      width=ans_width,
      pptime=ans_pptime,
      nnodes=ans_nnodes,
      nupatt=ans_nupatt,
      matchtime=ans_matchtime,
      totalcount=ans_totalcount
    )
  }
  stats <- lapply(8:25,
               function(width)
                   getPDictStats(DNAStringSet(dict0, end=width), chr3R))
  stats <- data.frame(do.call(rbind, stats))
  stats
}

## ---------------------------------------------------------------------
## D. USING A NON-PREPROCESSED DICTIONARY
## ---------------------------------------------------------------------

dict3 <- DNAStringSet(mkAllStrings(DNA_BASES, 3))  # all trinucleotides
dict3
pdict3 <- PDict(dict3)

## The 3 following calls are equivalent (from faster to slower):
res3a <- countPDict(pdict3, chr3R)
res3b <- countPDict(dict3, chr3R)
res3c <- sapply(dict3,
             function(pattern) countPattern(pattern, chr3R))
stopifnot(identical(res3a, res3b))
stopifnot(identical(res3a, res3c))

## One reason for using a non-preprocessed dictionary is to get rid of
## all the constraints associated with preprocessing, e.g., when
## preprocessing with PDict(), the input dictionary must be DNA and a
## Trusted Band must be defined (explicitly or implicitly).
## See '?PDict' for more information about these constraints.
## In particular, using a non-preprocessed dictionary can be
## useful for the kind of inexact matching that can't be achieved
## with a PDict object (if performance is not an issue).
## See '?`matchPDict-inexact`' for more information about inexact
## matching.

dictD <- xscat(dict3, "N", reverseComplement(dict3))

## The 2 following calls are equivalent (from faster to slower):
resDa <- matchPDict(dictD, chr3R, fixed=FALSE)
resDb <- sapply(dictD,
                function(pattern)
                  matchPattern(pattern, chr3R, fixed=FALSE))
stopifnot(all(sapply(seq_len(length(dictD)),
                     function(i)
                       identical(resDa[[i]], as(resDb[[i]], "IRanges")))))

## A non-preprocessed dictionary can be of any base class i.e. BString,
## RNAString, and AAString, in addition to DNAString:
matchPDict(AAStringSet(c("DARC", "EGH")), AAString("KMFPRNDEGHSTTWTEE"))

## ---------------------------------------------------------------------
## E. vcountPDict()
## ---------------------------------------------------------------------

## Load Fly upstream sequences (i.e. the sequences 2000 bases upstream of
## annotated transcription starts):
dm3_upstream_filepath <- system.file("extdata",
                                     "dm3_upstream2000.fa.gz",
                                     package="Biostrings")
dm3_upstream <- readDNAStringSet(dm3_upstream_filepath)
dm3_upstream

subject <- dm3_upstream[1:100]
mat1 <- vcountPDict(pdict0, subject)
dim(mat1)  # length(pdict0) x length(subject)
nhit_per_probe <- rowSums(mat1)
table(nhit_per_probe)

## Without vcountPDict(), 'mat1' could have been computed with:
mat2 <- sapply(unname(subject), function(x) countPDict(pdict0, x))
stopifnot(identical(mat1, mat2))
## but using vcountPDict() is faster (10x or more, depending of the
## average length of the sequences in 'subject').

if (interactive()) {
  ## This will fail (with message "allocMatrix: too many elements
  ## specified") because, on most platforms, vectors and matrices in R
  ## are limited to 2^31 elements:
  subject <- dm3_upstream
  vcountPDict(pdict0, subject)
  length(pdict0) * length(dm3_upstream)
  1 * length(pdict0) * length(dm3_upstream)  # > 2^31
  ## But this will work:
  nhit_per_seq <- vcountPDict(pdict0, subject, collapse=2)
  sum(nhit_per_seq >= 1)  # nb of subject sequences with at least 1 hit
  table(nhit_per_seq)  # max is 74
  which.max(nhit_per_seq)  # 1133
  sum(countPDict(pdict0, subject[[1133]]))  # 74
}

## ---------------------------------------------------------------------
## F. RELATIONSHIP BETWEEN vcountPDict(), countPDict() AND
## vcountPattern()
## ---------------------------------------------------------------------
subject <- dm3_upstream

## The 4 following calls are equivalent (from faster to slower):
mat3a <- vcountPDict(pdict3, subject)
mat3b <- vcountPDict(dict3, subject)
mat3c <- sapply(dict3,
                function(pattern) vcountPattern(pattern, subject))
mat3d <- sapply(unname(subject),
                function(x) countPDict(pdict3, x))
stopifnot(identical(mat3a, mat3b))
stopifnot(identical(mat3a, t(mat3c)))
stopifnot(identical(mat3a, mat3d))

## The 3 following calls are equivalent (from faster to slower):
nhitpp3a <- vcountPDict(pdict3, subject, collapse=1)  # rowSums(mat3a)
nhitpp3b <- vcountPDict(dict3, subject, collapse=1)
nhitpp3c <- sapply(dict3,
                   function(pattern) sum(vcountPattern(pattern, subject)))
stopifnot(identical(nhitpp3a, nhitpp3b))
stopifnot(identical(nhitpp3a, nhitpp3c))

## The 3 following calls are equivalent (from faster to slower):
nhitps3a <- vcountPDict(pdict3, subject, collapse=2)  # colSums(mat3a)
nhitps3b <- vcountPDict(dict3, subject, collapse=2)
nhitps3c <- sapply(unname(subject),
                   function(x) sum(countPDict(pdict3, x)))
stopifnot(identical(nhitps3a, nhitps3b))
stopifnot(identical(nhitps3a, nhitps3c))

## ---------------------------------------------------------------------
## G. vwhichPDict()
## ---------------------------------------------------------------------
subject <- dm3_upstream

## The 4 following calls are equivalent (from faster to slower):
vwp3a <- vwhichPDict(pdict3, subject)
vwp3b <- vwhichPDict(dict3, subject)
vwp3c <- lapply(seq_len(ncol(mat3a)), function(j) which(mat3a[ , j] != 0L))
vwp3d <- lapply(unname(subject), function(x) whichPDict(pdict3, x))
stopifnot(identical(vwp3a, vwp3b))
stopifnot(identical(vwp3a, vwp3c))
stopifnot(identical(vwp3a, vwp3d))

table(sapply(vwp3a, length))
which.min(sapply(vwp3a, length))
## Get the trinucleotides not represented in upstream sequence 21823:
dict3[-vwp3a[[21823]]]  # 2 trinucleotides

## Sanity check:
tnf <- trinucleotideFrequency(subject[[21823]])
stopifnot(all(names(tnf)[tnf == 0] == dict3[-vwp3a[[21823]]]))

## ---------------------------------------------------------------------
## H. MAPPING PROBE SET IDS BETWEEN CHIPS WITH vwhichPDict()
## ---------------------------------------------------------------------
## Here we show a simple (and very naive) algorithm for mapping probe
## set IDs between the hgu95av2 and hgu133a chips (Affymetrix).
## 2 probe set IDs are considered mapped iff they share at least one
## probe.
## WARNING: This example takes about 10 minutes to run.
if (interactive()) {

  library(hgu95av2probe)
  library(hgu133aprobe)
  probes1 <- DNAStringSet(hgu95av2probe)
  probes2 <- DNAStringSet(hgu133aprobe)
  pdict2 <- PDict(probes2)

  ## Get the mapping from probes1 to probes2 (based on exact matching):
  map1to2 <- vwhichPDict(pdict2, probes1) 

  ## The following helper function uses the probe level mapping to induce
  ## the mapping at the probe set IDs level (from hgu95av2 to hgu133a).
  ## To keep things simple, 2 probe set IDs are considered mapped iff
  ## each of them contains at least one probe mapped to one probe of
  ## the other:
  mapProbeSetIDs1to2 <- function(psID)
    unique(hgu133aprobe$Probe.Set.Name[unlist(
      map1to2[hgu95av2probe$Probe.Set.Name == psID]
    )])

  ## Use the helper function to build the complete mapping:
  psIDs1 <- unique(hgu95av2probe$Probe.Set.Name)
  mapPSIDs1to2 <- lapply(psIDs1, mapProbeSetIDs1to2)  # about 3 min.
  names(mapPSIDs1to2) <- psIDs1

  ## Do some basic stats:
  table(sapply(mapPSIDs1to2, length))

  ## [ADVANCED USERS ONLY]
  ## An alternative that is slightly faster is to put all the probes
  ## (hgu95av2 + hgu133a) in a single PDict object and then query its
  ## 'dups0' slot directly. This slot is a Dups object containing the
  ## mapping between duplicated patterns.
  ## Note that we can do this only because all the probes have the
  ## same length (25) and because we are doing exact matching:

  probes12 <- DNAStringSet(c(hgu95av2probe$sequence, hgu133aprobe$sequence))
  pdict12 <- PDict(probes12)
  dups0 <- pdict12@dups0

  mapProbeSetIDs1to2alt <- function(psID)
  {
    ii1 <- unique(togroup(dups0, which(hgu95av2probe$Probe.Set.Name == psID)))
    ii2 <- members(dups0, ii1) - length(probes1)
    ii2 <- ii2[ii2 >= 1L]
    unique(hgu133aprobe$Probe.Set.Name[ii2])
  }

  mapPSIDs1to2alt <- lapply(psIDs1, mapProbeSetIDs1to2alt)  # about 5 min.
  names(mapPSIDs1to2alt) <- psIDs1

  ## 'mapPSIDs1to2alt' and 'mapPSIDs1to2' contain the same mapping:
  stopifnot(identical(lapply(mapPSIDs1to2alt, sort),
                      lapply(mapPSIDs1to2, sort)))
}

anandhupresannan/biostrings documentation built on Nov. 2, 2019, 1:47 p.m.