Sequence Distance Computation

Description

This function computes and returns the auto-distance matrix between the vectors of a list or between the character strings of a vector treating them as sequences of symbols, as well as the cross-distance matrix between two such lists or vectors.

Usage

sdists(x, y = NULL, method = "ow", weight = c(1, 0, 2), 
       exclude = c(NA,NaN,Inf,-Inf))

Arguments

Details

This function provides a common interface to different methods for computation of distances between sequences, such as the edit a.k.a. Levenshtein distance. Conversely, in the context of sequence alignment the similarity of the maximizing alignment is computed.

Note that negative similarities are returned as distances. So be careful to use a proper weighting (scoring) scheme.

The following methods are currently implemented:

ow:

operation-weight edit distance. Weights have to be specified for insertion/deletion, match, and replacement.

aw:

alphabet-weight sequential alignment similarity. A matrix of weights (scores) for all possible symbol replacements needs to be specified with the convention that the first row/column defines the replacement with the empty (space) symbol. The colnames of this matrix are used as the levels argument for the encoding as factor. Consequently, unspecified symbols are mapped to NA.

awl:

alphabet-weight local sequential alignment similarity. The weight matrix must be as described above. However, note that zero acts as threshold for a 'restart' of the search for a local alignment and at the same time indicates that the solution is the empty substring. Thus, you normally would use non-negative scores for matches and non-positive weights otherwise.

Missing (and non-finite) values should be avoided, i.e. either be removed or recoded (and appropriately weighted). By default they are excluded when coercing to factor and therefore mapped to NA. The result is then defined to be NA as we cannot determine a match!

The time complexity is O(n*m) for two sequences of length n and m.

Note that in the case of auto-distances the weight matrix must be (exactly) symmetric. Otherwise, for asymmetric weights y must not be NULL. For instance, x may be supplied twice (see the examples).

Value

Auto distances are returned as an object of class dist and cross-distances as an object of class matrix.

Warning

The interface is experimental and may change in the future

Author(s)

Christian Buchta

References

D. Gusfield (1997) Algorithms on Strings, Trees, and Sequences. Cambridge University Press, Chapter 11.

See Also

sdists.trace for computation of edit transcripts and sequence alignments, dists for computation of common distances, agrep for searches for approximate matches.

Examples

### numeric data
sdists(list(c(2,2,3),c(2,4,3)))                 # 2
sdists(list(c(2,2,3),c(2,4,3)),weight=c(1,0,1)) # 1
### character data
w <- matrix(-1,nrow=8,ncol=8)                   # weight/score matrix for
diag(w) <- 0                                    # longest common subsequence
colnames(w) <- c("",letters[1:7])
x <- sapply(rbinom(3,64,0.5),function(n,x)
    paste(sample(x,n,rep=TRUE),collapse=""),
    colnames(w)[-1])
x
sdists(x,method="aw",weight=w)
sdists(x,x,method="aw",weight=w)                # check
diag(w) <- seq(0,7)
sdists(x,method="aw", weight=w)                 # global alignment
sdists(x,method="awl",weight=w)                 # local alignment
### asymmetric weights
w[] <- matrix(-sample(0:5,64,TRUE),ncol=8)
diag(w) <- seq(0,7)
sdists(x,x,method="aw", weight=w)
sdists(x,x,method="awl",weight=w)
### missing values
sdists(list(c(2,2,3),c(2,NA,3)),exclude=NULL)   # 2 (include anything)
sdists(list(c(2,2,3),c(2,NA,3)),exclude=NA)     # NA