DECIPHER (version 2.0.2)

IdClusters: Cluster Sequences By Distance or Sequence

Description

Groups the sequences represented by a distance matrix into clusters of similarity.

Usage

IdClusters(myDistMatrix = NULL, method = "UPGMA", cutoff = -Inf, showPlot = FALSE, asDendrogram = FALSE, myXStringSet = NULL, model = MODELS, processors = 1, verbose = TRUE)

Arguments

myDistMatrix
A symmetric $N$ x $N$ distance matrix with the values of dissimilarity between $N$ sequences, or NULL if method is "inexact".
method
An agglomeration method to be used. This should be (an abbreviation of) one of "complete", "single", "UPGMA", "WPGMA", "NJ", "ML", or "inexact". (See details section below.)
cutoff
A vector with the maximum edge length separating the sequences in the same cluster. Multiple cutoffs may be provided in ascending or descending order. If asDendrogram=TRUE or showPlot=TRUE then only one cutoff may be specified. (See details section below.)
showPlot
Logical specifying whether or not to plot the resulting dendrogram. Not applicable if method='inexact'.
asDendrogram
Logical. If TRUE then the object returned is of class dendrogram. Not applicable if method='inexact'.
myXStringSet
If method is "ML", the DNAStringSet or RNAStringSet used in the creation of myDistMatrix. If method is "inexact", the DNAStringSet, RNAStringSet, or AAStringSet to cluster. Not applicable for other methods.
model
One or more of the available MODELS of DNA evolution. Only applicable if method is "ML".
processors
The number of processors to use, or NULL to automatically detect and use all available processors.
verbose
Logical indicating whether to display progress.

Value

If asDendrogram=FALSE (the default), then a data.frame is returned with a column for each cutoff specified. This data.frame has dimensions $N*M$, where each one of $N$ sequences is assigned to a cluster at the $M$-level of cutoff. The row.names of the data.frame correspond to the dimnames of myDistMatrix. If asDendrogram=TRUE, returns an object of class dendrogram that can be used for further manipulation and plotting. Leaves of the dendrogram are randomly colored by cluster number.

Details

IdClusters groups the input sequences into clusters using a set dissimilarities representing the distance between $N$ sequences. Initially a phylogenetic tree is formed using the specified method. Then each leaf (sequence) of the tree is assigned to a cluster based on its edge lengths to the other sequences. The available clustering methods are described as follows:

Ultrametric methods: The method complete assigns clusters using complete-linkage so that sequences in the same cluster are no more than cutoff percent apart. The method single assigns clusters using single-linkage so that sequences in the same cluster are within cutoff of at least one other sequence in the same cluster. UPGMA (the default) or WPGMA assign clusters using average-linkage which is a compromise between the sensitivity of complete-linkage clustering to outliers and the tendency of single-linkage clustering to connect distant relatives that do not appear to be closely related. UPGMA produces an unweighted tree, where each leaf contributes equally to the average edge lengths, whereas WPGMA produces a weighted result.

Additive methods: NJ uses the Neighbor-Joining method proposed by Saitou and Nei that does not assume lineages evolve at the same rate (the molecular clock hypothesis). The NJ method is typically the most phylogenetically accurate of the above distance-based methods. ML creates a neighbor-joining tree and then iteratively maximizes the likelihood of the tree given the aligned sequences (myXStringSet). This is accomplished through a combination of optimizing edge lengths with Brent's method and improving tree topology with nearest-neighbor interchanges (NNIs). When method="ML", one or more MODELS of DNA evolution must be specified. Model parameters are iteratively optimized to maximize likelihood, except base frequencies which are empirically determined. If multiple models are given, the best model is automatically chosen based on BIC calculated from the likelihood and the sample size (defined as the number of variable sites in the DNA sequence).

Sequence-only method: inexact uses a heuristic algorithm to directly assign sequences to clusters without a distance matrix. First the sequences are ordered by length and the longest sequence becomes the first cluster seed. If the second sequence is less than cutoff percent distance then it is added to the cluster, otherwise it becomes a new cluster representative. The remaining sequences are matched to cluster representatives based on their k-mer distribution and then aligned to find the closest sequence. This approach is repeated until all sequences belong to a cluster. In the vast majority of cases, this process results in clusters with members separated by less than cutoff distance, where distance is defined as the percent dissimilarity between the overlapping region of a ``glocal'' alignment.

Multiple cutoffs may be provided if they are in increasing or decreasing order. If cutoffs are provided in descending order then clustering at each new value of cutoff is continued within the prior cutoff's clusters. In this way clusters at lower values of cutoff are completely contained within their umbrella clusters at higher values of cutoff. This is useful for defining taxonomy, where lower level groups (e.g., genera) are expected not to straddle multiple higher level groups (e.g., families). If multiple cutoffs are provided in ascending order then clustering at each level of cutoff is independent of the prior level. This may result in fewer high-level clusters for NJ and ML methods, but will have no impact on ultrametric methods. Providing cutoffs in descending order makes inexact clustering faster, but has negligible impact on the other methods.

References

Felsenstein, J. (1981) Evolutionary trees from DNA sequences: a maximum likelihood approach. Journal of Molecular Evolution, 17(6), 368-376.

Ghodsi, M., Liu, B., & Pop, M. (2011) DNACLUST. BMC Bioinformatics, 12(1), 271. doi:10.1186/1471-2105-12-271.

Saitou, N. and Nei, M. (1987) The neighbor-joining method: a new method for reconstructing phylogenetic trees. Molecular Biology and Evolution, 4(4), 406-425.

See Also

DistanceMatrix, Add2DB, MODELS

Examples

Run this code
# using the matrix from the original paper by Saitou and Nei
m <- matrix(0,8,8)
m[2:8,1] <- c(7, 8, 11, 13, 16, 13, 17)
m[3:8,2] <- c(5, 8, 10, 13, 10, 14)
m[4:8,3] <- c(5, 7, 10, 7, 11)
m[5:8,4] <- c(8, 11, 8, 12)
m[6:8,5] <- c(5, 6, 10)
m[7:8,6] <- c(9, 13)
m[8,7] <- c(8)

# returns an object of class "dendrogram"
myClusters <- IdClusters(m, cutoff=10, method="NJ", showPlot=TRUE, asDendrogram=TRUE)

# example of specifying multiple cutoffs
IdClusters(m, cutoff=c(2,6,10,20)) # returns a data frame

# example of 'inexact' clustering
fas <- system.file("extdata", "50S_ribosomal_protein_L2.fas", package="DECIPHER")
dna <- readDNAStringSet(fas)
IdClusters(myXStringSet=dna, method="inexact", cutoff=0.05)

Run the code above in your browser using DataLab