Hierarchical clustering based on dissimilarity or beta-diversity

This function generates a hierarchical tree from a dissimilarity (beta-diversity) data.frame, calculates the cophenetic correlation coefficient, and can get clusters from the tree if requested by the user. The function implements randomization of the dissimilarity matrix to generate the tree, with two different methods to generate the final tree. Typically, the dissimilarity data.frame is a bioregion.pairwise.metric object obtained by running similarity or similarity and then similarity_to_dissimilarity.

Usage

hclu_hierarclust(
  dissimilarity,
  index = names(dissimilarity)[3],
  method = "average",
  randomize = TRUE,
  n_runs = 100,
  keep_trials = FALSE,
  optimal_tree_method = "iterative_consensus_tree",
  n_clust = NULL,
  cut_height = NULL,
  find_h = TRUE,
  h_max = 1,
  h_min = 0,
  consensus_p = 0.5,
  verbose = TRUE
)

Arguments

dissimilarity: the output object from dissimilarity() or similarity_to_dissimilarity(), or a dist object. If a data.frame is used, the first two columns represent pairs of sites (or any pair of nodes), and the next column(s) are the dissimilarity indices.
index: name or number of the dissimilarity column to use. By default, the third column name of dissimilarity is used.
method: name of the hierarchical classification method, as in hclust. Should be one of "ward.D", "ward.D2", "single", "complete", "average" (= UPGMA), "mcquitty" (= WPGMA), "median" (= WPGMC) or "centroid" (= UPGMC).
randomize: a boolean indicating if the dissimilarity matrix should be randomized, to account for the order of sites in the dissimilarity matrix.
n_runs: number of trials to randomize the dissimilarity matrix.
keep_trials: a boolean indicating if all random trial results. should be stored in the output object (set to FALSE to save space if your dissimilarity object is large). Note that it cannot be set to TRUE if optimal_tree_method = "iterative_consensus_tree"
optimal_tree_method: a character indicating how the final tree should be obtained from all trials. Possible values are iterative_consensus_tree (default), best and consensus. We recommend iterative_consensus_tree. See details
n_clust: an integer or an integer vector indicating the number of clusters to be obtained from the hierarchical tree, or the output from partition_metrics. Should not be used at the same time as cut_height.
cut_height: a numeric vector indicating the height(s) at which the tree should be cut. Should not be used at the same time as n_clust.
find_h: a boolean indicating if the height of cut should be found for the requested n_clust.
h_max: a numeric indicating the maximum possible tree height for the chosen index.
h_min: a numeric indicating the minimum possible height in the tree for the chosen index.
consensus_p: a numeric, (only if optimal_tree_method = "consensus"), indicating the threshold proportion of trees that must support a region/cluster for it to be included in the final consensus tree.
verbose: a boolean. (only if optimal_tree_method = "iterative_consensus_tree"), Set to FALSE if you want to disable the progress message

Value

A list of class bioregion.clusters with five slots:

name: character containing the name of the algorithm
args: list of input arguments as provided by the user
inputs: list of characteristics of the clustering process
algorithm: list of all objects associated with the clustering procedure, such as original cluster objects
clusters: data.frame containing the clustering results

In the algorithm slot, users can find the following elements:

trials: a list containing all randomization trials. Each trial contains the dissimilarity matrix, with site order randomized, the associated tree and the cophenetic correlation coefficient (Spearman) for that tree
final.tree: a hclust object containing the final hierarchical tree to be used
final.tree.coph.cor: the cophenetic correlation coefficient between the initial dissimilarity matrix and final.tree

Details

The function is based on hclust. The default method for the hierarchical tree is average, i.e. UPGMA as it has been recommended as the best method to generate a tree from beta diversity dissimilarity Kreft2010bioregion.

Clusters can be obtained by two methods:

Specifying a desired number of clusters in n_clust
Specifying one or several heights of cut in cut_height

To find an optimal number of clusters, see partition_metrics()

It is important to pay attention to the fact that the order of rows in the input distance matrix influences the tree topology as explained in Dapporto2013bioregion. To address this, the function generates multiple trees by randomizing the distance matrix. Two methods are available to obtain the final tree:

optimal_tree_method = "iterative_consensus_tree": The Iterative Hierarchical Consensus Tree (IHCT) method reconstructs a consensus tree by iteratively splitting the dataset into two subclusters based on the pairwise dissimilarity of sites across n_runs trees based on n_runs randomizations of the distance matrix. At each iteration, it identifies the majority membership of sites into two stable groups across all trees, calculates the height based on the selected linkage method (method), and enforces monotonic constraints on node heights to produce a coherent tree structure. This approach provides a robust, hierarchical representation of site relationships, balancing cluster stability and hierarchical constraints.
optimal_tree_method = "best": This method selects one tree among with the highest cophenetic correlation coefficient, representing the best fit between the hierarchical structure and the original distance matrix.
optimal_tree_method = "consensus": This method constructs a consensus tree using phylogenetic methods with the function consensus. When using this option, you must set the consensus_p parameter, which indicates the proportion of trees that must contain a region/cluster for it to be included in the final consensus tree. Consensus trees lack an inherent height because they represent a majority structure rather than an actual hierarchical clustering. To assign heights, we use a non-negative least squares method (nnls.tree) based on the initial distance matrix, ensuring that the consensus tree preserves approximate distances among clusters.

We recommend using the "iterative_consensus_tree" as all the branches of this tree will always reflect the majority decision among many randomized versions of the distance matrix. This method is inspired by Dapporto2015bioregion, which also used the majority decision among many randomized versions of the distance matrix, but it expands it to reconstruct the entire topology of the tree iteratively.

We do not recommend using the basic consensus method because in many contexts it provides inconsistent results, with a meaningless tree topology and a very low cophenetic correlation coefficient.

For a fast exploration of the tree, we recommend using the best method which will only select the tree with the highest cophenetic correlation coefficient among all randomized versions of the distance matrix.

References

Kreft2010bioregion Dapporto2013bioregion Dapporto2015bioregion

Author

Boris Leroy (leroy.boris@gmail.com), Pierre Denelle (pierre.denelle@gmail.com) and Maxime Lenormand (maxime.lenormand@inrae.fr)

Examples

comat <- matrix(sample(0:1000, size = 500, replace = TRUE, prob = 1/1:1001),
20, 25)
rownames(comat) <- paste0("Site",1:20)
colnames(comat) <- paste0("Species",1:25)

dissim <- dissimilarity(comat, metric = "Simpson")

# User-defined number of clusters
tree1 <- hclu_hierarclust(dissim, 
                          n_clust = 5)
#> Building the iterative hierarchical consensus tree... Note that this process can take time especially if you have a lot of sites.
#> 
#> Final tree has a 0.4777 cophenetic correlation coefficient with the initial dissimilarity
#>       matrix
#> Determining the cut height to reach 5 groups...
#> --> 0.109375
tree1
#> Clustering results for algorithm : hclu_hierarclust 
#> 	(hierarchical clustering based on a dissimilarity matrix)
#>  - Number of sites:  20 
#>  - Name of dissimilarity metric:  Simpson 
#>  - Tree construction method:  average 
#>  - Randomization of the dissimilarity matrix:  yes, number of trials 100 
#>  - Method to compute the final tree:  Iterative consensus hierarchical tree 
#>  - Cophenetic correlation coefficient:  0.478 
#>  - Number of clusters requested by the user:  5 
#> Clustering results:
#>  - Number of partitions:  1 
#>  - Number of clusters:  5 
#>  - Height of cut of the hierarchical tree: 0.109 
plot(tree1)

str(tree1)
#> List of 6
#>  $ name        : chr "hclu_hierarclust"
#>  $ args        :List of 14
#>   ..$ index              : chr "Simpson"
#>   ..$ method             : chr "average"
#>   ..$ randomize          : logi TRUE
#>   ..$ n_runs             : num 100
#>   ..$ optimal_tree_method: chr "iterative_consensus_tree"
#>   ..$ keep_trials        : logi FALSE
#>   ..$ n_clust            : num 5
#>   ..$ cut_height         : NULL
#>   ..$ find_h             : logi TRUE
#>   ..$ h_max              : num 1
#>   ..$ h_min              : num 0
#>   ..$ consensus_p        : num 0.5
#>   ..$ verbose            : logi TRUE
#>   ..$ dynamic_tree_cut   : logi FALSE
#>  $ inputs      :List of 7
#>   ..$ bipartite      : logi FALSE
#>   ..$ weight         : logi TRUE
#>   ..$ pairwise       : logi TRUE
#>   ..$ pairwise_metric: chr "Simpson"
#>   ..$ dissimilarity  : logi TRUE
#>   ..$ nb_sites       : int 20
#>   ..$ hierarchical   : logi FALSE
#>  $ algorithm   :List of 6
#>   ..$ final.tree         :List of 5
#>   .. ..- attr(*, "class")= chr "hclust"
#>   ..$ final.tree.coph.cor: num 0.478
#>   ..$ final.tree.msd     : num 0.00245
#>   ..$ output_n_clust     : int 5
#>   ..$ output_cut_height  : Named num 0.109
#>   .. ..- attr(*, "names")= chr "k_5"
#>   ..$ trials             : chr "Trials not stored in output"
#>  $ clusters    :'data.frame':	20 obs. of  2 variables:
#>   ..$ ID : chr [1:20] "Site1" "Site10" "Site11" "Site12" ...
#>   ..$ K_5: chr [1:20] "1" "2" "3" "4" ...
#>  $ cluster_info:'data.frame':	1 obs. of  4 variables:
#>   ..$ partition_name   : chr "K_5"
#>   ..$ n_clust          : int 5
#>   ..$ requested_n_clust: num 5
#>   ..$ output_cut_height: num 0.109
#>  - attr(*, "class")= chr [1:2] "bioregion.clusters" "list"
tree1$clusters
#>            ID K_5
#> Site1   Site1   1
#> Site10 Site10   2
#> Site11 Site11   3
#> Site12 Site12   4
#> Site13 Site13   4
#> Site14 Site14   5
#> Site15 Site15   2
#> Site16 Site16   4
#> Site17 Site17   1
#> Site18 Site18   3
#> Site19 Site19   4
#> Site2   Site2   4
#> Site20 Site20   4
#> Site3   Site3   1
#> Site4   Site4   5
#> Site5   Site5   4
#> Site6   Site6   2
#> Site7   Site7   2
#> Site8   Site8   1
#> Site9   Site9   4

# User-defined height cut
# Only one height
tree2 <- hclu_hierarclust(dissim, 
                          cut_height = .05)
#> Building the iterative hierarchical consensus tree... Note that this process can take time especially if you have a lot of sites.
#> 
#> Final tree has a 0.5232 cophenetic correlation coefficient with the initial dissimilarity
#>       matrix
tree2
#> Clustering results for algorithm : hclu_hierarclust 
#> 	(hierarchical clustering based on a dissimilarity matrix)
#>  - Number of sites:  20 
#>  - Name of dissimilarity metric:  Simpson 
#>  - Tree construction method:  average 
#>  - Randomization of the dissimilarity matrix:  yes, number of trials 100 
#>  - Method to compute the final tree:  Iterative consensus hierarchical tree 
#>  - Cophenetic correlation coefficient:  0.523 
#>  - Heights of cut requested by the user:  0.05 
#> Clustering results:
#>  - Number of partitions:  1 
#>  - Number of clusters:  14 
#>  - Height of cut of the hierarchical tree: 0.05 
tree2$clusters
#>        ID K_14
#> 1   Site1    1
#> 2  Site10    2
#> 3  Site11    3
#> 4  Site12    4
#> 5  Site13    4
#> 6  Site14    5
#> 7  Site15    6
#> 8  Site16    7
#> 9  Site17    8
#> 10 Site18    9
#> 11 Site19   10
#> 12  Site2   10
#> 13 Site20   11
#> 14  Site3   12
#> 15  Site4   13
#> 16  Site5    4
#> 17  Site6   14
#> 18  Site7    2
#> 19  Site8    8
#> 20  Site9    4

# Multiple heights
tree3 <- hclu_hierarclust(dissim, 
                          cut_height = c(.05, .15, .25))
#> Building the iterative hierarchical consensus tree... Note that this process can take time especially if you have a lot of sites.
#> 
#> Final tree has a 0.5232 cophenetic correlation coefficient with the initial dissimilarity
#>       matrix

tree3$clusters # Mind the order of height cuts: from deep to shallow cuts
#>            ID K_1_1 K_1_2 K_14
#> Site1   Site1     1     1    1
#> Site10 Site10     1     1    2
#> Site11 Site11     1     1    3
#> Site12 Site12     1     1    4
#> Site13 Site13     1     1    4
#> Site14 Site14     1     1    5
#> Site15 Site15     1     1    6
#> Site16 Site16     1     1    7
#> Site17 Site17     1     1    8
#> Site18 Site18     1     1    9
#> Site19 Site19     1     1   10
#> Site2   Site2     1     1   10
#> Site20 Site20     1     1   11
#> Site3   Site3     1     1   12
#> Site4   Site4     1     1   13
#> Site5   Site5     1     1    4
#> Site6   Site6     1     1   14
#> Site7   Site7     1     1    2
#> Site8   Site8     1     1    8
#> Site9   Site9     1     1    4
# Info on each partition can be found in table cluster_info
tree3$cluster_info
#>        partition_name n_clust requested_cut_height
#> h_0.25          K_1_1       1                 0.25
#> h_0.15          K_1_2       1                 0.15
#> h_0.05           K_14      14                 0.05
plot(tree3)


# Recut the tree afterwards
tree3.1 <- cut_tree(tree3, n = 5)
#> Determining the cut height to reach 5 groups...
#> --> 0.109375

# Make multiple cuts
tree4 <- hclu_hierarclust(dissim, 
                          n_clust = 1:19)
#> Building the iterative hierarchical consensus tree... Note that this process can take time especially if you have a lot of sites.
#> 
#> Final tree has a 0.5232 cophenetic correlation coefficient with the initial dissimilarity
#>       matrix
#> Warning: The requested number of cluster could not be found
#>                          for k = 7. Closest number found: 6
#> Warning: The requested number of cluster could not be found
#>                          for k = 12. Closest number found: 11
#> Warning: The requested number of cluster could not be found
#>                          for k = 15. Closest number found: 14
#> Warning: The requested number of cluster could not be found
#>                          for k = 19. Closest number found: 18

# Change the method to get the final tree 
tree5 <- hclu_hierarclust(dissim,
                          optimal_tree_method = "best",
                          n_clust = 10)
#> Randomizing the dissimilarity matrix with 100 trials
#>  -- range of cophenetic correlation coefficients among trials: 0.3954 - 0.5233
#> 
#> Final tree has a 0.5233 cophenetic correlation coefficient with the initial dissimilarity
#>       matrix
#> Determining the cut height to reach 10 groups...
#> --> 0.0703125