ClusterJudge
Adrian Pasculescu
2018-04-30
Introduction
The quality of a cluster analysis can be judged based on external information. For example when the clustered entities possess attributes, the total mutual information between the clustering and each of the important and non-redundant attributes of entities might be helpful. Gene Ontology is an example of a system of attributes that can help in assessing gene clustering resulting from genomics or proteomics experiments.
A clustering that agrees well with a system of attributes should have a pronounced decrease of the total mutual information when the number of random swaps of entities between clusters is increased (Gibbons and Roth 2002).
The following simple steps must be performed to evaluate a clustering using ClusterJudge:
- Obtain the clusters as a structure containing entities and their cluster IDs;
- Obtain the entity-attribute associations and verify the entity-attribute associations for consistency;
- Often, the number of attributes can be consolidated (reduced) by using a set of uncertainty levels;
- ‘Judge’ the original clustering versus clustering of random changes that are applied to the original data.
Obtaining the clustering
In most of the cases the clustering is obtained from experimental data after applying some statistical methods such as k-means or cutting a hierarchical clustering at a predefined ‘height’.
The following example extracts the Yeast cell cycle clustering of genes related to the Yeast cell cycle (Cho 1998) (Tavazoie 1999):
## Cluster Distance Y.name SGDID GENE Chromsome Start End Introns
## 1 1 0.38 YBL072C S0000168 RPS8A II 89116 88514 0
## 2 1 0.75 YBL083C S0000179 II 70128 69703 0
## 3 1 0.47 YBR009C S0000213 HHF1 II 255638 255327 0
## 4 1 0.45 YBR010W S0000214 HHT1 II 256285 256695 0
## 5 1 0.47 YBR031W S0000235 RPL4A II 300122 301210 0
## 6 1 0.53 YBR093C S0000297 PHO5 II 430905 429502 0
## Exons Description
## 1 1-603 Ribosomal protein S8A (S14A) (rp19) (YS9)
## 2 1-426
## 3 1-312 Histone H4 (HHF1 and HHF2 code for identical proteins)
## 4 1-411 Histone H3 (HHT1 and HHT2 code for identical proteins)
## 5 1-1089 Ribosomal protein L4A (L2A) (rp2) (YL2)
## 6 1-1404 Acid phosphatase, repressible
The clustering should not contain any ambiguities ( i.e. one entity should be assigned to just one cluster) and no NA should be present. For this package, the clusters must be organized as a named vector:
## Named int [1:2885] 1 1 1 1 1 1 1 1 1 1 ...
## - attr(*, "names")= chr [1:2885] "S000000168" "S000000179" "S000000213" "S000000214" ...
Note that, in this example, we modified the names of the genes to match the names of entities in the Gene Ontology data (see next section).
Obtaining Entity Attribute associations
For genomics or proteomics data, one source of attributes is the Gene Ontology (http://www.geneontology.org/). For Yeast Gene Ontology attributes can also be downloaded from the [Saccharomyces Genome Database (SGD)] (http://www.yeastgenome.org/) In the ClusterJudge R package there is already a downloaded mapping of Yeast genes to attributes:
## 'data.frame': 70487 obs. of 2 variables:
## $ SGDID: chr "S000007287" "S000007287" "S000007287" "S000007288" ...
## $ GOID : chr "GO:0005763" "GO:0032543" "GO:0003735" "GO:0005762" ...
## SGDID GOID
## 1 S000007287 GO:0005763
## 2 S000007287 GO:0032543
## 3 S000007287 GO:0003735
## 4 S000007288 GO:0005762
## 5 S000007288 GO:0032543
## 6 S000007288 GO:0003735
## [1] "Validation OK!"
## [1] TRUE
The download_Yeast_GO_mapping() function is also available to download a fresh copy of the mapping.
For other species one can use specialized R and Bioconductor packages such as biomaRt as in the following toy example.
library(biomaRt)
rn <- useDataset("rnorvegicus_gene_ensembl", mart=useMart("ensembl"))
rgd.symbol=c("As3mt", "Borcs7", "Cyp17a1", "Wbp1l", "Sfxn2", "Arl3") ### exemplify for a limited set of genes
entity.attr <- getBM(attributes=c('rgd_symbol','go_id'), filters='rgd_symbol', values=rgd.symbol, mart=rn)
The validate_association() function quickly verifies if the input structure has two columns, if there are no NAs or NULLs and if there are no duplicated associations.
Consolidating Entity-Attribute associations
In some cases, the attributes may not bring additional useful information. For example, an attribute may be too specific as it is only assigned to one or two genes (e.g. it may simply be a new name of a gene). In other cases, two or more attributes may be very correlated and appear on almost the same set of genes.
For the first case a consolidated association will ignore those attributes that belong to very few genes (min.entities.per.attr)
## [1] 3
## [1] 70487 2
## [1] 64738 2
For the second case we use the pre-calculated mutual information (Cover 1991) between attributes and impose a maximum uncertainty level where uncertainty is defined as mutual information divided by the maximum of mutual information.
## ....
## [1] "For the uncertainty limit 0.1 removed: 74 attributes"
## .........................
## [1] "For the uncertainty limit 0.001 removed: 1939 attributes"
## List of 2
## $ 0.1 :'data.frame': 43899 obs. of 2 variables:
## ..$ SGDID: chr [1:43899] "S000007287" "S000007287" "S000007288" "S000007288" ...
## ..$ GOID : chr [1:43899] "GO:0005763" "GO:0032543" "GO:0005762" "GO:0032543" ...
## $ 0.001:'data.frame': 2517 obs. of 2 variables:
## ..$ SGDID: chr [1:2517] "S000004660" "S000000289" "S000003916" "S000005275" ...
## ..$ GOID : chr [1:2517] "GO:0005471" "GO:0005471" "GO:0018456" "GO:0018456" ...
The lower the Uncertainty limit the more attributes are ignored and the faster the judging of clusters will be performed.
Judging the clustering for selected uncertainty levels
The clusters named vector prepared above and one of the consolidated entity-attribute associations (for example the one at 0.001 level of Uncertainty) are needed at this step. The ClusterJudge will swap at random more and more entities between clusters and will calculate and plot the new total mutual information obtained at each iteration. If the total mutual information decreases by increasing the number of random swaps we can conclude that there is a good agreement between the clustering of experimental data and the ensemble of information provided by the entity-attribute associations.
## randomizing clusters ..............
## png
## 2
In most of the cases mutual information should have a decreasing trend when the number of shuffles increases. However, in rare cases, a consistent increasing trend might suggest novel discoveries; only when both experimental data and clustering are very reliable. In these rare cases a high confident clustering might complement or even contradict information accumulated in time into entity-attribute associations.
ClusterJudge can also be seen as a complement to many of the Bioconductor packages related to evaluation and comparison of Clustering (such as clustComp, clusterExperiment, clusterSignificance, GOpro, multiClust ).
References
Cho, R.C. et all. 1998. “A Genome-Wide Transcriptional Analysis of the Mitotic Cell Cycle.” Molecular Cell. Science Direct.
Cover, T.M. et all. 1991. “Elements of Information Theory.” Book. Wiley-Interscience.
Gibbons, F.D, and F.P. Roth. 2002. “Judging the Quality of Gene Expression-Based Clustering Methods Using Gene Annotation.” Genome Research. Genome Research.
Tavazoie, S. et all. 1999. “Systematic Determination of Genetic Network Architecture.” Nature Genetics. Nature Publishing Group.