A systematic comparison of genome-scale clustering algorithms

Abstract

BackgroundA wealth of clustering algorithms has been applied to gene co-expression experiments. These algorithms cover a broad range of approaches, from conventional techniques such as k-means and hierarchical clustering, to graphical approaches such as k-clique communities, weighted gene co-expression networks (WGCNA) and paraclique. Comparison of these methods to evaluate their relative effectiveness provides guidance to algorithm selection, development and implementation. Most prior work on comparative clustering evaluation has focused on parametric methods. Graph theoretical methods are recent additions to the tool set for the global analysis and decomposition of microarray co-expression matrices that have not generally been included in earlier methodological comparisons. In the present study, a variety of parametric and graph theoretical clustering algorithms are compared using well-characterized transcriptomic data at a genome scale from Saccharomyces cerevisiae.MethodsFor each clustering method under study, a variety of parameters were tested. Jaccard similarity was used to measure each cluster's agreement with every GO and KEGG annotation set, and the highest Jaccard score was assigned to the cluster. Clusters were grouped into small, medium, and large bins, and the Jaccard score of the top five scoring clusters in each bin were averaged and reported as the best average top 5 (BAT5) score for the particular method.ResultsClusters produced by each method were evaluated based upon the positive match to known pathways. This produces a readily interpretable ranking of the relative effectiveness of clustering on the genes. Methods were also tested to determine whether they were able to identify clusters consistent with those identified by other clustering methods.ConclusionsValidation of clusters against known gene classifications demonstrate that for this data, graph-based techniques outperform conventional clustering approaches, suggesting that further development and application of combinatorial strategies is warranted.

Genetic Analysis of Mps3 SUN Domain Mutants inSaccharomyces cerevisiaeReveals an Interaction with the SUN-Like Protein Slp1

“…From this analysis, it is clear that mps3-A540D , mps3-Y502H , and mps3-F592S result in similar genetic signatures, whereas mps3Δ75-150 has virtually no similarity with any of the SUN mutants. A comparison of datasets using Jaccard distance, which takes into account the intersection and the union of the datasets (Jay et al 2012), showed an index of 0.61 for mps3-A540D and mps3-Y502H mutants, 0.59 for mps3-F592S and mps3-A540D , and 0.63 for mps3-F592S and mps3-Y502H . In contrast, an index of 0.93-0.98 was observed between mps3Δ75-150 an each of the mps3 SUN mutant alleles (Figure 3C).…”

“…(B) Synthetic lethal, synthetic sick, or negative genetic interactions were extracted for yeast gene pairs from BioGRID 3.1.88 and integrated with our data. To characterize the similarity of interactions found with mps3-SUN domain mutants to interactions identified with other deletions, we used the Jaccard distance, which takes into account the intersection and union of datasets (Jay et al 2012). A total of 32 genes have a Jaccard distance <0.95 to all three mps3-SUN mutants, indicating a similar genetic signature.…”

“…In this matrix, 1 represents a negative interaction or synthetic growth/lethal defect between two gene pairs whereas gene pairs that do not interact are represented with 0. Jaccard distance was used to measure the similarity of our mps3 mutant profiles and the similarity between mps3 mutants and other mutants in the BioGRID dataset (Jay et al 2012). A gene with a distance of less than 0.95 to the query gene (one of our mps3 mutants) was considered to have a similar genetic signature since the majority (95%) of budding yeast ORFs in BioGRID fall above this cut off (data not shown).…”

Lower bounds on paraclique density

“…To the best of our knowledge, other than the elementary result from [4], these are the first formal density limits for what have come to be popularly known as network community methods. This gives paraclique another potential practical endorsement, in addition to those due to biological enrichment as discussed in [5]. …”

“…The availability of high-throughput data has prompted interest in noise-abatement relaxations, most notably k -clique communities [3] (more recently also called clique percolation) and paraclique [4]. These algorithms have been used for biological data clustering, and been found superior to traditional methods [5]. Although similar in objective, k -clique communities is hampered in practice by its bottom up approach relying on an exhaustive enumeration of maximal cliques.…”