Online Transitivity Clustering of Biological Data with Missing Values

Abstract

Motivation: Equipped with sophisticated biochemical measurement techniques we generate a massive amount of biomedical data that needs to be analyzed computationally. One long-standing challenge in automatic knowledge extraction is clustering. We seek to partition a set of objects into groups such that the objects within the clusters share common traits. Usually, we have given a similarity matrix computed from a pairwise similarity function. While many approaches for biomedical data clustering exist, most methods neglect two important problems: (1) Computing the similarity matrix might not be trivial but resource-intense. (2) A clustering algorithm itself is not sufficient for the biologist, who needs an integrated online system capable of performing preparative and follow-up tasks as well.

Results: Here, we present a significantly extended version of Transitivity Clustering. Our first main contribution is its’ capability of dealing with missing values in the similarity matrix such that we save time and memory. Hence, we reduce one main bottleneck of computing all pairwise similarity values. We integrated this functionality into the Weighted Graph Cluster Editing model underlying Transitivity Clustering. By means of identifying protein (super)families from incomplete all-vs-all BLAST results we demonstrate the robustness of our approach. While most tools concentrate on the partitioning process itself, we present a new, intuitive web interface that aids with all important steps of a cluster analysis: (1) computing and post-processing of a similarity matrix, (2) estimation of a meaningful density parameter, (3) clustering, (4) comparison with given gold standards, and (5) fine-tuning of the clustering by varying the parameters. 

Availability: Transitivity Clustering, the new Cost Matrix Creator, all used data sets as well as an online documentation are online available at http://transclust.mmci.uni-saarland.de/.