Large scale clustering of protein sequences with FORCE -A layout based heuristic for weighted cluster editing

Abstract

Background: Detecting groups of functionally related proteins from their amino acid sequence alone has been a long-standing challenge in computational genome research. Several clustering approaches, following different strategies, have been published to attack this problem. Today, new sequencing technologies provide huge amounts of sequence data that has to be efficiently clustered with constant or increased accuracy, at increased speed.

Results: We advocate that the model of weighted cluster editing, also known as transitive graph projection is well-suited to protein clustering. We present the FORCE heuristic that is based on transitive graph projection and clusters arbitrary sets of objects, given pairwise similarity measures. In particular, we apply FORCE to the problem of protein clustering and show that it outperforms the most popular existing clustering tools (Spectral clustering, TribeMCL, GeneRAGE, Hierarchical clustering, and Affinity Propagation). Furthermore, we show that FORCE is able to handle huge datasets by calculating clusters for all 192 187 prokaryotic protein sequences (66 organisms) obtained from the COG database. Finally, FORCE is integrated into the corynebacterial reference database CoryneRegNet.

Conclusion: FORCE is an applicable alternative to existing clustering algorithms. Its theoretical foundation, weighted cluster editing, can outperform other clustering paradigms on protein homology clustering. FORCE is open source and implemented in Java. The software, including the source code, the clustering results for COG and CoryneRegNet, and all evaluation datasets are available at http://gi.cebitec.uni-bielefeld.de/comet/force/.

Figure 1

Graphical summary of the obtained clustering results of FORCE for the two similarity functions (A) BeH and (B) SoH, and dataset ASTRAL95_1_161. We used MATLAB scripts provided by Paccanaro to create images similar to those of Figure 3 in [7]. Each row corresponds to a cluster. Green bars represent a protein assignment to a cluster; each protein is present in only one of the clusters. Boundaries between superfamilies are shown by vertical red lines, and boundaries between families within each superfamily are shown by dotted blue lines.

Figure 2

Comparison of the running times of FORCE against the exact fixed-parameter algorithm described in [2]. Plotted is the running time (y-axis in seconds) for different graph sizes (x-axis). Solely for visualization purposes, we describe the size of a graph on the x-axis as |V|·|E|. All graphs have been constructed from prokaryotic COG protein sequence comparisons using BeH as scoring function. Note that both axes are scaled logarithmically. The red points correspond to FORCE running times, and the blue points to the FP algorithm, respectively.

Figure 3

Relative cost deviations (y-axis in %) of the FORCE solutions from the optimal solutions found by the exact fixed-parameter algorithm described in [2]. The x-axis is as in Figure 2 (logarithmically scaled).