Background: Clustering sequences into groups of putative homologs (families) is a critical first step in many areas of comparative biology and bioinformatics. The performance of clustering approaches in delineating biologically meaningful families depends strongly on characteristics of the data, including content bias and degree of divergence. New, highly scalable methods have recently been introduced to cluster the very large datasets being generated by next-generation sequencing technologies. However, there has been little systematic investigation of how characteristics of the data impact the performance of these approaches.
Results: Using clusters from a manually curated dataset as reference, we examined the performance of a widely used graph-based Markov clustering algorithm (MCL) and a greedy heuristic approach (UCLUST) in delineating protein families coded by three sets of bacterial genomes of different G+C content. Both MCL and UCLUST generated clusters that are comparable to the reference sets at specific parameter settings, although UCLUST tends to under-cluster compositionally biased sequences (G+C content 33% and 66%). Using simulated data, we sought to assess the individual effects of sequence divergence, rate heterogeneity, and underlying G+C content. Performance decreased with increasing sequence divergence, decreasing among-site rate variation, and increasing G+C bias. Two MCL-based methods recovered the simulated families more accurately than did UCLUST. MCL using local alignment distances is more robust across the investigated range of sequence features than are greedy heuristics using distances based on global alignment.
Conclusions: Our results demonstrate that sequence divergence, rate heterogeneity and content bias can individually and in combination affect the accuracy with which MCL and UCLUST can recover homologous protein families. For application to data that are more divergent, and exhibit higher among-site rate variation and/or content bias, MCL may often be the better choice, especially if computational resources are not limiting.