Cluster Utility: A new metric to guide sequence clustering algorithms

Motivation: Automatic sequence clustering has become very important in analyzing the ever increasing number of biological sequences. Unresolved issues in sequence clustering include limiting generation of incorrect and fragmented clusters, and handling different levels of cluster specificity. Graph theoretical approaches are often used in clustering problem. Sequences are represented as nodes and the similarity between two sequences is represented as edges. Depending on edge cutoff selection, different graph structures result. Since cutoff setting is a fundamental part of a clustering process, clustering results should be evaluated with cutoff setting. However, few studies exist on evaluating clustering results in relation to cutoff setting. Results: We propose cluster utility (CU), a metric that is based on consideration of similarity within a cluster and difference between clusters. It accurately reflects the fitness of a clustering result to the underlying class structure of data. CU can be used in two ways; to guide sequence clustering algorithms and to evaluate clustering results. It outperformed other indices that have been suggested in general clustering applications. Availability: Programs to measure the cluster utility for a clustering result are available upon request. Contact: [email protected]