A general paradigm for fast, adaptive clustering of biological sequences

Abstract

There are numerous methods that compute clusterings of biological sequences based on pairwise distances. This necessitates the computation of O(n2) sequence comparisons. Users usually want to apply the most sensitive distance measure which normally is the most expensive in terms of runtime. This poses a problem if the number of sequences is large or the computation of the measure is slow. In this paper we present a general heuristic to speed up distance based clustering methods considerably while compromising little on the accuracy of the results. The speedup comes from using fast comparison methods to perform an initial ‘top-down’ split into relatively homogeneous clusters, while the slower measures are used for smaller groups. Then profiles are computed for the final groups and the resulting profiles are used in a bottom-up phase to compute the final clustering. The algorithm is general in the sense that any sequence comparison method can be employed (e.g. for DNA, RNA or amino acids). We test our algorithm using a prototypical imple- mentation for agglomerative RNA clustering and show its effectiveness.