An Improved Method for Clustering Fragmented Protein and Coding Sequences

We present a new method for clustering fragmented protein and coding sequences (CDS).  Clustering sequences into groups of suspected homologs is a useful step in the analysis of most novel genomic and transcriptomic data.  The reciprocal best BLAST hits approach is intuitive, and remains one of the most popular methods for orthology assessment.  This general approach often works well when the sequences are full-length and few of them have been recently duplicated.  A prominent alternative is the OrthoMCL pipeline, which utilizes a Markov Clustering (MCL) algorithm that is robust to gene duplication and can accommodate sequences from an arbitrarily large number of organisms.  All of these approaches rely on expectation values (E-values), which drop off very rapidly and in a non-linear fashion when the sequences are fragmented.  This can be a serious problem for researchers analyzing sequences obtained from de novo assemblies of short read sequence data obtained from novel or under-studied organisms.  Our method addresses this issue by replacing E-values with bit score ratios (BSR), which are calculated by dividing the bit score obtained from aligning two sequences to each other by the minimum of the two bit scores obtained when aligning each of the sequences to itself.  We find that each metric has unique strengths, and compare the results from several test cases using the CEGMA set of reference sequences.