Evaluation of BLAST-based edge-weighting metrics used for homology inference with the Markov Clustering algorithm

BMC Bioinformatics

Research Article

Theodore R. Gibbons¹  Endymion D. Cooper¹  Stephen M. Mount²  Charles F. Delwiche³ 

[1] Department of Cell Biology and Molecular Genetics, University of Maryland, 20742, College Park, Baltimore, Maryland, USA;Department of Cell Biology and Molecular Genetics, University of Maryland, 20742, College Park, Baltimore, Maryland, USA;Center for Bioinformatics and Computational Biology, University of Maryland, 20742, College Park, Baltimore, Maryland, USA;Department of Cell Biology and Molecular Genetics, University of Maryland, 20742, College Park, Baltimore, Maryland, USA;Maryland Agricultural Experiment Station, University of Maryland, 20742, College Park, Baltimore, Maryland, USA;

关键词: MCL;    Protein clustering;    Sequence clustering;    Homology prediction;    Graph;    Genomics;    Bioinformatics;    Transcriptomics;    Short-read sequencing;    High-throughput sequencing;   

DOI  :  10.1186/s12859-015-0625-x

 received in 2014-10-16, accepted in 2015-05-20,  发布年份 2015

来源: Springer

【 摘 要 】

BackgroundClustering protein sequences according to inferred homology is a fundamental step in the analysis of many large data sets. Since the publication of the Markov Clustering (MCL) algorithm in 2002, it has been the centerpiece of several popular applications. Each of these approaches generates an undirected graph that represents sequences as nodes connected to each other by edges weighted with a BLAST-based metric. MCL is then used to infer clusters of homologous proteins by analyzing these graphs. The various approaches differ only by how they weight the edges, yet there has been very little direct examination of the relative performance of alternative edge-weighting metrics. This study compares the performance of four BLAST-based edge-weighting metrics: the bit score, bit score ratio (BSR), bit score over anchored length (BAL), and negative common log of the expectation value (NLE). Performance is tested using the Extended CEGMA KOGs (ECK) database, which we introduce here.ResultsAll metrics performed similarly when analyzing full-length sequences, but dramatic differences emerged as progressively larger fractions of the test sequences were split into fragments. The BSR and BAL successfully rescued subsets of clusters by strengthening certain types of alignments between fragmented sequences, but also shifted the largest correct scores down near the range of scores generated from spurious alignments. This penalty outweighed the benefits in most test cases, and was greatly exacerbated by increasing the MCL inflation parameter, making these metrics less robust than the bit score or the more popular NLE. Notably, the bit score performed as well or better than the other three metrics in all scenarios.ConclusionsThe results provide a strong case for use of the bit score, which appears to offer equivalent or superior performance to the more popular NLE. The insight that MCL-based clustering methods can be improved using a more tractable edge-weighting metric will greatly simplify future implementations. We demonstrate this with our own minimalist Python implementation: Porthos, which uses only standard libraries and can process a graph with 25 m + edges connecting the 60 k + KOG sequences in half a minute using less than half a gigabyte of memory.

【 授权许可】

CC BY   
© Gibbons et al. 2015