Searching sequence databases for functional homologs using profile HMMs: how to set bit score thresholds?

Abstract

UniProt and BFD databases together have 2.5 billion protein sequences. A large majority of these proteins have been electronically annotated. Automated annotation pipelines, vis-à-vis manual curation, have the advantage of scale and speed but are fraught with relatively higher error rates. This is because sequence homology does not necessarily translate to functional homology, molecular function specification is hierarchic and not all functional families have the same amount of experimental data that one can exploit for annotation. Consequently, customization of annotation workflow is inevitable to minimize annotation errors. In this study, we illustrate possible ways of customizing the search of sequence databases for functional homologs using profile HMMs. Choosing an optimal bit score threshold is a critical step in the application of HMMs. We illustrate ways in which an optimal bit score can be arrived at using four Case Studies. These are the single domain nucleotide sugar 6-dehydrogenase and lysozyme-C families, and SH3 and GT-A domains which are typically found as a part of multi-domain proteins. We also discuss the limitations of using profile HMMs for functional annotation and suggests some possible ways to partially overcome such limitations.

Clues to reaction specificity in PLP‐dependent fold type I aminotransferases of monosaccharide biosynthesis

“…Bit score thresholds for NSAT HMM and NSD HMM , profile HMMs of aminotransferases and dehydratases, respectively, were set based on Receiver Operator Characteristic (ROC) curves generated using hits from TrEMBL under the assumption that annotations provided in TrEMBL are correct. The procedure used to generate ROC curves is described in detail elsewhere 6,44 . Briefly, ROC curves were generated by calculating the number of true positives, false positives and false negatives for various values of bit score threshold (Figure S3).…”

“…The procedure used to generate ROC curves is described in detail elsewhere. 6,44 profiles was obtained and on the basis of this plot, thresholds for both C3_NSAT HMM and C4_NSAT HMM were set to 400 bits (Figure 4A).…”

Clues to reaction specificity in PLP-dependent fold type I aminotransferases of monosaccharide biosynthesis

“…TrEMBL under the assumption that annotations provided in TrEMBL are correct. The procedure used to generate ROC curves is described in detail elsewhere [6], [35]. Briefly, ROC curves were generated by calculating the number of true positives, false positives and false negatives for various values of bit score threshold (Figure S3).…”

“…Bit score thresholds for NSATHMM and NSDHMM, profile HMMs of aminotransferases and dehydratases, respectively, were set based on Receiver Operator Characteristic (ROC) curves generated using hits from TrEMBL under the assumption that annotations provided in TrEMBL are correct. The procedure used to generate ROC curves is described in detail elsewhere 6,34 . Briefly, ROC curves were generated by calculating the number of true positives, false positives and false negatives for various values of bit score threshold (Figure S3).…”