Challenges in homology search: HMMER3 and convergent evolution of coiled-coil regions.

Mistry J, Finn RD, Eddy SR, Bateman A, Punta M. · 2013

Read Challenges in homology search: HMMER3 and convergent evolution of coiled-coil regions. by Mistry J, Finn RD, Eddy SR, Bateman A, Punta M. (2013) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/europmc-pmc3695513.txt.

Introduction

Mistry and colleagues stress-tested HMMER3, the main software for finding protein relatives by sequence, against the hand-curated Pfam family database. Profile methods are powerful, but two traps persist: alignments that stretch past the truly related region, and regions in unrelated proteins that converged on similar patterns.

Running the software against more than thirteen thousand curated families, they pinpointed problem cases where one region matched several supposedly unrelated families. Error estimates proved least trustworthy for families with repeating, biased patterns, such as coiled coils. The authors recommend keeping human-curated thresholds in Pfam, especially for the flagged families, rather than trusting automatic cutoffs alone.

Catalog record

CREATORMistry J, Finn RD, Eddy SR, Bateman A, Punta M.
YEAR2013
COLLECTIONPAPERS
FORMATTXT
FILE SIZE41 KB
LICENSECC BY 4.0
RIGHTSCC BY
SOURCEEurope PMC

Source: https://europepmc.org/articles/PMC3695513

Browse the full library · About the Tower of Babel

Full text

The opening of the work, from the library's hosted text:

Challenges in homology search: HMMER3 and convergent evolution of coiled-coil regions By Jaina Mistry, Robert D Finn, Sean R Eddy, Alex Bateman, Marco Punta Published in: Nucleic Acids Research DOI: 10.1093/nar/gkt263 ABSTRACT Abstract Detection of protein homology via sequence similarity has important applications in biology, from protein structure and function prediction to reconstruction of phylogenies. Although current methods for aligning protein sequences are powerful, challenges remain, including problems with homologous overextension of alignments and with regions under convergent evolution. Here, we test the ability of the profile hidden Markov model method HMMER3 to correctly assign homologous sequences to >13 000 manually curated families from the Pfam database. We identify problem families using protein regions that match two or more Pfam families not currently annotated as related in Pfam. We find that HMMER3 E-value estimates seem to be less accurate for families that feature periodic patterns of compositional bias, such as the ones typically observed in coiled-coils. These results support the continued use of manually curated inclusion thresholds in the Pfam database, especially on the subset of families that have been identified as problematic in experiments such as these. They also highlight the need for developing new methods that can correct for this particular type of compositional bias. INTRODUCTION Known problems in detecting homology via sequence similarity Homology, or evolutionary descent from a common ancestor, is widely used as a basis to transfer structural and functional annotation between proteins (1–3) and to study protein evolution (4). Homology is most often inferred by establishing statistical significance of observed similarity in protein sequence alignments. Several methods based on sequence–sequence, sequence–profile or profile–profile alignments have been developed to identify evolutionary-related protein sequences (5–10). Current challenges in homology detection via sequence similarity include problems with homologous overextension of alignments (11) and regions of amino acid compositional bias under convergent evolution (12). Note that the expression ‘compositional bias’ is often used to indicate a bias in amino acid frequencies within a protein sequence region, that is, a deviation from the frequencies typically observed in globular soluble proteins. An example of this bias is the highly polar, highly charged intrinsically disordered regions (13). Convergent evolution, however, may also lead to a different type of bias, which is reflected not in amino acid frequencies but rather in the way amino acids are distributed along the sequence. The heptad repeats that characterize canonical coiled-coils represent a typical example of this bias (14). Benchmarking HMMER3 E-values estimates using Pfam Here, we test the ability of the profile hidden Markov model (profile-HMM) method HMMER3 [http://hmmer.janelia.org/) to correctly infer homology. To this end, we use a library of >13 000 profile-HMMs taken from the Pfam database of protein families (15). Pfam’s comprehensiveness (the fact that Pfam models annotate a large proportion of all residues in UniProt Knowledgebase, UniProtKB (16)] allows us in many cases to identify sources of discrepancies in HMMER3 E-value estimation without using a benchmark. For example, we can look at cases where Pfam annotation, at a given significance threshold, would classify the same region of a sequence into two or more families thought to be evolutionarily unrelated. We call this ‘overlap analysis’. Benchmarks are typically derived from protein structure database classifications, but solved protein structures are depleted for some of the most challenging issues in sequence analysis, such as disordered and/or highly biased sequence composition. Overlap analysis in the entire sequence database reduces this important bias, as Pfam has models representing many such regions, although it has the disadvantage that an overlap does not necessarily imply a false positive (it could be an as-yet unannotated true homology). Nonetheless, manual examination of particular Pfam families that cause an unusual number of overlaps can be used to identify systematic failure modes in HMMER3’s inference of homology. Pfam overlaps We define overlaps as sequence regions that at a given significance threshold simultaneously match two or more Pfam families found in different clans (i.e. families that are not annotated as evolutionary related in the Pfam database). Additionally, we require that ≥50% of the region that matches the profile-HMM of one family also matches the profile-HMM of the other (for at least one of the two families). Overlaps are likely to point to some form of erroneous or incomplete annotation. In particular: (i) the profile-HMMs of one or both of the two families might have been built from a set of non-homologous sequences (incorrect Pfam model); (ii) the overlapping families might be evolutionary related and the relationship not yet annotated in Pfam (incomplete Pfam annotation); and (iii) the profile-HMMs might have incorrectly assigned the sequence to one or both of the overlapping families (false positive). We calculate overlaps between families at E-values that are typically used to suggest homology. Although a number of these overlaps will fall into cases (i) to (ii), we expect this set to be enriched in false positives (iii) with respect to all Pfam matches. MATERIALS AND METHODS Pfam profile-HMM database Pfam (15) is a database that uses HMMER3 (http://hmmer.janelia.org/) to group homologous protein regions into families. Pfam 26.0 comprises 13 672 families. Each family is represented by a profile-HMM generated with HMMER3 using a manually curated ‘seed’ alignment. The profile-HMM is searched against UniProtKB (16), (the Pfam protein sequence database of choice), to find additional homologous sequences. Pfam curators define two separate family-specific significance thresholds for including matches into families, one for domains and one for sequences [a domain can match the same sequence multiple times; see (15) for a detailed explanation of what the different thresholds represent]. Each family is annotated with functional and/or structural information where available. In this study, to simplify our overlap analysis that is based on family-independent significance thresholds, we exclude the 316 families where sequence and domain thresholds differ. These families mostly represent repeats in which the domain threshold is set to be lower than the sequence threshold to capture the highest number of occurrences along individual sequences. Definition of overlap As outlined in the ‘Introduction’ section, overlaps are sequence regions that match two or more Pfam families that belong to different clans, with families in different clans not annotated as related in the Pfam database. To exclude cases that could be resolved by small changes in family boundaries, we consider only overlapping regions that cover ≥50% of the alignment co-ordinates of at least one of the two families involved. Note: alignment coordinates represent the region of the sequence for which HMMER3 can produce a reliable alignment in contrast to envelope coordinates that define the region for which there is substantial probability mass to support a homologous match. Thus, the envelope coordinates span a region that is equal or wider than the alignment coordinates. Calculating Pfam family overlaps A subset of Pfam 26.0 profile-HMMs (13 356, see earlier) was searched against UniProtKB release 2011_06 using HMMER3.0 hmmsearch with default parameters. E-values are as reported by HMMER3 hmmsearch, with no additional correction for having performed multiple profile-HMM searches. We define a match as a domain alignment that has an E-value less than or equal to a specified significance threshold. We calculate overlaps between different Pfam families’ matches, as defined earlier in the text, for three different E-value thresholds: 0.001, 0.01 and 0.1. Calculating expected

Read the complete text