Identification of homologs in insignificant blast hits by exploiting extrinsic gene properties
© Boekhorst and Snel; licensee BioMed Central Ltd. 2007
Received: 18 June 2007
Accepted: 21 September 2007
Published: 21 September 2007
Homology is a key concept in both evolutionary biology and genomics. Detection of homology is crucial in fields like the functional annotation of protein sequences and the identification of taxon specific genes. Basic homology searches are still frequently performed by pairwise search methods such as BLAST. Vast improvements have been made in the identification of homologous proteins by using more advanced methods that use sequence profiles. However additional improvement could be made by exploiting sources of genomic information other than the primary sequence or tertiary structure.
We test the hypothesis that extrinsic gene properties gene length and gene order can be of help in differentiating spurious sequence similarity from homology in the gray zone. Sharing gene order and similarity in size dramatically increase the chance of a query-hit pair being homologous: gray zone query-hit pairs of similar size and with conserved gene order are homologous in 99% of all cases, while for query-hit pairs without gene order conservation and with different sizes this is only 55%.
We have shown that using gene length and gene order drastically improves the detection of homologs within the BLAST gray zone. Our findings suggest that the use of such extrinsic gene properties can also improve the performance of homology detection by more advanced methods, and our study thereby underscores the importance of true data integration for fully exploiting genomic information.
Homology is a key concept in both evolutionary biology and genomics. Homology designates a relationship of common descent between entities. In genomics, homologs are genes or genomics regions sharing a common origin, related through speciation, duplication or a combination of both. Orthology is a specific case of homology, in which genes in different species evolved from a common ancestral gene through speciation . The identification of homologous proteins is an important step in predicting the function of proteins that have not been studied experimentally and is crucial in comparative genomics studies. Identification of taxon-specific genes and estimation of the rate of gene genesis all rely on the detection of orthology (and thus homology). Rapid sequence divergence can obscure the real evolutionary relationship between genes , a scenario that could result in an overestimation of the contribution of the emergence of taxon-specific genes.
Algorithms like BLAST detect sequence similarity . Many insignificant (by e-value) BLAST hits are nevertheless homologs. Increased sensitivity can be obtained by using more sophisticated methods, such as more accurate search algorithms like PSI-BLAST  and hidden Markov models (HMMs) . These methods have allowed for large improvements in the identification of homologous proteins compared to BLAST, yet detection of homology might also be improved by using information not derived from the primary sequence of proteins. From here on such information is referred to as label information. We test the hypothesis that label information can make an important contribution in the detection of homology.
In order to estimate the significance of label information, we need both a method for the initial detection of sequence similarity as well as a golden standard against which we can measure the improvements made by the use of label information. The use of BLAST provides us with a commonly used method for the detections of homology, while at the same time leaving room for improvement, allowing us to use methods like HMMs as a golden standard to benchmark gray zone BLAST hits.
The first label we investigate in this study is gene order. In the absence of changes in gene order (i.e. if no rearrangement, deletion or duplication of genes were to occur), even genes with no detectable sequence but sharing gene order would be homologous. Gene order has been proven to be a useful tool in the identification of homologs with low sequence similarity, as illustrated for non-coding RNA genes . We test the hypothesis that proteins sharing gene order and identified in a BLAST search as being weakly similar in sequence (gray zone hits) are more likely to be homologous than proteins with the same sequence similarity that do not share gene order. The reasoning for the second label we study, protein length, is similar. Unless gene fusion or fission events have taken place, gene sharing a common ancestor and retaining the same fold should be similar in length. A substantial difference in length between two proteins suggests a difference in three dimensional structure, in turn suggesting a different evolutionary origin; two proteins with similar sizes are more likely to be homologs than two proteins with a relatively large difference in size.
Here we test whether label information can play an important role in the identification of homologs in the gray zone. We use the prokaryotic protein sequences from the COG database  as dataset in an all-against-all BLAST, and the PFAM database as a golden standard. We discuss the implications of our findings for the detection of sequence homology with more advanced methods, the identification of taxon specific genes, and the importance of true data integration.
The value of label information
Here we test the expectation that label information can play an important role in the identification of homologs in the gray zone by comparing the results of an all-against-all BLAST search of the proteins from the COG database to data from the PFAM database. The gray-zone of BLAST is not systematically defined; normally, it represents a range of values starting directly above some e-value threshold up to BLAST's default maximum e-value of 10. We here refrain from directly designating a grey zone and instead vary the e-value threshold. For values which are normally considered grey-zone, we then observe as expected that a substantial portion of hits are not homologous according to PFAM: 65% of the BLAST hits with an e-value above 1e-03 but below 10 are homologous according to PFAM clans, and only 43% of the hits are homologous at e-values between 1 and 10. In contrast, over 99% of the hits below a threshold of 1e-03 are homologous according to PFAM clans, also as expected.
Combining gene order and gene size (Figure 1C) further increases the differentiation between homologous query-hit pairs and spurious BLAST hits. Query-hit pairs with gene order conservation and with similar sizes (80%–100%) are homologous in over 99% of all cases, while for query-hit pairs of different sizes (40%–60%) and without conserved gene order this is 55% (Figure 1C). Expressed as a reduction in the fraction of false positives, this improvement is quite high: from 0.45 to 0.01 is a 45 fold reduction. Note that because of the relatively small number of genes with conserved gene order (more on this below) the inclusion of gene order does not have a big impact on the precision when looking at the fraction of homologous pairs with a large size difference (56% for all pairs, 55% for the pairs without conserved gene order).
Comparing the contributions of conserved gene order and size similarity
Estimating the number of additional homologs that can be detected
A well-known problem in the detection of homology using sequence profiles is the threshold effect: creating a sequence profile and setting a threshold that excludes all non-homologous sequences while at the same still recognizing all homologs can be very difficult. To illustrate the value of gene order and gene length in such scenarios, we mined our data set for gray-zone query-hit pairs of similar size and with conserved gene order, but where only one of the two proteins is hit by a PFAM model. These pairs were left out of our original analysis, as for these pairs we can not reliably automatically establish if they are homologous according to our golden standard. The protein SA0534 from Staphylococcus aureus hits the protein AGl121 from Agrobacterium tumefacie (e-value: 0.6). SA0534 is member of protein family PF00108 (Thiolase), while AGl121 is not a member of any protein family. However, conservation of gene order and similarity in length suggest that the proteins are homologous. Again we could confirm this using PSI-BLAST: after 3 iterations SA0534 hits AGl121 with an e-value of 2e-52. A multiple sequence alignment of AGl121 with the members of PF00108 shows the cysteine residue experimentally proven to be essential for thiolase activity  to be conserved.
The value of label information
We have shown that query-hit pairs of similar size and sharing gene order are much more likely to be homologous than pairs less similar in size and not sharing gene order. Obviously, this effect is most dramatic for gray-zone hits: query-hit pairs without gene order conservation and with different sizes (the length of the smallest protein between 40% and 60% of that of the largest protein) are homologous in only 55% of all cases, while for query-hit pairs of similar size (80%–100%) and conserved gene order this is 99% (Figure 1). For an e-value threshold between 1e-02 and 1e-12, the use of the labels gene size and gene order provides an average increase in the number of hits ranging from 6% to 12% (Figure 3B). Using label information can mean the difference between ignoring a gray zone BLAST hit as spurious and recognizing a distant homolog.
The small fraction of gray zone query-hit pairs with conserved gene order compared to gray zone hits of similar size (Table 1B,C) means that conserved gene order has less impact on the overall picture. In addition, an a priori advantage of length as a label over conserved gene context is that the use of conserved gene order is in practice limited to prokaryotes, because in eukaryotes gene order evolves relatively fast compared to protein sequences .
Note that the use of gene order in this context is not so obvious, because it does not equal the use of conserved gene order for function prediction, such as it is applied in genomic context methods [10, 11]. In fact conserved gene order would not even have been detected if only blast were used (as is often the case), as one of the genes in a pair of neighbouring genes does not have a significant blast hit in the other species. The use of gene order as a label is here thus the reverse approach of the genomic context method: we suppose the conservation of gene order as a clue.
Fusion of proteins
As the COG database offers an orthology prediction we could use for the detection of conserved gene order, and because of the summary of an all-against-all BLAST provided with the database, we used the proteins from COG as a starting point. A potential problem with this dataset is the presence of fusion proteins in the COG database as separate entries . We investigated the effect of this lack of fusion proteins on size similarity as a relevant label by concatenating the relevant protein entries our analysis with this alternative set. The results show proteins size to be a relevant label for recognizing gray zone homologs also in datasets containing fusion proteins.
Methods other than BLAST
BLAST is applied extensively, yet in many scenarios more advanced methods are more suitable. The observation that label information can be of great value in interpreting the results of a BLAST search could be extrapolated to more advanced methods. A somewhat trivial example of this is given in the results section, where we show via label information that two sequences likely belong to the same PFAM family, even though only one of them is hit by the PFAM model above the inclusion bitscore threshold. In related work , the label taxonomy is shown to enhance the detection of PFAM domains.
In the field of distant homology detection, the challenge has become the detection of homologous domains (e.g. profile vs. profile alignment and fold prediction [13, 14]). The applicability of the labels for this field might be difficult, but we think the label protein length could perhaps be fairly directly translated to the length of conserved domains. It is very unlikely to find such a direct analogy for conserved gene order, but more tenuous analogs could be found among the co-occurrence  and relative order of domains in a protein, or to the more general concept of functional links, such as co-expression or participating in the same pathway (thus implicitly assuming the duplication of functional modules). Note also that investigating the value of label information for more advanced methods such as PSI-BLAST provides a practical challenge in the sense that it requires an even more reliable method or database to provide a golden standard.
One of the scenarios in which improved homology detection and thus label information can have a big impact is the estimation of percentage of so-called novel genes: genes for which no homologs could be identified in other organisms. BLAST is often the method of choice for the detection of genes specific to a species or taxon [16–18], as BLAST can make use of all available sequences, while many methods based on sequence models like PFAM and SCOP [19, 20] do not cover the full spectrum of available sequence data. Almost two-thirds of the proteins encoded by the genome of Plasmodium falciparum did not have sufficient similarity to proteins in other organisms to justify provision of functional assignment, which led the authors to state that two-thirds of the proteins appear to be unique to the organism . The use of label information would allow a more accurate identification of proteins that in fact do have homologs in the set of proteins with only gray zone BLAST hits, in turn leading to a decrease in the fraction of unique proteins.
There has been a call for so-called true data integration : tools and databases should not only combine different types of biological data in the sense that they provide access to different types of analysis or data sources through a single interface, they should in fact combine different types of data to come to a single answer taking into account as much of the available data as possible. This would, in the case of BLAST and the labels gene size and gene order, be a tool that combines a similarity search, gene size, and information on genome organization into a statistical framework that produces a single score representing the likelihood of a hit being homologous to the query. The use of other labels like co-expression and phylogeny requires an extensive integration of databases and tools, yet as we have shown the gain is potentially substantial.
Sequence information and analysis
Protein sequences were taken from the COG database . As one of the two labels we are interested in is gene order conservation, and gene order evolves more slowly in bacteria , we excluded the three genomes of eukaryotes, leaving us with 63 bacterial genomes with a total of 192987 protein sequence. Gene order information was taken from Genbank  after mapping COG identifiers to Genbank gene identifiers. Gene order was considered to be conserved when two genes have one or more neighbours belonging to the same COG. We used BLAST  for the detection of sequence similarity. Where possible we took advantage of the pre-calculated BLAST data provided by the COG database. Information on protein families was taken from version 20.0 of the PFAM database . In the cases where an exact match of a protein in our data set could not be identified in the PFAM database, we assigned protein families using HMMer  and the models provided by PFAM; in all other cases we used the pre-calculated results distributed with the database. Multiple sequence alignments were made with MUSCLE , phylogenetic trees were constructed using PhyML .
Measuring the value of label information
We measured the value of the label information in identifying homologous by binning all query-hit pairs from the all-against-all BLAST provided with the COG database by e-value and similarity in protein length (defined as the relative size of the smallest protein of a query-hit pair compared the largest in %). For each bin we counted the homologous pairs sharing gene order, the homologous pairs not sharing gene order, the not-homologous pairs sharing gene order and the not-homologous pairs not sharing gene order. Numbers of hits were normalized by dividing by the number of hits in that category of a query protein by the total number of hits of that query in order to prevent large protein families (which have many very similar hits) from skewing the results. The chance that the two proteins of a query-hit pair in a specific score bin and with specific label properties are homologous (also called the positive predictive value (PPV) or precision) was calculated by dividing the number of homologous pairs (true positives) by sum of the number of homologous pairs and the number of non-homologous pairs (true positives + false positives).
Our methodology requires both a golden standard for defining homologous proteins as well as a standard for defining non-homologous proteins. Pairs were considered homologous when the proteins belong to the same protein family (as defined by the PFAM database) or to protein families belonging to the same clan , while pairs were considered to be non-homologous when they were part of protein families not belonging to the same clan. The number of false negatives was reduced by excluding pairs belonging to different PFAM clans while at the same time belonging to the same orthologous group of the COG database. Query-hit pairs formed by one or two proteins not belonging to any protein family were left out of the analysis; this keeps protein pairs that are in fact homologous, but that are not covered by any of the models in the PFAM database, as well as proteins that are hit by a PFAM model but fall just below the gathering threshold established by PFAM, from being counted as non-homologous.
This work was supported by BioRange project SP 2.3.1
- Koonin EV: Orthologs, paralogs, and evolutionary genomics. Annu Rev Genet 2005, 39: 309–338. 10.1146/annurev.genet.39.073003.114725View ArticlePubMedGoogle Scholar
- Copley RR, Goodstadt L, Ponting C: Eukaryotic domain evolution inferred from genome comparisons. Curr Opin Genet Dev 2003, 13(6):623–628. 10.1016/j.gde.2003.10.004View ArticlePubMedGoogle Scholar
- Altschul SF, Gish W, Miller W, Myers EW, Lipman DJ: Basic local alignment search tool. J Mol Biol 1990, 215(3):403–410.View ArticlePubMedGoogle Scholar
- Altschul SF, Madden TL, Schaffer AA, Zhang J, Zhang Z, Miller W, Lipman DJ: Gapped BLAST and PSI-BLAST: a new generation of protein database search programs. Nucleic Acids Res 1997, 25(17):3389–3402. 10.1093/nar/25.17.3389PubMed CentralView ArticlePubMedGoogle Scholar
- Eddy SR: Profile hidden Markov models. Bioinformatics 1998, 14(9):755–763. 10.1093/bioinformatics/14.9.755View ArticlePubMedGoogle Scholar
- Sridhar J, Rafi ZA: Small RNA identification in Enterobacteriaceae using synteny and genomic backbone retention. Omics 2007, 11(1):74–99. 10.1089/omi.2006.0006View ArticlePubMedGoogle Scholar
- Tatusov RL, Fedorova ND, Jackson JD, Jacobs AR, Kiryutin B, Koonin EV, Krylov DM, Mazumder R, Mekhedov SL, Nikolskaya AN, Rao BS, Smirnov S, Sverdlov AV, Vasudevan S, Wolf YI, Yin JJ, Natale DA: The COG database: an updated version includes eukaryotes. BMC Bioinformatics 2003, 4: 41. 10.1186/1471-2105-4-41PubMed CentralView ArticlePubMedGoogle Scholar
- Kursula P, Ojala J, Lambeir AM, Wierenga RK: The catalytic cycle of biosynthetic thiolase: a conformational journey of an acetyl group through four binding modes and two oxyanion holes. Biochemistry 2002, 41(52):15543–15556. 10.1021/bi0266232View ArticlePubMedGoogle Scholar
- Huynen MA, Snel B, Bork P: Inversions and the dynamics of eukaryotic gene order. Trends Genet 2001, 17(6):304–306. 10.1016/S0168-9525(01)02302-2View ArticlePubMedGoogle Scholar
- Dandekar T, Snel B, Huynen M, Bork P: Conservation of gene order: a fingerprint of proteins that physically interact. Trends Biochem Sci 1998, 23(9):324–328. 10.1016/S0968-0004(98)01274-2View ArticlePubMedGoogle Scholar
- Overbeek R, Fonstein M, D'Souza M, Pusch GD, Maltsev N: The use of gene clusters to infer functional coupling. Proc Natl Acad Sci U S A 1999, 96(6):2896–2901. 10.1073/pnas.96.6.2896PubMed CentralView ArticlePubMedGoogle Scholar
- Coin L, Bateman A, Durbin R: Enhanced protein domain discovery using taxonomy. BMC Bioinformatics 2004, 5: 56. 10.1186/1471-2105-5-56PubMed CentralView ArticlePubMedGoogle Scholar
- Sadreyev R, Grishin N: COMPASS: a tool for comparison of multiple protein alignments with assessment of statistical significance. J Mol Biol 2003, 326(1):317–336. 10.1016/S0022-2836(02)01371-2View ArticlePubMedGoogle Scholar
- Soding J: Protein homology detection by HMM-HMM comparison. Bioinformatics 2005, 21(7):951–960. 10.1093/bioinformatics/bti125View ArticlePubMedGoogle Scholar
- Coin L, Bateman A, Durbin R: Enhanced protein domain discovery by using language modeling techniques from speech recognition. Proc Natl Acad Sci U S A 2003, 100(8):4516–4520. 10.1073/pnas.0737502100PubMed CentralView ArticlePubMedGoogle Scholar
- Griffiths E, Ventresca MS, Gupta RS: BLAST screening of chlamydial genomes to identify signature proteins that are unique for the Chlamydiales, Chlamydiaceae, Chlamydophila and Chlamydia groups of species. BMC Genomics 2006, 7: 14. 10.1186/1471-2164-7-14PubMed CentralView ArticlePubMedGoogle Scholar
- Hartman H, Fedorov A: The origin of the eukaryotic cell: a genomic investigation. Proc Natl Acad Sci U S A 2002, 99(3):1420–1425. 10.1073/pnas.032658599PubMed CentralView ArticlePubMedGoogle Scholar
- Makarova K, Slesarev A, Wolf Y, Sorokin A, Mirkin B, Koonin E, Pavlov A, Pavlova N, Karamychev V, Polouchine N, Shakhova V, Grigoriev I, Lou Y, Rohksar D, Lucas S, Huang K, Goodstein DM, Hawkins T, Plengvidhya V, Welker D, Hughes J, Goh Y, Benson A, Baldwin K, Lee JH, Diaz-Muniz I, Dosti B, Smeianov V, Wechter W, Barabote R, Lorca G, Altermann E, Barrangou R, Ganesan B, Xie Y, Rawsthorne H, Tamir D, Parker C, Breidt F, Broadbent J, Hutkins R, O'Sullivan D, Steele J, Unlu G, Saier M, Klaenhammer T, Richardson P, Kozyavkin S, Weimer B, Mills D: Comparative genomics of the lactic acid bacteria. Proc Natl Acad Sci U S A 2006, 103(42):15611–15616. 10.1073/pnas.0607117103PubMed CentralView ArticlePubMedGoogle Scholar
- Finn RD, Mistry J, Schuster-Bockler B, Griffiths-Jones S, Hollich V, Lassmann T, Moxon S, Marshall M, Khanna A, Durbin R, Eddy SR, Sonnhammer EL, Bateman A: Pfam: clans, web tools and services. Nucleic Acids Res 2006, 34(Database issue):D247–51. 10.1093/nar/gkj149PubMed CentralView ArticlePubMedGoogle Scholar
- Murzin AG, Brenner SE, Hubbard T, Chothia C: SCOP: a structural classification of proteins database for the investigation of sequences and structures. J Mol Biol 1995, 247(4):536–540. 10.1006/jmbi.1995.0159PubMedGoogle Scholar
- Gardner MJ, Hall N, Fung E, White O, Berriman M, Hyman RW, Carlton JM, Pain A, Nelson KE, Bowman S, Paulsen IT, James K, Eisen JA, Rutherford K, Salzberg SL, Craig A, Kyes S, Chan MS, Nene V, Shallom SJ, Suh B, Peterson J, Angiuoli S, Pertea M, Allen J, Selengut J, Haft D, Mather MW, Vaidya AB, Martin DM, Fairlamb AH, Fraunholz MJ, Roos DS, Ralph SA, McFadden GI, Cummings LM, Subramanian GM, Mungall C, Venter JC, Carucci DJ, Hoffman SL, Newbold C, Davis RW, Fraser CM, Barrell B: Genome sequence of the human malaria parasite Plasmodium falciparum. Nature 2002, 419(6906):498–511. 10.1038/nature01097View ArticlePubMedGoogle Scholar
- Koonin EV: Eugene V. Koonin (Interview). Curr Biol 2004, 14(3):R96–7. 10.1016/S0960-9822(04)00025-9View ArticlePubMedGoogle Scholar
- Benson DA, Karsch-Mizrachi I, Lipman DJ, Ostell J, Wheeler DL: GenBank. Nucleic Acids Res 2006, 34(Database issue):D16–20. 10.1093/nar/gkj157PubMed CentralView ArticlePubMedGoogle Scholar
- Edgar RC: MUSCLE: multiple sequence alignment with high accuracy and high throughput. Nucleic Acids Res 2004, 32(5):1792–1797. 10.1093/nar/gkh340PubMed CentralView ArticlePubMedGoogle Scholar
- Guindon S, Gascuel O: A simple, fast, and accurate algorithm to estimate large phylogenies by maximum likelihood. Syst Biol 2003, 52(5):696–704. 10.1080/10635150390235520View ArticlePubMedGoogle Scholar
- Page RD: TreeView: an application to display phylogenetic trees on personal computers. Comput Appl Biosci 1996, 12(4):357–358.PubMedGoogle Scholar
This article is published under license to BioMed Central Ltd. This is an Open Access article distributed under the terms of the Creative Commons Attribution License (http://creativecommons.org/licenses/by/2.0), which permits unrestricted use, distribution, and reproduction in any medium, provided the original work is properly cited.