Iterative refinement of structure-based sequence alignments by Seed Extension
© Kim et al; licensee BioMed Central Ltd. 2009
Received: 02 February 2009
Accepted: 09 July 2009
Published: 09 July 2009
Accurate sequence alignment is required in many bioinformatics applications but, when sequence similarity is low, it is difficult to obtain accurate alignments based on sequence similarity alone. The accuracy improves when the structures are available, but current structure-based sequence alignment procedures still mis-align substantial numbers of residues. In order to correct such errors, we previously explored the possibility of replacing the residue-based dynamic programming algorithm in structure alignment procedures with the Seed Extension algorithm, which does not use a gap penalty. Here, we describe a new procedure called RSE (Refinement with Seed Extension) that iteratively refines a structure-based sequence alignment.
RSE uses SE (Seed Extension) in its core, which is an algorithm that we reported recently for obtaining a sequence alignment from two superimposed structures. The RSE procedure was evaluated by comparing the correctly aligned fractions of residues before and after the refinement of the structure-based sequence alignments produced by popular programs. CE, DaliLite, FAST, LOCK2, MATRAS, MATT, TM-align, SHEBA and VAST were included in this analysis and the NCBI's CDD root node set was used as the reference alignments. RSE improved the average accuracy of sequence alignments for all programs tested when no shift error was allowed. The amount of improvement varied depending on the program. The average improvements were small for DaliLite and MATRAS but about 5% for CE and VAST. More substantial improvements have been seen in many individual cases. The additional computation times required for the refinements were negligible compared to the times taken by the structure alignment programs.
RSE is a computationally inexpensive way of improving the accuracy of a structure-based sequence alignment. It can be used as a standalone procedure following a regular structure-based sequence alignment or to replace the traditional iterative refinement procedures based on residue-level dynamic programming algorithm in many structure alignment programs.
In searching for protein functions and in building homology models, it is desirable to have accurate sequence motifs and profiles [1–3], which are obtained from sequence alignments of homologous proteins. However, it is often difficult to obtain accurate sequence alignments based on sequence similarity alone when sequence similarity is low.
Therefore, structural alignments, when available, have been used to guide sequence alignments. Such structure-based sequence alignments have been used as the gold standard to evaluate pure sequence alignment methods [4, 5] and to derive structural environment-specific substitution matrices which have been shown to be useful for detection of remote homologs and for sequence-structure alignments [6–9].
However, structure-based sequence alignments produced by different programs can be different even when the structures are similar [10, 11]. There are a large number of instances wherein all or parts of the structure are shifted by 2 or 4 residues or even by an odd number of residues . Some methods are probably quite good at detecting structural similarity, yet relatively poor in terms of the accuracy of the sequence alignment they produce .
DaliLite and VAST use a Monte-Carlo procedure after initial structural alignment [13, 14], FATCAT and MATT adopt AFP (aligned fragment pair)-based dynamic programming without constructing initial structural alignments [15, 16], and other programs mostly rely on residue-level dynamic programming algorithm according to various scoring schemes with or without initial rigid-body superposition [17–20].
We previously developed the SE (Seed Extension) algorithm which generates a sequence alignment from a superimposed structure pair without changing the superposition . A number of other programs [22–25] also provide a similar function, but these programs use the dynamic programming algorithm and a gap penalty. We have shown that SE, which is not based on the dynamic programming algorithm and does not use a gap penalty, generates a more accurate alignment on average than programs that use a dynamic programming algorithm.
In this study, we report on the development of a fast refinement procedure, which can be used to improve an existing structure-based sequence alignment. The procedure, which we call RSE (Refinement with SE), is an iterative procedure that uses SE in its core. Using CDD (Conserved Domain Database)  "root node set" as the reference alignment , we show that appending the RSE procedure to a structure-based sequence alignment program improves the accuracy of the alignment for all 9 programs tested.
Improvement of the overall alignment accuracy
Composition of the CDD root node set
Number of CDs†
Number of structure pairs
The nature of the improvement varied among different methods. For CE, MATT and TM-align, RSE improved <FCAR(0)> but not <FCAR(8)> (Figure 1), which indicates that it is mostly alignment shift error that was reduced by the RSE procedure. For FAST and SHEBA-4, the improvements appear to be mainly correction of under-alignments, presumably by reducing the number of gaps, since <FCAR(8)> increased almost as much as <FCAR(0)> by the refinement.
Dependence on structural types
Refinement of good and not-so-good initial input alignments
Average performance of the control methods
RSE also improved the accuracy of the alignments from the pure sequence alignment program SSEARCH by 19% to about 67% and from the profile-profile alignment program SALIGN by 20% to 75% (Table 2). This shows that RSE improves even a poor alignment. But the final accuracy attained was substantially lower than those from any structure comparison programs.
Comparison of improvements between SE and RSE
The most improved case in immunoglobulin superfamily (cd00096) for each method
RSE could correct these alignments, unlike SE (panels B and C). Since SE just derives a sequence alignment from a given structural superposition without changing it, it cannot correct a bad superposition. In contrast, RSE iteratively adjusts the structural superposition, which can result in a large improvement.
Quality of the CDD alignments as the standard
In order to better understand the nature of the changes of the CDD alignments by the RSE procedure, many cases were visually inspected. There were 136 pairs (3.4% of all pairs) from 21 different superfamilies for which the fCAR(0) in RSE-refined CDD alignment decreased by more than 20%. As expected, some of these structure pairs were from the cd00531 (7 pairs) and cd01984 (6 pairs) superfamilies, for which our previous study  indicated that the CDD alignments were in error. For some pairs from two other superfamilies (cd00198 and cd00385), RSE again appeared to produce more reasonable alignments than CDD, in terms of the distances and orientations of side chains between aligned residues. Fourteen pairs including the worst three cases were from cd00688, which are made of α/α toroid structures (a barrel made of two layers of alpha-helices). Not all helices in these structures could be superposed simultaneously without ambiguity and RSE produced tilted alignments. There were 47 pairs from the three superfamilies having the (β/α)8 TIM-barrel structure (cd01292, cd00415 and cd00945), for which the inner layer of beta-strands were reasonably alignable but the outer helices were not. There were other helix-containing superfamilies (cd00389, cd00397, cd00198, and cd00385), for which at least one pair of alpha-helices was not unambiguously alignable. For some pairs in cd00158, CDD has pairs of residues aligned, which RSE could not align because they were too far apart from each other in an irregularly shaped region of the superposed structures. These were aligned in CDD presumably by sequence similarity.
Structure-based sequence alignments are not as robust as one would like. In some cases, they can be inherently ambiguous. But more frequently different structure alignment programs generate alignments that contain errors that can be easily recognized by human experts. We showed in a previous study  that, the overall average accuracy of structure-based sequence alignments, as measured by <FCAR(0)> with the CDD root node set as the reference, ranged from 81% to 89% depending on the program used. When the five outlier superfamilies  are excluded, it ranges from 84% to 92% (Figure 1). The two newly included methods, TM-align  and MATT , are not exceptional in this regard.
The RSE procedure reported here was designed to improve the structure-based sequence alignments. It uses the previously reported SE algorithm  to obtain a refined sequence alignment from an input alignment. SE is a heuristic algorithm that produces an alignment from two superimposed structures without using a gap penalty. Figure 1 shows that the average accuracy improved for all structure alignment programs tested by adding the RSE refinement procedure. Notably, alignments from MATT, which is a program that considers structural flexibility, could also be improved significantly by the RSE procedure, which does not explicitly consider structural flexibility. RSE reduced the shift error for most programs since the refinement increases FCAR(0) more than FCAR(8). For FAST and SHEBA-4, RSE seems to lengthen the alignment also since FCAR(8) and FCAR(0) increased to a similar extent. The alignments improved for structure pairs from all SCOP classes for most of the programs tested (Figure 5).
Impressively, the alignments from FAST, one of the fastest programs, could be improved to about the same level of accuracy as those from DaliLite, the best performer without RSE (Figure 1). The accuracies of MATT and SHEBA-4 also increased to similar levels. These improvements were achieved with nearly negligible increase in overall processing times (Figures 2 and 3). Therefore structure alignments can be done with substantially reduced computational cost without compromising accuracy by combining RSE with one of the fastest programs. Alternatively, the RSE procedure can be implemented to replace the traditional residue-based dynamic programming algorithm in a structure comparison program that uses it to improve both the accuracy and computing time.
An ideal refinement procedure will fix incorrectly aligned regions without degrading the correctly aligned ones (Figure 7). Unfortunately, RSE seems to degrade some alignments when compared to the CDD alignments (Figure 2). When the CDD alignment itself was used as the initial alignment for an RSE procedure, <FCAR(0)> and <fCAR(0)> decreased to about 95% and 96%, respectively (Table 2). According to our visual inspection of a number of cases for which fCAR(0) fell to a value below 80%, the RSE procedure appears to have found an alternate alignment or to have corrected CDD errors in most cases. We expect that similar causes are at work for at least some of the cases seen in Figure 2 for which there is an apparent degradation of alignment accuracy.
RSE greatly improves the alignments from SSEARCH and SALIGN, which are non-structure-based, pure sequence-based alignment procedures (Table 2). This is to be expected since use of the structural information should improve the sequence alignment. One notes, however, that the average accuracy attained after the refinement is far below those of any of the structure alignment methods (Compare the numbers in Table 2 and the bar heights in Figure 1). This indicates that the outcome of the RSE procedure does depend on the quality of the input alignment. One can also note that there are about 7 to 11% error left after the RSE refinement of the alignments of all methods (Figure 1) and that no method reached the accuracy of refined CDD alignments (about 95% in Table 2). These observations imply that RSE could not correct certain errors of the input alignments. This could happen because some needed seed alignments could not be found from a poor initial superimposed structures and/or because of the constraints imposed by the inflexible, rigid body superposition of structures.
We devised a refinement procedure for structure-based sequence alignments, called RSE. It uses the SE algorithm, which produces a sequence alignment without using a gap penalty. When applied to the structure-based sequence alignments generated by various structure comparison/alignment programs, the average accuracy increased for all programs tested. This refinement procedure is fast enough to be routinely used as a supplemental procedure following a regular structure-based sequence alignment or to replace the traditional dynamic programming algorithm-based refinement procedure which is a part of many structural alignment programs.
The RSE procedure
Find seeds and seed segments.
Find aligned segments by extending seed segments.
Find the consistent set of aligned segments with the best score.
Discard all other aligned segments.
Extend the surviving aligned segments after discarding the inconsistent aligned segments.
Change tied seeds to extended pairs if they do not overlap with already aligned residue pairs.
Repeat steps 3 to 5.
The reason for introducing step 5 is that there may be room for extension after removal of inconsistent segments. The additional steps 6 and 7 were used only in the last refinement cycle (see below) to pick up isolated pairs of alignable residue pairs.
For RSE, the sequence alignment by SE without steps 6 and 7 was followed by a rigid body superposition routine KABSCH [27, 28]. This two-step process was repeated for up to 10 times until the alignment converged (until the last two alignments were the same). In the rigid-body superposition step, each aligned residue pair was weighted according to the distance d ij between Cα atoms of the aligned residues: , where d0 and n are adjustable parameters with default values of 3.0 Å and 2, respectively. Several combinations of d0 (= 2.5 to 4.0 Å in 0.5 Å steps) and n (= 1 to 4) were tested, but the RSE procedure was rather insensitive to these parameters. During the iteration, the transformation matrix of the superposition that generated the best alignment, in terms of the number of aligned residues, was selected. The final sequence alignment was produced by an additional round of SE that included steps 6 and 7 after two structures were superimposed according to the chosen transformation matrix.
The RSE procedure accepts as input two superimposed structures or two independent structures with a sequence alignment, in which case a superposition is obtained through KABSCH procedure with unit residue weights. In this work, the RSE was run in the latter mode, since some structure alignment programs did not generate superimposed structures. Different programs produced sequence alignments in different formats, which had to be converted into a standard format (the FASTA format). The iterative refinement steps can be skipped by giving -norefine command line option, in which case the input superposition is used directly to generate the sequence alignment output. The program is downloadable from the following web site: http://lmbbi.nci.nih.gov/.
Reference alignments, structure alignment programs and time measure
We used the CDD (v.2.07) "root node set" introduced in our previous work  as the reference sequence alignments with corresponding SCOP domains. We chose this dataset because it is manually procured and because it includes many sequences that are sufficiently dissimilar that structure is needed for their accurate alignment. The 5 'outlier' superfamilies (cd00651, cd01345, cd02156, cd01284, and cd02688) were excluded, for which the CDD alignments were judged questionable as reference alignments . The composition of the dataset is described in Table 1.
We included CE (Algorithm 1.0, Alignment calculator 1.02) , DaliLite_2.4.1 , LOCK2 , FAST , MATRAS (version 1.2) , MATT , SHEBA-4.0 , TM-align  and VAST (directly from Dr. Gibrat) . We also included SSEARCH from FASTA3 package for pure sequence alignment  and SALIGN from Modeller (mod9v6) for profile-profile alignment . The input multiple alignments for SALIGN were prepared from PSI-BLAST alignments (BLASTPGP  in blast-2.2.20 package), allowing up to 20 iterations with e-value cutoff of 0.0005 against nr database (as of 04/19/2009). Up to 1,000 sequences with most significant e-values were retained in the multiple sequence alignment. The parameter settings for PSI-BLAST were as described in Marti-Renom et.al. . Otherwise, default values were used for all the programs.
In order to measure the execution times for the methods including the RSE procedure, time-stamps were recorded before and after system calls for the executables. For the CPU times per refinement cycle with CDD alignments, the elapsed time from after the initial structure superposition to the end of refinement cycles, which did not include the file I/O time, was divided by the number of refinement cycles. The CPU times for each alignment were averaged over three independent runs.
This research was supported by the Intramural Research Program of the NIH, National Cancer Institute, Center for Cancer Research.
- Yang AS, Honig B: An integrated approach to the analysis and modeling of protein sequences and structures. III. A comparative study of sequence conservation in protein structural families using multiple structural alignments. J Mol Biol 2000, 301(3):691–711. 10.1006/jmbi.2000.3975View ArticlePubMedGoogle Scholar
- Eidhammer I, Jonassen I, Taylor WR: Structure comparison and structure patterns. J Comput Biol 2000, 7(5):685–716. 10.1089/106652701446152View ArticlePubMedGoogle Scholar
- Chakrabarti S, Lanczycki CJ, Panchenko AR, Przytycka TM, Thiessen PA, Bryant SH: State of the art: refinement of multiple sequence alignments. BMC Bioinformatics 2006, 7: 499. 10.1186/1471-2105-7-499PubMed CentralView ArticlePubMedGoogle Scholar
- Lassmann T, Sonnhammer EL: Automatic assessment of alignment quality. Nucleic Acids Res 2005, 33(22):7120–7128. 10.1093/nar/gki1020PubMed 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
- Gelly JC, Chiche L, Gracy J: EvDTree: structure-dependent substitution profiles based on decision tree classification of 3D environments. BMC Bioinformatics 2005, 6: 4. 10.1186/1471-2105-6-4PubMed CentralView ArticlePubMedGoogle Scholar
- Goonesekere NC, Lee B: Context-specific amino acid substitution matrices and their use in the detection of protein homologs. Proteins 2008, 71(2):910–919. 10.1002/prot.21775View ArticlePubMedGoogle Scholar
- Yang AS: Structure-dependent sequence alignment for remotely related proteins. Bioinformatics 2002, 18(12):1658–1665. 10.1093/bioinformatics/18.12.1658View ArticlePubMedGoogle Scholar
- Bennett-Lovsey RM, Herbert AD, Sternberg MJ, Kelley LA: Exploring the extremes of sequence/structure space with ensemble fold recognition in the program Phyre. Proteins 2008, 70(3):611–625. 10.1002/prot.21688View ArticlePubMedGoogle Scholar
- Godzik A: The structural alignment between two proteins: is there a unique answer? Protein Sci 1996, 5(7):1325–1338. 10.1002/pro.5560050711PubMed CentralView ArticlePubMedGoogle Scholar
- Mayr G, Domingues FS, Lackner P: Comparative analysis of protein structure alignments. BMC Struct Biol 2007, 7: 50. 10.1186/1472-6807-7-50PubMed CentralView ArticlePubMedGoogle Scholar
- Kim C, Lee B: Accuracy of structure-based sequence alignment of automatic methods. BMC Bioinformatics 2007, 8: 355. 10.1186/1471-2105-8-355PubMed CentralView ArticlePubMedGoogle Scholar
- Holm L, Park J: DaliLite workbench for protein structure comparison. Bioinformatics 2000, 16(6):566–567. 10.1093/bioinformatics/16.6.566View ArticlePubMedGoogle Scholar
- Madej T, Gibrat JF, Bryant SH: Threading a database of protein cores. Proteins 1995, 23(3):356–369. 10.1002/prot.340230309View ArticlePubMedGoogle Scholar
- Ye Y, Godzik A: FATCAT: a web server for flexible structure comparison and structure similarity searching. Nucleic Acids Res 2004, (32 Web Server):W582–585. 10.1093/nar/gkh430Google Scholar
- Menke M, Berger B, Cowen L: Matt: local flexibility aids protein multiple structure alignment. PLoS Comput Biol 2008, 4(1):e10. 10.1371/journal.pcbi.0040010PubMed CentralView ArticlePubMedGoogle Scholar
- Zhu J, Weng Z: FAST: a novel protein structure alignment algorithm. Proteins 2005, 58(3):618–627. 10.1002/prot.20331View ArticlePubMedGoogle Scholar
- Shapiro J, Brutlag D: FoldMiner and LOCK 2: protein structure comparison and motif discovery on the web. Nucleic Acids Res 2004, (32 Web Server):W536–541. 10.1093/nar/gkh389Google Scholar
- Kawabata T: MATRAS: A program for protein 3D structure comparison. Nucleic Acids Res 2003, 31(13):3367–3369. 10.1093/nar/gkg581PubMed CentralView ArticlePubMedGoogle Scholar
- Zhang Y, Skolnick J: TM-align: a protein structure alignment algorithm based on the TM-score. Nucleic Acids Res 2005, 33(7):2302–2309. 10.1093/nar/gki524PubMed CentralView ArticlePubMedGoogle Scholar
- Tai CH, Vincent JJ, Kim C, Lee B: SE: An algorithm for deriving sequence alignment from a pair of superimposed structures. BMC Bioinformatics 2009, 10(Suppl 1):S4. 10.1186/1471-2105-10-S1-S4PubMed CentralView ArticlePubMedGoogle Scholar
- Pettersen EF, Goddard TD, Huang CC, Couch GS, Greenblatt DM, Meng EC, Ferrin TE: UCSF Chimera – a visualization system for exploratory research and analysis. J Comput Chem 2004, 25(13):1605–1612. 10.1002/jcc.20084View ArticlePubMedGoogle Scholar
- Yang AS, Honig B: An integrated approach to the analysis and modeling of protein sequences and structures. I. Protein structural alignment and a quantitative measure for protein structural distance. J Mol Biol 2000, 301(3):665–678. 10.1006/jmbi.2000.3973View ArticlePubMedGoogle Scholar
- Kleywegt GJ, Jones TA: A super position. CCP4/ESF-EACBM Newsletter on Protein Crystallography 1994, 31: 9.Google Scholar
- Russell RB, Barton GJ: Multiple protein sequence alignment from tertiary structure comparison: assignment of global and residue confidence levels. Proteins 1992, 14(2):309–323. 10.1002/prot.340140216View ArticlePubMedGoogle Scholar
- Marchler-Bauer A, Anderson JB, Derbyshire MK, DeWeese-Scott C, Gonzales NR, Gwadz M, Hao L, He S, Hurwitz DI, Jackson JD, et al.: CDD: a conserved domain database for interactive domain family analysis. Nucleic Acids Res 2007, (35 Database):D237–240. 10.1093/nar/gkl951Google Scholar
- Kabsch W: A solution for the best rotation to relate two sets of vectors. Acta Crystallographica Section A 1976, 32: 922–923. 10.1107/S0567739476001873View ArticleGoogle Scholar
- Kabsch W: A discussion of the solution for the best rotation to relate two sets of vectors. Acta Crystallographica Section A 1978, 34: 827–828. 10.1107/S0567739478001680View ArticleGoogle Scholar
- Gerstein M, Levitt M: Comprehensive assessment of automatic structural alignment against a manual standard, the scop classification of proteins. Protein Sci 1998, 7(2):445–456.PubMed CentralView ArticlePubMedGoogle Scholar
- Shindyalov IN, Bourne PE: Protein structure alignment by incremental combinatorial extension (CE) of the optimal path. Protein Eng 1998, 11(9):739–747. 10.1093/protein/11.9.739View ArticlePubMedGoogle Scholar
- Jung J, Lee B: Protein structure alignment using environmental profiles. Protein Eng 2000, 13(8):535–543. 10.1093/protein/13.8.535View ArticlePubMedGoogle Scholar
- Pearson WR: Searching protein sequence libraries: comparison of the sensitivity and selectivity of the Smith-Waterman and FASTA algorithms. Genomics 1991, 11(3):635–650. 10.1016/0888-7543(91)90071-LView ArticlePubMedGoogle Scholar
- Marti-Renom MA, Madhusudhan MS, Sali A: Alignment of protein sequences by their profiles. Protein Sci 2004, 13(4):1071–1087. 10.1110/ps.03379804PubMed CentralView 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
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.