 Research article
 Open access
 Published:
Testing statistical significance scores of sequence comparison methods with structure similarity
BMC Bioinformatics volumeÂ 7, ArticleÂ number:Â 444 (2006)
Abstract
Background
In the past years the SmithWaterman sequence comparison algorithm has gained popularity due to improved implementations and rapidly increasing computing power. However, the quality and sensitivity of a database search is not only determined by the algorithm but also by the statistical significance testing for an alignment. The evalue is the most commonly used statistical validation method for sequence database searching. The CluSTr database and the Protein World database have been created using an alternative statistical significance test: a Zscore based on MonteCarlo statistics. Several papers have described the superiority of the Zscore as compared to the evalue, using simulated data. We were interested if this could be validated when applied to existing, evolutionary related protein sequences.
Results
All experiments are performed on the ASTRAL SCOP database. The SmithWaterman sequence comparison algorithm with both evalue and Zscore statistics is evaluated, using ROC, CVE and AP measures. The BLAST and FASTA algorithms are used as reference. We find that two out of three SmithWaterman implementations with evalue are better at predicting structural similarities between proteins than the SmithWaterman implementation with Zscore. SSEARCH especially has very high scores.
Conclusion
The compute intensive Zscore does not have a clear advantage over the evalue. The SmithWaterman implementations give generally better results than their heuristic counterparts. We recommend using the SSEARCH algorithm combined with evalues for pairwise sequence comparisons.
Background
Sequence comparison is still one of the most important methodologies in the field of computational biology. It enables researchers to compare the sequences of genes or proteins with unknown functions to sequences of wellstudied genes or proteins. However, due to a significant increase in whole genome sequencing projects, the amount of sequence data is nowadays very large and rapidly increasing. Therefore, pairwise comparison algorithms should not only be accurate and reliable but also fast. The SmithWaterman algorithm [1] is one of the most advanced and sensitive pairwise sequence comparison algorithms currently available. However, it is theoretically about 50 times slower than other popular algorithms [2], such as FASTA [3] and BLAST [4]. All three algorithms generate local alignments, but the SmithWaterman algorithm puts no constraints on the alignment it reports other than that it has a positive score in terms of the similarity table used to score the alignment. BLAST and FASTA put additional constraints on the alignments that they report in order to speed up their operation: only sequences above a certain similarity threshold are reported, the rest is used for the estimation of certain parameters used in the alignment calculation. Because of this SmithWaterman is more sensitive than BLAST and FASTA. The SmithWaterman algorithm finds the best matching regions in the same pair of sequences. However, BLAST and FASTA are still far more popular because of their speed and the addition of a statistical significance value, the Expectvalue (or simply evalue), whereas the original SmithWaterman implementation relies only on the SWscore without any further statistics. The newer SmithWaterman implementations of Paracel [5], SSEARCH [6] and ParAlign [7] do include the evalue as a measure of statistical significance, which makes the SmithWaterman algorithm more usable as the engine behind a similarity search tool. The evalue is far more useful than the SWscore, because it describes the number of hits one can expect to see by chance when searching a database of a certain size. An evalue threshold can be used easily to separate the 'interesting' results from the background noise. However, a more reliable statistical estimate is still needed [8]. The Zscore, based on MonteCarlo statistics, was introduced by Doolittle [9] and implemented by GeneIT [10] in its sequence comparison suite Biofacet [11]. The Zscore has been used in the creation of the sequence annotation databases CluSTr [12] and Protein World [13] and was used in orthology studies [14]. The Zscore has also been implemented in algorithms other than SmithWaterman, such as FASTA [15]. It is calculated by performing a number (e.g., 100) of shuffling randomizations of both sequences that are compared, completed by an estimation of the SW score significance as compared to the original pairwise alignment. This makes the Zscore very useful for doing allagainstall pairwise sequence comparisons: Zscores of different sequence pairs can be compared to each other, because they are only dependent on the sequences itself and not on the database size, which is one of the parameters used to calculate the evalue. However, this independency of the database size makes the Zscore unsuitable for determining the probability that an alignment has been obtained by chance. The randomizations make the Zscore calculation quite slow, but theoretically it is more sensitive and more selective than evalue statistics [16, 17]. Unfortunately, this has never been validated experimentally.
Some methods have been used to combine the sensitivity and selectivity of a sequence comparison algorithm into one single score [18]. Receiver operating characteristic (ROC) is a popular measure of search accuracy [19]. For a perfect search algorithm, all true positives for these queries should appear before any false positive in the ranked output list, which gives an ROC score of 1. If the first n items in the list are all false positives, the ROC_{n} score is 0. Although researchers have devised many ways to merge ROC scores for a set of queries [20], one simple and popular method is to 'pool' search results so as to get an overall ROC score [21]. Another method to evaluate different methods is the errors per query (EPQ) criterion and the 'coverage versus error' plots [2]. EPQ is a selectivity indicator based on allagainstall comparisons, and coverage is a sensitivity measure. The assumption for EPQ is that the search algorithm can yield a 'normalized similarity score' rather than a lengthdependent one, so that results from queries are comparable. Like ROC, the coverage versus error plot can give an overall performance comparison for search algorithms. A third method, the average precision (AP) criterion, is adopted from information retrieval research [22]. The method defines two values: the recall (true positives divided by the number of homologs) and the precision (true positives divided by the number of hits), which are plotted in a graph. The AP then is an approximate integral to calculate the area under this recallprecision curve. These methods were used to compare several sequence comparison algorithms, but we use them to compare the evalue and Zscore statistics. Analyses of BLAST and FASTA are also included as reference material.
Here we show that two out of the three SmithWaterman implementations with evalue statistics are more accurate than the SmithWaterman implementation of Biofacet with Zscore statistics. Furthermore, the comparison of BLAST and FASTA with the four SmithWaterman implementations shows that FASTA is a more reliable algorithm when using the ASTRAL SCOP structural classification as a benchmark. The SmithWaterman implementation of Paracel even has lower scores than both BLAST and FASTA. SSEARCH, the SmithWaterman implementation in the FASTA package, scores best.
Results
We used a nonredundant proteindomain sequence database derived from PDB as the target database. It is automatically generated using the ASTRAL system [23]. According to the structural classification of proteins (SCOP release 1.65), it includes 9498 sequences and 2326 families. True positives are those in the same family as the query sequence. SCOP as an independent and accurate source for evaluating database search methods has been used by other researchers [2, 24]. ASTRAL SCOP sets with different maximal percentage identity thresholds (10%, 20%, 25%, 30%, 35%, 40%, 50%, 70%, 90% and 95%) were downloaded from the ASTRAL SCOP website [25]. Their properties (number of families, number of members, etc.) are shown in table 1. Three different statistical measures were applied: receiver operating characteristic (ROC), coverage versus error (CVE) and mean average precision (AP). We compared six different pairwise sequence comparison algorithms, which are listed in table 2, together with the parameters used in this study.
Receiver operating characteristic
The mean ROC_{50} scores increase if more structurally identical proteins are included, for both the evalue and the Zscore measurements (Fig. 1). The ROC_{50} scores of the PDB010 set show a large difference between the several SmithWaterman implementations: 0.19 for Paracel, 0.23 for Biofacet (with Zscore), 0.27 for ParAlign and 0.31 for SSEARCH. The advantage of ParAlign over Biofacet decreases with increasing inclusiveness of the ASTRAL SCOP set that is used. The ROC_{50} scores of the PDB095 set are 0.28 for Paracel, 0.35 for both ParAlign and Biofacet (with Zscore) and 0.46 for SSEARCH. SSEARCH scores best of all studied methods, regardless of which ASTRAL SCOP set is used. The reference methods FASTA and BLAST give quite different results: FASTA is a good second and BLAST has scores similar to Paracel and Biofacet.
Coverage versus error
This method differs from the ROC analysis on one crucial point: instead of looking at the first 100 hits, we varied the threshold at which a hit was seen as a positive. Hence the results are somewhat dissimilar: the differences between the several algorithms in the coverage versus error plots (Fig. 2) are not as obvious as they are in the ROC_{50} graph (Fig. 1). Figure 2A shows the coverage versus error plot for the smallest ASTRAL SCOP set (PDB010), figure 2B shows the plot for the largest ASTRAL SCOP set (PDB095) and figure 2C shows the plot for the intermediate set PDB035. An ideal algorithm would have a very high coverage but not many errors per query, which places it in the lower right corner of the graph. SSEARCH has the best scores when using the PDB010 set, followed by ParAlign and FASTA, with the latter scoring best in the lowestcoverage range (<0.02). Biofacet with Zscore has the lowest scores. The PDB095 plot shows some differences between the lowcoverage range (<0.25) and the highcoverage range (>0.50). In the low coverage range, FASTA and Paracel have the highest scores, whereas SSEARCH and ParAlign have the highest scores in the lowcoverage range. It should be noted that the highcoverage range might statistically be more reliable because of the larger number of hits. The PDB035 set gives similar results.
Average precision
The average precision graph (Fig. 3) shows some minor differences from the ROC_{50} graph (Fig. 1): for the PDB020, PDB025 and PDB030 set, Paracel (evalue) scores better than Biofacet (Zscore). However, the advantage of the Biofacet SmithWaterman with Zscore increases from that point on (PDB035, Paracel: 0.16, Biofacet: 0.17) to the right side (PDB095, Paracel: 0.19, Biofacet: 0.24). The Zscore seems to score better when more similar proteins are compared. Once more, SSEARCH has the highest scores for all structural identity percentages, with FASTA as the second best.
Case studies
We included two examples of our statistical analysis, which show how the ROC and mean AP measures differ from each other and how results can be different for each studied protein. We choose two wellstudied proteins: enoylACP reductase and the progesterone receptor, the first from a prokaryote (E. coli) and the second from a eukaryote (H. sapiens). Both case studies were done using the PDB095 set, which is the most complete ASTRAL SCOP PDB set used in our study.
Bacterial enoylACP reductase
Table 3 shows the results of our analysis of the ASTRAL SCOP entry of E. coli enoylACP reductase chain A, d1qg6a_, using the PDB095 set. One way of testing the reliability of a sequence comparison method is by looking at the first false positive (FFP) in the list of top 100 hits (Table S.1 [see Additional file 1]). The c.2.1.2 structural family has 46 members within the PDB095 set, so the perfect sequence comparison algorithm would return its first false positive at the 46^{th} hit (the hit containing the query protein is discarded). For the Paracel SmithWaterman implementation, this is already the twentyfirst hit. Four algorithms score best with the first false positive at 24^{th} place. A second testing method is counting the total number of true positives (NTP), of which the perfect algorithm would return all 45. BLAST has the highest score here: 27 out of the top 100 hits are true positives. FASTA and Paracel are at the second place with 25 true positives. Biofacet has the lowest score: only 23 true positives. Note that differences are very small, which is a reason to look at the ROC and mean AP scores. FASTA and SSEARCH have both the highest ROC_{50} scores and the highest mean APs. ParAlign and BLAST are third and fourth, followed by Paracel and Biofacet. The ROC and mean AP scores give a clearer view of the differences between the algorithms than the FFP or NTP scores, because they take into account the ranking of all hits instead of just the first false positive or just the true positives.
Human progesterone receptor
Table 4 shows our analysis of ASTRAL SCOP entry d1a28a_, using again the PDB095 set. The structural family a.123.1.1 has 29 members, so the perfect algorithm should have the first false positive at the 29^{th} hit. Surprisingly, BLAST scores best here with its first false positive at the 25^{th} hit (Table S.2 [see Additional file 1]), although the differences are quite small. BLAST is, together with Biofacet, the only algorithm that does not have all the 28 family members of d1a28a_ in its top 100 list; d1n83a_ is missing here. The ROC_{50} and mean AP analysis of d1a28a_ shows again that SSEARCH and FASTA give the best results. Paracel and Biofacet have the lowest scores once more. The differences are not large enough to put any definite conclusions to the results of this example, but by combining all ROC and mean AP scores for all ASTRAL SCOP entries, we created a reliable comparison between all sequence comparison methods.
Timing
Table 5 shows the time that each of the six algorithms needs to perform an allagainstall sequence comparison of the ASTRAL SCOP PDB095 set. The BLAST algorithm is clearly the fastest, followed by the other heuristic algorithm FASTA. Of the SmithWaterman algorithms, ParAlign is by far the fastest. The Biofacet algorithm needs much time to calculate 2 Ã— 100 randomizations and is therefore the slowest sequence comparison algorithm.
Discussion
The theoretical advantage of the Zscore over the evalue appears to be rejected by our results. Our results show that the evalue calculation gives an advantage over the computationally intensive Zscore, at least when looking only at the results from the SmithWaterman algorithm. Some caution should be taken however, drawing any definite conclusions. First, the Zscore was designed to make a distinction between significant hits and nonsignificant hits that have high SW scores. It might have an advantage over the evalue when applied to the top hits only, but might have less advantage for the hits with lower SW scores. This idea is supported by the fact that the Zscore is better at scoring highsimilarity sequence pairs. This is also reflected in the different ROC and AP scores for the PDB010 set and the PDB095 set: the difference between Zscore and evalue increases when structurally more similar protein pairs are being included. Second, the Zscore can differ for each run, because of its different randomizations [17]. The standard deviation of the Zscore increases almost proportionally with the Zscore itself, i.e. for higher Zscores the variance will be larger [16]. However, the Zscore increases its precision when more randomizations are calculated (2 Ã— 100 in this study). Third, the PDB set is somewhat biased: it only contains crystallized proteins, and it contains no hypothetical proteins and membrane proteins. The crystallized proteins in the PDB are on average smaller than proteins included in large sequence databases such as the UniProt [26] database (Figure 4), whereas the amino acid distribution is approximately the same for these databases (Figure 5).
Figure 6 shows that the bias in sequence length is not the reason for the difference in scores: if we only look at proteins with a sequence length of 500 or more, the scores are similar. Other studies have shown that FASTA performs better than BLAST [18, 27], but these did not include several SmithWaterman implementations. The SSEARCH algorithm, an implementation of SmithWaterman, was analyzed in these studies, but this algorithm differs from other SmithWaterman algorithms used in this study due to the use of length regression statistics [7, 28]. A difference can also be found by comparing the SW scores of Biofacet, ParAlign and SSEARCH: Biofacet and ParAlign have the same SW scores, but the SSEARCH SW scores are different. We calculated the ROC_{50} and mean AP for these three SW scores and found that the SSEARCH SW scores gives slightly worse results than the other two SW scores (Figure 7). Another problem is that protein sequences within a certain ASTRAL SCOP family usually have equivalent lengths, since the ASTRAL SCOP database consists of protein domains and not of whole proteins. Results might vary when whole proteins, with different lengths, are studied. Unfortunately, the composition of the ASTRAL SCOP database does not allow us to confirm this statement.
Finally, we would like to stress that the results from the CVE analysis might be more reliable than those from the ROC and mean AP analyses. ROC and mean AP make use of a ranking system based on the evalue or Zscore, instead of looking at the evalue or Zscore directly. This means that in some cases, especially the smaller protein families, a large number of very lowscoring hits (e.g. e>100 or Z<3) is still used for the calculation of the scores. This is not the case for the CVE plots, because we varied the evalue and Zscore thresholds above which a hit is seen as a true positive, instead of relying on a ranking system. However, because the results from the CVE plots are similar to the results from the ROC and mean AP graphs, the use of a ranking system does not seem to give a large disadvantage.
Conclusion
For a complete analysis we need a less biased database, having a wide range of proteins classified by structure similarity. Until such a database is available, it will be difficult to pinpoint the reasons for the different results between FASTA, BLAST and SmithWaterman, and the theoretical advantages of the Zscore. Regardless of all these theoretical assumptions, the computational disadvantage of the Zscore is smaller for larger databases. Zscores do not have to be recalculated when sequences are added to the database, in contrast to evalues, which are dependent on database size. For very large databases containing allagainstall comparisons, this is an important advantage of the Zscore. Although recalculating the evalues does not take much time when the alignments and SW scores are already available, this may cause a change in research results that were obtained earlier. Despite these considerations, we recommend using SSEARCH with evalue statistics for pairwise sequence comparisons.
Methods
Sequence comparisons
For the SmithWaterman evalue calculation, the ASTRAL SCOP files were loaded onto the Paracel file system as protein databases and subsequently used as queries against these databases: the set with 10% maximal identity (PDB010) against itself, the set with 20% maximal identity (PDB020) against itself, etc. The matrix used for all sequence comparisons was the BLOSUM62 matrix [29]. This is the default scoring matrix for most alignment programs. For all sequence comparisons in this article, the gap open penalty was set to 12 and the gap extension penalty was set to 1. These are the averages of the default penalties over the six studied methods. Both the matrix and gap penalties used are suited for comparing protein sets with a broad spectrum of evolutionary distances, like the PDB set [30, 31]. Per query sequence, the best 100 hits were kept [see section Data availability], discarding the match of each query sequence with itself.
Receiver operating characteristic calculation
For each query, the 100 best hits were marked as true positives or false positives, i.e. the hit being in the same or in a different SCOP family than the query. For each of the first 50 false positives that were found, the number of true positives with a higher similarity score was calculated. The sum of all of these numbers was then divided by the number of false positives (50), and finally divided by the total number of possible true positives in the database (i.e. the total number of members in the SCOP family minus 1), giving an ROC_{50} score for each query sequence. The average of these ROC_{50} scores gives the final ROC score for that specific statistical value and that specific ASTRAL SCOP set. Mean ROC_{50} scores were calculated for all ten different ASTRAL SCOP sets.
Coverage versus error calculation
Instead of taking the first 100 hits for each query, like in the ROC analysis, we varied the threshold at which a certain hit was seen as a positive. For the evalue analysis, we created a list of 49 thresholds in the range of 10^{50} to 100. For Zscore, we created a list of 58 thresholds in the range of 0 to 100. Then, for each threshold, two parameters were measured: the coverage and the errors per query (EPQ). The coverage is the number of true hits divided by the total number of sequence pairs that are in the same SCOP family, for that specific ASTRAL SCOP set. The EPQ is the number of false hits divided by the number of queries. We used the most inclusive ASTRAL SCOP set (PDB095), the least inclusive set (PDB010) and an intermediate set (PDB035) to create the coverage versus error plots.
Average precision calculation
For the calculation of the average precision (AP), the 100 best hits per query were marked again as either true positives or false positives. Subsequently for each true positive found by the search algorithm, the true positive rank of this hit (i.e. the number of true positives with a higher score + 1) was divided by the positive rank (i.e. the number of hits with a higher score + 1). These numbers were all added up and then divided by the total number of hits (i.e. 100), giving one AP value per query. The mean AP is the average of all these APs. Mean APs were calculated for all ten different ASTRAL SCOP sets.
Bacterial enoylACP reductase
The ASTRAL SCOP entry for E. coli enoylACP reductase chain A, d1qg6a_, was picked as an example for our methodology. The 100 best hits of this entry on the PDB095 set were calculated using each of the six algorithms and sorted by ascending evalue and descending Zscore. Then they were marked as either true positives or false positives, depending on if the hit was in the same structural family (c.2.1.2) or not. Furthermore, the ROC_{50} scores and mean APs were calculated.
Human progesterone receptor
A second example is the analysis of d1a28a_, the H. sapiens progesterone receptor chain A. Once more, the 100 best hits of this entry on the PDB095 set were calculated using each of the six algorithms and sorted by ascending evalue and descending Zscore. These hits were marked as either true positives or false positives, depending on if the hit was in the same structural family (a.123.1.1) or not. Finally, the mean AP and ROC_{50} scores were calculated.
Timing
We measured the speed of the sequence comparison algorithms, by doing an allagainstall comparison of the ASTRAL SCOP PDB095 set and using the 'time' command provided by UNIX. All calculations were performed on the same machine, except for the Paracel calculation which could only be performed on the Paracel machine. The Paracel calculation time had to be estimated because of the unaivailability of the Paracel machine at the time of performing this analysis.
Data availability
All raw sequence comparison output files (containing the top 100 hits per query sequence) are available through our website [32]. The top 100 hits for the two case studies of the bacterial enoylACP reductase (i.e. Table S.1) and the human progesterone receptor (i.e. Table S.2) can be found in the additional files section [see Additional file 1].
Abbreviations
 AP:

Average Precision
 bf z:

Biofacet (Zscore)
 BLAST:

Basic Local Alignment Search Tool
 bl e:

BLAST (evalue)
 BLOSUM:

BLOcks SUbstitution Matrix
 CluSTr:

Clusters of SWISSPROT and TrEMBL
 CVE:

Coverage Versus Error
 EPQ:

Errors Per Query
 fa e:

FASTA (evalue)
 FFP:

First False Positive
 NTP:

Number of True Positives
 pa e:

ParAlign (evalue)
 pc e:

Paracel (evalue)
 PDB:

Protein Data Bank
 ROC:

Receiver Operating Characteristic
 SCOP:

Structural Classification Of Proteins
 ss e:

SSEARCH (evalue)
 SW:

SmithWaterman
References
Smith TF, Waterman MS: Identification of common molecular subsequences. J Mol Biol 1981, 147(1):195â€“197. 10.1016/00222836(81)900875
Brenner SE, Chothia C, Hubbard TJ: Assessing sequence comparison methods with reliable structurally identified distant evolutionary relationships. Proc Natl Acad Sci U S A 1998, 95(11):6073â€“6078. 10.1073/pnas.95.11.6073
Pearson WR, Lipman DJ: Improved tools for biological sequence comparison. Proc Natl Acad Sci U S A 1988, 85(8):2444â€“2448. 10.1073/pnas.85.8.2444
Altschul SF, Madden TL, Schaffer AA, Zhang J, Zhang Z, Miller W, Lipman DJ: Gapped BLAST and PSIBLAST: a new generation of protein database search programs. Nucleic Acids Res 1997, 25(17):3389â€“3402. 10.1093/nar/25.17.3389
Paracel[http://www.paracel.com]
Pearson WR: Searching protein sequence libraries: comparison of the sensitivity and selectivity of the SmithWaterman and FASTA algorithms. Genomics 1991, 11(3):635â€“650. 10.1016/08887543(91)90071L
Rognes T: ParAlign: a parallel sequence alignment algorithm for rapid and sensitive database searches. Nucleic Acids Res 2001, 29(7):1647â€“1652. 10.1093/nar/29.7.1647
Pearson WR, Sierk ML: The limits of protein sequence comparison? Curr Opin Struct Biol 2005.
Doolittle RF: Of URFs and ORFs: a primer on how to analyze derived amino acid sequences. Mill Valley California, University Science Books; 1986.
GeneIT[http://www.geneit.com]
Codani JJ, Comet JP, Aude JC, GlÃ©met E, Wozniak A, Risler JL, HÃ©naut A, Slonimski PP: Automatic Analysis of LargeScale Pairwise Alignments of Protein Sequences. Methods in Microbiology 1999, 28: 229â€“244.
Kriventseva EV, Servant F, Apweiler R: Improvements to CluSTr: the database of SWISSPROT+TrEMBL protein clusters. Nucleic Acids Res 2003, 31(1):388â€“389. 10.1093/nar/gkg035
Protein World[http://www.cmbi.ru.nl/pw/]
Hulsen T, Huynen MA, de Vlieg J, Groenen PM: Benchmarking ortholog identification methods using functional genomics data. Genome Biol 2006, 7(4):R31. 10.1186/gb200674r31
Booth HS, Maindonald JH, Wilson SR, Gready JE: An efficient Zscore algorithm for assessing sequence alignments. J Comput Biol 2004, 11(4):616â€“625. 10.1089/cmb.2004.11.616
Comet JP, Aude JC, Glemet E, Risler JL, Henaut A, Slonimski PP, Codani JJ: Significance of Zvalue statistics of SmithWaterman scores for protein alignments. Comput Chem 1999, 23(3â€“4):317â€“331. 10.1016/S00978485(99)00008X
Bastien O, Aude JC, Roy S, Marechal E: Fundamentals of massive automatic pairwise alignments of protein sequences: theoretical significance of Zvalue statistics. Bioinformatics 2004, 20(4):534â€“537. 10.1093/bioinformatics/btg440
Chen Z: Assessing sequence comparison methods with the average precision criterion. Bioinformatics 2003, 19(18):2456â€“2460. 10.1093/bioinformatics/btg349
Gribskov M, Robinson NL: Use of receiver operating characteristic (ROC) analysis to evaluate sequence matching. Compu Chem 1996, 20: 25â€“33. 10.1016/S00978485(96)800040
Kester AD, Buntinx F: Metaanalysis of ROC curves. Med Decis Making 2000, 20(4):430â€“439.
Schaffer AA, Aravind L, Madden TL, Shavirin S, Spouge JL, Wolf YI, Koonin EV, Altschul SF: Improving the accuracy of PSIBLAST protein database searches with compositionbased statistics and other refinements. Nucleic Acids Res 2001, 29(14):2994â€“3005. 10.1093/nar/29.14.2994
Salton G: Developments in automatic text retrieval. Science 1991, 253: 974â€“980. 10.1126/science.253.5023.974
Brenner SE, Koehl P, Levitt M: The ASTRAL compendium for protein structure and sequence analysis. Nucleic Acids Res 2000, 28(1):254â€“256. 10.1093/nar/28.1.254
Park J, Karplus K, Barrett C, Hughey R, Haussler D, Hubbard T, Chothia C: Sequence comparisons using multiple sequences detect three times as many remote homologues as pairwise methods. J Mol Biol 1998, 284(4):1201â€“1210. 10.1006/jmbi.1998.2221
ASTRAL SCOP release 1.65[http://astral.berkeley.edu/scopseq1.65.html]
Apweiler R, Bairoch A, Wu CH, Barker WC, Boeckmann B, Ferro S, Gasteiger E, Huang H, Lopez R, Magrane M, Martin MJ, Natale DA, O'Donovan C, Redaschi N, Yeh LS: UniProt: the Universal Protein knowledgebase. Nucleic Acids Res 2004, 32(Database issue):D115â€“9. 10.1093/nar/gkh131
Agarwal P, States DJ: Comparative accuracy of methods for protein sequence similarity search. Bioinformatics 1998, 14(1):40â€“47. 10.1093/bioinformatics/14.1.40
Pearson WR: Comparison of methods for searching protein sequence databases. Protein Sci 1995, 4(6):1145â€“1160.
Henikoff S, Henikoff JG: Amino acid substitution matrices from protein blocks. Proc Natl Acad Sci U S A 1992, 89(22):10915â€“10919. 10.1073/pnas.89.22.10915
Reese JT, Pearson WR: Empirical determination of effective gap penalties for sequence comparison. Bioinformatics 2002, 18(11):1500â€“1507. 10.1093/bioinformatics/18.11.1500
Price GA, Crooks GE, Green RE, Brenner SE: Statistical evaluation of pairwise protein sequence comparison with the Bayesian bootstrap. Bioinformatics 2005, 21(20):3824â€“3831. 10.1093/bioinformatics/bti627
Supplementary data[http://www.cmbi.ru.nl/~timhulse/ezcomp/]
Acknowledgements
This work was supported financially by NV Organon and the Netherlands Organization for Scientific Research (NWO). The authors like to thank Scott Lusher for critically reading this manuscript.
Author information
Authors and Affiliations
Corresponding author
Additional information
Authors' contributions
TH participated in the design of the study, carried out the calculations and statistical analysis and drafted the manuscript
JdV participated in the design of the study
JL gave some technical and scientific advice and helped to draft the manuscript
PG participated in the design and coordination of the study and helped to draft the manuscript
All authors read and approved the final manuscript
Electronic supplementary material
12859_2006_1183_MOESM1_ESM.doc
Additional File 1: Supplementary tables: Table S.1. Top 100 hits of bacterial enoylACP reductase. Table S.2. Top 100 hits of human progesterone receptor. (DOC 640 KB)
Authorsâ€™ original submitted files for images
Below are the links to the authorsâ€™ original submitted files for images.
Rights and permissions
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.
About this article
Cite this article
Hulsen, T., de Vlieg, J., Leunissen, J.A. et al. Testing statistical significance scores of sequence comparison methods with structure similarity. BMC Bioinformatics 7, 444 (2006). https://doi.org/10.1186/147121057444
Received:
Accepted:
Published:
DOI: https://doi.org/10.1186/147121057444