arXiv:q-bio/0603007v2 [q-bio.

QM] 20 Oct 2006

Compression ratios based on the Universal Similarity Metric still yield protein distances far from CATH distances
F. Rossell o J. Rocha J. Segura
Dept. of Mathematics and Computer Science, University of the Balearic Islands, 07122 Palma de Mallorca (Spain) E-mail: {jairo,cesc.rossello}@uib.es

August 2, 2011
Abstract Kolmogorov complexity has inspired several alignment-free distance measures, based on the comparison of lengths of compressions, which have been applied successfully in many areas. One of these measures, the socalled Universal Similarity Metric (USM), has been used by Krasnogor and Pelta to compare simple protein contact maps, showing that it yielded good clustering on four small datasets. We report an extensive test of this metric using a much larger and representative protein dataset: the domain dataset used by Sierk and Pearson to evaluate seven protein structure comparison methods and two protein sequence comparison methods. One result is that Krasnogor-Pelta method has less domain discriminant power than any one of the methods considered by Sierk and Pearson when using these simple contact maps. In another test, we found that the USM based distance has low agreement with the CATH tree structure for the same benchmark of Sierk and Pearson. In any case, its agreement is lower than the one of a standard sequential alignment method, SSEARCH. Finally, we manually found lots of small subsets of the database that are better clustered using SSEARCH than USM, to conrm that Krasnogor-Peltas conclusions were based on datasets that were too small.

Introduction

Krasnogor and Pelta (2004) represented a protein structure by means of the list of contacts (position pairs) and used a compress algorithm to approximate the Kolmogorov complexity on which the denition of the Universal Similarity Metric is based. To test the metric they applied it independently to four protein datasets comprising 122 proteins and clustered each set using an o-the-shelf hierarchical clustering software applied to the resulting inter-distance matrices. In each case they obtained an almost correct classication of the proteins. They inferred that the Universal Similarity Metric captures protein structure similarity. Since each aminoacid has its own sequential position in a protein, and the contact pairs are pairs of these positions, we were surprised that the method could nd similarities between two dierent proteins each with its own numbering. But the authors conclude the following in the discussion section of the above mentioned paper: This new metric can be used as a robust measure of similarity of domains where either, there is no enough modeling information, or there is no consensus on which aspects are to be modeled. We gave mathematical and experimental evidence that USM can be used to successfully assess protein structure similarity. The USM seems to be capable of capturing protein similarities that encompasses a variety of other, more heuristic, criteria on a fully automated way. It seems that the Universal Similarity Metric is so robust that even with a rough guess of parameters it is still possible to deliver good results. One disadvantage of using USM on its own is that, although it can dierentiate between protein families and sub-families and measure similarity based on a rigorous mathematical denition, it does not give indications of where these (di)similarities come from. Given these strong conclusions on the robustness of their method to dierentiate between protein sub-families, rst, we downloaded their algorithms and data-sets and conrmed exactly their results using this simple protein format. Then, we set o for verifying the results on a bigger set because if they are satisfactory we could abandon our search for sophisticated comparison methods. We present in this paper a bigger test on almost 2800 protein domains already used to assess the performance of the most important algorithms used nowadays for protein comparison, namely the benchmark prepared by Sierk and Pearson (2004). When we apply the same Sierk-Pearson test, a though one, that was applied to the other comparison algorithms, the method revels to be useless for helping to decide whether two proteins are similar or not; the result is that the Universal Similarity Metric using this contact map format and compression approximation performs quite worse than any method considered by Sierk and 2

Pearson, even the ones that only use as input the protein sequences. Therefore, the method as it stands does not encompasses other criteria for comparison as the authors claimed. Therefore, a global analysis of the method is needed to nd out under which conditions it is reliable, if any. One way to undertake this analysis is to compare the matrix generated by USM for the domains in the database used above with respect to the metrics associated to the CATH tree that represents the family relationships of the domains. We found the distance on the tree that best approximates the USM matrix. Then, we calculate the correlation between the USM matrix and the tree distance. A low correlation means that the USM matrix does not represent correctly the tree structure. As we explain in the Results Section below, we found higher correlations for the distance generated by SSEARCH (Smith and Waterman, 1981) than to for USM. In short, we have a global evidence that clusterings produced by USM are farther from the CATH tree than the ones produced by a sequential method that does not uses 3D information. We also conrm manually this result. We found several (more than ten) easily classiable datasets contained in the benchmark set where the proposed method does not classify correctly the domains. Since the Krasnagor-Pelta method only has three small real datasets as evidence to argue that it can be used to successfully assess protein structural similarity but we can nd several easy subsets in the benchmark where it performs poorly, the robustness of the method is open to doubt. We have found subsets of proteins in very dierent families that the method cannot classify correctly but SSEARCH can, even thought this one uses only sequence information. All data and programs used for the preparation of this paper are available upon request. See dmi.uib.es/people/jairo/USM/ for cluster examples. In the rest of the introduction, we review the use and notation of the Kolmogorov complexity on pattern comparison. As an alternative to general sequence and structure comparison methods based on alignments, several metrics have been proposed based on Kolmogorov complexity (Benedetto et al., 2002; Bennett et al., 1998; Keogh et al., 2004; Li et al., 2001, 2004; Varre et al., 1999). Roughly speaking, the conditional Kolmogorov complexity K(x|y) of two sequences x, y is the length of the shortest binary program P that computes x with input y (Kolmogorov, 1965). Thus, K(x|y) represents the minimal amount of information required to generate x by any eective computation when y is furnished as an input to the computation. The Kolmogorov complexity K(x) of a sequence x is dened as K(x|), where stands for the empty sequence. Given a string x, let x denote the shortest binary pro3

gram that produces x on an empty input; if there are more than one shortest program, we take as x the rst in alphabetic order. The Kolmogorov complexity K(x, y) of a pair of objects x, y is the length of the shortest binary program that produces x and y and a way to tell them apart. More formal denitions of all these concepts and their main properties can be found in the textbook by Li and Vitnyi (2nd ed. 1997). a The most outstanding metric based on Kolmogorov complexity is the Universal Similarity Metric proposed by Li et al. (2004), du (x, y) = max{K(x|y ), K(y|x)} . max{K(x), K(y)} (1)

These authors proved that this metric (actually, it only satises the axioms of metrics up to a certain additive precision) renes any other computable similarity metric, like for instance eective versions of Hamming distance, Euclidean distance, edit distances or alignment distances (Li et al., 2004, Thm. VI.2). This Universal Similarity Metric has been used successfully for instance to compute phylogenetic trees based on whole mitochondrial genomes (Li et al., 2004; Cilibrasi and Vitnyi, 2005), cluster SARS virus (Cilibrasi and Vitnyi, 2005), a a classify languages (Li et al., 2004), musical pieces (Cilibrasi et al., 2004; Cilibrasi and Vitnyi, a 2005; Li and Sleep, 2004), and images (Svangard and Nordin, 2004), detect plagiarism is student assignments (Chen et al., 2004), and cluster Russian literature (Cilibrasi and Vitnyi, 2005). But it has failed to compare TOPS diagrams a (Gilbert et al., 2006). Actually, the Universal Similarity Metric was not used in these applications as it stands, but approximations of it. The reason is that Kolmogorov complexities are non-computable in the Turing sense, and therefore they must be heuristically approximated in practice. Since K(x) is intuitively the minimal amount of information required to generate x, i.e., the shortest length of a compressed binary version of x, Kolmogorov complexities are approximated by means of lengths of compressions, and then the formula (1) given above is simplied using suitable properties of Kolmogorov complexity, so that it no longer involves conditional Kolmogorov complexities. In this way, and once xed a compression algorithm, the Kolmogorov complexity K(x) of an object x is replaced by the length C(x) of the compression of it using this algorithm. Furthermore, since according to Li et al. (2004) K(x, y) = K(xy) up to additive logarithmic precision,

K(x, y) can be replaced by the length C(xy) of a compression of the concatenation of x and y. Finally, and since K(x, y) = K(x) + K(y|x ) = K(y) + K(x|y ) up to constant additive precision (Li and Vitnyi, 2nd ed. 1997), the conditional a complexity K(x|y ) can be approximated by C(xy) C(y), and K(y|x) can be approximated by C(xy) C(x) or by C(yx) C(x). This lead Li et al. (2004) to approximate the Universal Similarity Metric by the Normalized Information Distance NCD(x, y) = C(xy) min{C(x), C(y)} ; max{C(x), C(y)}

this distance has been thoroughly studied by Cilibrasi and Vitnyi (2005). A a methodological study of its application to protein sequence classication has been published recently by Kocsor et al. (2006); they show that a compression based distance combined with a BLAST score has a performance even slightly better than that of the Smith-Waterman algorithm (SSEARCH). Krasnogor and Pelta (2004); Pelta et al. (2005) have used a slightly more general approximation of the Universal Similarity Metric to compare protein structures. The formula they use (and we use in this paper) is USM(x, y) = max{C(xy) C(x), C(yx) C(y)} . max{C(x), C(y)}

Methods

A subset of CATH 2.3 database (which stands for Class, Architecture, Topology, and Hierarchy (Orego et al., 1997)) was selected by Sierk and Pearson (2004) to obtain a non-redundant sample of the entire database. A 2771 subset of CATH domains and 86 prototype domains1 were selected to test the following algorithms: Dali, Structal, CE, VAST, Matras, SGM, PRIDE, SSEARCH and PS-Blast. This domain set was carefully screened in order to be considered as a valid benchmark for testing protein comparison algorithms. Of the 2771 domains, 1120 belong to the 86 families of the 86 prototypes. Therefore, when comparing each of the 2771 domains against each of the 86 prototypes, there is a maximum of 1120 correct hits. The prototypes are part of these 1120 domains.
1

Available at the FASTA repository ftp.virginia.edu:/pub/fasta/prot sci 04

For our experiments, the le representation is the one used by Krasnogor and Pelta: a protein is represented by a list of adjacent aminoacids, where two aminoacids are adjacent if the distance of the corresponding C atoms are below 6.5 , a threshold also used by them (we also tested other thresholds as we A explain below). Each line in the le contains two numbers with the rst one smaller than the second one. The le has a header that consists of two lines: in the rst one, the number of aminoacids is followed by the comment # Number of Residues; in the second one, the number of contacts is followed by the comment # Number of Contacts at 6.5 Angstroms . Therefore, two concatenated les could be trivially separated from each other. The Unix compress program was used. Other representations of protein structures (binary adjacency matrix with and without end of line characters) and other compression programs were tried but are not reported because the results obtained with them were similar or even slightly worse. We wanted to carry out two tests: The rst one is to test its selectivity on the whole database. The second one is to nd out why much agreement there is between the USM based distance among the domains and the tree that CATH associates to them. With respect to the rst test, the best algorithms to decide whether two proteins are similar have been tested with the 2771 domain database. Therefore, we tested if the USM based method is useful on deciding this demanding task and compare the results with the other methods. Krasnogor-Peltas USM approximation of the Universal Similarity Metric was applied carrying out a pairwise comparison of the 86 prototypes versus the 2771 domains. The values returned were examined from the best score to the worst. For each pair considered in this order whose domains belong to the same family, a coverage value was increased. Otherwise, an error value was increased. This is the method used by Sierk and Pearson to evaluate the other algorithms with the same database. A perfect classier would arrive to 100% coverage before the rst error arrives. To assess the sensitivity and selectivity of the Universal Similarity Metric, we plot Errors per Query versus Coverage curves for both approximations. These curves show how much coverage is obtained at a given error level, i.e., the number of true positives detected at a given number of false positive detected. With respect to the second test, we aim to compare the hierarchical clusterings produced by the USM based distance and the x clustering dened by CATH for the same domains. To do so, we nd the distance on the tree that best approximates the USM distance, and calculate the linear correlation between the two distances. For each tree edge, a weight variable is introduced. Let D be 6

Errors USM 0.1 0.09 1 0.4 10 1.7

Dali Worst Method 19 10 43 17 71 28

Table 1: Coverage performace for a given error level for dierent protein distances. the domain set and x be the vector of edge weight variables. We minimize the squares sum of the distances in the tree and in the known USM matrix: min
i,jD

(C(x, i, j) di,j )2 ,

where di,j is the constant USM based distance between domains i and j and C(x, i, j) is the sum of the variables associated to the known edge path between i and j in the CATH tree. Under the positivity constrains for the variables, the t is resolved using standard methods for quadratic programming. Finally, the tree distance ti,j between the domains i and j is calculated using the edge weights that produce the best t. The cophenetic correlation between (di,j ) and (ti,j ) is calculated, and compared to the correlation of the distance generated by SSEARCH.

3
3.1

Results
Selectivity test

According to Krasnogor and Pelta (2004), their method sometimes clusters well proteins that belong to a small number of groups. However, the most important question that needs to be answered when a new 3D structure of a protein is discovered is whether it is similar to one or more better known proteins. The problem of clustering proteins is not a crucial one, since lots of structures are not discovered at once, and clusters are already dened in public databases using lots of information sources. The Error-Coverage curve generated with the Universal Similarity Metric approximations is quite worse than those corresponding to the methods considered by Sierk and Pearson, as seen in Figure 1 with respect to Dali. Table 1 summarizes the results. 7

10 9 8 7 6 5 4 3 2 1 0 0.0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8

Figure 1: Error-Coverage curves. Left, USM method, right, Dali.

The Errors column shows the number of errors per prototype (i.e., 10 corresponds to 860 errors) and all other values are coverage percentages with respect to 1120 (i.e, 1% means 11 hits). The column USM refers to the results corresponding to the approximation explained in the introduction, the Dali column displays the results corresponding to the structural alignment program Dali (Holm and Sander, 1996) and the Worst Method gives the worst result obtained by any method considered by Sierk and Pearson for the corresponding error rate. So, for instance, at 0.1 errors per query on average, i.e., when 8 errors have been found, USM has only correctly covered 0.09% of the domains, i.e., 1 domain, while the Dali program at the same error rate reaches 19% of correctly covered domains (209 domains) and the worst method at this error rate (VAST) reaches a coverage of 10% (112 domains). At a rate of 1.0 errors per query (86 errors), USM covers correctly 0.4% while Dali covers 43% and the worst method in this case (SGM) covers 17%. At a rate of 10.0 errors per query (860 errors), the evaluated method covers only 1.7% while Dali covers 71% and the worst method (SGM) covers 28%. Notice that his 1.7% means 19 hits, i.e., after 860 false positives, USM has not even detected the identity of all 86 prototypes with themselves. Other tests, where the contact map is calculated when the threshold ranges from 5 to 8 gave the same negative results. A

3.2

Cophenetic correlation

When comparing a distance to its closest one on the CATH tree, the correlation coecient for USM is 0.60 and for SSEARCH is 0.87. This fact conrms that USM yields weaker clusterings with respect to CATH than a method that does not uses 3D information. We also found manually small subsubsets of domains that are wrongly clustered by USM, but that SSEARCH classies almost perfectly, using the same clustering algorithm used in Krasnogor and Pelta (2004). Due to paper limitations, they are not shown here but in the web page dmi.uib.es/people/jairo/USM/. In the st one in the web, for instance, there are proteins that belong to two dierent classes (the wider tree level) that are wrongly clustered by USM.

Conclusion

The USM based method under a simple contact map format clusters poorly protein groups easily discriminable and, in general, their clusters are father form CATH than the ones generated by some methods based on sequences; the three 9

small real sets that it clusters almost correctly are by no means an evidence that the method is robust for clustering. In addition, the method is very far from becoming a reliable protein comparison method from the point of view of deciding if two given protein structures are similar or not, one of the most important procedures to detect protein function similarity.

Acknowledgements
We thank the following people: Kevin Karplus for bringing out the idea on the assignment of lengths to the tree edges in order to nd how far a distance is from an additive one; Ming Li for discussing with the authors the rst version of this paper, and Joe Mir for his comments on an early draft. We acknowledge o the nancial support of the Spanish and UE grant BFM2003-00771 and the UE grant INTAS IT 04-77-7178.

References
Benedetto, D., Caglioti, E., Loreto, V., 2002. Language trees and zipping. Physical Review Letters 88 (4), 048702. Bennett, C., Gcs, P., Li, M., Vitnyi, P. M. B., Zurek, W., 1998. Information distance. IEEE Transactions on a a Information Theory 44 (7), 14071423. Chen, X., Francia, B., Li, M., Mckinnon, B., Seker, A., 2004. Shared information and program plagiarism detection. IEEE Transactions on Information Theory 50 (7), 15451551. Cilibrasi, R., Vitnyi, P., 2005. Clustering by compression. IEEE Transactions on Information Theory 51 (4), 1523 a 1545. Cilibrasi, R., Vitnyi, P., de Wolf, R., 2004. Algorithmic clustering of music based on string compression. Computer a Music Journal 28 (4), 4967. Gilbert, D., Rossell, F., Valiente, G., Veeramalai, M. Alignment-Free Comparison of TOPS Strings. In London Algoo rithmics and Stringology. Jackie Daykin and Manal Mohamed and Kathleen Steinhfel ed., College Publications, o 2006, in press. Holm, L., Sander, C., 1996. Mapping the protein universe. Science 273, 595602. Keogh, E., Lonardi, S., Rtanamahatana, C. A., 2004. Toward parameter-free data mining. In: Proc. 10th ACM Int. Conf. Knowledge Discovery and Data Mining. ACM, pp. 206215. Kocsor, A., Kertsz-Farkas, A., Kajn, L., Pongor, S., 2006. Application of compression-based distance measures to e a protein sequence classication: a methodological study. Bioinformatics 22 (4), 407412. Kolmogorov, A., 1965. Three approaches to the quantitative denition of information. Probl. Inform. Transm. 1 (1), 17. Krasnogor, N., Pelta, D., 2004. Measuring the similarity of protein structures by means of the universal similarity metric. Bioinformatics 20 (7), 10151021.

10

Li, M., Badger, J., Chen, X., Kwong, S., Kearney, P., Zhang, H., 2001. An information-based sequence distance and its application to whole mitochondrial genome phylogeny. Bioinformatics 17 (2), 149154. Li, M., Chen, X., Li, X., Ma, B., Vitnyi, P., 2004. The similarity metric. IEEE Transactions on Information Theory a 50 (12), 32503264. Li, M., Sleep, M., 2004. Melody classication using a similarity metric based on kolmogorov complexity. In: Proc. Conference on Sound and Music Computing SMC04. IRCAM, pp. 126129. Li, M., Vitnyi, P., 2nd ed. 1997. An Introduction to Kolmogorov Complexity and its Applications. Springer-Verlag. a Orego, C., et al., 1997. CATH: A hierarchic classication of protein domain structures. Structure 5, 10931108. Pelta, D., Gonzales, J., Krasnogor, N., 2005. Protein structure comparison through fuzzy contact maps and the universal similarity metric. In: Proc. Fourth Conference of the European Society for Fuzzy Logic and Technology and 11 Rencontres Francophones sur la Logique Floue et ses Applications (EUSFLAT-LFA, 2005). pp. 11241129. Sierk, M., Pearson, W., 2004. Sensitivity and selectivity in protein structure comparison. Protein Science 13, 773785. Smith, T. F., Waterman, M. S., 1981. Identication of common molecular subsequences. J. Mol. Biol., (147):195197. Svangard, N., Nordin, P., 2004. Automated aesthetic selection of evolutionary art by distance based classication of genomes and phenomes using the universal similarity metric. In: Proc. Applications of Evolutionary Computing: EvoWorkshops 2004. Vol. 3005 of Lecture Notes in Computer Science. Springer-Verlag, pp. 447456. Varre, J.-S., Delahaye, J.-P., Rivals, E., 1999. The transformation distance: A dissimilarity measure based on movements of segments. Bioinformatics 15 (3), 194202.

11