In order to scan whole proteomes for repeat-containing proteins, we created the tool Reptile. It uses a "brute-force" algorithm that detects all perfect repeats and enables direct calculation of a P-value. For each input sequence, Reptile generates all possible substrings from length 2 to a user-defined maximum (the default is 20) and counts their occurrences. After removing redundant repeats that are contained within longer ones, the repeated sequences are returned by ascending P-value. The probability P to find at least n repeats of length r in a random sequence of length L (with nr ≤ L ≤ n20^r) equals the number of possible sequences that contain the desired repeat, divided by the total number of possible sequences (20^L).

P ∗ ( n,r,L ) = 20 r 20 L-nr 20 L ( L-nr+n n ) = 20 -r ( n- 1 ) ( L-n ( r- 1 ) n ) MathType@MTEF@5@5@+=feaagaart1ev2aaatCvAUfKttLearuWrP9MDH5MBPbIqV92AaeXatLxBI9gBaebbnrfifHhDYfgasaacPC6xNi=xI8qiVKYPFjYdHaVhbbf9v8qqaqFr0xc9vqFj0dXdbba91qpepeI8k8fiI+fsY=rqGqVepae9pg0db9vqaiVgFr0xfr=xfr=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@6477@

Where 20^r is the number of possible repeat sequences, 20^L-nr the number of possible sequences around the repeats, and the binomial equals the number of ways to place the n repeats in L. P* is an overestimate because the sequences with more than n repeats are counted too often. Taking this into account gives the correct formula for P:

Where i counts from n to the maximal number of repeats (L/r), switching signs with every increment according to the inclusion-exclusion principle 36. For practical purposes calculation of P*, the first summand of P, is sufficient since further summands decrease rapidly with increasing number of repeats. Reptile returns all repeats below a user-defined cut-off P-value (the default is 10^-5, corresponding to an expectancy of one false positive in 100'000 sequences). Direct repeats are marked. The P-value being independent of the actual sequence of a repeat, Reptile also returns a measure of whether a detected repeat consists of rare or frequent amino acids. This "Amino acid abundance measure" (AM) was defined as follows:

AM ( repeat ) = log ⁡ 10 ( 20 r ∏ i=1 r f i ) MathType@MTEF@5@5@+=feaagaart1ev2aaatCvAUfKttLearuWrP9MDH5MBPbIqV92AaeXatLxBI9gBaebbnrfifHhDYfgasaacPC6xNi=xI8qiVKYPFjYdHaVhbbf9v8qqaqFr0xc9vqFj0dXdbba91qpepeI8k8fiI+fsY=rqGqVepae9pg0db9vqaiVgFr0xfr=xfr=xc9adbaqaaeGacaGaaiaabeqaaeqabiWaaaGcbaGaeeyqaeKaeeyta0KaeiikaGIaeeOCaiNaeeyzauMaeeiCaaNaeeyzauMaeeyyaeMaeeiDaqNaeiykaKIaeyypa0JagiiBaWMaei4Ba8Maei4zaC2aaSbaaSqaaiabigdaXiabicdaWaqabaGccqGGOaakcqaIYaGmcqaIWaamdaahaaWcbeqaaiabbkhaYbaakmaarahabaGaemOzay2aaSbaaSqaaiabbMgaPbqabaaabaGaeeyAaKMaeeypa0JaeeymaedabaGaeeOCaihaniabg+GivdGccqGGPaqkaaa@4E4D@

Where r is the length of the repeat and f_i is the frequency in the corresponding proteome – respectively set of sequences submitted by the user – of the amino acid at position i of the repeat. AM is symmetric to zero, negative values indicating that a repeat predominantly consists of rare amino acids (and vice versa). Reptile is running on-line 37 and accepts batch input of up to 50,000 sequences in any of the commonly used formats.

Compared to other repeat-prediction programs (Table 1) the main strengths of Reptile are its quantitative assessment of the detected repeats and its infallibility regarding short perfect repeats, such as they occur in antigens from parasites. Reptile will spot in a given protein all recurring subsequences from length two to twenty, even if they are dispersed. In contrast to programs implementing self-alignment, however, Reptile does not properly recognize degenerate repeats. Though proteins harbouring degenerate repeats also exhibit low P-values and will not go unnoticed, Reptile will not identify the basic repetitive unit but several shorter ones contained within. Other programs (Table 1) should be used when studying large repeat regions or imperfect, diverging repeats.

We defined highly repetitive proteins as proteins that contain perfect repeats of a P-value below 10^-10. Reptile was used to screen for such proteins in predicted proteomes from fully sequenced genomes. The median proportion of highly repetitive proteins was 2.7% in eukaryote proteomes and 0.43% in prokaryotes, confirming the notion 1 that eukaryotes possess more repetitive proteins than bacteria (p < 0.0001, Mann-Whitney test). The more repeats a protein has, the longer it becomes. In eukaryotic proteomes the percentage of highly-repetitive proteins correlated to some degree with the mean protein length (Spearman coefficient r_S = 0.51, p = 0.011). When distinguishing free-living from (endo)parasitic eukaryotes (Table 2), it was evident that the correlation was caused entirely by the latter. Obligate parasites exhibited a good correlation between highly-repetitive proteins and mean protein length (r_S = 0.82, p = 0.003) while free-living eukaryotes showed no correlation at all (Figure 1). The finding that the percentage of highly repetitive proteins predicts average protein length only in parasite proteomes reflects the significance of repeat-containing proteins for survival in the host, possibly counterbalanced by a selective pressure on parasites for shorter proteins 38.

The eukaryote with the largest proportion of highly repetitive proteins, Plasmodium falciparum with 28%, and that with the smallest one, Encephalitozoon cuniculi with 0.42%, were both obligate parasites. The same applied to prokaryotes, where the highest proportions of highly repetitive proteins were exhibited by Mycobacterium bovis (3.0%), M. tuberculosis (2.9%) and Parachlamydia sp. (2.7%), and the lowest ones by Bacillus anthracis (Porton strain, 0.02%) and Streptococcus pyogenes (SSI strain, 0.05%) – however, it must be noted that with bacteria, the available genome sequences are biased towards pathogenic species. The most repetitive protein from eukaryotes was a hypothetical protein from the sleeping sickness parasite T. brucei, followed by the 11-1 gene product from P. falciparum, a known malaria antigen of more than 1 MD size 39. The most repetitive prokaryotic protein was a predicted cell wall surface anchor family member from Streptococcus pneumoniae, the leading cause of pneumonia. Table 3 summarizes these and other highly repetitive proteins identified from pathogens, emphasizing on sequences with experimentally verified expression. The genome-wide surveys yielded other known virulence factors such as proteophosphoglycans of Leishmania 40 or PGRS (polymorphic GC-rich repetitive sequence) proteins of Mycobacterium, an antituberculosis vaccine candidate 41. The presence of avirulence proteins from phytopathogenic bacteria among the most repetitive proteins indicates that repeats also serve to specifically trigger host defence responses. Remarkably repetitive are also the ice nucleation proteins of plant pathogens. Table 3 also shows examples of previously undescribed proteins. The complete datasets on repeat-containing proteins from 49 eukaryotes and 193 prokaryotes are accessible on-line in the archive REPository 37.

To further characterize the repeats, we investigated which amino acids are over- or underrepresented in repeats of P < 10^-10 compared to the rest of the respective proteome. Overall, the amino acid composition of the repeats was more biased in eukaryotes than in bacteria (Figure 2). Small amino acids occurred more frequently in the repeats than large ones in both eukaryotes and prokaryotes. Hydrophobic residues were underrepresented in the repeats, with the exception of leucine, which in bacterial repeats was even overrepresented (p < 0.0001, two-tailed Wilcoxon signed rank test). Strongly overrepresented in the repeats were alanine (p < 0.0001) in bacteria and serine (p = 0.0001) in eukaryotes (Figure 2). Thus "cheap" amino acids seem to be preferred over energetically expensive ones. Interestingly, asparagine was overrepresented in the repeats from eukaryotes (p = 0.057) but not from bacteria (Figure 2), suggesting that asparagines might be preferentially glycosylated in repeats. Contrary to expectation though, the probability of an asparagine to be in N-glycosylation consensus was significantly lower in repeats than in non-repetitive sequences (Figure 3). This was the case for free-living eukaryotes (p = 0.004) as well as for parasites (p = 0.027). The only exception was T. brucei, where the likelihood of an asparagine to be in N-glycosylation consensus was three-fold higher in repetitive than in non-repetitive sequences (Figure 3).

In order to predict which of the repeat-containing proteins are at the cell surface, Reptile was combined with Phobius 42, a program for prediction of transmembrane domains and N-terminal export signals, and GPI-SOM 43, a program that predicts C-terminal GPI (glycosylphosphatidyl-inositol) anchor attachment sites. The three programs were run over all available proteomes predicted from completely sequenced genomes. The identified repeats were scanned for potential N-glycosylation sites. The combined output was stored in a relational database called Dora, the

d

o

r

a

New specific and robust tests are urgently needed for the diagnosis of sleeping sickness, malaria, tuberculosis, and other neglected diseases 4546. PCR not being applicable in the field, serology (i.e. the detection of parasite-specific antibodies) remains the principal method of detection for many tropical diseases. Dora provides a convenient portal for identification of candidate antigens for serological tests. In addition, it can be helpful for the selection of vaccine candidates. Dora returns the hits in Fasta format, which is suitable for subsequent bioinformatic analyses.

Predicted proteome files were obtained from the Integr8 database of the European Bioinformatics Institute 47. The download was automated with a Python script that periodically checks for newly available proteomes, respectively for updates to previous proteome files.

Statistical tests were performed with Prism 4.0 (GraphPad Software). Since the percentages of repeats in proteomes as well as the frequencies of amino acids were not normally distributed, non-parametric tests were used: Mann-Whitney test 48, Wilcoxon signed rank test 49, and Spearman correlation 50.

The repeat detection algorithm is described under Results. The program is written in C++ and the web-interface in Perl-CGI. Reptile uses sreformat from the HMMer package 51 to convert different input formats (Fasta, GenBank, EMBL, Swiss-Prot, PIR, GCG) to Fasta. Reptile runs on a vmware (virtual infrastructure) server. Availability and requirements:

Project name: Reptile

Project home page: http://genomics.unibe.ch/software/reptile.tar.gz

Operating systems: Linux, Unix

Programming language: C++

Licence: GNU GPL

A Python script periodically runs Reptile, GPI-SOM, and Phobius over all new or updated proteome files of Integr8. The results are stored in a MySQL database. For sake of simplicity, for each protein only the repeat with the lowest P-value is stored. A Perl script is used to interconvert Fasta format and SQL. The web interface of Dora is written in PhP. The database and all the programs run on the vmware server of the Informatics Services of the University of Bern.