Skip to main content
NCBI home page
NIHPA Author Manuscripts logoLink to NIHPA Author Manuscripts
. Author manuscript; available in PMC: 2009 Apr 8.
Published in final edited form as: Microbiology (Reading). 2006 Jan;152(Pt 1):113–121. doi: 10.1099/mic.0.28317-0

Lipoprotein computational prediction in spirochaetal genomes

João C Setubal 1, Marcelo Reis 2, James Matsunaga 3,4, David A Haake 3,4
PMCID: PMC2667199  NIHMSID: NIHMS97687  PMID: 16385121

Abstract

Lipoproteins are of great interest in understanding the molecular pathogenesis of spirochaetes. Because spirochaete lipobox sequences exhibit more plasticity than those of other bacteria, application of existing prediction algorithms to emerging sequence data has been problematic. In this paper a novel lipoprotein prediction algorithm is described, designated SpLip, constructed as a hybrid of a lipobox weight matrix approach supplemented by a set of lipoprotein signal peptide rules allowing for conservative amino acid substitutions. Both the weight matrix and the rules are based on a training set of 28 experimentally verified spirochaetal lipoproteins. The performance of the SpLip algorithm was compared to that of the hidden Markov model-based LipoP program and the rules-based algorithm Psort for all predicted protein-coding genes of Leptospira interrogans sv. Copenhageni, L. interrogans sv. Lai, Borrelia burgdorferi, Borrelia garinii, Treponema pallidum and Treponema denticola. Psort sensitivity (13–35 %) was considerably less than that of SpLip (93–100 %) or LipoP (50–84 %) due in part to the requirement of Psort for Ala or Gly at the −1 position, a rule based on E. coli lipoproteins. The percentage of false-positive lipoprotein predictions by the LipoP algorithm (8–30 %) was greater than that of SpLip (0–1 %) or Psort (4–27 %), due in part to the lack of rules in LipoP excluding unprecedented amino acids such as Lys and Arg in the −1 position. This analysis revealed a higher number of predicted spirochaetal lipoproteins than was previously known. The improved performance of the SpLip algorithm provides a more accurate prediction of the complete lipoprotein repertoire of spirochaetes. The hybrid approach of supplementing weight matrix scoring with rules based on knowledge of protein secretion biochemistry may be a general strategy for development of improved prediction algorithms.

Introduction

Lipoproteins are universal components of eubacterial membranes. Anchoring of lipoproteins to lipid bilayers occurs via fatty acids that covalently modify the amino-terminal Cys of the mature protein (Braun & Wolff, 1970). The ability of lipoproteins to decorate bacterial membranes provides for a wide variety of essential structural and functional roles. Murein lipoprotein anchors the inner face of the Gram-negative outer membrane to the peptidoglycan cell wall. Other lipoproteins function as adhesins, enzymes, transporters, binding proteins, toxins and in a variety of other capacities essential for virulence (Madan Babu & Sankaran, 2002). Given their broad distribution among bacteria and their unique structural features, it is not surprising that at least one of the toll-like receptors (TLR2) is designed to detect lipoproteins as an innate response to the presence of bacteria (Aliprantis et al., 1999; Brightbill et al., 1999).

In spirochaetes, lipoproteins are the most prominent proteins in the total membrane protein profile. Examples of highly abundant spirochaetal proteins are OspA of Borrelia burgdorferi, the causative agent of Lyme disease, Tpp47 of the syphilis spirochaete Treponema pallidum and LipL32, the major outer-membrane protein of pathogenic Leptospira species (Chamberlain et al., 1988; Haake et al., 2000; Howe et al., 1985). Research on spirochaetal lipoproteins has been essential to the understanding of spirochaetal physiology and pathogenesis. In Borrelia and Leptospira species, differential expression of lipoproteins is a hallmark of the transition to life inside the mammalian host (Barnett et al., 1999; Schwan et al., 1995). Several lipoproteins have been shown to be involved in the interaction of pathogenic spirochaetes with host molecules. Adherence of B. burgdorferi to extracellular matrix proteins is mediated by DbpA (Guo et al., 1998) and Bbk32 (Probert & Johnson, 1998). B. burgdorferi has been shown to evade activation of the complement cascade by binding of Factor H to lipoprotein OspE and related proteins (Hellwage et al., 2001). A number of spirochaetal lipoproteins have been shown to be targets of a protective immune response (Haake, 2000), confirming the importance of lipoproteins in the pathogenesis of spirochaetal diseases.

Since spirochaetes form a deep branch of the phylogenetic tree, it is not surprising that spirochaetal lipoproteins frequently have no homologues in the sequence databases. The divergence of spirochaetal lipoprotein sequences from those of other bacteria includes the signal peptide ‘lipobox’ region recognized by lipid modification enzymes. Lipid modification has been demonstrated experimentally for a relatively large number of spirochaetal lipoproteins (Haake, 2000). From these sequences it is possible to conclude that spirochaetal lipoproteins are secreted across the cytoplasmic membrane via a signal peptide, but that the lipobox at the carboxy terminus of the signal peptide differs significantly from that of other bacteria. Differences in the lipobox sequence presumably result from differences in the active site substrate specificities of the glyceryl transferase and type II signal peptidase that transfer a diacylglyceryl group to Cys and remove the signal peptide, respectively (Paetzel et al., 2002).

The von Heijne consensus lipobox pattern is based on lipoprotein sequences of Escherichia coli and similar Gram-negative bacteria (von Heijne, 1989). The Psort program (Nakai & Horton, 1999), which is based on the von Heijne consensus lipobox pattern, fails to recognize 43 % of experimentally verified spirochaetal lipoprotein sequences. In general, the inaccuracy of Psort reflects the increased plasticity of the spirochaetal lipobox. Recent application of the hidden Markov model approach in the LipoP program (Juncker et al., 2003) showed improved accuracy for lipoprotein recognition in general for bacteria other than E. coli, but includes many non-spirochaetal lipoproteins in its training set (TS). The emergence of data from spirochaetal genome sequencing efforts has resulted in the need for tools to accurately and efficiently identify lipoprotein genes (Fraser et al., 1997, 1998; Glockner et al., 2004; Nascimento et al., 2004; Ren et al., 2003; Seshadri et al., 2004). We set out to design an algorithm specifically tailored to identify spirochaetal lipoproteins; our program is designated SpLip. Application of the SpLip program to the six available spirochaetal genomes shows improved accuracy over existing generic lipoprotein prediction algorithms.

Methods

SpLip

SpLip uses a position-specific scoring matrix, also known as a weight matrix (WM) (Durbin et al., 1998; Mount, 2001). The SpLip WM uses a TS consisting of 28 spirochaetal proteins with experimental evidence of lipidation (Table 1). The TS includes 26 sequences described by Haake (2000) plus LipL21 (Cullen et al., 2003b) and LigB (Matsunaga et al., 2003) from Leptospira interrogans. Analysis of the TS was used to define the three regions of the spirochaetal lipoprotein signal peptide shown in Fig. 1: the carboxy-terminal region (C-region or lipobox), the hydrophobic (H-) region and amino-terminal (N-) region. The SpLip WM is focused on the lipobox because this is the most conserved region among lipoproteins. The lipobox is an ungapped motif with 4 or 5 positions, as will be seen below. Additionally, the program determines the H-region length and hydrophobicity, and net charge of the N-terminal region.

Table 1.

A set of 28 spirochaetal proteins with experimental evidence for lipidation used as the TS for the SpLip program

Gene/product N-region sequence H-region sequence Lipobox (C-region)
Borrelia afzelii
NlpH MK IINILFCLFLI MLSGC
Borrelia burgdorferi
OppA-1 MKKENPMKYIK IALMLIIF SLIAC
OppA-2 MKLQR SLFLIIFFL TFLCC
OppA-3 MSFNKTKKIGKKIK IVTLLMLAV SLIAC
LA7 MYKNGFFK NYLSLFLIF LVIAC
OspA MKK YLLGIGLIL ALIAC
OspB MR LLIGFALAL ALIGC
DbpA MIKCNNKTFNNLLK LTILVN LLISC
OspC MKK NTLSAILMTLF LFISC
OspD MKKLIK ILLLSLFLL LSISC
OspE MNKKMK MFIICAVFI LIGAC
OspF MNKKIK MFIICAIFM LISSC
Borrelia hermsii
Vmp7 MRKRISAIINK LNISIIIMTVV LMIGC
Vmp33 MKK NTLSAILMTLF LFISC
Brachyspira hyodysenteriae
SmpA MNKK IFTLFLVVAASAI FAVSC
Leptospira interrogans
LigB MKK IFCISIFLSM FFQSC
LipL21 MINR LIALSLATM IFAAC
LipL32 MKK LSILAISVALFA SITAC
LipL36 MRRNIMK IAAVAALTV ALTAC
LipL41 MRK LSSLISVLVLLM FLGNC
Treponema pallidum
GlpQ MR GTYCVTLWGGVFAA LVAGC
MglB-2 MKENSCTACSRR LALFVGAAV LVVGC
TmpA MNAH TLVYSGVALACAA MLGSC
TmpC MREKWVR AFAGVFCAM LLIGC
Tp47 MKVK YALLSAGALQL LVVGC
Tpd MKR VSLLGSAAIFAL VFSAC
Tpp15 MVKR GGAFALCLAV LLGAC
Tpp17 MKGSVR ALCAFLGVGALGSA LCVSC
Open in a new tab

Fig. 1.

Fig. 1

Open in a new tab

Definition of spirochaetal lipoprotein signal peptide regions. Analysis of the TS and paralogous sequences of probable lipoprotein sequences was used to define signal peptide regions for use in the SpLip algorithm. The amino-terminal N-region extends from the start methionine to the last charged residue. The hydrophobic H-region extends from the last charged residue to the −5 position and must be at least 6 aa in length. The carboxy-terminal C-region, or lipobox, is 4 aa in length. At least one Leu, Phe, Val, Ile or Tyr must occur in positions −3 or −4. Position −1 must contain Ser, Asn, Ala, Gly, Cys, Thr or Gln. Charged residues (i.e. Lys, Arg, Asp, Glu and His) are forbidden in the H- and C-regions.

Characterization of the spirochaetal lipobox

The C-terminal region, or lipobox, of the spirochaetal lipoprotein signal peptide is defined in principle as the four positions (−1, −2, −3 and −4) upstream of the cleavage site (position −5 is also considered, see below). A Cys residue is always found in position +1. This is an absolute requirement as there is no precedent that we are aware of for amino-terminal lipid modification of membrane proteins at amino acids other than Cys.

The TS was used to characterize the spirochaetal lipobox in three steps:

  1. Analysis of the TS yielded a set of lipobox rules which are a refinement of the spirochaetal lipobox described by Haake (2000).

    1. Position −1 – only Ala, Gly, Ser, Asn or Cys are allowed;

    2. Positions −3 or −4 – at least one of these positions should contain at least one of Leu, Ile, Val or Phe;

    3. The charged amino acids Lys, Arg, Asp, Glu and His are forbidden anywhere in the lipobox.

    A predicted protein-coding gene (PPCG) that has a lipobox conforming to these rules (and to other constraints not pertaining to the lipobox itself, to be described below) is considered a probable lipoprotein.

  2. The WM was built following standard procedures (Durbin et al., 1998; Mount, 2001) (see below)

  3. Lipoboxes in all PPCGs of Leptospira interrogans sv. Copenhageni were scored according to the WM. An analysis of the high-scoring PPCGs and TS members together with the multiple alignments of significantly similar pairs of PPCGs (see below), resulted in modification of the lipobox rules, as follows.

    1. Position −1 – in addition to Ala, Gly, Ser, Asn and Cys, the related amino acids Gln and Thr are also allowed;

    2. Position −5 is also considered to be part of the lipobox; rule (1b) above is extended to this position;

    3. In addition to Leu, Ile, Val and Phe, the hydrophobic amino acids Tyr and Met are also included as possible amino acids required in positions −3, −4 or −5.

    A PPCG that has a putative lipobox conforming to these modified rules (and to other constraints not pertaining to the lipobox itself, to be described below), and not to rules in step (1), is considered a possible lipoprotein.

Characterization of the hydrophobic (H-) region

Based on analysis of the TS and the requirement for a hydrophobic signal peptide, charged residues Lys, Arg, Asp, Glu and His were forbidden in the H-region. The H-region should be at least 7 aa long for probable lipoproteins and 6 aa long for possible lipoproteins.

Characterization of the amino-terminal (N-) region

In a lipoprotein signal peptide the N-terminal region should be positively charged. The N-terminal region is considered to extend from the first residue to the last charged residue (i.e. Lys, Arg, Asp, Glu or His). The residue following the last charged residue defines the start of the H-region.

Length of the mature lipoprotein

Both probable and possible lipoproteins have to follow this additional rule: the PPCG should have at least 50 residues downstream of the +1 position.

WM construction

The standard procedure (Durbin et al., 1998; Mount, 2001) for building a WM is as follows. Background frequencies for the residues are determined for each organism separately, thus obtaining a different matrix for each organism. The background frequencies are given by the first 50 residues in all PPCGs. Then residue frequencies in the TS are determined. Entry (i,j) in the matrix (where i is a residue and j is a position in the sequence) is given by log2(Fi,j/Gi,j), where Fi,j is the observed frequency of residue i in position j in the TS, and Gi,j is the observed background frequency of residue i in position j.

The standard WM procedure was adapted to accommodate the rules described above. The lipobox for each sequence in the TS was known. Matrix entries for which Fi,j=0 (the amino acid i was not observed in position j in the TS, but is not forbidden either) were changed to −2 [this is the largest integer less than the lowest value for log2(Fi,j/Gi,j) observed], except for Gln and Thr at position −1 or Met and Tyr at positions −3, −4 and −5; these residues received a score of zero at those positions. Matrix entries for which the corresponding amino acid is forbidden were set to −100.

C-region scoring

Given the WM, scoring was done in the standard way as follows. For each PPCG, all Cys residues in the first 50 residues are located, and for each Cys the four positions −1 to −4 are evaluated according to the WM. That is, if residue j is found at position i, it gets the score given by entry (i,j) in the WM. The sum of the scores for the four positions −1 to −4 is the C-region score. The putative lipobox is taken as the one with highest score (in case there is more than one Cys in the first 50 residues). If no score is positive, position −5 is also included in the analysis. If the highest score is still negative or zero, the PPCG is rejected.

H-region scoring

The H-region is defined as the region from position −5 (or −6, depending on the start position of the C-region) upstream of the putative lipobox to the position after the first charged residue defined as Lys, Arg, Asp, Glu or His. The Kyte–Doolittle hydrophobicity matrix (Kyte & Doolittle, 1982) was used to score the H-region. PPCGs with a negative H-region score or a length less than 6 residues are rejected.

N-region scoring

The SpLip algorithm calculates the net charge of the N-region according to the formula #Lys+#Arg+#His>#Asp+#Glu. PPCGs without a net positive charge in the N-region are rejected.

The final predictions are those that achieve positive scores for regions C, H and N. The pseudocode summarizing the SpLip algorithm is shown in Fig. 2.

Fig. 2.

Fig. 2

Open in a new tab

Pseudocode illustrating the SpLip algorithm. This pseudocode is executed for every PPCG in each of the six genomes.

Methodology of prediction evaluation

We compared the performance of the SpLip, Psort (Nakai & Horton, 1999) and LipoP (Juncker et al., 2003) algorithms. The primary input for all three prediction algorithms was the complete set of PPCGs of L. interrogans sv. Copenhageni (Nascimento et al., 2004), L. interrogans sv. Lai (Ren et al., 2003), B. burgdorferi (Fraser et al., 1997), Borrelia garinii (Glockner et al., 2004), T. pallidum (Fraser et al., 1998) and Treponema denticola (Seshadri et al., 2004). Positive predictions in the case of Psort were those for which the output contained the words ‘may be a lipoprotein’. Positive predictions in the case of LipoP were those with a positive score. In the case of SpLip, positive predictions included those with a positive score using rules for both probable and possible lipoproteins. The ‘base set’ comprised all PPCGs that were predicted to be a lipoprotein by at least one program. The base set of results was manually curated as ‘true-positive’ or ‘false-positive’ lipoproteins. We refer to the complete set of true-positive lipoproteins, including members of the TS, as the ‘Liposet’ for that organism. It should be noted that except for the TS, Liposet members identified in this paper have not been confirmed experimentally. True-positive lipoproteins in the base set that were not predicted to be lipoproteins by a given program were considered ‘false-negative’ sequences for that program. False-negative rates for each program and each genome were computed with the formula #fn/(#fn+#tp), where #fn is the number of false-negatives and #tp is the number of true-positives found by the program. The denominator #fn+#tp corresponds to the size of the Liposet for that genome and it is the same number for all programs. Sensitivity is given by 1–false-negative rate. False-positive rates for each program and each genome were computed with the formula #fp/(#fp+#tp), where #fp corresponds to the number of false-positives for the given program and #tp is the same as above. Note that the denominator #fp+#tp corresponds to all positive predictions of a given program and may be different for different programs. As an additional test of false-positive rates, all three algorithms were run on a set of 298 transmembrane cytoplasmic proteins of spirochaetes (known to be non-lipoproteins) available from SWISS-PROT (http://us.expasy.org) (Bairoch & Apweiler, 2000).

Results

The total number of true-positive lipoproteins for each genome is given in Table 2. Note that for B. burgdorferi and T. pallidum the number of lipoproteins listed is higher than what has been originally reported. For B. burgdorferi, 105 lipoproteins were originally reported (Fraser et al., 1997); 127 were identified in the present study. For T. pallidum, 22 lipoproteins were previously reported (Fraser et al., 1998); 46 were identified in the present study. To a large extent, these higher numbers reflect the greater sensitivity of the SpLip algorithm relative to previously available methods.

Table 2.

Prediction results of all three programs

L. interrogans sv. Copenhageni L. interrogans sv. Lai B. burgdorferi B. garinii T. pallidum T. denticola
PPCGs 3660 4727* 1637 832 1031 2767
Total predictions (all 3 algorithms) 217 206 140 36 58 181
Liposet† 164 157 127 30 46 166
SpLip
Probable 134 125 112 24 36 160
Possible 23 21 8 6 9 2
Total predictions 157 146 120 30 45 162
False-positive rate (%)‡ 2 (1.3) 0 (0.0) 0 (0.0) 0 (0.0) 0 (0.0) 0 (0.0)
False-negative rate (%)‡ 9 (5.5) 11 (7.0) 7 (5.5) 0 (0.0) 1 (2.2) 4 (2.4)
LipoP
Predictions 163 149 116 25 31 151
False-positive rate (%)‡ 47 (28.8) 44 (29.5) 13 (11.2) 5 (20.0) 8 (25.8) 12 (7.9)
False-negative rate (%)‡ 48 (29.2) 52 (33.1) 24 (18.9) 10 (33.3) 23 (50.0) 27 (26.7)
Psort
Predictions 47 46 26 5 22 48
False-positive rate (%)‡ 10 (21.3) 12 (26.1) 1 (3.8) 1 (20.0) 6 (27.3) 3 (6.3)
False-negative rate (%)‡ 127 (77.4) 123 (78.3) 102 (80.3) 26 (86.7) 30 (65.2) 121 (72.9)
Open in a new tab
*

Errors in the L. interrogans sv. Lai genome annotation led to a major overestimation in the number of PPCGs.

†

The Liposet is defined as the total number of true lipoproteins for each genome as determined by expert review of all lipoprotein predictions by the SpLip, LipoP, and Psort algorithms. Lipoproteins not predicted by any of the three algorithms would not have been captured using this method.

‡

See Methods for definitions. The SpLip algorithm had the lowest false-positive and false-negative rates for all six genomes.

The T. denticola genome was found to encode 166 predicted lipoproteins, the largest number of any of the six spirochaete genomes. However, as a fraction of the total number of PPCGs, the B. burgdorferi sequence contains the highest percentage (7.8 %) of lipoproteins (Fig. 3). The relatively low percentage of lipoproteins in the B. garinii genome is due in large measure to the incomplete sequencing of the seven highly redundant lipoprotein-rich cp32 plasmids (Glockner et al., 2004).

Fig. 3.

Fig. 3

Open in a new tab

Percentage of lipoprotein sequences in spirochaetal genomes. The total number of predicted lipoprotein sequences as a percentage of the total number of PPCGs was calculated for L. interrogans sv. Copenhageni (Lic), L. interrogans sv. Lai (Lil), B. burgdorferi (Bb), B. garinii (Bg), T. pallidum (Tp) and T. denticola (Td).

Leptospiral lipobox

The lipoboxes of the L. interrogans sv. Copenhageni Liposet were compared to those of experimentally confirmed E. coli lipoproteins (Gonnet et al., 2004). Considerable differences were noted. As shown in Table 3, there is increased amino acid sequence variability in the leptospiral lipobox. For example, while Leu is the most common amino acid in both the −3 and −4 positions of the E. coli lipobox, Leu and Phe occurred with similar frequency at these positions in leptospiral lipoproteins. There was also a difference in the preferred amino acids, particularly at the −1 and −2 positions before Cys. Seventy-eight of 81 (96 %) aa at position −1 in E. coli lipoproteins are Ala or Gly. In contrast, the most common amino acids at position −1 of leptospiral lipoproteins are Ser and Asn. While the most common amino acid at position −2 in E. coli lipoproteins is Ala (29/81=36 %), seven other amino acids occur more frequently than Ala at this position in leptospiral lipoproteins.

Table 3.

Amino acid counts and frequencies (%) for the lipoboxes (C-regions) of the 164-member L. interrogans sv. Copenhageni Liposet identified in the current study and 81 experimentally confirmed E. coli lipoproteins (Gonnet et al., 2004)

Amino acid 164 L. interrogans sv. Copenhageni lipoproteins 81 E. coli lipoproteins
−4 −3 −2 −1 −4 −3 −2 −1
Phe 44 (26.8) 55 (33.5) 14 (8.5) 0 (0.0) 5 (6.2) 1 (1.2) 0 (0.0) 0 (0.0)
Ser 16 (9.8) 8 (4.9) 22 (13.4) 44 (26.8) 5 (6.2) 0 (0.0) 18 (22.2) 3 (3.7)
Tyr 0 (0.0) 1 (0.6) 5 (3.1) 0 (0.0) 0 (0.0) 0 (0.0) 0 (0.0) 0 (0.0)
Cys 3 (1.8) 0 (0.0) 0 (0.0) 2 (1.2) 0 (0.0) 1 (1.2) 2 (2.5) 0 (0.0)
Trp 2 (1.2) 0 (0.0) 4 (2.4) 0 (0.0) 1 (1.2) 0 (0.0) 0 (0.0) 0 (0.0)
Leu 46 (28.1) 53 (32.3) 22 (13.4) 0 (0.0) 38 (46.9) 67 (82.7) 0 (0.0) 0 (0.0)
Pro 1 (0.6) 1 (0.6) 0 (0.0) 0 (0.0) 0 (0.0) 0 (0.0) 0 (0.0) 0 (0.0)
Gln 0 (0.0) 2 (1.2) 9 (5.5) 10 (6.1) 0 (0.0) 0 (0.0) 2 (2.5) 0 (0.0)
Ile 12 (7.3) 16 (9.8) 23 (14.0) 0 (0.0) 2 (2.5) 5 (6.2) 2 (2.5) 0 (0.0)
Met 0 (0.0) 3 (1.8) 2 (1.2) 0 (0.0) 9 (11.1) 2 (2.5) 1 (1.2) 0 (0.0)
Thr 8 (4.9) 3 (1.8) 15 (9.2) 5 (3.1) 6 (7.4) 1 (1.2) 18 (22.2) 0 (0.0)
Asn 1 (0.6) 1 (0.6) 5 (3.1) 41 (25.0) 0 (0.0) 0 (0.0) 0 (0.0) 0 (0.0)
Lys 0 (0.0) 0 (0.0) 0 (0.0) 0 (0.0) 1 (1.2) 0 (0.0) 0 (0.0) 0 (0.0)
Val 15 (9.2) 15 (9.2) 21 (12.8) 0 (0.0) 3 (3.7) 4 (4.9) 9 (11.1) 0 (0.0)
Ala 7 (4.3) 4 (2.4) 10 (6.1) 24 (14.6) 9 (11.1) 0 (0.0) 29 (35.8) 22 (27.2)
Gly 9 (5.5) 2 (1.2) 12 (7.3) 38 (22.6) 2 (2.5) 0 (0.0) 0 (0.0) 56 (69.1)
Open in a new tab

Amino acids His, Arg, Asp, and Glu are not listed because they were not found to occur in any L. interrogans or E. coli lipoboxes.

Sensitivity

As shown in Fig. 4, the SpLip and LipoP algorithms had much better sensitivity than Psort for identification of spirochaetal lipoprotein sequences. The primary reason for the poor sensitivity of Psort is the requirement of Ala or Gly in the −1 position, based on sequences of E. coli lipoproteins (von Heijne, 1989). Psort sensitivity ranged from 35 % in the case of T. pallidum to as low as 20 and 13 % for the B. burgdorferi and B. garinii genomes, respectively. These results indicate a higher diversity of lipobox amino acids in spirochaetal genomes. It should be noted that Psort failed to predict 12 out of 28 experimentally confirmed lipoprotein sequences in the SpLip TS (a false-negative rate of 43 % for members of the SpLip TS) and LipoP failed to predict one (false-negative rate of 3.6 % for members of the SpLip TS).

Fig. 4.

Fig. 4

Open in a new tab

Sensitivity of algorithms for spirochaetal lipoprotein sequences. Sensitivity (see Methods for definition) was calculated for SpLip (white bars), LipoP (black bars) and Psort (grey bars) for L. interrogans sv. Copenhageni (Lic), L. interrogans sv. Lai (Lil), B. burgdorferi (Bb), B. garinii (Bg), T. pallidum (Tp) and T. denticola (Td). For all three spirochaetal genomes, algorithm sensitivity was SpLip>LipoP>Psort.

False-positive predictions

The SpLip algorithm produced no false-positive lipoprotein predictions for any spirochaetal genome except for L. interrogans sv. Copenhageni (LIC). Two LIC sequences identified by SpLip as possible lipoproteins, DsbD and FlaB5, were judged to be false-positives because of their close homology with proteins that are known not to be lipoproteins. In E. coli, DsbD is a disulfide interchange protein located in the cytoplasmic membrane with nine transmembrane loops (Stewart et al., 1999). Rather than being conjugated with fatty acids, the Cys residue at position 30 is involved in the transfer of disulfide reducing potential from the cytoplasm to the periplasm. FlaB5 is one of five leptospiral FlaB flagellin proteins. All other known bacterial flagellin proteins are secreted without a signal peptide by the flagellar type III secretion system (Minamino & Namba, 2004). FlaB5 is unique in that it has a signal peptide, indicating sec-dependent secretion. For FlaB to be incorporated into the flagellar filament, FlaB should be secreted into the periplasm. Since lipoproteins generally remain membrane bound and are not secreted, we judged that FlaB5 is likely to be a false-positive lipoprotein prediction.

LipoP had a higher mean rate of false-positive lipoprotein sequences than either Psort or SpLip (Table 2). The percentage of false-positive LipoP predictions was highest for leptospiral sequences. Psort false-positive predictions were slightly lower than those for LipoP. The percentage of false-positive Psort predictions was lowest for B. burgdorferi sequences. The LipoP and Psort algorithms tended to have different patterns of false-positive errors (see supplementary data available with the online version of this paper). The most frequent cause of LipoP errors was due to allowing unprecedented amino acids in the −1 position, frequently including charged amino acids Asp, Glu, Arg and Lys. Psort does not allow charged amino acids in the −1 position, but did allow them in other positions in the carboxy-terminal and hydrophobic regions of the lipoprotein signal peptide. Psort had a higher frequency than LipoP of unacceptably short hydrophobic regions and amino-terminal regions without a net positive charge.

As an additional control for false-positive predictions, we also tested all three algorithms using a set of 298 SWISS-PROT (Bairoch & Apweiler, 2000) spirochaetal proteins identified as having a cytoplasmic subcellular localization and that typically lack a signal peptide. The following results were obtained: SpLip had zero false-positive predictions, Psort had two and LipoP had seven. These results show that all three algorithms had low false-positive lipoprotein predictions when queried with cytoplasmic protein sequences.

Brachyspira lipoproteins

Brachyspira is a spirochaetal genus for which there are no complete genomes. However, there are reports of Brachyspira proteins with experimental evidence of lipidation. One of these, SmpA, is included in the SpLip TS (Table 1). In addition, published reports indicate that two additional proteins, MglB (Zhang et al., 2000) of Brachyspira pilosicoli and BlpA (Cullen et al., 2003a) of Brachyspira hyodysenteriae are also lipidated. Because Leptospira is the closest phylogenetic relative of Brachyspira, we tested the SpLip algorithm on the MglB and BlpA sequences using the L. interrogans sv. Copenhageni WM. SpLip correctly predicted that both MglB and BlpA are ‘probable’ lipoproteins with lipobox scores of 2.82 and 2.57, respectively. These results provide additional confirmatory evidence for the SpLip algorithm's accuracy.

Discussion

We have described a novel algorithm, designated SpLip, for the prediction of spirochaetal lipoproteins. SpLip uses a hybrid approach incorporating both lipoprotein signal peptide rules and a statistical WM based on a TS consisting of experimentally verified spirochaetal lipoproteins. The SpLip algorithm appeared to be more accurate than either the rules-based Psort algorithm or the more general hidden Markov model approach represented by the LipoP algorithm. The rules governing the Psort algorithm do not take into account the higher variability of the spirochaetal lipobox. LipoP has improved sensitivity for detection of spirochaetal lipoproteins compared to Psort, but allows charged amino acids in the lipobox, which is not consistent with a lipoprotein sequence. Analysis of all available spirochaetal genome sequences using the SpLip algorithm made it possible to produce an accurate and virtually complete set of lipoproteins for six spirochaetes, including L. interrogans sv. Copenhageni, L. interrogans sv. Lai, B. burgdorferi, B. garinii, T. pallidum and T. denticola. T. denticola had 166 predicted lipoproteins, the largest number of lipoproteins for any of the six organisms. When the sequences of its plasmids are included, 7.8 % of B. burgdorferi PPCGs encoded predicted lipoproteins, the highest fraction of PPCGs of any organism that we are aware of. Although Brachyspira genome sequences are not yet available and strict application of the SpLip methodology requires the complete PPCG set for each genome to be tested, we found that the SpLip algorithm correctly classified the sequences of two experimentally confirmed Brachyspira lipoprotein sequences, MglB and BlpA.

The SpLip algorithm is supplemented with rules based both on precedent and on an understanding of the biochemistry of lipoprotein signal peptides. Aside from Cys in the +1 position and the start Met, the most constrained position in the lipoprotein signal peptide is the −1 position. The structural constraints on the −1 position are largely imposed by the substrate specificity of the diacylglyceryl transferase, which transfers the initial lipid to Cys and to a lesser extent by the lipoprotein signal peptidase, which removes the signal peptide from the preprotein (Paetzel et al., 2002). In many bacteria, the −1 position consists exclusively of the small non-polar amino acids Ala or Gly, as reflected in the von Heijne consensus pattern (von Heijne, 1989). In spirochaetes, a high percentage of lipoprotein signal peptides contain Ser at the −1 position, which largely accounts for the failure of Psort to correctly predict 12 of the 28 sequences in the SpLip TS of experimentally verified lipoproteins. The lipid modification of spirochaetal lipoproteins with Ser at −1 has been well documented experimentally in 9/28 sequences in the SpLip TS, including B. burgdorferi proteins OspC, OspD and the decorin-binding protein, DbpA. Lipoproteins with Ser at −1 are uncommon in non-spirochaetes, but examples do exist, including MltC and YddW of E. coli (Gonnet et al., 2004), β-lactamase III of Bacillus cereus and VirB7 of Agrobacterium tumefaciens.

In contrast to Ser, the occurrence of Asn and Cys in the −1 position of lipoproteins may be unique to spirochaetes. Asn occurs in the −1 position in the leptospiral lipoprotein LipL41 (Shang et al., 1996). The experimental evidence for lipidation of LipL41 includes 3H-palmitate intrinsic labelling studies and inhibition of labelling with globomycin. Cys occurs in the −1 position in the B. burgdorferi oligopeptide-binding protein OppA-2, which has been shown to be a lipoprotein by intrinsic labelling with 3H-palmitate (Kornacki & Oliver, 1998). The SpLip algorithm was initially run allowing only Ala, Gly, Ser, Asn and Cys in the −1 position. When the algorithm was re-run allowing Thr and Gln (conservative amino substitutions for Ser and Asn, respectively) in the −1 position, a number of additional lipoproteins were identified that would have otherwise received an unfavourable score. Because there is no experimental evidence available at this time to confirm that spirochaetal proteins with Thr or Gln in the −1 position would be lipidated, we refer to predicted lipoproteins with Thr or Gln in the −1 position as ‘possible’ rather than ‘probable’ lipoproteins.

In contrast to the conservative amino acid substitutions allowed by the SpLip algorithm, the lipoproteins predicted by the LipoP algorithm contain a wide variety of amino acids in the −1 position. We observed that the LipoP algorithm predicts as lipoprotein sequences that are lipoprotein-like at other positions but have unprecedented amino acids at the −1 position. When we analysed spirochaetal genomes with the LipoP algorithm, a large number of lipoproteins were predicted with charged amino acids (Asp, Glu, Lys, Arg or His) in the −1 position and elsewhere in the hydrophobic or carboxy-terminal regions of the signal peptide of spirochaetal lipoproteins (see supplementary data available with the online version of this paper). We interpret such sequences as false-positive lipoprotein predictions because there is no precedent or justification that we are aware of for charged amino acids to occur in regions other than the amino-terminal portion of the lipoprotein signal peptide. Another measure of the inaccuracy of the LipoP algorithm is the finding that 7/298 spirochaetal sequences designated cytoplasmic membrane proteins by the SWISS-PROT database are predicted by LipoP to be lipoproteins. In contrast, SpLip correctly rejected all 298 of the SWISS-PROT cytoplasmic membrane protein sequences.

There were important differences between the TSs used to construct the models used for scoring in the LipoP and SpLip algorithms. Only 17/63 lipoproteins from the LipoP TS are spirochaetal in origin. Because there is evidence of substantial differences in the preferred amino acids in the lipobox of spirochaetes relative to other bacteria, the non-spirochaetal lipoproteins in the LipoP algorithm TS would reduce the accuracy of the LipoP algorithm for correctly scoring spirochaetal lipoproteins. LipoP incorrectly predicts that one of the sequences in the SpLip TS, the 17-kDa TpN17 T. pallidum lipoprotein, has a signal peptidase 1 cleavage site rather than a lipoprotein signal peptidase II cleavage site. TpN17 is included in the SpLip TS because of intrinsic labelling studies in T. pallidum have demonstrated lipidation (Akins et al., 1993). Another problem is that the LipoP TS was obtained by searching the SWISS-PROT database for ‘probable’ or ‘potential’ lipoproteins. Consequently, many of the 17 spirochaetal lipoproteins in the LipoP TS, such as the Borrelia Bmp proteins, lack experimental evidence of lipidation. In contrast, the SpLip algorithm relied exclusively on lipoprotein sequences for which there was experimental evidence of lipidation. Homologous lipoprotein sequences were excluded from the LipoP TS but retained by the SpLip TS. For example, the B. burgdorferi oligopeptide-binding proteins OppA-1 and OppA-3 share the same lipobox sequence, SLIAC. One justification for retaining both sequences in the TS is that the signal peptide sequences of these homologous proteins are otherwise dissimilar. Another justification is that because the SpLip TS is large enough to be a representative sampling of all spirochaetal lipoproteins, sequences that occur more frequently should be given more weight in the WM.

In summary, we have developed a novel lipoprotein prediction algorithm that is a hybrid approach using a combination of rules based on a biochemical knowledge of lipoprotein signal peptides and a statistical WM based on a lipoprotein sequence TS. Application of the SpLip algorithm to six spirochaetal genome sequences resulted in a more accurate set of predicted lipoproteins, with significantly improved sensitivity and specificity than either of the previously existing programs. The lipoprotein databases provided by this study will be useful in the search for spirochaetal virulence factors and vaccine candidates. In addition, the SpLip program will be useful for analysis of emerging spirochaetal genome sequence data. More importantly, we believe that our hybrid approach, by taking advantage of the strengths of rules grounded in biochemistry and statistical empiricism, can be generalized not only to lipoprotein identification in non-spirochaetal bacteria but also to other types of sequence prediction algorithms.

SpLip program

The SpLip program was encoded in perl, runs on any operating system that supports the perl interpreter and is available upon request by contacting the authors.

Supplementary Material

B. burgdorferi
B. garinii
L. int sv Cop
L. int sv Lai
T. denticola
T. pallidum

Acknowledgments

This work was supported in part by VA Medical Research Funds (to J. M. and D. A. H.); Public Health Service grant AI-34431 (to D. A. H.) from the National Institute of Allergy and Infectious Diseases; FAPESP (Brazil) Leptospira grant (to J. C. S.) and fellowship (to M. S. R.); and CNPq (Brazil) fellowship (to J. C. S.).

Abbreviations

PPCG

predicted protein-coding gene

TS

training set

WM

weight matrix

Footnotes

Supplementary data tables with the complete listing of all lipoprotein predictions described in this paper are available with the online version of this paper. The SpLip program was encoded in perl, runs on any operating system that supports the perl interpreter and is available upon request by contacting the authors.

References

  1. Akins DR, Purcell BK, Mitra MM, Norgard MV, Radolf JD. Lipid modification of the 17-kilodalton membrane immunogen of Treponema pallidum determines macrophage activation as well as amphiphilicity. Infect Immun. 1993;61:1202–1210. doi: 10.1128/iai.61.4.1202-1210.1993. [DOI] [PMC free article] [PubMed] [Google Scholar]
  2. Aliprantis AO, Yang RB, Mark MR, Suggett S, Devaux B, Radolf JD, Klimpel GR, Godowski P, Zychlinsky A. Cell activation and apoptosis by bacterial lipoproteins through toll-like receptor-2. Science. 1999;285:736–739. doi: 10.1126/science.285.5428.736. [DOI] [PubMed] [Google Scholar]
  3. Bairoch A, Apweiler R. The SWISS-PROT protein sequence database and its supplement TrEMBL in 2000. Nucleic Acids Res. 2000;28:45–48. doi: 10.1093/nar/28.1.45. [DOI] [PMC free article] [PubMed] [Google Scholar]
  4. Barnett JK, Barnett D, Bolin CA, Summers TA, Wagar EA, Cheville NF, Hartskeerl RA, Haake DA. Expression and distribution of leptospiral outer membrane components during renal infection of hamsters. Infect Immun. 1999;67:853–861. doi: 10.1128/iai.67.2.853-861.1999. [DOI] [PMC free article] [PubMed] [Google Scholar]
  5. Braun V, Wolff H. The murein-lipoprotein linkage in the cell wall of Escherichia coli. Eur J Biochem. 1970;14:387–391. doi: 10.1111/j.1432-1033.1970.tb00301.x. [DOI] [PubMed] [Google Scholar]
  6. Brightbill HD, Libraty DH, Krutzik SR, et al. Host defense mechanisms triggered by microbial lipoproteins through toll-like receptors. Science. 1999;285:732–736. doi: 10.1126/science.285.5428.732. [DOI] [PubMed] [Google Scholar]
  7. Chamberlain NR, Radolf JD, Hsu PL, Sell S, Norgard MV. Genetic and physicochemical characterization of the recombinant DNA-derived 47-kilodalton surface immunogen of Treponema pallidum subsp. pallidum. Infect Immun. 1988;56:71–78. doi: 10.1128/iai.56.1.71-78.1988. [DOI] [PMC free article] [PubMed] [Google Scholar]
  8. Cullen PA, Coutts SA, Cordwell SJ, Bulach DM, Adler B. Characterization of a locus encoding four paralogous outer membrane lipoproteins of Brachyspira hyodysenteriae. Microbes Infect. 2003a;5:275–283. doi: 10.1016/s1286-4579(03)00027-3. [DOI] [PubMed] [Google Scholar]
  9. Cullen PA, Haake DA, Bulach DM, Zuerner RL, Adler B. LipL21 is a novel surface-exposed lipoprotein of pathogenic Leptospira species. Infect Immun. 2003b;71:2414–2421. doi: 10.1128/IAI.71.5.2414-2421.2003. [DOI] [PMC free article] [PubMed] [Google Scholar]
  10. Durbin R, Eddy SR, Krogh A, Mitchison G. Biological Sequence Analysis Probabilistic Models of Proteins and Nucleic Acids. Cambridge: Cambridge University Press; 1998. [Google Scholar]
  11. Fraser CM, Casjens S, Huang WM, et al. Genomic sequence of a Lyme disease spirochaete, Borrelia burgdorferi. Nature. 1997;390:580–586. doi: 10.1038/37551. [DOI] [PubMed] [Google Scholar]
  12. Fraser CM, Norris SJ, Weinstock GM, et al. Complete genome sequence of Treponema pallidum, the syphilis spirochete. Science. 1998;281:375–388. doi: 10.1126/science.281.5375.375. [DOI] [PubMed] [Google Scholar]
  13. Glockner G, Lehmann R, Romualdi A, Pradella S, Schulte-Spechtel U, Schilhabel M, Wilske B, Suhnel J, Platzer M. Comparative analysis of the Borrelia garinii genome. Nucleic Acids Res. 2004;32:6038–6046. doi: 10.1093/nar/gkh953. [DOI] [PMC free article] [PubMed] [Google Scholar]
  14. Gonnet P, Rudd KE, Lisacek F. Fine-tuning the prediction of sequences cleaved by signal peptidase II: a curated set of proven and predicted lipoproteins of Escherichia coli K-12. Proteomics. 2004;4:1597–1613. doi: 10.1002/pmic.200300749. [DOI] [PubMed] [Google Scholar]
  15. Guo BP, Brown EL, Dorward DW, Rosenberg LC, Hook M. Decorin-binding adhesins from Borrelia burgdorferi. Mol Microbiol. 1998;30:711–723. doi: 10.1046/j.1365-2958.1998.01103.x. [DOI] [PubMed] [Google Scholar]
  16. Haake DA. Spirochaetal lipoproteins and pathogenesis. Microbiology. 2000;146:1491–1504. doi: 10.1099/00221287-146-7-1491. [DOI] [PMC free article] [PubMed] [Google Scholar]
  17. Haake DA, Chao G, Zuerner RL, Barnett JK, Barnett D, Mazel M, Matsunaga J, Levett PN, Bolin CA. The leptospiral major outer membrane protein LipL32 is a lipoprotein expressed during mammalian infection. Infect Immun. 2000;68:2276–2285. doi: 10.1128/iai.68.4.2276-2285.2000. [DOI] [PMC free article] [PubMed] [Google Scholar]
  18. Hellwage J, Meri T, Heikkila T, Alitalo A, Panelius J, Lahdenne P, Seppala IJ, Meri S. The complement regulator factor H binds to the surface protein OspE of Borrelia burgdorferi. J Biol Chem. 2001;276:8427–8435. doi: 10.1074/jbc.M007994200. [DOI] [PubMed] [Google Scholar]
  19. Howe TR, Mayer LW, Barbour AG. A single recombinant plasmid expressing two major outer surface proteins of the Lyme disease spirochete. Science. 1985;227:645–646. doi: 10.1126/science.3969554. [DOI] [PubMed] [Google Scholar]
  20. Juncker AS, Willenbrock H, Von Heijne G, Brunak S, Nielsen H, Krogh A. Prediction of lipoprotein signal peptides in Gram-negative bacteria. Protein Sci. 2003;12:1652–1662. doi: 10.1110/ps.0303703. [DOI] [PMC free article] [PubMed] [Google Scholar]
  21. Kornacki JA, Oliver DB. Lyme-disease-causing Borrelia species encode multiple lipoproteins homologous to peptide-binding proteins of ABC-type transporters. Infect Immun. 1998;66:4115–4122. doi: 10.1128/iai.66.9.4115-4122.1998. [DOI] [PMC free article] [PubMed] [Google Scholar]
  22. Kyte J, Doolittle RF. A simple method for displaying the hydropathic character of a protein. J Mol Biol. 1982;157:105–132. doi: 10.1016/0022-2836(82)90515-0. [DOI] [PubMed] [Google Scholar]
  23. Madan Babu M, Sankaran K. DOLOP – database of bacterial lipoproteins. Bioinformatics. 2002;18:641–643. doi: 10.1093/bioinformatics/18.4.641. [DOI] [PubMed] [Google Scholar]
  24. Matsunaga J, Barocchi MA, Croda J, et al. Pathogenic Leptospira species express surface-exposed proteins belonging to the bacterial immunoglobulin superfamily. Mol Microbiol. 2003;49:929–945. doi: 10.1046/j.1365-2958.2003.03619.x. [DOI] [PMC free article] [PubMed] [Google Scholar]
  25. Minamino T, Namba K. Self-assembly and type III protein export of the bacterial flagellum. J Mol Microbiol Biotechnol. 2004;7:5–17. doi: 10.1159/000077865. [DOI] [PubMed] [Google Scholar]
  26. Mount DW. Bioinformatics: Sequence and Genome Analysis. Cold Spring Harbor, NY: Cold Spring Harbor Laboratory; 2001. [Google Scholar]
  27. Nakai K, Horton P. Psort: a program for detecting sorting signals in proteins and predicting their subcellular localization. Trends Biochem Sci. 1999;24:34–36. doi: 10.1016/s0968-0004(98)01336-x. [DOI] [PubMed] [Google Scholar]
  28. Nascimento AL, Ko AI, Martins EA, et al. Comparative genomics of two Leptospira interrogans serovars reveals novel insights into physiology and pathogenesis. J Bacteriol. 2004;186:2164–2172. doi: 10.1128/JB.186.7.2164-2172.2004. [DOI] [PMC free article] [PubMed] [Google Scholar]
  29. Paetzel M, Karla A, Strynadka NC, Dalbey RE. Signal peptidases. Chem Rev. 2002;102:4549–4580. doi: 10.1021/cr010166y. [DOI] [PubMed] [Google Scholar]
  30. Probert WS, Johnson BJ. Identification of a 47 kDa fibronectin-binding protein expressed by Borrelia burgdorferi isolate B31. Mol Microbiol. 1998;30:1003–1015. doi: 10.1046/j.1365-2958.1998.01127.x. [DOI] [PubMed] [Google Scholar]
  31. Ren SX, Fu G, Jiang XG, et al. Unique physiological and pathogenic features of Leptospira interrogans revealed by whole-genome sequencing. Nature. 2003;422:888–893. doi: 10.1038/nature01597. [DOI] [PubMed] [Google Scholar]
  32. Schwan TG, Piesman J, Golde WT, Dolan MC, Rosa PA. Induction of an outer surface protein on Borrelia burgdorferi during tick feeding. Proc Natl Acad Sci U S A. 1995;92:2909–2913. doi: 10.1073/pnas.92.7.2909. [DOI] [PMC free article] [PubMed] [Google Scholar]
  33. Seshadri R, Myers GS, Tettelin H, et al. Comparison of the genome of the oral pathogen Treponema denticola with other spirochete genomes. Proc Natl Acad Sci U S A. 2004;101:5646–5651. doi: 10.1073/pnas.0307639101. [DOI] [PMC free article] [PubMed] [Google Scholar]
  34. Shang ES, Summers TA, Haake DA. Molecular cloning and sequence analysis of the gene encoding LipL41, a surface-exposed lipoprotein of pathogenic Leptospira species. Infect Immun. 1996;64:2322–2330. doi: 10.1128/iai.64.6.2322-2330.1996. [DOI] [PMC free article] [PubMed] [Google Scholar]
  35. Stewart EJ, Katzen F, Beckwith J. Six conserved cysteines of the membrane protein DsbD are required for the transfer of electrons from the cytoplasm to the periplasm of Escherichia coli. EMBO J. 1999;18:5963–5971. doi: 10.1093/emboj/18.21.5963. [DOI] [PMC free article] [PubMed] [Google Scholar]
  36. von Heijne G. The structure of signal peptides from bacterial lipoproteins. Protein Eng. 1989;2:531–534. doi: 10.1093/protein/2.7.531. [DOI] [PubMed] [Google Scholar]
  37. Zhang P, Cheng X, Duhamel GE. Cloning and DNA sequence analysis of an immunogenic glucose-galactose MglB lipoprotein homologue from Brachyspira pilosicoli, the agent of colonic spirochetosis. Infect Immun. 2000;68:4559–4565. doi: 10.1128/iai.68.8.4559-4565.2000. [DOI] [PMC free article] [PubMed] [Google Scholar]

Associated Data

This section collects any data citations, data availability statements, or supplementary materials included in this article.

Supplementary Materials

B. burgdorferi
NIHMS97687-supplement-B__burgdorferi.pdf (342.8KB, pdf)
B. garinii
NIHMS97687-supplement-B__garinii.pdf (141.1KB, pdf)
L. int sv Cop
NIHMS97687-supplement-L__int_sv_Cop.pdf (625.2KB, pdf)
L. int sv Lai
NIHMS97687-supplement-L__int_sv_Lai.pdf (475.5KB, pdf)
T. denticola
NIHMS97687-supplement-T__denticola.pdf (363.9KB, pdf)
T. pallidum
NIHMS97687-supplement-T__pallidum.pdf (149.5KB, pdf)

RESOURCES