Computational comparative study of tuberculosis proteomes using a model learned from signal peptide structures.

Abstract

Secretome analysis is important in pathogen studies. A fundamental and convenient way to identify secreted proteins is to first predict signal peptides, which are essential for protein secretion. However, signal peptides are highly complex functional sequences that are easily confused with transmembrane domains. Such confusion would obviously affect the discovery of secreted proteins. Transmembrane proteins are important drug targets, but very few transmembrane protein structures have been determined experimentally; hence, prediction of the structures is essential. In the field of structure prediction, researchers do not make assumptions about organisms, so there is a need for a general signal peptide predictor.To improve signal peptide prediction without prior knowledge of the associated organisms, we present a machine-learning method, called SVMSignal, which uses biochemical properties as features, as well as features acquired from a novel encoding, to capture biochemical profile patterns for learning the structures of signal peptides directly.We tested SVMSignal and five popular methods on two benchmark datasets from the SPdb and UniProt/Swiss-Prot databases, respectively. Although SVMSignal was trained on an old dataset, it performed well, and the results demonstrate that learning the structures of signal peptides directly is a promising approach. We also utilized SVMSignal to analyze proteomes in the entire HAMAP microbial database. Finally, we conducted a comparative study of secretome analysis on seven tuberculosis-related strains selected from the HAMAP database. We identified ten potential secreted proteins, two of which are drug resistant and four are potential transmembrane proteins.SVMSignal is publicly available at http://bio-cluster.iis.sinica.edu.tw/SVMSignal. It provides user-friendly interfaces and visualizations, and the prediction results are available for download.

Full text

Page 1

Computational Comparative Study of Tuberculosis
Proteomes Using a Model Learned from Signal Peptide
Structures
Jhih-Siang Lai, Cheng-Wei Cheng, Ting-Yi Sung*, Wen-Lian Hsu*
Institute of Information Science, Academia Sinica, Taipei, Taiwan
Abstract
Secretome analysis is important in pathogen studies. A fundamental and convenient way to identify secreted proteins is to
first predict signal peptides, which are essential for protein secretion. However, signal peptides are highly complex
functional sequences that are easily confused with transmembrane domains. Such confusion would obviously affect the
discovery of secreted proteins. Transmembrane proteins are important drug targets, but very few transmembrane protein
structures have been determined experimentally; hence, prediction of the structures is essential. In the field of structure
prediction, researchers do not make assumptions about organisms, so there is a need for a general signal peptide
predictor. To improve signal peptide prediction without prior knowledge of the associated organisms, we present a
machine-learning method, called SVMSignal, which uses biochemical properties as features, as well as features acquired
from a novel encoding, to capture biochemical profile patterns for learning the structures of signal peptides directly. We
tested SVMSignal and five popular methods on two benchmark datasets from the SPdb and UniProt/Swiss-Prot databases,
respectively. Although SVMSignal was trained on an old dataset, it performed well, and the results demonstrate that
learning the structures of signal peptides directly is a promising approach. We also utilized SVMSignal to analyze proteomes
in the entire HAMAP microbial database. Finally, we conducted a comparative study of secretome analysis on seven
tuberculosis-related strains selected from the HAMAP database. We identified ten potential secreted proteins, two of which
are drug resistant and four are potential transmembrane proteins. SVMSignal is publicly available at http://bio-cluster.iis.
sinica.edu.tw/SVMSignal. It provides user-friendly interfaces and visualizations, and the prediction results are available for
download.
Citation: Lai J-S, Cheng C-W, Sung T-Y, Hsu W-L (2012) Computational Comparative Study of Tuberculosis Proteomes Using a Model Learned from Signal Peptide
Structures. PLoS ONE 7(4): e35018. doi:10.1371/journal.pone.0035018
Editor: Bin Xue, University of South Florida, United States of America
Received November 9, 2011; Accepted March 8, 2012; Published April 9, 2012
Copyright: ? 2012 Lai et al. This is an open-access article distributed under the terms of the Creative Commons Attribution License, which permits unrestricted
use, distribution, and reproduction in any medium, provided the original author and source are credited.
Funding: This work was supported in part by the thematic program of Academia Sinica under grant AS95ASIA03 and the National Science Council under grant
NSC 100-2319-B-010 -002. The funders had no role in study design, data collection and analysis, decision to publish, or preparation of the manuscript.
Competing Interests: The authors have declared that no competing interests exist.
* E-mail: tsung@iis.sinica.edu.tw (TYS); hsu@iis.sinica.edu.tw (WLH)
Introduction
Signal peptides are short sequences that start from the N-
terminus and control protein secretion. They are related to drug
targets, protein production, and even biomarker discovery [1–4].
Normally, signal peptides in proteins are recognized and cleaved
by their corresponding proteases, and then the cleaved proteins
are secreted [5]. In some cases, however, instead of being cleaved,
the signal peptides form signal anchors, which are a type of
transmembrane protein [6]. Moreover, as shown by Gierasch [7]
signal peptides are interchangeable as well as highly tolerant, i.e.,
they allow some mutations. Thus it is important to identify signal
peptides in proteins.
Proteins targeting to organelles or outside of the cell sometimes
need a cleavable signal peptide. Signal peptide has its systematic
structure, an amino-terminal positively charged region (n-region),
followed by a central, hydrophobic region (h-region), then
followed by a more polar carboxy-terminal region (c-region) [6].
The hydrophobic core of h-region could be recognized by the SRP
(signal recognition particle). C-region usually contains a motif
before the cleavage site that can be cleaved by appropriate
protease. For example, the bacterial signal peptides consist of
positive charge residues, hydrophobic core, and a motif such as
Ala-X-Ala, just before the cleavage site to direct the protein going
through the Sec pathway. The Tat signal peptide also has the
above (n, h, c)-regions structure, particularly having consecutive
arginines in n-region, to direct the protein going through the Tat
pathway. The lipoprotein signal peptide also has the above
structure, and particularly a cysteine follows the cleavage site for
lipid modification.
Modifying the residues of these cleavable signal peptides may
affect protein secretion. For example, the secretion efficiency may
be mediated by hydrophobicity in h-region and charge in n-region
[8,9]. Modification at cleavage site may extend or shorten the
mature protein sequence, and then may slightly alter the protein
structure. Completely removing c-region may yield the signal
peptide uncleaved and form a signal anchor for transmembrane
proteins.
Since the pairwise sequence similarity of signal peptides is
usually low, they cannot be detected simply by sequence alignment
analysis [10]. To predict signal peptides, rules have been devised
for the analysis of signal peptide cleavage sites [11]. Combined the
rules with the signal peptide structure, i.e., (n, h, c)-regions, can
PLoS ONE | www.plosone.org1 April 2012 | Volume 7 | Issue 4 | e35018

Page 2

lead to accurate signal peptide prediction. However, a major
difficulty with this technique is that signal peptides may be
misclassified as transmembrane domains, and vice versa, because
both regions contain hydrophobic cores, and hydrophobicity is a
key feature of signal peptide prediction methods [6]. Transmem-
brane proteins are important drug targets, but very few
transmembrane protein structures have been determined exper-
imentally. Accurate prediction of transmembrane protein struc-
tures is essential. If the transmembrane domain of a transmem-
brane protein is misclassified as a signal peptide, or vice versa, it
would lead to incorrect transmembrane protein structure predic-
tion and also inaccurate secretion analysis. Several methods have
been developed for signal peptide prediction based on three
domains of a signal peptide structure. For example, SignalP uses
neural networks and hidden Markov models to construct the (n, h,
c)-regions and improve the disambiguation between transmem-
brane proteins and signal peptides [12–14]; and PrediSi exploits a
position weight matrix to predict cleavage sites [15]. Phobius,
which uses a hidden Markov model, was the first predictor to
predict both the topologies of transmembrane proteins and signal
peptides [16]. RPSP is a neural network-based method designed
for proteomic analysis [17]; and Philius uses dynamic Bayesian
networks to model transmembrane protein topology and signal
peptide [18].
Although the structure of (n, h, c)-regions provides good clues
for signal peptide prediction and clearer rules have been defined
for the c-region, the n-region and h-region have ambiguous
boundaries and are diversified in terms of sequences and
organisms. Thus in this paper we present a machine learning
approach based on support vector machines (SVMs), called
SVMSignal, to learn the structures of signal peptides and classify
signal peptides from transmembrane proteins. SVMSignal uses
basic biochemical profiles as features and also defines a novel
feature to capture the inter-profile relationships that describe the
structural properties of signal sequences.
We compared the performance of SVMSignal with that of
existing methods on two benchmark datasets, the signal peptide
database SPdb [19] and a hybrid dataset compiled from UniProt/
Swiss-Prot [20] and PDBTM [21] (Protein Data Bank of
Transmembrane Proteins), a transmembrane protein structure
database. We used experimentally determined signal peptide
sequences from SPdb and UniProt/Swiss-Prot as the signal
peptide benchmark to demonstrate the sensitivities of various
signal peptide predictors. In order to evaluate the specificities and
classification abilities of different predictors, we collected soluble
proteins from UniProt/Swiss-Prot and transmembrane proteins
from both UniProt/Swiss-Prot and PDBTM as benchmark. Our
method SVMSignal achieved good performance on the bench-
mark datasets.
Signal peptides are crucial to protein secretion, and secretome
analysis is important in pathogen studies because it has been
shown that some hosts are affected by proteins secreted from
bacteria [22,23]. There are at least six known secretion systems,
i.e., Types I to VI, as well as the Sec and Tat pathways in Gram-
negative bacteria [5]. Signal peptides are not responsible for all of
the secretions. Although proteins containing signal peptides (called
signal peptide proteins hereafter) cannot characterize the full
secretome, it is still necessary to use sequence information to
discover potential secreted proteins for further research [24,25].
HAMAP (High-quality Automated and Manual Annotation of
microbial Proteomes) microbial database [26] provides 1292
curated microbial proteomes and description of microbial
pathogens, and proteomes in the database are also integrated
with UniProt/Swiss-Prot. It is considered as an important
database for pathogen studies [27,28]. We applied our method
to HAMAP to predict signal peptides in proteomes so that signal
peptide-dependent secreted proteins could be identified.
From the pathogens recorded in the HAMAP database, we
selected tuberculosis to conduct secretome analysis because it is a
historical human interactive pathogen, and it is still difficult to
treat in some cases, e.g., drug-resistant strains and immunocom-
promised HIV patients. Recently, Walzl et al. presented a review
paper that suggests using an ‘‘omics’’ approach to find potential
immunological biomarkers of tuberculosis [29]. To this end, we
made an in-depth analysis of proteins with signal peptides in
Mycobacterium tuberculosis and Mycobacterium bovis, which can
cause tuberculosis in humans and cattle, respectively. It is also
possible for Mycobacterium bovis to infect humans via certain
foods, such as unpasteurized milk. We selected seven strains from
the two pathogens for study, including virulent strains, attenuated
strain, and drug-resistant strains. Due to their close lineage, the
comparative study of signal peptide-dependent secreted proteins
belonging to different strains can provide clues about the
tuberculosis mechanism, virulence factors or biomarkers. The
signal peptide prediction results of the entire HAMAP database
are provided for further research.
Results
Performance evaluation of SVMSignal
We compared the performance of SVMSignal with that of five
existing predictors, namely, Phobius [16], RPSP [17], Philius [18],
SignalP [12–14] (specifically SignalP 3.0), and PrediSi [15]. Each
of the last two predictors provides three models specifically
developed for Gram-positive bacteria, Gram-negative bacteria and
eukaryotes. We included all of the models in the performance
evaluation.
To avoid over-estimating SVMSignal’s performance in com-
parison with the other predictors, we used the original Phobius
2004 dataset to train our model. We tested all of the methods on
the two benchmark datasets mentioned earlier, i.e., the signal
peptide database SPdb [19] (called the SPDB dataset hereafter),
and a hybrid dataset compiled from UniProt/Swiss-Prot [20] and
PDBTM [21] in 2010. Note that we filtered each of the
benchmark datasets by removing sequences with over 30%
similarity to sequences in the training dataset and within the
dataset itself. Both datasets were decomposed by organism into
mixed, eukaryotes, and bacteria for analysis; archaea and viruses
were omitted because there was insufficient data for them. We
further decomposed the second benchmark dataset into SP
(proteins with signal peptides), TM (proteins with transmembrane
domains only), and G (proteins without signal peptides or
transmembrane domains) datasets. Note that the SP dataset
contains very few proteins with both transmembrane domains and
signal peptides. All of the above datasets are provided in the
Supporting Information (Dataset S1).
To evaluate the performance of various predictors, we define
sequences with signal peptides as positive (P) data, and those
without signal peptides as negative data (N). We use the following
metrics to evaluate the performance sensitivity~TP= TPzFN
specificity~TN= TNzFP
ð
FPzTNzFNÞ, where T and F denote ‘‘true’’ and ‘‘false’’,
respectively. Accuracy alone is insufficient to reflect a global
perspective of good sensitivity in signal peptides and speci-
ficity in non-signal peptides, especially specificity in transmem-
brane domains required by a good predictor, due to the data
imbalance. Therefore, to evaluate the performance from a
global perspective, we use MCC (Matthew’s correlation coe-
ðÞ,
Þ and accuracy~ TNzTP
ðÞ= TPz
ð
Tuberculosis Analysis by Signal Peptide Prediction
PLoS ONE | www.plosone.org2 April 2012 | Volume 7 | Issue 4 | e35018

Page 3

fficient), which is formulated as MCC~ TP ? TN{FP ? FN
ffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
Evaluating SVMSignal’s ability to disambiguate proteins
with and without signal peptides
Predictors sometimes confuse signal peptides with transmem-
brane domains due to hydrophobic composition. We evaluated
SVMSignal’s ability to classify signal peptides and non-signal
peptides, i.e., to disambiguate the SP dataset from the TM and
TM+G datasets, denoted by SP/TM and SP/(TM+G), respec-
tively. The MCCs of SVMSignal and the other predictors on
different organism datasets are detailed in Table 1. SVMSignal
outperformed the other predictors on the mixed organisms and
eukaryotes in SP/TM and SP/(TM+G), except on the eukaryotes
in SP/TM as the second best, with 2.24% slightly worse than the
best SignalP model. Notably, in the mixed organisms, SVMSignal
outperformed the second best predictor by 0.54% and 3.31% on
SP/TM and SP/(TM+G), respectively. It also achieved the second
best performance on the bacteria in SP/TM and the third best
performance on the bacteria in SP/(TM+G).
As shown in Table 2, the accuracy of SVMSignal’s classification
with respect to different organisms ranged from 90% to 92% for
SP/TM and exceeded 96% for SP/(TM+G). SVMSignal achieved
the best accuracy in classifying SP/(TM+G) of the mixed
organisms and eukaryote datasets. It achieved the second best
accuracy in SP/TM of mixed organisms and SP/(TM+G) of
bacteria with at most 0.4% difference to the best accuracy, and the
third best accuracy in the remaining datasets with at most 2% gap
from the best accuracy.
ðÞ=
TPzFP
ðÞ TPzFN
ðÞ TNzFN
ðÞ TNzFP
ðÞ
p
??.
Sensitivity and Specificity of SVMSignal’s predictions on
the signal peptide benchmark datasets
The sensitivities of SVMSignal and the other predictors on the
signal peptide benchmark datasets, SPDB and SP, are shown in
Table 3. Although SVMSignal was trained on an old dataset and it
was not trained on different organisms, it performed well
compared to most of the predictors. As SignalP was trained
specifically on different organisms, reflected by variations in the
sensitivities of different models, SVMSignal performed slightly
inferior to SignalP by at most 0.92% on the mixed organism
datasets, at most 2.88 on the eukaryotes datasets, and at most
2.58% on the bacteria datasets.
We compared the specificities of SVMSignal and the other
predictors on proteins without signal peptides in the TM and
TM+G datasets. As shown in Table 4, for the TM classification,
SVMSignal achieved the best performance on the mixed organism
and eukaryote datasets, and was only outperformed by Philius on
the bacteria dataset. For the TM+G classification, SVMSignal’s
performance differed from the best performance by at most
1.48%; however, on the three organism datasets of TM+G, it
achieved over 97% specificity, which was very close to the
specificity achieved by SignalP.
Application of SVMSignal to tuberculosis pathogen study
We applied signal peptide prediction to secretome analysis in
tuberculosis pathogen since tuberculosis is a well-known infectious
disease that can be fatal. Notably, signal peptides of tuberculosis
proteomes are related to its virulence [30–32]. Since very few
proteins of tuberculosis strains have been annotated, we used
SVMSignal to predict signal peptides and performed a compar-
ative study of the predicted signal peptide proteins of tuberculosis
Table 1. MCCs (%) of various predictors on classification of proteins with and without signal peptides.
Species
Data set
Cleavable
SVMSignal
Phobius
Philius
RPSP
SignalP
NN+ +Euk
SignalP
HMM+ +Euk
SignalP
NN,G+ +
SignalP
HMM,G+ +
SignalP
NN,G2 2
SignalP
HMM,G2 2
PrediSi
Euk
PrediSi
G+ +
PrediSi
G2 2
Mixed
SP/TM
12.83%
66.24%
65.70%
64.36%
33.23%
42.22%
57.12%
42.88%
28.42%
46.54%
37.98%
20.96%
15.53%
18.73%
SP/(TM+G)
32.73%
86.25%
82.94%
76.93%
82.21%
81.03%
78.65%
63.08%
71.21%
64.90%
75.75%
77.07%
54.51%
58.35%
Euk
SP/TM
12.68%
58.11%
54.91%
46.60%
39.09%
60.35%
48.79%
22.97%
19.58%
29.38%
26.71%
28.30%
6.24%
6.38%
SP/(TM+G)
35.48%
87.83%
83.13%
75.34%
86.53%
84.54%
78.15%
56.47%
63.87%
58.80%
70.72%
80.31%
45.17%
50.18%
Bac
SP/TM
9.81%
72.14%
71.80%
75.05%
27.79%
26.11%
59.73%
65.18%
59.78%
64.14%
66.21%
11.51%
40.06%
47.02%
SP/(TM+G)
29.57%
84.72%
83.92%
80.76%
74.45%
75.34%
80.57%
79.17%
86.07%
79.95%
85.89%
71.13%
72.97%
73.81%
doi:10.1371/journal.pone.0035018.t001
Tuberculosis Analysis by Signal Peptide Prediction
PLoS ONE | www.plosone.org3 April 2012 | Volume 7 | Issue 4 | e35018

Page 4

Table 2. Accuracies (%) of various predictors on classification of proteins with and without signal peptides.
Species
Data set
Cleavable
SVMSignal
Phobius
Philius
RPSP
SignalP
NN+ +Euk
SignalP
HMM+ +Euk
SignalP
NN,G+ +
SignalP
HMM,G+ +
SignalP
NN,G2 2
SignalP
HMM,G2 2
PrediSi
Euk
PrediSi
G+ +
PrediSi
G2 2
Mixed
SP/TM
88.56%
91.14%
91.54%
90.85%
78.61%
87.56%
89.85%
82.79%
68.36%
85.07%
77.11%
81.29%
56.82%
63.98%
SP/(TM+G)
56.43%
96.72%
95.71%
93.80%
96.04%
95.18%
94.38%
89.28%
93.77%
89.64%
94.51%
94.30%
90.12%
90.40%
Euk
SP/TM
94.05%
92.02%
92.33%
90.77%
82.94%
94.05%
92.18%
79.81%
58.69%
83.10%
70.89%
86.85%
46.64%
55.40%
SP/(TM+G)
57.71%
96.79%
95.26%
92.48%
96.64%
95.61%
93.41%
85.46%
91.53%
85.99%
92.78%
94.61%
87.14%
87.54%
Bac
SP/TM
79.82%
90.21%
90.50%
91.10%
71.51%
76.56%
86.05%
88.43%
86.94%
88.43%
89.32%
72.11%
75.67%
80.42%
SP/(TM+G)
53.66%
96.38%
96.10%
95.20%
94.39%
94.01%
95.34%
94.72%
96.76%
94.91%
96.62%
92.86%
93.86%
93.67%
doi:10.1371/journal.pone.0035018.t002
Table 3. Sensitivities (%) of various predictors on the signal peptide protein benchmark datasets.
Species
Dataset
Cleavable
SVMSignal
Phobius
Philius
RPSP
SignalP
NN+ +Euk
SignalP
HMM+ +Euk
SignalP
NN,G+ +
SignalP
HMM,G+ +
SignalP
NN,G2 2
SignalP
HMM,G2 2
PrediSi
Euk
PrediSi
G+ +
PrediSi
G2 2
Mixed
SPDB
99.24%
95.27%
93.90%
94.05%
85.82%
96.19%
96.04%
80.34%
58.38%
84.76%
67.53%
91.46%
49.54%
58.38%
SP
97.67%
91.12%
92.01%
91.12%
79.91%
91.12%
91.34%
83.68%
67.26%
86.24%
76.80%
86.13%
55.27%
63.71%
Euk
SPDB
100.00%
95.97%
94.43%
94.43%
85.99%
98.85%
97.50%
76.97%
50.48%
82.15%
61.61%
92.71%
42.80%
52.98%
SP
98.52%
91.76%
92.59%
91.43%
82.54%
94.40%
93.08%
80.72%
57.17%
83.86%
70.02%
88.47%
45.47%
55.19%
Bac
SPDB
96.55%
93.97%
92.24%
93.10%
84.48%
86.21%
90.52%
95.69%
93.10%
96.55%
94.83%
87.07%
80.17%
82.76%
SP
95.99%
90.51%
91.61%
90.88%
74.45%
84.31%
87.96%
90.51%
90.15%
91.24%
92.34%
81.75%
76.64%
82.48%
doi:10.1371/journal.pone.0035018.t003
Tuberculosis Analysis by Signal Peptide Prediction
PLoS ONE | www.plosone.org4April 2012 | Volume 7 | Issue 4 | e35018

Page 5

strains. Our objective was to identify some critical proteins that
would be worth further investigation.
For our study, we selected seven interesting strains related to
tuberculosis: four strains from Mycobacterium tuberculosis and
three strains from Mycobacterium bovis in the HAMAP database
[26]. Specifically, the four strains of Mycobacterium tuberculosis
are MYCTU, MYCTA, MYCTF, MYCTK, where MYCTU is
the virulent H37Rv strain, MYCTA is the highly attenuated
H37Ra strain, MYCTF is the virulent strain family 11, and
MYCTK is the drug resistant KZN 1435 strain. The other three
strains are one strain of Mycobacterium bovis coded as MYCBO,
and two strains of Mycobacterium bovis bacillus Calmette–Gue ´rin
(BCG) coded as MYCBP and MYCBT, respectively. MYCBP is a
BCG strain developed at the Pasteur Institute in Paris and
MYCBT is another BCG strain used in Japan. In the following
discussion, we use the species code instead of the full name to
represent each of the strains and the UniProt/Swiss-Prot accession
number to represent a protein.
To identify signal peptide proteins that characterize each
pathogen strain, we first applied SVMSignal to the seven
proteomes of the strains, and we were only interested in the
predicted signal peptide proteins. The statistics of the signal peptide
proteins in the seven proteomes, i.e., four strains from Mycobac-
terium tuberculosis and three strains from Mycobacterium bovis,
are shown in Table 5. There are 334 to 345 signal peptide
proteins, i.e., 8.33% to 8.73% of the total proteins, in each
proteome. Signal peptide proteins predicted to contain transmem-
brane domains account for 1.59% to 1.88% of the proteins in a
proteome. Furthermore, the number of multi-spanning trans-
membrane proteins is double that of single-spanning transmem-
brane proteins. The average length of signal peptides is 28 to 29
residues, and the average length of the entire bacteria pathogen in
the HAMAP database is 26.02. See the Discussion section and
Supporting Information (Datasets S2 and S3) for more details.
Comparative Study on the selected tuberculosis strains
To compare the seven tuberculosis strains, we used CD-HIT
[33] and the 30% similarity threshold to cluster the signal peptide
proteins by their sequences with signal peptides removed, i.e., their
secreted sequences. Since secreted proteins affect their hosts and
are of interest to us, we used them to cluster the corresponding
signal peptide proteins. The process generated 313 clusters, of
which 10 contained only one protein, whose similarity to all the
other sequences was less than 30%. We call such a signal peptide
protein a unique protein because it does not occur in any of the other
six proteomes. Table 6 lists all the unique proteins found in the
seven proteomes. MYCBP and MYCBT do not contain any
unique proteins, while each of the other five strains has at least one
unique protein. Interestingly, the human tuberculosis drug
resistant strain MYCTK has six unique proteins. Furthermore,
we removed signal peptides from the unique proteins and used
TMHMM [34,35] to predict whether their secreted sequences
contain transmembrane domains. Four of the unique proteins,
O06239, A5WU15, C6DWG6 and Q7U0W0, were predicted as
multi-spanning transmembrane proteins and also annotated as
transmembrane proteins in the UniProt/Swiss-Prot database.
Next, for each unique protein, we used BLAST to search
against the seven proteomes for similar proteins and the results are
shown in Table 6. Note that the homologous protein of each
unique protein is not signal peptide protein. The alignment results
of the homologous protein pairs are provided in Supporting
Information (Text S1). Six of the homologous protein pairs share a
very high sequence similarity (over 95%), namely, pairs O06239/
C6DPS4 (97.87%), A5U3R8/O07733 (95.54%), A5WU15/
Table 4. Specificities (%) of various predictors on the non-signal peptide protein benchmark datasets.
Species
Data set
Cleavable
SVMSignal
Phobius
Philius
RPSP
SignalP
NN+ +Euk
SignalP
HMM+ +Euk
SignalP
NN,G+ +
SignalP
HMM,G+ +
SignalP
NN,G2 2
SignalP
HMM,G2 2
PrediSi
Euk
PrediSi
G+ +
PrediSi
G2 2
Mixed
TM
9.62%
91.35%
87.50%
88.46%
67.31%
56.73%
76.92%
75.00%
77.88%
75.00%
79.81%
39.42%
70.19%
66.35%
TM+G
50.09%
97.58%
96.28%
94.21%
98.52%
95.80%
94.85%
90.13%
97.85%
90.17%
97.24%
95.56%
95.48%
94.50%
Euk
TM
9.38%
96.88%
87.50%
78.13%
90.63%
87.50%
75.00%
62.50%
87.50%
68.75%
87.50%
56.25%
68.75%
59.38%
TM+G
50.38%
97.69%
95.74%
92.67%
99.17%
95.83%
93.47%
86.31%
97.69%
86.37%
96.87%
95.71%
94.62%
93.35%
Bac
TM
9.52%
88.89%
85.71%
92.06%
58.73%
42.86%
77.78%
79.37%
73.02%
76.19%
76.19%
30.16%
71.43%
71.43%
TM+G
47.32%
97.26%
96.77%
95.84%
97.37%
95.46%
96.44%
95.35%
97.76%
95.46%
97.26%
94.53%
96.44%
95.35%
doi:10.1371/journal.pone.0035018.t004
Tuberculosis Analysis by Signal Peptide Prediction
PLoS ONE | www.plosone.org5April 2012 | Volume 7 | Issue 4 | e35018