Computational prediction of human proteins that can be secreted into the bloodstream.
ABSTRACT We present a novel computational method for predicting which proteins from highly and abnormally expressed genes in diseased human tissues, such as cancers, can be secreted into the bloodstream, suggesting possible marker proteins for follow-up serum proteomic studies. A main challenging issue in tackling this problem is that our understanding about the downstream localization after proteins are secreted outside the cells is very limited and not sufficient to provide useful hints about secretion to the bloodstream. To bypass this difficulty, we have taken a data mining approach by first collecting, through extensive literature searches, human proteins that are known to be secreted into the bloodstream due to various pathological conditions as detected by previous proteomic studies, and then asking the question: 'what do these secreted proteins have in common in terms of their physical and chemical properties, amino acid sequence and structural features that can be used to predict them?' We have identified a list of features, such as signal peptides, transmembrane domains, glycosylation sites, disordered regions, secondary structural content, hydrophobicity and polarity measures that show relevance to protein secretion. Using these features, we have trained a support vector machine-based classifier to predict protein secretion to the bloodstream. On a large test set containing 98 secretory proteins and 6601 non-secretory proteins of human, our classifier achieved approximately 90% prediction sensitivity and approximately 98% prediction specificity. Several additional datasets are used to further assess the performance of our classifier. On a set of 122 proteins that were found to be of abnormally high abundance in human blood due to various cancers, our program predicted 62 as blood-secreted proteins. By applying our program to abnormally highly expressed genes in gastric cancer and lung cancer tissues detected through microarray gene expression studies, we predicted 13 and 31 as blood secreted, respectively, suggesting that they could serve as potential biomarkers for these two cancers, respectively. Our study demonstrated that our method can provide highly useful information to link genomic and proteomic studies for disease biomarker discovery. Our software can be accessed at http://csbl1.bmb.uga.edu/cgi-bin/Secretion/secretion.cgi.
- SourceAvailable from: Minesh Mehta[Show abstract] [Hide abstract]
ABSTRACT: A comparative analysis of genome-scale transcriptomic data of two types of skin cancers, melanoma and basal cell carcinoma in comparison with other cancer types, was conducted with the aim of identifying key regulatory factors that either cause or contribute to the aggressiveness of melanoma, while basal cell carcinoma generally remains a mild disease. Multiple cancer-related pathways such as cell proliferation, apoptosis, angiogenesis, cell invasion and metastasis, are considered, but our focus is on energy metabolism, cell invasion and metastasis pathways. Our findings include the following. (a) Both types of skin cancers use both glycolysis and increased oxidative phosphorylation (electron transfer chain) for their energy supply. (b) Advanced melanoma shows substantial up-regulation of key genes involved in fatty acid metabolism (β-oxidation) and oxidative phosphorylation, with aerobic metabolism being far more efficient than anaerobic glycolysis, providing a source of the energetics necessary to support the rapid growth of this cancer. (c) While advanced melanoma is similar to pancreatic cancer in terms of the activity level of genes involved in promoting cell invasion and metastasis, the main metastatic form of basal cell carcinoma is substantially reduced in this activity, partially explaining why this cancer type has been considered as far less aggressive. Our method of using comparative analyses of transcriptomic data of multiple cancer types focused on specific pathways provides a novel and highly effective approach to cancer studies in general.PLoS ONE 01/2012; 7(1):e30750. · 3.53 Impact Factor
- 02/2012; , ISBN: 978-953-307-812-0
- [Show abstract] [Hide abstract]
ABSTRACT: Proteins can move from blood circulation into salivary glands through active transportation, passive diffusion or ultrafiltration, some of which are then released into saliva and hence can potentially serve as biomarkers for diseases if accurately identified. We present a novel computational method for predicting salivary proteins that come from circulation. The basis for the prediction is a set of physiochemical and sequence features we found to be discerning between human proteins known to be movable from circulation to saliva and proteins deemed to be not in saliva. A classifier was trained based on these features using a support-vector machine to predict protein secretion into saliva. The classifier achieved 88.56% average recall and 90.76% average precision in 10-fold cross-validation on the training data, indicating that the selected features are informative. Considering the possibility that our negative training data may not be highly reliable (i.e., proteins predicted to be not in saliva), we have also trained a ranking method, aiming to rank the known salivary proteins from circulation as the highest among the proteins in the general background, based on the same features. This prediction capability can be used to predict potential biomarker proteins for specific human diseases when coupled with the information of differentially expressed proteins in diseased versus healthy control tissues and a prediction capability for blood-secretory proteins. Using such integrated information, we predicted 31 candidate biomarker proteins in saliva for breast cancer.PLoS ONE 01/2013; 8(11):e80211. · 3.53 Impact Factor
BIOINFORMATICS ORIGINAL PAPER
Vol. 24 no. 20 2008, pages 2370–2375
Computational prediction of human proteins that can be secreted
into the bloodstream
Juan Cui1,∗, Qi Liu1,2, David Puett1, Ying Xu1,3,∗
1Department of Biochemistry and Molecular Biology, University of Georgia, Athens, GA 30602, USA,
2Zhejiang-California International Nanosystems Institute, Zhejiang University, Hangzhou 310029, China and
3Institute of Bioinformatics, University of Georgia, Athens, GA 30602, USA
Received on April 9, 2008; revised on August 6, 2008; accepted on August 7, 2008
Advance Access publication August 12, 2008
Associate Editor: Limsoon Wong
We present a novel computational method for predicting which
proteins from highly and abnormally expressed genes in diseased
human tissues, such as cancers, can be secreted into the
bloodstream, suggesting possible marker proteins for follow-up
serum proteomic studies. A main challenging issue in tackling this
problem is that our understanding about the downstream localization
after proteins are secreted outside the cells is very limited and not
sufficient to provide useful hints about secretion to the bloodstream.
To bypass this difficulty, we have taken a data mining approach by
first collecting, through extensive literature searches, human proteins
that are known to be secreted into the bloodstream due to various
pathological conditions as detected by previous proteomic studies,
and then asking the question: ‘what do these secreted proteins
have in common in terms of their physical and chemical properties,
amino acid sequence and structural features that can be used to
predict them?’ We have identified a list of features, such as signal
peptides, transmembrane domains, glycosylation sites, disordered
regions, secondary structural content, hydrophobicity and polarity
measures that show relevance to protein secretion. Using these
features, we have trained a support vector machine-based classifier
to predict protein secretion to the bloodstream. On a large test set
containing 98 secretory proteins and 6601 non-secretory proteins
of human, our classifier achieved ∼90% prediction sensitivity
and ∼98% prediction specificity. Several additional datasets are
used to further assess the performance of our classifier. On a
set of 122 proteins that were found to be of abnormally high
abundance in human blood due to various cancers, our program
predicted 62 as blood-secreted proteins. By applying our program
to abnormally highly expressed genes in gastric cancer and lung
cancer tissues detected through microarray gene expression studies,
we predicted 13 and 31 as blood secreted, respectively, suggesting
that they could serve as potential biomarkers for these two cancers,
respectively. Our study demonstrated that our method can provide
highly useful information to link genomic and proteomic studies
for disease biomarker discovery. Our software can be accessed at
∗To whom correspondence should be addressed.
Supplementary information: Supplementary data are available at
Alterations in gene and protein expression provide important clues
about the physiological states of a tissue or an organ. During
malignant transformation, genetic alterations in tumor cells can
disrupt autocrine and paracrine signaling networks, leading to the
over-expression of some classes of proteins such as growth factors,
cytokines and hormones that may be secreted outside the cancerous
cells (Hanahan and Weinberg, 2000; Sporn and Roberts, 1985).
through various complex secretion pathways and can potentially be
used as marker proteins for blood or urine tests. Recent genomic
studies on various cancer specimens have identified numerous genes
that are consistently over-expressed and some of these genes encode
secreted proteins (Buckhaults et al., 2001; Welsh et al., 2001,
2003). For example, prostasin and osteopontin genes have elevated
expression levels in ovarian cancer while MIC1 gene is over-
expressed in colorectal, breast and prostate cancers. The increased
abundance of these secretory proteins has been detected in the
serum of patients harboring these cancers compared to the healthy
individuals (Kim et al., 2002; Mok et al., 2001; Welsh et al., 2003).
It has also been found that some of the secreted proteins have
shown varying levels of concentration increases in serum associated
with different developmental stages of cancers, suggesting that they
could possibly be used as markers of both cancer typing and staging
(Huang et al., 2006).
The human serum proteome is a very complex mixture of highly
abundant native serum proteins such as albumin, immunoglobulins,
transferrin, haptoglobin and lipoproteins as well as proteins and
peptides that are secreted from different tissues, diseased or normal
or leak from cells throughout the human body (Adkins et al.,
2002; Schrader and Schulz-Knappe, 2001). A challenging issue
when working with the human serum proteome is that most of
the circulating native blood proteins are orders of magnitude more
abundant than those of the putative proteins of interest. Hence, it
is very difficult to experimentally detect such secreted proteins,
or possibly more native blood proteins without knowing what
proteins we are looking for in blood a priori. It is thus desirable
© The Author 2008. Published by Oxford University Press. All rights reserved. For Permissions, please email: email@example.com
Prediction of human proteins secreted into bloodstream
to investigate computational approaches to predict proteins that are
both abnormally highly expressed in cancer tissues and can get
secreted into the bloodstream, providing a target list for targeted
Numerous studies have been carried out to predict proteins
that can get secreted to cell surface or into the extracellular
prediction servers are available (Guda, 2006; Horton et al., 2007;
Menne et al., 2000; Nair and Rost, 2005). Most of these methods
have been developed based on our general understanding about
protein subcellular localization—localization of most proteins is
done through a cascade of sorting events that are directed by short
(signal) peptides or motifs that enable site-specific uptake, retention
and transport (Doudna and Batey, 2004;Tjalsma et al., 2000).These
programs have been developed using various statistical learning
methods, based on information such as amino acid composition,
co-occurrence of protein domains and annotated protein functions
(Guda, 2006; Mott et al., 2002). A fundamental difference between
our work and the previous work is that while they all are concerned
about whether a protein gets secreted outside of a cell, previous
studies are not concerned about where the proteins will end up but
In order to predict proteins that can get secreted into the
bloodstream, we have collected 305 non-native blood proteins
that have been experimentally found in the blood under various
physiological conditions from the published literature. We have
analyzed these proteins carefully to derive various sequence
and structural features commonly shared by these proteins or
by some subsets of these proteins, including signal peptides,
transmembrane domains, glycosylation sites, disordered regions,
secondary structural content, radius of gyration of a protein
tertiary structure, hydrophobicity and polarity measures that show
relevance to protein secretion. Using these features, we have trained
support vector machine (SVM)-based classifier to distinguish these
experimentally verified blood-secreted proteins from the rest of the
are highly promising, providing very useful information that could
detected through microarray experiments and proteomic studies of
blood serum for effective marker–protein identification.
2.1 Collection of blood-secreted proteins and
We have collected a total of 1620 human proteins that are annotated as
secretory proteins from the Swissprot and SPD database (Chen et al.,
2005), and then determined if any of these proteins have been detected
experimentally in blood by previous studies. We have done this by checking
the 1620 proteins against the known serum protein dataset compiled by the
Plasma Proteome Project (PPP) (Omenn et al., 2005) and a few additional
datasets generated by other serum proteomic studies (Adkins et al., 2002;
Pieper et al., 2003), which consist of a total of ∼16000 proteins. We
found that 305 of the 1620 proteins match at least two peptides with the
∼16000 proteins, and hence we consider that these 305 proteins are secreted
into blood—a common practice for protein identification based on mass
spectrometry data. To ensure the good quality of our dataset, it is should be
noted that we only chose these 305 proteins which meet two criteria (both
secreted and serum/plasma detected), as the positive dataset and did not
myoglobin released into plasma after a heart attack).
To generate a negative dataset of proteins for the classification, we
selected representatives from non-blood-secreted proteins, which should
include both proteins unrelated to secretory pathway and secreted proteins
from each of the Pfam protein families (Bateman et al., 2002) that contain
no previously mentioned blood-secreted proteins as the negative set.
In order to obtain a non-redundant dataset for a final independent
validation step, we used BLAST (Altschul et al., 1997) to remove the
redundant proteins using 20% sequence identity as the cutoff, giving rise to
56 positive and 13716 negative proteins. We then divided the remains, 249
positive and 13246 negative proteins into separate training and testing sets,
respectively, using the following procedure. All the proteins in the positive
set were divided into clusters based on the similarity of our selected features
(see Section 2.2), measured by the Euclidean distance, using a hierarchical
clustering method. A total of 151 clusters were obtained with the ratio
between the maximum intra-cluster distance and the minimum inter-cluster
distance for each cluster, ranging from 0.27 to 0.51. From each cluster, one
representative protein was chosen randomly to form the positive training set.
We do the same for the negative training set. The training set was selected
in this way to ensure it is sufficiently diverse and broadly distributed in the
feature space. The remaining proteins are used as the test set. We repeated
this process to construct five different datasets to train the classifier, which
can be used to assess the stability of our data generation strategy.
We have examined a number of features computed based on protein
sequences and secondary structures that are possibly relevant to the
classification of proteins being blood secreted or not. Some features are
included because they are known to be relevant to protein secretion, while
others are included because of their statistical relevance to our classification
problem. For example, signal peptides and transmembrane domains are
known to be important factors to prediction of extracellularly secreted
proteins. Twin-arginine (TAT) signal peptides, only observed in prokaryotes
so far, are known to be used to export proteins into the periplasmic
compartment or extracellular environment independent of the well-studied
sec-dependent translocation pathway (Bendtsen et al., 2005; Taylor et al.,
2006). We included this motif information in our study to check if it might
be relevant to transporting folded proteins across the human cell membrane.
In addition, it is known that the structures of the capillaries determine that
only proteins under a certain size can diffuse through their walls and get into
the bloodstream. For example, blood proteins are expected to be larger than
size that is up to 400nm in diameter (under some tumor conditions), for their
Hence, we have included information about the protein size and shape in our
initial feature list.Another important feature is the glycosylation sites. It has
been observed that most blood-secreted proteins are glycosylated (Bosques
et al., 2006), including important tumor biomarkers such as prostate-specific
antigen (PSA) and the ovarian cancer marker CA125.
In addition, we have included a number of general features in our
initial feature list, derived from protein sequence, secondary structural and
physicochemical properties widely used in various protein classification
studies, such as protein function prediction and protein–protein interaction
prediction, as reviewed in Cui et al. (2007b), which might be relevant to our
prediction of blood-secreted proteins. Supplementary Table 1 summarizes
the features discussed above. The actual relevance of these features to
our classification problem is assessed using a feature-selection algorithm
presented in the following section.
Features in Supplementary Table 1 can be roughly grouped into four
categories: (i) general sequence features such as amino acid composition,
sequence length and di-peptide composition (Bhasin and Raghava, 2004;
Reczko and Bohr, 1994); (ii) physicochemical properties such as solubility,
J.Cui et al.
unfoldability, disordered regions, hydrophobicity, normalized Van der Waals
volume, polarity, polarizability and charges, (iii) structural properties such
as secondary structural content, solvent accessibility and radius of gyration
and (iv) domains/motifs such as signal peptides, transmembrane domains
and twin-arginine signal peptides motif (TAT). In total, 25 properties are
included in the initial list, which give rise to a 1521-dimensional feature
vector for each protein sequence. Note that for each included property,
different amount of information is needed to encode it in our feature vector
representation of the properties. For example, amino acid composition and
dipeptide composition are represented as a 20- and 400(20×20)-dimensional
feature vector, respectively. The feature vector of the secondary structural
content is a four-dimensional vector, including alpha-helix content, beta-
strand content, coil content and the assigned class by the SSCP program
(Eisenhaber et al., 1996). Our encoding of physicochemical properties is
illustrated by the example of hydrophobicity feature vector: amino acids can
be divided into hydrophobic (CVLIMFW), neutral (GASTPHY) and polar
(RKEDQN) groups. Then three descriptors, composition (C), transition (T)
and distribution (D), are used to describe the global composition with C
being the number of amino acids of a particular group (such as hydrophobic)
divided by the total number of amino acids in the protein sequence (Cai
et al., 2003; Cui et al., 2007; Dubchak et al., 1995); T being the relative
frequency in changing amino acid groups along the protein sequence, and D
denoting the chain length within which the first, 25%, 50%, 75% and 100%
of the amino acids of a particular group is located, respectively. Overall, 21
elements are used to represent these three descriptors: 3 for C, 3 for T and
15 for D. By following these procedures, the feature vector of a protein is
constructed using a total of 1521 feature elements.
We have trained an SVM-based classifier to distinguish the positive from
the negative training data, using a Gaussian kernel (Keerthi et al., 2001;
Platt, 1999). SVM has been successfully applied to a wide range of
pattern recognition problems in data mining and bioinformatics, such as
protein function prediction (Cui et al., 2007), protein–protein interaction
prediction (Ben-Hur and Noble, 2005) and protein subcellular location
prediction (Su et al., 2007). The Gaussian radial basis function kernel
has been extensively used in those studies with good results, which
consistently shows superior performance to other kernels used in SVM
such as linear and polynomial kernels (Ben-Hur and Noble, 2005; Burbidge
et al., 2001; Su et al., 2007). Thus, Gaussian kernel SVM is used in
our classification study. The inputs to the SVM are the aforementioned
1521 features for each protein in the training set, and the output of the
classifier is an assignment of the input protein to be blood secreted or
not. An independent evaluation set is used to estimate the accuracy of
the overall protein assignment for the whole dataset. The classification
performance is measured using the prediction sensitivity SE=TP/(TP+FN),
prediction specificity SP=TN/(TN+FP), the overall prediction accuracy
Q=(TP+TN)/N, precision=TP/(TP+FP), AUC (Mason and Graham,
2002) and Matthews correlation coefficient MCC=(TP×TN−FP×
FN)/?(TP+FN)(TP+FP)(TN+FP)(TN+FN). Here TP, TN, FP and FN
negative, respectively, and N =TP+FN+TN+FP is the total number of
proteins in the training set. We have used a reliability score, R-value, to
assess the reliability for each of our predictions, shown as follows:
whered is the distance between the position of a target protein in the
feature space and the optimal separating hyperplane derived through our
SVM training. There is a strong correlation between the R-value and the
classification accuracy (probability of correct classification) (Hua and Sun,
2001). Thus, a P-value is introduced to indicate the expected classification
accuracy, derived from the statistical relationship between the R-value and
Classification and feature selection
are the number of true positive, true negative, false positive and false
Fig. 1. Statistical relationship between the R-value and P-value (probability
negative samples of proteins.
the actual classification accuracy based on the analysis of 305 positive and
26962 negative proteins, as shown in Figure 1.
Based on the performance of each initially trained SVM, a feature
selection process, named recursive feature elimination (RFE) (Tang et al.,
2007), is used to remove features irrelevant or negligible to our classification
goal. For example, Moreau–Broto autocorrelation descriptors defined as:
were reported to be useful to the prediction of membrane proteins based
on the hydrophobic index of amino acids (Feng and Zhang, 2000), but our
classification study shows that it does not contribute to the accuracy of our
classification, where d is the lag of the autocorrelation, andPiand Pi+dare
the hydrophobicity of the amino acids at position i and i+d, respectively.
Hence, it is removed from our initial feature list, by the RFE procedure. The
a consensus-scoring scheme and gene-ranking consistency evaluation (Tang
et al., 2007). Specifically, in each iteration, features with the lowest score
(least ranked) given by RFE based on randomly sampled training data are
eliminated from the feature list. Essentially a majority-rule voting scheme is
used to overcome possible discrepancies among different randomly chosen
samples. This process continues until a minimal set of features, without
losing the classification performance, is obtained.
Using the initial list of 1521 features, we trained an SVM classifier
based on the provided positive and negative training sets. We then
evaluated the performance of the best classifier, measured by the
overall accuracy as defined in Section 2.3, using an independent
evaluation set containing 47 positive and 3296 negative samples.
We found that the prediction performance of this classifier gives
only ∼40% accuracy, a clearly undesirable result. We believe that
this is mostly due to the reason that the classifier used a number of
features that are irrelevant to our classification and only add noises
to the training of the SVM classifier; in addition, over-fitting the
data by this large classifier with many parameters may be another
cause for the underpar performance. Hence, it is desirable to remove
some of the less relevant features by carrying out feature selection
to optimize the performance of the SVM-based classifier.
Using the feature selection procedure outlined in Section 2.3,
we have selected a total of 85 features, which gives the best
cross-validation performance of our SVM classifier (Tang et al.,
2007), as shown in Supplementary Table 2. We found that the
Prediction of human proteins secreted into bloodstream
Table 1. Performance statistics of our classifier on prediction of blood-
secreted protein and non-blood-secreted proteins in the training, testing and
independent evaluation sets
TPFNTN FP SE (%) SP (%) Q (%) MCC AUC
following features are the most important ones for our classification,
transmembrane domains, charges, TatP motif, solubility, polarity,
signal peptides, hydrophobicity, O-linked glycosylation motif and
secondary structural content, which rank among the top 20 features.
This observation is consistent with our general understanding of
secretory proteins, except that we found the TatP motif contributes
substantially to our prediction result, which ranks among the top
three features in our prediction, where TatP is known to be used to
export proteins into the periplasmic compartment or extracellular
environment in prokaryotes (Bendtsen et al., 2005; Taylor et al.,
2006). To the best of our knowledge, this represents a novel finding
linking the TatP motifs to protein secretion in eukaryotes.
Based on the 85 selected features, we have trained five new SVM-
based classifiers and tested their performance using the reduced
level of performance by these five classifiers is generally consistent,
ranging from 87.2% to 93.7% for the blood-secreted proteins and
from 98.2% to 98.6% for non-blood-secreted proteins, which is
detailed in Supplementary Table 3. We have also calculated the
precision, MCC and AUC values of our prediction performance,
which have average values 44.6%, 0.63 and 0.94, respectively.
Clearly, the AUC value is consistent with the earlier performance
low. It should be noted that MCC may fluctuate substantially on
previously (Klee and Sosa, 2007; Smialowski et al., 2007). In our
case, the relatively low precision and MCC value are partially due to
the skewed sizes between the positive and negative evaluation sets,
which cause underestimation of the system performance. This could
possibly be improved by increasing the size of positive set. Our
goal here is to include as many previously unknown blood-secreted
proteins as possible, while keeping the specificity high, so we have
chosen one of the classifiers with the best sensitivity, as shown in
When applying WolF PSORT (Horton et al., 2007) to the same
evaluation set, the most cited method for protein extracellular
secretion prediction, it gives 81.0% prediction accuracy with MCC
value 0.37. This is not surprising since all the previous protein-
secretion prediction methods, including WolF PSORT, are not
designed for solving our problem as we are interested in both
extracellular secretion and secretion to the bloodstream.
Our classifier has been further evaluated through a screening test
against all human proteins in the Swissprot database, which can
applied to large datasets. For this test, we collected 20832 human
Table 2. Results of screening all human proteins in Swissprot for blood-
No.of human proteins in Swissprot 20832
No. of proteins annotated as secreted 1563
No. of potentially secreted proteins based on signal
peptide and location (Welsh et al., 2003)
No. of blood-proteins
No. of SVM predicted blood-secreted proteins 4063
proteins, among which 1563 are annotated as secreted proteins and
additional ∼750 proteins are considered to be relevant to secretion
based on their signal peptides and annotated subcellular locations
(Welsh et al., 2003). As shown in Table 2, our classifier predicted
largely agrees with the total (estimated and reported) numbers of
secreted proteins and blood proteins (Welsh et al., 2003). All these
proteins shows good representation of the relevant proteins across
the whole protein space.
search of published proteomics studies and compiled a list of 240
differentially expressed proteins in human blood due to various
diseases. These studies cover multiple cancers in 14 types of human
tissues such as pancreas, ovary, melanoma, lung, prostate, stomach,
liver, colon, nasopharynx, kidney, uterine cervix, brain, breast and
bladder.Among the 240 proteins, 122 are not included in our initial
in Supplementary Table 4. The main reasons for not including these
122 proteins in our initial collection of blood-secreted proteins are
(i) mis-annotation of these proteins in Swissprot and (ii) failing to
detect them by the proteomics studies, from which we collected
this initial list of proteins. As indicated in their respective studies,
all these 122 proteins can be used as potential biomarkers in blood
of a particular cancer to discriminate the normal from the tumor
tissues or distinguish different developmental stages of a particular
cancer, e.g. heat shock protein beta-1 for breast cancer (Rui et al.,
2003), cathepsin D for melanoma (Pardo et al., 2007), l-lactate
dehydrogenase for renal cancer (Unwin et al., 2003) and PSA for
prostate cancer (Bradford et al., 2006). We predicted 97 out of
122 (79.5%) proteins correctly, while the remaining 25 proteins
have prediction results inconsistent with the published literature (the
names of these 122 proteins are given in Supplementary Table 4).
The following gives a few examples of our predictions on the
122 proteins and relevant evidence as reported in the literature.
Among the correct predictions with supporting evidence from the
literature, the tumor necrosis factor, tenascin, C–C motif chemokine
3 and the insulin-like growth factor-binding protein 7 are detected
with elevated gene expression levels in cancer patients’ serum and
are annotated as secreted proteins in Swissprot and SPD database
(Chen et al., 2005). Some membrane proteins like calsyntenin-1,
immunoglobulin alpha chain C and hepatocyte growth factor
receptor are predicted as secreted proteins but these predictions
can only be considered having partial supporting evidence in the
published literature since there is evidence that these proteins are
J.Cui et al.
found outside the cells, through secretion or other means, e.g.
proteolytic cleavage of membrane-associated proteins. Besides,
some predictions can also be partially supported by the annotated
protein functions. For example, thrombospondin 1 precursor is
described as an adhesive glycoprotein that mediates cell-to-cell and
cell-to-matrix interactions, thus it is expected to function outside
the cells. We consider those proteins annotated as secreted proteins
but predicted as non-blood-secreted or as blood-secreted proteins
but without any evidence showing relevance to secretion as ‘not
consistent with the literature’, such as profilin-1 and carbonic
One key planned application of our SVM-based classifier is
to predict if abnormally and highly expressed genes, detected by
microarray gene expression experiments, will have their proteins
secreted into the bloodstream. Previous studies have identified
a number of such genes that show abnormally high expression
levels in patients of various cancers. For examples, a total of 26
and 57 genes were found to have abnormal expression levels,
including both up-regulated and down-regulated in comparison with
the normal cells (Supplementary Table 5), from studies on gastric
cancer (Kim et al., 2005) and lung cancer (Lo et al., 2008). All
these genes have been considered as potential markers for cancer
diagnosis or for distinguishing different cancer stages, as shown in
Supplementary Figure 1 from Lo et al. (2008). We have run our
classifier on each of these genes to check if its encoded protein
is predicted to be blood-secreted and thus can possibly serve as
biomarkers for the corresponding cancer. Our prediction results
show that 13 and 31 proteins out of the 26 and 57 proteins,
respectively, can be secreted into the bloodstream. For example,
complement factor D is encoded by the CFD gene. According to
a quantitative analysis of factor D secretion by gastric cancer cells
(Kitano and Kitamura, 2002), factor D secreted by gastric tissues
is considered to likely contribute to the factor D level in blood
is the multi-drug and toxin extrusion protein 2, encoded by gene
MATE1 with elevated expression in gastric cancer patients. It is
a solute transporter for tetraethylammonium (TEA), 1-methyl-4-
phenylpyridinium (MPP), cimetidine and ganciclovir, and directly
transports toxic organic cations OCs into urine and bile (Otsuka
et al., 2005). Members of the MATE families are observed on the
surface of various tissue cells including endothelial cells of blood
vessels (Pardo et al., 2007). Thus, our prediction of this protein as
being blood secreted is consistent with the previous knowledge.
Based on the results on multiple datasets presented above, we can
see that the overall prediction accuracy of our SVM-based classifier
ranges from 79.5% to 98.1%, with at least 80% of known blood-
test and the extra blood proteins test. From the independent negative
evaluation test, the false positive rate is found to be ∼10%, a
reasonable percentage of misclassified non-blood-secreted proteins,
which is helpful in alleviating the doubts of the low precision.
The prediction accuracies show a good level of consistency across
It should be noted that several factors may affect the accuracy
of our prediction. One is the diversity of protein samples used
for training the SVM-based classifier. It is likely that not all
possible types of blood-secreted proteins are adequately represented
in our training set. The current limitations in the proteomic
technologies for precise separation, detection and identification
of relevant proteins might explain why some of the proteins
with relatively low abundance (lower than ng/ml in serum) are
not detected when in the presence of the high-abundance native
blood proteins (greater than mg/ml in serum). This apparent
discrepancy can overcome with the accumulation of more proteins
abundance in blood. Another problem is that the protein secretion
mechanisms may not be sufficiently represented by the structural
and physicochemical descriptors currently used in our classifier,
(features) may be needed to alleviate this problem.
We have developed a novel sequence-based approach to the
classification of proteins into ‘blood-secreted’ and ‘non-blood-
secreted’ proteins. Global and local characteristics of sequence-
derived properties have been studied for their dominance and
usefulness in predicting proteins that are likely to be possibly
secreted into the bloodstream. A number of proteins encoded by
abnormally and highly expressed genes in tumor cells are predicted
can evaluate the potential of those proteins as serum biomarkers.We
expect that this strategy will prove to be highly useful for marker
protein identification for various human diseases including cancers.
The authors would like to thank Yunmei Lu, Zhongbo Cao from
Jilin University of China and Zhiyi Tong from CSBL (UGA) for
their helpful discussions and their help in data collection.
Funding: National Science Foundation (DBI-0354771, ITR-IIS-
0407204, CCF-0621700, DBI-0542119); National Institutes of
Health (1R01GM075331); ‘Distinguished Scholar’ grant from the
Georgia Cancer Coalition.
Conflict of Interest: none declared.
Adkins,J.N. et al. (2002) Toward a human blood serum proteome: analysis by
database search programs. Nucleic Acids Res., 25, 3389–3402.
Anderson,N.L. and Anderson,N.G. (2002) The human plasma proteome: history,
character, and diagnostic prospects. Mol. Cell Proteomics, 1, 845–867.
Bateman,A. et al. (2002) The Pfam protein families database. Nucleic Acids Res., 30,
Bendtsen,J.D. et al. (2005) Prediction of twin-arginine signal peptides. BMC
Bioinformatics, 6, 167.
Ben-Hur,A. and Noble,W.S. (2005) Kernel methods for predicting protein-protein
interactions. Bioinformatics, 21 (Suppl. 1), i38–i46.
Bhasin,M. and Raghava,G.P. (2004) Classification of nuclear receptors based on amino
acid composition and dipeptide composition. J. Biol. Chem., 279, 23262–23266.
Bosques,C.J. et al. (2006) The sweet side of biomarker discovery. Nat. Biotechnol., 24,
Bradford,T.J. et al. (2006) Molecular markers of prostate cancer. Urol. Oncol., 24,
Brown,J.M. and Giaccia,A.J. (1998) The unique physiology of solid tumors:
opportunities (and problems) for cancer therapy. Cancer Res., 58, 1408–1416.
Buckhaults,P. et al. (2001) Secreted and cell surface genes expressed in benign and
malignant colorectal tumors. Cancer Res., 61, 6996–7001.
Prediction of human proteins secreted into bloodstream
Burbidge,R. et al. (2001) Drug design by machine learning: support vector machines
for pharmaceutical data analysis. Comput. Chem., 26, 5–14.
Cai,C.Z. et al. (2003) SVM-Prot: web-based support vector machine software for
functional classification of a protein from its primary sequence. Nucleic Acids Res.,
Chen,Y. et al. (2005) SPD – a web-based secreted protein database. Nucleic Acids Res.,
Cui,J. et al. (2007a) Computer prediction of allergen proteins from sequence-derived
protein structural and physicochemical properties. Mol. Immunol., 44, 514–520.
functional class and interaction profiles of proteins and peptides irrespective of
sequence homology. Curr. Bioinformatics, 2, 95–112.
Doudna,J.A. and Batey,R.T. (2004) Structural insights into the signal recognition
particle. Annu. Rev. Biochem., 73, 539–557.
Dubchak,I. et al. (1995) Prediction of protein folding class using global description of
amino acid sequence. Proc. Natl Acad. Sci. USA, 92, 8700–8704.
Eisenhaber,F. et al. (1996) Prediction of secondary structural content of proteins from
their amino acid composition alone. I. New analytic vector decomposition methods.
Proteins, 25, 157–168.
Feng,Z.P. and Zhang,C.T. (2000) Prediction of membrane protein types based on the
hydrophobic index of amino acids. J. Protein Chem., 19, 269–275.
Guda,C. (2006) pTARGET: a web server for predicting protein subcellular localization.
Nucleic Acids Res., 34, W210–W213.
Hanahan,D. and Weinberg,R.A. (2000) The hallmarks of cancer. Cell, 100, 57–70.
Horton,P. et al. (2007)WoLF PSORT: protein localization predictor. NucleicAcids Res.,
Huang,L.J. et al. (2006) Proteomics-based identification of secreted protein dihydrodiol
dehydrogenase as a novel serum markers of non-small cell lung cancer. Lung
Cancer, 54, 87–94.
Keerthi,S.S. et al. (2001) Improvements to Platt’s SMO algorithm for SVM classifier
design. Neural Comput., 13, 637–649.
Kim,J.H. et al. (2002) Osteopontin as a potential diagnostic biomarker for ovarian
cancer. J. Am. Med. Assoc., 287, 1671–1679.
Kim,J.M. et al. (2005) Identification of gastric cancer-related genes using a cDNA
microarray containing novel expressed sequence tags expressed in gastric cancer
cells. Clin. Cancer Res., 11, 473–482.
Kitano,E. and Kitamura,H. (2002) Synthesis of factor D by gastric cancer-derived cell
lines. Int. Immunopharmacol., 2, 843–848.
Klee,E.W. and Sosa,C.P. (2007) Computational classification of classically secreted
proteins. Drug Discov. Today, 12, 234–240.
Lo,K.C. et al. (2008) Identification of genes involved in squamous cell carcinoma of
the lung using synchronized data from DNAcopy number and transcript expression
profiling analysis. Lung Cancer, 59, 315–331.
Meteorol. Soc., 128, 2145–2166.
Menne,K.M. et al. (2000) A comparison of signal sequence prediction methods using
a test set of signal peptides. Bioinformatics, 16, 741–742.
Mok,S.C. et al. (2001) Prostasin, a potential serum marker for ovarian cancer:
identification through microarray technology. J. Natl Cancer Inst., 93, 1458–1464.
Mott,R. et al. (2002) Predicting protein cellular localization using a domain projection
method. Genome Res., 12, 1168–1174.
Nair,R. and Rost,B. (2005) Mimicking cellular sorting improves prediction of
subcellular localization. J. Mol. Biol., 348, 85–100.
Omenn,G.S. et al. (2005) Overview of the HUPO plasma proteome project: results
from the pilot phase with 35 collaborating laboratories and multiple analytical
Proteomics, 5, 3226–3245.
Otsuka,M. et al. (2005) A human transporter protein that mediates the final excretion
step for toxic organic cations. Proc. Natl Acad. Sci. USA, 102, 17923–17928.
Pardo,M. et al. (2007) Biomarker discovery from uveal melanoma secretomes:
identification of gp100 and cathepsin D in patient serum. J. Proteome Res., 6,
Pieper,R. et al. (2003) The human serum proteome: display of nearly 3700
chromatographically separated protein spots on two-dimensional electrophoresis
gels and identification of 325 distinct proteins. Proteomics, 3, 1345–1364.
Platt,J.C. (1999) Fast training of support vector machines using sequential minimal
optimization. InAdvances in Kernel Methods: SupportVector Learning. MITPress,
Cambridge, MA, USA, pp. 185–208.
Reczko,M. and Bohr,H. (1994) The DEF data base of sequence based protein fold class
predictions. Nucleic Acids Res., 22, 3616–3619.
Rui,Z. et al. (2003) Use of serological proteomic methods to find biomarkers associated
with breast cancer. Proteomics, 3, 433–439.
Schrader,M. and Schulz-Knappe,P. (2001) Peptidomics technologies for human body
fluids. Trends Biotechnol., 19, S55–S60.
Smialowski,P. et al. (2007) Protein solubility: sequence based prediction and
experimental verification. Bioinformatics, 23, 2536–2542.
Sporn,M.B. and Roberts,A.B. (1985)Autocrine growth factors and cancer. Nature, 313,
Su,E.C. et al. (2007) Protein subcellular localization prediction based on compartment-
specific features and structure conservation. BMC Bioinformatics, 8, 330.
evaluation. Cancer Res., 67, 9996–10003.
Taylor,P.D. et al. (2006) TATPred: a Bayesian method for the identification of twin
arginine translocation pathway signal sequences. Bioinformation, 1, 184–187.
Tjalsma,H. et al. (2000) Signal peptide-dependent protein transport in Bacillus subtilis:
a genome-based survey of the secretome. Microbiol. Mol. Biol. Rev., 64, 515–547.
Unwin,R.D. et al. (2003) Serological and proteomic evaluation of antibody responses
in the identification of tumor antigens in renal cell carcinoma. Proteomics, 3,
Welsh,J.B. et al. (2001) Analysis of gene expression profiles in normal and neoplastic
ovarian tissue samples identifies candidate molecular markers of epithelial ovarian
cancer. Proc. Natl Acad. Sci. USA, 98, 1176–1181.
Welsh,J.B. et al. (2003) Large-scale delineation of secreted protein biomarkers
overexpressed in cancer tissue and serum. Proc. Natl Acad. Sci. USA, 100,