Vol.1, No.2, 93-106 (2009) Natural Science
http://dx.doi.org/10.4236/ns.2009.12012
Sequence-Based Protein Crystallization Propensity
Prediction for Structural Genomics: Review and
Comparative Analysis
Lukasz Kurgan*, Marcin J. Mizianty
Department of Electrical and Computer Engineering, University of Alberta, Edmonton, Alberta, Canada.
*University of Alberta, ECERF, 9107 116 Street, Edmonton, Alberta, Canada;
Received 6 August 2009; revised 28 August 2009; accepted 30 August 2009.
ABSTRACT
Structural genomics (SG) is an international
effort that aims at solving three-dimensional
shapes of important biological macro-molecules
with primary focus on proteins. One of the main
bottlenecks in SG is the ability to produce dif-
fraction quality crystals for X-ray crystallogra-
phy based protein structure determination. SG
pipelines allow for certain flexibility in target
selection which motivates development of in-
silico methods for sequence-based prediction/
assessment of the protein crystallization pro-
pensity. We overview existing SG databanks
that are used to derive these predictive models
and we discuss analytical results concerning
protein sequence properties that were discov-
ered to correlate with the ability to form crystals.
We also contrast and empirically compare mo-
dern sequence-based predictors of crystalliza-
tion propensity including OB-Score, ParCrys,
XtalPred and CRYSTALP2. Our analysis shows
that these methods provide useful and compli-
mentary predictions. Although their average ac-
curacy is similar at around 70%, we show that
application of a simple majority-vote based en-
semble improves accuracy to almost 74%. The
best improvements are achieved by combining
XtalPred with CRYSTALP2 while OB-Score and
ParCrys methods overlap to a larger extend,
although they still complement the other two
predictors. We also demonstrate that 90% of the
protein chains can be correctly predicted by at
least one of these methods, which suggests that
more accurate ensembles could be built in the
future. We believe that current protein crystalli-
zation propensity predictors could provide
useful input for the target selection procedures
utilized by the SG centers.
Keywords: Structural Genomics; X-Ray
Crystallography; Crystallization Propensity Prediction;
Protein Structure; Protein Crystallization
1. INTRODUCTION
Proteins are organic compounds composed of amino
acids arranged in a linear chain polymer with the help of
peptide bonds. Proteins implement a wide variety of
functions such as transportation, signalling, catalysis of
chemical reactions, formation of the cell cytoskeleton,
immune responses, regulation of cell processes, etc. etc.
They are so versatile due to their ability to adopt an im-
mense variety of shapes. Knowledge of the tertiary
(three-dimensional) structure of proteins is vitally im-
portant for understanding and manipulating their bio-
chemical and cellular functions. For instance, this know-
ledge is exploited in rational drug design via virtual
screening [1-3], provides insights into various diseases
[4] and it is used in deciphering interactions of proteins
with other macro molecules and smaller ligands [5-7].
1.1. Structural Genomics
As of July 2009 we know close to 8.2 million nonre-
dundant protein chains which can be found in SeqRef
database [8] but the corresponding structure is know for
“only” about 55 thousand proteins which are deposited
into the Protein Data Bank (PDB) database [9]. This
wide and continually widening sequence-structure gap
calls for new and efficient efforts that would help in ac-
quiring protein structures. This resulted in creation of
structural genomics (SG) which is an international effort
to find the three-dimensional shapes of important bio-
logical macro-molecules, primarily focusing on proteins
[10]. In contrast to a traditional approach used by struc-
tural biologists who often work with a given protein that
they try to solve for many years, the structural genomics
efforts frequently concern ”unknown” proteins. More-
over, SG focuses on development and usage of high

L. Kurgan et al. / Natural Science 1 (2009) 93-106
94
throughput and cost-effective methods for protein pro-
duction and determination of the corresponding structure
which are implemented with the help of dedicated SG
centers. In the United States one of the first SG efforts,
which was undertaken around year 2000, was the crea-
tion of a multi-center, including four large-scale centers
and six specialized centers, Protein Structure Initiative.
Similar SG projects were also carried out in Canada,
Israel, Japan, and Europe. For example, Structural Ge-
nomics Consortium which was formed in 2004 spans
centers at the Oxford University, University of Toronto
and Karolinska Institute. Analysis shows that in 2004/
2005 about half of protein structures were solved at a SG
centers rather than in the traditional laboratory [11]. Also,
at that time the cost of solving a structure at the most
efficient SG center in the United States was equal to
about 25% of the estimated cost when using the tradi-
tional methods [11]. Another more recent study shows
that the production-line approach taken at the Protein
Structure Initiative centers reduced the cost of solving
structures from ~$250,000 apiece in 2000 to ~$66,000 in
2008 [12]. Most importantly, from our point of view,
these SG initiatives shifted the focus from one-by-one
determination of individual protein structures, which is
being pursued by structural biologists, to protein fam-
ily-directed structure analyses in which a group of pro-
teins is targeted and structure(s) of representative mem-
bers are determined and used to represent the entire
group [13]. The corresponding process of choosing rep-
resentative proteins is known as target selection and it
encompasses a computational process of restricting can-
didate proteins to those that are tractable and of un-
known structure and prioritizing them according to ex-
pected interest and accessibility [14]. In the case of the
Protein Structure Initiative, the target selection concen-
trates on representatives from large, structurally unchar-
acterized protein domain families, and from structurally
uncharacterized subfamilies in very large and diverse
families with incomplete structural coverage [15]. We
note that this approach allows for some flexibility in the
selection of the targets.
1.2. X-ray Crystallography and Protein
Crystallization
The protein structures are being determined with the
help of experimental methods including X-ray crystal-
lography [16], NMR spectroscopy [17], electron mi-
croscopy [18], and (more recently) by application of
computational approaches such as homology modelling
[19, 20]. The most popular method, which accounts for
approximately 86% of the solved and deposited protein
structures, is the X-ray crystallography; see Figure 1. At
the same time, the other approaches play a strong com-
plementary role for some protein types, such as mem-
brane proteins [21, 22].
One of the main challenges the SG initiative faces it
that only about 2-10% of protein targets pursued in the
context of the second step of the Protein Structure Initia-
tive yield high-resolution protein structures [23]. We
further investigated these estimates based on data pub-
lished in the TargetDB database [24] in July 2009. Tar-
getDB is a world-wide database that provides informa-
tion on the experimental progress and status of targets
0
5000
10 0 0 0
15 0 0 0
20000
25000
30000
35000
40000
45000
50000
55000
1990199119921993199419951996199719981999 2000200120022003 2004 2005 20062007 2008 2009
yea
r
cumulative number of
solved structures
X-ray crystallography
NMR spectroscopy
Electron microscopy
Figure 1. The growth in the number of protein structures deposited into PDB by that were solved by X-ray crystallography, NMR
spectroscopy and electron microscopy (source http://www.rcsb.org/).

L. Kurgan et al. / Natural Science 1 (2009) 93-106
95
95
selected for structure determination. Among 150,727
cloned targets that were deposited into TargetDB, only
37,398 (24.8%) were reported to be successfully purified,
12,923 (8.6%) to be successfully crystallized, and 6,942
(4.6%) gave diffraction quality crystals. Moreover, some
estimates show that more than 60% of the cost of struc-
ture determination is consumed by the failed attempts
[25] while crystallization is characterized by a signifi-
cant rate of attrition and is among the most complex and
least understood problems in structural biology [26]. The
above provides a strong motivation for further research
and development in this area. Several strategies have
been proposed to improve the success rate including
obtaining one representative structure per protein family
and working with multiple orthologues [14, 26, 27, 28].
In spite of advances made in the context of protein crys-
tallization [29], the above numbers and insights from
some researchers [30-32] demonstrate that the produc-
tion of high-quality crystals is one of the major bottle-
necks in the protein structure determination. The crystals
should be sufficiently large (> 50 micrometres), pure in
composition, regular in structure and with no significant
internal imperfections. The problem of production of
diffraction-quality crystals is usually tackled using an
empirical approach based mainly on trial and error (also
called the “art” of crystallisation), in which a large
number of experiments is brute-forced to find a suitable
setup, and through understanding of the fundamental
principles that govern crystallisation [30]. The latter is
used to design new (and improved) experimental meth-
odologies that would produce high-quality crystals.
1.3. Databases
One of the early steps taken to alleviate the abovemen-
tioned difficulties in resolving the structures via X-ray
crystallography was to create databanks that record in-
formation concerning both successful and failed attempts
to produce the structures. The importance of these efforts
was advocated in 2000 by Raymond Stevens who said
that “industrial-scale efforts will lead to the generation
of knowledge bases that will be mined to expand our
understanding of the techniques used in protein crystal-
lography. These efforts will act as ‘learning factories’, in
which successes and failures will be used to continually
improve the technology for high-throughput protein cry-
stallography“ [33].
These words were echoed in 2003 by Rodrigues and
Hubbard who said “as structural genomics projects
evolve, valuable experimental data will be accumulated,
thus presenting researchers with a unique opportunity to
establish improved predictive methods for a protein’s
chemical and physical behaviour based on its amino acid
sequence. It is essential for laboratories producing such
data to keep track of both ‘successful’ and ‘unsuccessful’
results, so that these can be fed back into the structural
determination pipeline through the improvement of the
target selection procedures” [34]. The development of the
databases was fuelled by generation of large and well an-
notated experiments by SG centers, such as one for the
Thermotoga maritime proteome [35]. To the best of our
knowledge, the first such initiative was the PRESAGE
database which included annotations indicating current
experimental status, structural predictions and suggestions
[36]. Some of the SG consortia have established on-line
progress reports which contain details and current ex-
perimental status of their targets. Examples include Inte-
grated Consortium Experimental Database [37], ZebaView
(http://www-nmr.cabm.rutgers.edu/bioinformatics/ZebaView/),
ReportDB (http://www.secsg.org/cgi-bin/report.pl) and
SPINE (Structural Proteomics in the NorthEast) [38, 39].
SPINE, which was developed in early 2000 and reengi-
neered in 2003, integrates a tracking database and a data
mining method for identifying feasible targets. Each
protein deposited in this database is described with in-
formation related to the experimental progress (e.g., ex-
pression level, solubility, ability to crystallize) and 42
descriptors of the underlying protein sequence (amino
acid composition, secondary structure, etc.). The largest
and most comprehensive TargetDB [24] was launched
July 2001 and it builds upon the work on the PRESAGE
database. TargetDB serves as a primary target registra-
tion database for structural SG project worldwide. It
consolidates data from 28 SG centers in USA, Canada,
Germany, Isreal, Japan, France and UK, including 9
Protein Structure Initiative centers. PepcDB (Protein
Expression Purification and Crystallization DataBase),
which was created around 2004, was established as an
extension to TargetDB to collect more detailed status
information and the experimental details of each step in
the protein structure production pipeline [40]. This da-
tabase stores a complete history of the experimental
steps in each production trial besides describing the cur-
rent target production status. PepcDB records status his-
tory, stop conditions, reusable text protocols and contact
information collected from 15 SG centers in USA. The
interested readers are directed to two recent articles by
Helen Berman that introduce a wealth of resources con-
cerning the SG initiative [41] and a knowledgebase de-
veloped by the Protein Structure Initiative [42].
1.4. Computational Models in Protein
Crystallization
The problems with the protein crystallization and the
availability of the suitable databases motivated the de-
velopment of analytical and predictive models that can
be used to either support or directly predict protein crys-
tallization [43]. These models were often developed by
researchers at certain SG centers who used their own
data to draw conclusions. In one of the first attempts, a

L. Kurgan et al. / Natural Science 1 (2009) 93-106
96
decision tree that predicts solubility from protein se-
quence was developed [44]. The SPINE system, which
was developed at the Northeast Structural Genomics
Consortium, incorporates decision tree-based classifiers
for solubility and crystallization propensity. This system
was used to extract a few interesting rules such that
soluble proteins tend to have more acidic residues and
fewer hydrophobic segments [38]. The SG project on
Plasmodium falciparum has lead to an analysis of pro-
tein characteristics, such as the presence of transmem-
brane helices, low-complexity regions, and coiled-coil
regions, in the context of the crystallization propensity
[34]. Another decision tree-based predictive model de-
veloped by Goh and colleagues in 2004 using data from
TargetDB has revealed several protein features that in-
fluence the feasibility of using a given target protein
chain for a high-throughput structure determination [45].
They include conservation of the sequence across organ-
isms, composition of charged residues, occurrence of
hydrophobic patches in the sequence, number of binding
partners, and chain length. Based on the data from the
Thermotoga maritime proteome [35], the researchers at
the Joint Center for Structural Genomics discovered a
few features, which include isoelectric point, sequence
length, average hydropathy, existence of low complexity
regions, presence of signal peptides and trans-membrane
helices, that correlate with crystallization [46]. The
isoelectric point calculated from the protein sequence
was also used to develop a method that suggests optimal
pH ranges for crystallization screening [47, 48]. Ex-
perimental work by Derewenda’s group shows that crys-
tallization can be improved by application of surface
entropy reduction approach in which clusters of two or
three exposed amino acids with high conformational
entropy side chains (such as Lys, Glu and Gln) are re-
placed with lower-entropy residues (like Ala) [49-54].
One drawback of this method is that it may decrease
protein solubility which hinders crystallization screening
[50, 52]. The surface entropy reduction approach was
recently implemented as a web server [55]. This server
utilizes information concerning conformational entropy
and solvent exposure indices, predicted secondary struc-
ture, residues conservation scores, and close homologues
to propose crystallization enhancing mutations for a
given protein sequence. Another study, which was con-
ducted at the Center for Eukaryotic Structural Genomics,
used disorder prediction algorithms to analyze the im-
pact of intrinsic protein disorder on crystallization effi-
ciency [56]. The Berkeley Structural Genomics Center
has utilized several protein features including length of
the sequence and predicted transmembrane helices,
coiled coils, and low-complexity regions to eliminate
targets predicted to be intractable for the high-through-
put structure determination [57]. The most recent study
that was performed at the Northeast Structural Genomics
Consortium shows that crystallization propensity de-
pends primarily on the prevalence of well-ordered sur-
face epitopes [58]. More specifically, the authors show
that crystallization propensity can be computed from the
knowledge of predicted disordered regions, side-chain
entropy of predicted exposed residues, the amount of
predicted buried Gly and the fraction of Phe in the input
sequence.
2. SEQUENCE-BASED METHODS FOR
PREDICTION OF PROTIEN
CRYSTALLIZATIONPROPENSITY
The SG efforts allow for certain flexibility in selection
of the chains for the crystallization and the subsequent
structure determination and this motivates development
of methods that aim at the prediction/assessment of the
crystallization propensity for a given input sequence.
Such methods could be incorporated into target selection
pipelines that are utilized by SG centers. Their develop-
ment is often supported and motivated by the described
above computational analyses/models. We also note that
numerous studies have already demonstrated that se-
quence-based prediction approaches, which may address
a variety of structural and functional properties of pro-
teins, provide useful information and insights for both
basic research and drug design and hence are widely
welcome by the scientific community [59-63].
Crystallization propensity prediction methods incor-
porate predictive models that are extracted from larger
datasets that span data coming from multiple SG centers
and they take the protein sequence as their only input.
The underlying principle is that the predictive models
summarize/describe patterns (similarities) hidden in the
data from databases such as TargetDB. This is done by
generating a set of patterns that describe sequences that
can be crystallized (crystallizable proteins) and another
set of patterns for sequences that were shown to be im-
possible to crystallize (noncrystallizable proteins). The
two sets of patterns should describe the two correspond-
ing sets of protein chain and, at the same time, each of
them should exclude sequences from the other set. The
existing crystallization propensity predictors include
SECRET [64] that was developed by Frishman’s group,
OB-Score [65] and ParCrys [66] that were produced by
the Barton’s group, XtalPred [67, 68] that came from
Godzik’s group, and CRYSTALP [69] and most recent
CRYSTALP2 [70] that were developed by Kurgan’s
group. These methods perform the prediction in two
steps: (1) the input sequence is converted into a set of
numerical features that describe certain characteristics of
the sequence; and (2) the feature values are fed into a
predictive model that outputs the outcome that quantifies
propensity for crystallization. The predictive model en-
capsulates the patterns that are computed from the in-
formation encoded by the features. Table 1 shows a

L. Kurgan et al. / Natural Science 1 (2009) 93-106
97
97
Table 1. A side-by-side comparison of existing methods for sequence-based protein crystallization propensity prediction.
Input features
Methods
[reference]
Source of
data description #
Predictive
model Web server/page Notes
SECRET
[64]
Deposition
from PDB
assuming that
NMR only
solved protein
are diffi-
cult/impossibl
e to crystal-
lize; Deposi-
tions in Tar-
getDB
Content of mono-,
di-, and tripeptides
represented by
20-letter amino acid
alphabet and by
several reduced
alphabets grouped
by physicochemical
and structural prop-
erties of amino
acids
103
Two-layered
structure where
output of sev-
eral support
vector machine
classifiers are
combined by a
second-level
Naive Bayes
classifier
http://mips.helmholtz-muenchen.de/secret/
Limited
to se-
quences
between
46 and
200
amino
acids
OB-Score
[65]
Depositions in
TargetDB
Isoelectric point
and average hydro-
phobicity
2
Z-score
(two-dimen-
sional
lookup-table)
http://www.compbio.dundee.ac.uk/xtal/
CRYS-
TALP
[69]
Deposition
from PDB
assuming that
NMR only
solved protein
are diffi-
cult/impossibl
e to crystal-
lize;
Content of selected
mono-, di- and
collocated dipep-
tides
46 Naive Bayes N/A
Limited
to se-
quences
between
46 and
200
amino
acids
XtalPred
[67, 68]
Depositions in
TargetDB
Protein length,
molecular mass,
gravy and instabil-
ity indices, extinc-
tion coefficient,
isoelectric point,
content of Cys,
Met, Trp, Tyr, and
Phe residues, inser-
tions in the align-
ment compared to
homologs in
non-redundant
protein sequences
database, predicted
secondary structure,
predicted disor-
dered,
low-complexity and
coiled-coil regions,
predicted trans-
membrane helices
and signal peptides.
9 Normalized
product http://ffas.burnham.org/XtalPred/
Outputs
1 of 5
crystal-
lization
classes:
optimal,
subopti-
mal,
average,
difficult,
and very
difficult
ParCrys
[66]
Depositions in
TargetDB and
PepcDB
Isoelectric point
and average hydro-
phobicity, content
of Ser, Cys, Gly,
Phe, Tyr, and Met
residues
8
Kernel-based
classifier using
Parzen window
http://www.compbio.dundee.ac.uk/xtal/
CRYS-
TALP2
[70]
Depositions in
TargetDB and
PepcDB
Isoelectric point,
average hydropho-
bicity, content of
selected mono-, di-
and collocated di-
and tripeptides
88
Normalized
Gaussian radial
basis function
network
http://biomine.ece.ualberta.ca/CRYSTALP2/CRYSTALP2.html

L. Kurgan et al. / Natural Science 1 (2009) 93-106
98
side-by-side comparison of the six existing methods
based on the data source that was used to generate pre-
dictive model and the applied input features and predic-
tive models. It also provides URLs of the corresponding
web servers or web pages.
Two early methods, namely SECRET and CRYSTALP,
accept only sequences between 46 and 200 amino acids
in length. This limitation is due to the composition of
datasets used to generate these prediction models. Al-
though OB-Score predictor does not impose a limit on
sequence size, it considers only two predictive features,
i.e., isoelectric point and hydrophobicity. This method
was developed for the Scottish Structural Proteomics
Facility [65]. The ParCrys method extends OB-score by
using an advanced kernel-based classification algorithm
and by adding information concerning content of several
amino acids including Ser, Cys, Gly, Phe, Tyr, and Met
to the set of predictive features. Similarly, CRYSTALP2
improves upon CRYSTALP by applying a more ad-
vanced kernel-based classifier and by introducing new
predictive features that are based on the collocation of
amino acids in the sequence, isoelectric point and hy-
drophobicity. The motivation for the application of the
collocation based features comes from their applications
in related fields [71-74] and the fact that they consider
local neighbourhood information in the protein chain,
which was also utilized in a recent method for surface
entropy reduction based design of crystallizable protein
variants [55]. A significant majority of the collocations
used by CRYSTALP2 incorporate residues with high
conformational entropy, or with low entropy and high
potential to mediate crystal contacts, and these residues
are utilized by the surface entropy reduction methods [51,
52].
The above five methods are built using black-box (not
readable by a human) classification models, which are
inductively learned from a set of protein chains which
are annotated as crystallizable and noncrystallizable. By
contrast, the XtalPred is a white-box (human readable)
approach that combines probabilities of successful crys-
tallization calculated from several protein features. This
method, which was developed based on experiences at
the Joint Center for Structural Genomics, which is one of
the large centers in the Protein Structure Initiative, mim-
ics the work performed by structural biologists. XtalPred
utilizes nine biochemical and biophysical features of an
input protein with probability distributions estimated
from data from TargetDB. The individual probabilities
concerning each input feature are combined into a single
crystallization score which is used to assign one of five
crystallization classes: optimal, suboptimal, average,
difficult, and very difficult. The design of XtalPred
shows that medium sequence length and hydrophobicity
combined with acidic character improve the success in
protein production. It also demonstrates that very short,
very long, or very hydrophobic proteins are more diffi-
cult to crystallize under standard experimental setups.
This method also confirms the utility of predicted struc-
tural disorder, presence of transmembrane helices, insta-
bility, and high content of predicted loops, insertions,
and coiled-coil structures for the prediction of the crys-
tallization propensity [67]. Several methods, including
XtalPred, OB-Score, ParCrys and CRYSTALP2, utilize
information concerning isoelectric point which is esti-
mated from protein sequence. This agrees with prior
finding that indicate important role of this feature
[46-48].
We note that all investigated crystallization propensity
predictors take into account only intra-molecular factors
that are encoded in the protein chain. This means that
they may not provide reliable predictions when in-
ter-molecular factors such as protein-protein and/or pro-
tein-precipitant interactions, buffer composition, pre-
cipitant diffusion method, etc. must be considered. Also,
they are limited to predictions for non-redundant chains
and should not be used when assessing crystallization of
homologues. In the latter case we recommend the use of
the surface entropy reduction server [55].
3. COMPARATIVE ANALYSIS
Following we perform empirical comparison of the qual-
ity of predictions offered by the sequence-based protein
crystallization propensity predictors. Our analysis ex-
cludes CRYSTALP and SECRET methods since they are
limited to only relatively small chains and since their
quality was show to be inferior when compared with
other methods [66,70]. Our comparative analysis is per-
formed based on predictions performed for a dataset of
relatively recent depositions to TargetDB and PepcDB.
We analyze predictive power of individual methods and
we also investigate their complementarity.
3.1. Dataset
We use a dataset composed of 2000 protein chains
(hereafter TEST-NEW), which was originally introduced
in [70] and which was developed using procedure pro-
posed in [66]. The crystallizable proteins were extracted
from sequences deposited in TargetDB and they include
the last 1000 depositions as of December 2008. The non-
crystallizable sequences, which correspond to the actual
construct sequences used, were extracted from the trial
sequences stored in PepcDB. As in the case of crystal-
lizable chains, they include the last 1000 depositions as
of December, 2008. The selected sequences were also
processed to remove the N-terminal hexaHis tag and
LEHHHHHH tag at the C-terminus, which are intro-
duced to ease the purification. Duplicate sequences were
removed and thus the resulting dataset consists of non-
redundant chains. It can be freely downloaded from

L. Kurgan et al. / Natural Science 1 (2009) 93-106
99
99
Table 2. Summary of results for predictions performed with OB-Score, ParCrys, XtalPred and CRYSTALP2 methods on the
TEST-NEW dataset.
Accuracy MCC TPR TNR AROC
OB-Score1 69.8 0.42 0.86 0.54 0.74
ParCrys1 70.6 0.43 0.83 0.58 0.75
XtalPred2 70.0 0.40 0.76 0.64 0.76
CRYSTALP23 69.3 0.39 0.76 0.63 0.74
1Results computed using the ParCrys/OB-Score server at http://www.compbio.dundee.ac.uk/xtal/
2Results computed using the XtalPred server at http://ffas.burnham.org/XtalPred/
3Results based on [70]
http://biomine.ece.ualberta.ca/CRYSTALP2/CRYSTALP2.html.
3.2. Quality Measures
The annotations from TargetDB were stripped from the
input sequences, which in turn were inputted into the
corresponding predictors. The prediction outputs were
compared with the original annotations to assess the
prediction quality. Four potential prediction outcomes
are possible: TP (true positive) which corresponds to
crystallizable chains that were correctly predicted as
crystallizable, FN (false negative) which corresponds to
crystallizable chains that were incorrectly predicted as
noncrystallizable, FP (false positive) which indicates that
noncrystallizable chains were incorrectly predicted as
crystallizable, and TN (true negative) which denotes
cases where noncrystallizable chains were correctly pre-
dicted as noncrystallizable. The predictions were as-
sessed based on the following quality indices:
100



FN
FP
T
N
T
P
TNTP
accuracy
)()()()( FNTNFPTNFNTPFPTP
FNFPTNTP
MCC



FN
T
P
TP
TPR  
FP
T
N
TN
TNR  
The accuracy measures the fraction of correct predic-
tions among all predictions. The Matthews Correlation
Coefficient (MCC) is confined to <-1,1> interval. If the
MCC value is close to 0 then the prediction method is
not better than a random classification. Higher MCC
value corresponds to better performance of the predic-
tion method. These two measures provide an evaluation
of the prediction quality over the entire dataset. In con-
trast, TPR (true positive rate) and TNR (true negative
rate) evaluate the quality separately for crystallizable
(positive) and noncrystallizable (negative) proteins.
TPR/TNR quantifies the fraction of correctly predicted
crystallizable/noncrystallizable proteins. We also report
receiver-operator characteristics (ROC) curves that pre-
sent a graphical plot of the TP rate = TP / (TP + FN)
against FP rate = FP / (FP + TN). This is performed by
thresholding the confidence values (probabilities) that
are generated together with the predicted classes (crys-
tallizable vs. noncrystallizable). These plots are also
used to compute the area under the ROC curve (AROC).
The higher the AROC value is the better the predictive
power of the corresponding method.
3.3. Comparison of Existing Prediction
Methods
Results of application of the four crystallization propen-
sity predictors on the TEST-NEW dataset are summa-
rized in Table 2. In the case of XtalPred we assume a
prediction assignment in which optimal, suboptimal, and
average outcomes are categorized as crystallizable pro-
teins and difficult and very difficult as noncrystallizable.
The same assignment was used in [70] since it leads to
optimal results.
The comparison shows that the four methods are char-
acterized by relatively similar prediction quality with
MCC and accuracy values ranging between 0.39 and
0.43 and between 69.3 and 70.6%, respectively. We note
that since the dataset is balanced a random assignment of
the prediction outcomes would give accuracy of 50%.
This means that the accuracy of the existing methods is
better by about 20% than the random coin-toss approach.
At the same time we observe a considerable space for
improvement although we caution the reader that the
upper limit of the prediction accuracy should not be as-
sumed at 100%. This is since the input data likely in-
cludes mislabeled proteins. In particular, since data
comes from multiple SG centers, some proteins that
could not be crystallized in one center could be poten-
tially crystallized by another center that uses different
protocols and equipment. This means that some of the
proteins could be mislabeled as noncrystallizable, i.e.,
some of the FPs are in fact TPs. At this time we are not
able to estimate their number. We observe that OB-Score
and ParCrys are both strongly biased towards prediction
of crystallizable proteins, i.e., their TPR values are much
higher than TNR values and the TNR values are rela-
tively low. The XtalPred and CRYSTALP2 provide a

L. Kurgan et al. / Natural Science 1 (2009) 93-106
100
more balanced prediction for the two classes of proteins
and their TNR values are above 0.63. All four methods
provide better predictions for crystallizable proteins, i.e.,
they correctly predict a bigger fraction of crystallizable
proteins, when compared with the noncrystallizable pro-
teins. In other words, they are more likely to succeed in
confirming that a crystallizable chain can be crystallized
rather than in showing that a chain difficult to crystallize
cannot be crystallized; although in both cases all of the
considered methods work better than the coin-toss. Fig-
ure 2 shows the ROC curves for the four predictors. We
again observe that all considered methods behave simi-
larly, i.e., they provide comparable TP rates for the same
FP rates.
0
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
0.9
1
00.10.2 0.30.4 0.5 0.60.7 0.80.91
FP rate
TP rate
CRYSTALP2 on TEST-NEW
ParCrys on TEST-NEW
XtalPred on TEST-NEW
OB-Score on TEST-NEW
Figure 2. ROC curves for the tests performed with OB-Score, ParCrys, XtalPred and CRYS-
TALP2 methods on the TEST-NEW dataset.
0%
10%
20%
30%
40%
50%
60%
70%
80%
90%
100%
all proteinsnoncrystallizable
pr oteins
crystallizable
proteins
dataset
percentage of
predictions
no correct
predictions
1 method with
correct prediction
2 methods with
correct predictions
3 methods with
correct predictions
All 4 methods with
correct predictions
Figure 3. Analysis of the number of correct predictions produced by OB-Score, ParCrys,
XtalPred and CRYSTALP2 methods on the all proteins, only crystallizable and only
noncrystallizable proteins from the TEST-NEW dataset.

L. Kurgan et al. / Natural Science 1 (2009) 93-106
101
101
0
10 0
200
300
400
500
600
700
[38, 99][100, 149][150, 199][200, 299][300, 399][400, 1948]
protein length intervals
number of
pr oteins
all proteins
crystallizable proteins
noncrystallizable proteins
A
0.00
0.10
0.20
0.30
0.40
0.50
0.60
[38, 99][100, 149][150, 199][200, 299][300, 399][400, 1948]
protein length intervals
MCC
CRY STA LP2
ParCrys
XtalPred
OB-Score
B
Figure 4. Analysis the predictions and characteristics of the TEST-NEW dataset with respect to the input protein chain length. A)
Distribution of number of proteins (black bars), number of crystallizable (green bars) and noncrystallizable (red bars) proteins in the
considered protein length intervals. B) Prediction quality measured using MCC for OB-Score, ParCrys, XtalPred and CRYSTALP2
methods for each of the protein size intervals.
Figure 3 analyzes the predictions with respect to the
number of correct predictions produced by the four
methods for each input protein. Analysis of the results
obtained on the entire TEST-NEW set indicates that at
least three methods provide correct predictions simulta-
neously for two thirds of the test proteins. It also shows
that only 9.6% of the proteins cannot be correctly pre-
dicted by any of the considered methods. We again ob-
serve that predictions for crystallizable proteins are
characterized by higher quality than for the noncrystal-
lizable proteins. In particular, only 1.6% of crystallizable
proteins are never correctly predicted and 78.0% are
correctly predicted by at least 3 methods. In contrast, the
same numbers for the noncrystallizable proteins are
17.6% and 53.2%, respectively.
3.4. Analysis of Predictions for Varying Pro-
tein Sizes
The protein chain length was indicated as one of the
important factors related to the protein crystallization
propensity [45, 46, 57, 67]. It is also correlated with the
quality of the secondary structure prediction [75], which
is utilized in the prediction of protein crystallization [55,
67]. To this end, Figure 4 summarizes results that are
organized by binning the input protein chains into six
size-based intervals. Figure 4A shows, as expected [67],
uneven distribution of the crystallizable and noncrystal-
lizable proteins against the protein chain length. We ob-
serve that majority of short chains with less than 100
amino acids are difficult to crystallize while the crystal-
lization is more successful for longer chains. More im-
portantly, the XtalPred method stands out from the
competition as it provides better performing predictions
for short sequences of up to 150 amino acids. On the
other hand, a slight improvement over the competition is
observed for the OB-Score method when predicting long
chains with above 400 amino acids. Finally, the CRYS-
TALP2 method is characterized by the most even quality.
We also observe a generic trend that best results are on
average obtained for the average sized protein chains
between 100 and 200 amino acids.
3.5. Complementarity of Existing Methods
Although the above results indicate that the existing
methods are characterized by comparable prediction
quality, substantial differences in their underlying design
and results shown in Figures 3 and 4B suggest that their
results could be complementary with each other. In other
words, although on average they provide the same num-
ber of correct predictions, these prediction likely concern
different input proteins.
We investigate the complementarity by combining
multiple methods using OR operator, i.e., a given predic-
tion is assumed correct if at least one of the methods in
an ensemble provides a correct prediction. This approach
allows quantifying the amount of overlap in predictions
and it also estimates the upper boundary of a potential
meta-predictor that combines predictions from the indi-
vidual methods. Figure 5 shows summary of results, in
terms of achieved TPR, TNR and MCC values for all
combinations of two, three, and four predictors as well
as for the individual methods. We observe that certain
ensembles obtain higher quality of predictions indicating
a stronger complementarity. In particular combining
either OB-Score and XtalPred or CRYSTALP2 with
XtalPred gives better results than any other combination

L. Kurgan et al. / Natural Science 1 (2009) 93-106
102
OB PC XP C2 .82
OB PC XP .76
OB XP C2 .80
PC XP C2 .80
OB PC C2 .73
PC C2 .68
OB C2 .68
XP C2 .70
OB PC .58
PC XP .69
OB XP .70
OB .42
PC .43
XP .40
C2 .39
0.50
0.60
0.70
0.80
0.90
1.00
0.75 0.80 0.85 0.90 0.95 1.00
TPR
TNR
Figure 5. Analysis the complementarity of predictions for OB-Score (OB), ParCrys (PC), Xtal-
Pred (XP) and CRYSTALP2 (C2) methods on the TEST-NEW dataset. Each combination of 1,
2, 3, and 4 methods was applied using OR operator, i.e., a given prediction was assumed correct
if at least one of the predictors predicted it correctly. The x-axis/y-axis shows TPR/TNR values
(TPR values are scaled between 0.75 and 1 while TNR values are scaled between 0.5 and 1),
and the labels next to markers denote a particular combination of applied predictions together
with the MCC value (e.g., “PC XP C2 .80” means that combination of ParCrys, XtalPRed and
CRYSTALP2 obtained MCC of 0.8). Markers and labels in red denote the best results for a
given number of applied methods.
of two methods. Among the ensembles of three methods,
the combination of XtalPred and CRYSTALP2 with ei-
ther ParCrys or OB-Score works best. This observation
and the fact that OB-Score and ParCrys are the least
complimentary among all pairs of predictors indicate
that these two methods provide relatively overlapping
outputs. Finally, an ensemble of all four methods obtains
MCC of 0.82 which is not much higher than 0.80
achieved with just three methods, showing that addition
of the fourth predictor brings relatively minor improve-
ments. Finally, we again observe that results indicate that
both individual and ensemble-based predictions are
characterized by higher quality for crystallizable rather
than noncrystallizable proteins.
We also investigate a possibility of implementing a
simple, majority-vote based meta-predictor. Such met-
hod generates predictions which correspond to the most
frequent prediction of its member methods. We apply a
simple majority vote for the three members based
meta-predictors, while for ensemble of four methods we
resolve the tie-break (2 vs 2 split decisions from the
member methods) by applying the prediction of one se-
lected method. This leads to eight potential configura-
tions, i.e., three combinations of three out of four meth-
ods and four configurations with four member methods
each time using a different method as a tie-breaker. The
corresponding results are presented in Figure 6. The
results demonstrate that the best ensemble includes Xtal-
Pred, CRYSTALP2 and OB-Score. The runner-up con-
figurations include an ensemble of XtalPred, CRYS-
TALP2 and ParCrys and two ensembles of four methods
with tie-breakers as XtalPred and CRYSTALP2. These
results are consistent with the above complementarity
analysis and indicate beneficial overlap between Xtal-
Pred and CRYSTALP2. We also observe that application
of a majority-vote mechanism provides only moderate
improvements. More specifically, the best vote-based
ensemble obtains MCC of 0.49 while the MCC of best
individual method equals 0.43 and the MCC of best
combination of methods from Figure 5 gives MCC

L. Kurgan et al. / Natural Science 1 (2009) 93-106
103
103
XP .40
C2 .39
ALL tie-bkr OB .47
ALL tie-bkr PC .46
ALL tie-bkr XP .48
ALL tie-bkr C2 .48
PC XP C2 .48
OB PC XP .47
OB PC C2 .47
OB XP C2 .49
OB .42
PC .43
0.50
0.52
0.54
0.56
0.58
0.60
0.62
0.64
0.75 0.770.79 0.81 0.830.85 0.87 0.89
TPR
TNR
Figure 6. Analysis the performance of majority-vote based ensembles of OB-Score (OB), Par-
Crys (PC), XtalPred (XP) and CRYSTALP2 (C2) methods on the TEST-NEW dataset. The
x-axis/y-axis shows TPR/TNR values (TPR values are scaled between 0.75 and 0.9 while TNR
values are scaled between 0.5 and 0.65), and the labels next to markers denote a particular en-
semble together with the MCC value (e.g., “OB XP C2 .49” means that ensemble composed of
OB-Score, XtalPRed and CRYSTALP2 obtained MCC of 0.49). The prediction of the ensemble
corresponds to the most frequent prediction of its members. The tie-breaker for ensembles of 4
methods is chosen as the prediction of one specific method, i.e., “ALL tie-brk XP” corresponds
to an ensemble of all four methods in which a split 2 vs 2 decision is decided by the prediction
of XtalPred. Markers and labels in red/blue denote the best/second best results.
equal to 0.82. In terms of the corresponding accuracies,
this means that although the considered four methods
can correctly predict up to 90.4% of proteins, the simple
voting provides only 73.6% of correct predictions.
Overall, the analysis shows that the best improve-
ments, when compared with using individual predictors,
are achieved by combining XtalPred with CRYSTALP2.
The OB-Score and PareCrys methods overlap to a larger
extend although they also complement the other two
predictors. This can be explained by the use of very
similar input features in ParCrys and OB-Score and use
of larger numbers of more complementary features in
CRYSTALP2 and XtalPred. Finally, a simple voting
based meta-predictor is shown to provide some im-
provements although more complex designs should be
considered to better exploit complementarity between
the existing prediction methods. Such advanced hetero-
geneous (using diverse types of member methods)
meta-predictors were already successfully used in se-
quence-based prediction of other protein properties such
as fold type [76, 77], subcellular localization [78-80],
structural class [81], and solvent accessibility [82].
4. SUMMARY AND CONCLUSIONS
Structural genomics efforts have entered a mature stage
when a wealth of data that could be analyzed to build
useful supporting tools has been already accumulated.
One of most significant bottlenecks in the protein struc-
ture determination pipelines implemented by SG centers
is the ability to generate diffraction quality crystals. Al-
though some mechanisms were already implemented to
improve the corresponding success rates, our analysis
shows a significant room for further improvements. In
this context we have overviewed existing databases,
analytical results and predictive methods that aim at
supporting the protein crystallization task.
We show that analysis of data from certain SG centers
and community-wide databases such as TargetBD re-

L. Kurgan et al. / Natural Science 1 (2009) 93-106
104
vealed that certain factors, such as protein size, isoelec-
tric point, disorder regions, presence of transmembrane
helices, etc. were found to correlate with the ability to
produce quality protein crystals. We also contrasted and
compared several modern sequence-based predictors of
crystallization propensity including OB-Score, ParCrys,
XtalPred and CRYSTALP2. We demonstrate that these
methods provide useful predictions which are comple-
mentary to each other. Although their average success
rate is similar and at about 70%, we show that usage of a
simple majority-vote based combination of these meth-
ods can improve the success rate to almost 74%. Our
work also reveals that close to 90% of the protein chains
can be correctly predicted by at least one of these meth-
ods, which motivates development of more advanced
meta-predictors. The best predictions for short, under
100 amino acids, chains are produced by XtalPred and
the most accurate predictions, on average, are generated
for medium-sized chains of 100 to 200 amino acids. We
believe that these crystallization propensity predictors
could provide useful input for current SG efforts that
could be incorporated into the target selection procedure.
REFERENCES
[1] Guido, R.V., Oliva, G. and Andricopulo, A.D. (2008)
Virtual screening and its integration with modern drug
design technologies. Current Medicinal Chemistry, 15(1),
37-46.
[2] Norin, M. and Sundström, M. (2001) Protein models in
drug discovery. Current Opinion in Drug Discovery &
Development, 4, 284-290.
[3] Klebe, G. (2000) Recent developments in structure-based
drug design. Journal of Molecular Medicine, 78(5),
269-281.
[4] Fernàndez-Busquets, X., de Groot, N.S., Fernandez, D.
and Ventura, S. (2008) Recent structural and computa-
tional insights into conformational diseases. Current Me-
dicinal Chemistry, 15, 1336-1349.
[5] Luscombe, N.M., Laskowski, R.A. and Thornton, J.M.
(2001) Amino acid-base interactions: a three-dimensional
analysis of protein-DNA interactions at an atomic level.
Nucleic Acids Research, 29, 2860-2874.
[6] Ellis, J.J., Broom, M. and Jones, S. (2007) Protein-RNA
interactions: structural analysis and functional classes.
Proteins, 66, 903-911.
[7] Chen, K. and Kurgan, L. (2009) Investigation of atomic
level patterns in protein - small ligand interactions. PLoS
ONE, 4(2), e4473.
[8] Pruitt, K.D., Tatusova, T. and Maglott, D.R. (2007) NCBI
Reference Sequence (RefSeq): a curated non-redundant
sequence database of genomes, transcripts and proteins.
Nucleic Acids Research, 35(Database issue), D61-65.
[9] Berman, H.M., Westbrook, J., Feng, Z., Gilliland, G.,
Bhat, T.N., Weissig, H., Shindyalov, I.N., Bourne, P.E.
(2000) The Protein Data Bank. Nucleic Acids Research,
28, 235-242.
[10] Brenner, S.E. (2001) A tour of structural genomics. Na-
ture Reviews Genetics, 2(10), 801-809.
[11] Chandonia, J.M. and Brenner, S.E. (2006) The impact of
structural genomics: expectations and outcomes. Science,
311, 347-351.
[12] Service, R.F. (2008) Protein Structure Initiative: Phase 3
or Phase Out. Science, 319(5870), 1610-1613.
[13] Terwilliger, T.C., Waldo, G., Peat, T.S., Newman, J.M.,
Chu, K. and Berendzen, J. (1998) Class-directed struc-
ture determination: Foundation for a protein structure
initiative. Protein Science, 7(9), 1851-1856.
[14] Brenner, S.E. (2000) Target selection for structural ge-
nomics. Nature Structural Biology, 7, 967-969.
[15] Dessailly, B.H., Nair, R., Jaroszewski, L., Fajardo, J.E.,
Kouranov, A., Lee, D., Fiser, A., Godzik, A., Rost, B. and
Orengo, C. (2009) PSI-2: structural genomics to cover
protein domain family space. Structure, 17(6), 869-881.
[16] Ilari, A. and Savino, C. (2008) Protein structure determi-
nation by x-ray crystallography. Methods in Molecular
Biology, 452, 63-87.
[17] Wishart, D. (2005) NMR spectroscopy and protein
structure determination: applications to drug discovery
and development. Current Pharmaceutical Biotechnol-
ogy, 6(2), 105-120.
[18] Hite, R.K., Raunser, S. and Walz, T. (2007) Revival of
electron crystallography. Current Opinion in Structural
Biology, 17(4), 389-395.
[19] Fischer, D. (2006) Servers for protein structure prediction.
Current Opinion in Structural Biology, 16(2), 178-182.
[20] Xiang, Z. (2006) Advances in homology protein structure
modeling. Current Protein & Peptide Science, 7(3),
217-227.
[21] Lacapère, J.J., Pebay-Peyroula, E., Neumann, J.M. and
Etchebest, C. (2007) Determining membrane protein
structures: still a challenge! Trends in Biochemical Sci-
ences, 32(6), 259-270.
[22] Schnell, J.R. and Chou, J.J. (2008) Structure and mecha-
nism of the M2 proton channel of influenza A virus. Na-
ture, 451, 591-595.
[23] Service, R. (2005) Structural genomics, round 2. Science,
307, 1554-1558.
[24] Chen, L., Oughtred, R., Berman, H.M. and Westbrook, J.
(2004) TargetDB: a target registration database for struc-
tural genomics projects. Bioinformatics, 20(16),
2860-2862.
[25] Slabinski, L., Jaroszewski, L., Rychlewski, L., Wilson,
I.A., Lesley, S.A. and Godzik, A. (2007) XtalPred: a web
server for prediction of protein crystallizability. Bioin-
formatics, 23(24), 3403-3405.
[26] Hui, R. and Edwards, A. (2003) High-throughput protein
crystallization. Journal of Structural Biology, 142,
154-161.
[27] Savchenko, A., Yee, A., Khachatryan, A., Skarina, T.,
Evdokimova, E., Pavlova, M., Semesi, A., Northey, J.,
Beasley, S., Lan, N., Das, R., Gerstein, M., Arrowmith,
C.H. and Edwards, A.M. (2003) Strategies for structural
proteomics of prokaryotes: quantifying the advantages of
studying orthologous proteins and of using both NMR
and x-ray crystallography approaches. Proteins, 50,
392-399.
[28] Chandonia, J.M. and Brenner, S.E. (2005) Implications
of structural genomics target selection strategies:

L. Kurgan et al. / Natural Science 1 (2009) 93-106
105
105
Pfam5000, whole genome, and random approaches. Pro-
teins, 58, 166-179.
[29] McPherson, A. (2004) Protein crystallization in the
structural genomics era. Journal of Structural and Func-
tional Genomics, 5(1-2), 3-12.
[30] Chayen, N.E. (2004) Turning protein crystallisation from
an art into a science. Current Opinion in Structural Biol-
ogy, 14(5), 577-583.
[31] Biertumpfel, C., Basquin, J. and Suck, D. (2005) Practi-
cal implementations for improving the throughput in a
manual crystallization setup. Journal of Applied Crystal-
lography, 38, 568-570.
[32] Puesy, M., Liu, Z.J., Tempel, W., Praissman, J., Lin, D.,
Wang, B.C., Gavira, J.A. and Ng, J.D. (2005) Life in the
fast lane for protein crystallization and X-ray crystallog-
raphy. Progress in Biophysics and Molecular Biology, 88,
359-386.
[33] Stevens, R.C. (2000) High-throughput protein crystalli-
zation. Current Opinion in Structural Biology, 10(5),
558-63.
[34] Rodrigues, A. and Hubbard, R.E. (2003) Making deci-
sions for structural genomics. Briefings in Bioinformatics,
4, 150-167.
[35] Lesley, S.A., Kuhn, P., Godzik, A., Deacon, A.M.,
Mathews, I., Kreusch, A., Spraggon, G., Klock, H.E.,
McMullan, D., Shin, T., Vincent, J., Robb, A., Brinen,
L.S., Miller, M.D., McPhillips, T.M., Miller, M.A.,
Scheibe, D., Canaves, J.M., Guda, C., Jaroszewski, L.,
Selby, T.L., Elsliger, M.A., Wooley, J., Taylor, S.S.,
Hodgson, K.O., Wilson, I.A., Schultz, P.G., Stevens, R.C.
(2002) Structural genomics of the Thermotoga maritima
proteome implemented in a high-throughput structure
determination pipeline. Proceedings of the National
Academy of Sciences of USA, 99, 11664-11669.
[36] Brenner, S.E., Barken, D. and Levitt, M. (1999) The
PRESAGE database for structural genomics. Nucleic
Acids Research, 27(1), 251-253.
[37] Chance, M.R., Bresnick, A.R. Burley, S.K., Jiang, J.S.,
Lima, C.D., Sali, A., Almo, S.C., Bonanno, J.B., Buglino,
J.A., Boulton, S., Chen, H., Eswar, N., He, G., Huang, R.,
Ilyin, V., McMahan, L., Pieper, U., Ray, S., Vidal, M.,
Wang, L.K. (2002) Structural genomics: pipeline for pro-
viding structures for the biologist, Protein Science, 11(4) ,
723-738.
[38] Bertone, P., Kluger, Y., Lan, N., Zheng, D., Christendat,
D., Yee, A., Edwards, A.M., Arrowsmith, C.H., Mon-
telione, G.T. and Gerstein, M. (2001) SPINE: An inte-
grated tracking database and data mining approach for
identifying feasible targets in high-throughput structural
proteomics. Nucleic Acids Research, 29, 2884-2898.
[39] Goh, C.S., Lan, N., Echols, N., Douglas, S.M., Milburn,
D., Bertone, P., Xiao, R., Ma, L.C., Zheng, D., Wunder-
lich, Z., Acton, T., Montelione, G.T. and Gerstein, M.
(2003) SPINE 2: a system for collaborative structural
proteomics within a federated database framework. Nu-
cleic Acids Research, 31, 2833-2838.
[40] Kouranov, A., Xie, L., de la Cruz, J., Chen, L., West-
brook, J., Bourne, P.E. and Berman, H.M. (2006) The
RCSB PDB information portal for structural genomics.
Nucleic Acids Research, 4(Database issue), D302-305.
[41] Berman, H.M. (2008) Harnessing knowledge from
structural genomics. Structure, 16, 16-18.
[42] Berman, H.M., Westbrook, J.D., Gabanyi, M.J., Tao, W.,
Shah, R., Kouranov, A., Schwede, T., Arnold, K., Kiefer,
F., Bordoli, L., Kopp, J., Podvinec, M., Adams, P.D.,
Carter, L.G., Minor, W., Nair, R. and La Baer, J. (2008)
The protein structure initiative structural genomics
knowledgebase. Nucleic Acids Research, 37(Database
issue), D365-368.
[43] Rupp, B. and Wang, J.W. (2004) Predictive models for
protein crystallization. Methods, 34, 391-408.
[44] Christendat, D., Yee, A., Dharamsi, A., Kluger, Y.,
Savchenko, A., Cort, J.R., Booth, V., Mackereth, C.D.,
Saridakis, V., Ekiel, I., Kozlov, G., Maxwell, K.L., Wu,
N., McIntosh, L.P., Gehring, K., Kennedy, M.A., David-
son, A.R., Pai, E.F., Gerstein, M., Edwards, A.M., Ar-
rowsmith, C.H. (2000) Structural proteomics of an ar-
chaeon. Nature Structural Biology, 7, 903-909.
[45] Goh, C.S., Lan, N., Douglas, S.M., Wu, B., Echols, N.,
Smith, A., Milburn, D., Montelione, G.T., Zhao, H. and
Gerstein, M. (2004) Mining the structural genomics
pipeline: Identification of protein properties that affect
high-throughput experimental analysis. Journal of Mo-
lecular Biology, 336, 115-130.
[46] Canaves, J.M., Page, R., Wilson, I.A. and Stevens, R.C.
(2004) Protein biophysical properties that correlate with
crystallization success in Thermotoga maritima: Maxi-
mum clustering strategy for structural genomics. Journal
of Molecular Biology, 344, 977-991.
[47] Kantardjieff, K.A. and Rupp, B. (2004) Protein isoelec-
tric point as a predictor for increased crystallization
screening efficiency. Bioinformatics, 20, 2162-2168.
[48] Kantardjieff, K.A., Jamshidian, M. and Rupp, B. (2004)
Distributions of pI vs pH provide strong prior informa-
tion for the design of crystallization screening experi-
ments. Bioinformatics, 20, 2171-2174.
[49] Longenecker, K.L., Garrard, S.M., Sheffield, P.J. and
Derewenda, Z.S. (2001) Protein crystallization by ra-
tional mutagenesis of surface residues: Lys to Ala muta-
tions promote crystallization of RhoGDI. Acta Crystal-
lographica Section D: Biological Crystallography, 57,
679-688.
[50] Mateja, A., Devedjiev, Y., Krowarsch, D., Longenecker,
K., Dauter, Z., Otlewski, J., Derewenda, Z.S. (2002) The
impact of Glu-Ala and Glu-Asp mutations on the crystal-
lization properties of RhoGDI: the structure of RhoGDI
at 1.3 A resolution. Acta Crystallographica Section D:
Biological Crystallography, 58, 1983-1991.
[51] Derewenda, Z.S. (2004) The use of recombinant methods
and molecular engineering in protein crystallization.
Methods, 34, 354-363.
[52] Derewenda, Z.S. (2004) Rational protein crystallization
by mutational surface engineering. Structure, 12,
529-535.
[53] Derewenda, Z.S. and Vekilov, P.G. (2006) Entropy and
surface engineering in protein crystallization. Acta Crys-
tallographica Section D: Biological Crystallography, 62,
116-124.
[54] Cooper, D.R., Boczek, T., Grelewska, K., Pinkowska, M.,
Sikorska, M., Zawadzki, M. and Derewenda, Z. (2007)
Protein crystallization by surface entropy reduction: op-
timization of the SER strategy. Acta Crystallographica
Section D: Biological Crystallography, 63, 636-645.

L. Kurgan et al. / Natural Science 1 (2009) 93-106
106
[55] Goldschmidt, L., Cooper, D.R., Derewenda, Z. and
Eisenberg, D. (2007) Toward rational protein crystalliza-
tion: A Web server for the design of crystallizable protein
variants. Protein Science, 16, 1569-1576.
[56] Oldfield, C.J., Ulrich, E.L., Cheng, Y., Dunker, A.K. and
Markley, J.L. (2005) Addressing the intrinsic disorder
bottleneck in structural proteomics. Proteins, 59,
444-453.
[57] Chandonia, J.M., Kim, S.H. and Brenner, S.E. (2006)
Target selection and deselection at the Berkeley Struc-
tural Genomics Center. Proteins, 62, 356-370.
[58] Price, W.N. 2nd, Chen, Y., Handelman, S.K., Neely, H.,
Manor, P., Karlin, R., Nair, R., Liu, J., Baran, M., Everett,
J., Tong, S.N., Forouhar, F., Swaminathan, S.S., Acton, T.,
Xiao, R., Luft, J.R., Lauricella, A., DeTitta, G.T., Rost, B.,
Montelione, G.T. and Hunt, J.F.. (2009) Understanding
the physical properties that control protein crystallization
by analysis of large-scale experimental data. Nature Bio-
technology, 27(1), 51-57.
[59] Chou, K.C. (2004) Structural bioinformatics and its im-
pact to biomedical science. Current Medicinal Chemistry,
11, 2105-2134.
[60] Chou, K.C. (2005) Progress in protein structural class
prediction and its impact to bioinformatics and pro-
teomics. Current Protein & Peptide Science, 6, 423-436.
[61] Yang, Z. R., Wang, L., Young, N. and Chou, K.C. (2005)
Pattern recognition methods for protein functional site
prediction. Current Protein & Peptide Science, 6,
479-491.
[62] Chou, K.C. and Shen, H.B. (2007) Recent progresses in
protein subcellular location prediction. Analytical Bio-
chemistry, 370, 1-16.
[63] Kurgan, L., Cios, K.J., Zhang, H., Zhang, T., Chen, K.,
Shen, S. and Ruan, J. (2008) Sequence-based methods
for real value predictions of protein structure. Current
Bioinformatics, 3(3), 183-196.
[64] Smialowski, P., Schmidt, T., Cox, J., Kirschner, A. and
Frishman, D. (2006) Will my protein crystallize? A se-
quence-based predictor. Proteins, 62, 343-355.
[65] Overton, I.M. and Barton, G.J. (2006) A normalised scale
for structural genomics target ranking: the OB-Score.
FEBS Letters, 580, 4005-4009.
[66] Overton, I.M., Padovani, G., Girolami, M.A. and Barton,
G.J. (2008) ParCrys: a Parzen window density estimation
approach to protein crystallization propensity prediction.
Bioinformatics, 24, 901-907.
[67] Slabinski, L., Jaroszewski, L., Rodrigues, A.P.C.,
Rychlewski, L., Wilson, I.A., Lesley, S.A. and Godzik, A.
(2007) The challenge of protein structure determination -
lessons from structural genomics. Protein Science, 16(11),
2472-2482.
[68] Slabinski, L., Jaroszewski, L., Rychlewski, L., Wilson,
I.A., Lesley, S.A. and Godzik, A. (2007) XtalPred: a web
server for prediction of protein crystallizability. Bioin-
formatics, 23(24), 3403-3405.
[69] Chen, K., Kurgan, L. and Rahbari, M. (2007) Prediction
of protein crystallization using collocation of amino acid
pairs. Biochemical and Biophysical Research Communi-
cations, 355, 764-769.
[70] Kurgan, L., Razib, A.A., Aghakhani, S., Dick, S.,
Mizianty, M.J. and Jahandideh, S. (2009) CRYSTALP2:
sequence-based protein crystallization propensity predic-
tion. BMC Structural Biology, 9, 50.
[71] Campbell, K. and Kurgan, L. (2008) Sequence-only
based prediction of ß-turn location and type using collo-
cation of amino acid pairs. Open Bioinformatics Journal,
2, 37-49.
[72] Chen, K., Kurgan, L. and Ruan, J. (2007) Prediction of
flexible/rigid regions in proteins from sequences using
collocated amino acid pairs. BMC Structural Biology, 7,
25.
[73] Chen, Y.Z., Tang, Y.R., Sheng, Z.Y. and Zhang, Z. (2008)
Prediction of mucin-type O-glycosylation sites in mam-
malian proteins using the composition of k-spaced amino
acid pairs. BMC Bioinformatics, 9, 101.
[74] Chen, K., Jiang, Y., Du, L. and Kurgan, L. (2009) Predic-
tion of integral membrane protein type by collocated hy-
drophobic amino acid pairs. Journal of Computational
Chemistry, 30(1), 163-172.
[75] Kurgan L. (2008) On the relation between the predicted
secondary structure and the protein size. The Protein
Journal, 24(4), 234-239.
[76] Shen, H.B. and Chou, K.C. (2009) Predicting protein
fold pattern with functional domain and sequential evo-
lution information. Journal of Theoretical Biology,
256(3), 441-446.
[77] Chen, K. and Kurgan, L. (2007) PFRES: protein fold
classification by using evolutionary information and pre-
dicted secondary structure. Bioinformatics, 23(21),
2843-2850.
[78] Assfalg, J., Gong, J., Kriegel, H.P., Pryakhin, A., Wei, T.
and Zimek, A. (2009) Supervised ensembles of predic-
tion methods for subcellular localization. Journal of Bio-
informatics and Computational Biology, 7(2), 269-285.
[79] Shen, H.B. and Chou, K.C. (2007) Hum-mPLoc: an en-
semble classifier for large-scale human protein subcellu-
lar location prediction by incorporating samples with
multiple sites. Biochemical and Biophysical Research
Communications, 355(4), 1006-1011.
[80] Chou, K.C. and Shen, H. B. (2006) Hum-PLoc: A novel
ensemble classifier for predicting human protein subcel-
lular localization. Biochemical and Biophysical Research
Communications, 347, 150-157.
[81] Kedarisetti, K.D., Kurgan, L. and Dick, S. (2006) Classi-
fier ensembles for protein structural class prediction with
varying homology. Biochemical and Biophysical Re-
search Communications, 348(3), 981-988.
[82] Chen, H. and Zhou, H.X. (2005) Prediction of solvent
accessibility and sites of deleterious mutations from pro-
tein sequence. Nucleic Acids Research, 33(10),
3193-3199.