Content uploaded by Pascal Benkert
Author content
All content in this area was uploaded by Pascal Benkert on Mar 16, 2015
Content may be subject to copyright.
Protein structure homology modeling using
SWISS-MODEL workspace
Lorenza Bordoli, Florian Kiefer, Konstantin Arnold, Pascal Benkert, James Battey & Torsten Schwede1,2
1Biozentrum, University of Basel, Klingelbergstrasse 50-70, CH 4056 Basel, Switzerland. 2SIB Swiss Institute of Bioinformatics, Biozentrum, University of Basel,
Klingelbergstrasse 50/70, CH 4056 Basel, Switzerland. Correspondence should be addressed to T.S. (torsten.schwede@unibas.ch).
Published online 11 December 2008; doi:10.1038/nprot.2008.197
Homology modeling aims to build three-dimensional protein structure models using experimentally determined structures of related
family members as templates. SWISS-MODEL workspace is an integrated Web-based modeling expert system. For a given target
protein, a library of experimental protein structures is searched to identify suitable templates. On the basis of a sequence alignment
between the target protein and the template structure, a three-dimensional model for the target protein is generated. Model quality
assessment tools are used to estimate the reliability of the resulting models. Homology modeling is currently the most accurate
computational method to generate reliable structural models and is routinely used in many biological applications. Typically, the
computational effort for a modeling project is less than 2 h. However, this does not include the time required for visualization and
interpretation of the model, which may vary depending on personal experience working with protein structures.
INTRODUCTION
The three-dimensional structure of a protein provides important
information for understanding its biochemical function and inter-
action properties in molecular detail. However, the number of
known protein sequences is much larger than the number of
experimentally solved protein structures. As of August 2008,
more than 52,500 experimentally determined protein structures
were deposited in the Protein Data Bank (PDB)1. Yet, this number
appears relatively small compared with the more than 6 million
protein sequences held in the UniProt knowledge database2. For-
tunately, the number of different protein fold families occurring in
nature appears to be limited3, and within a protein family,
structural similarity between two homologous proteins can be
inferred from sequence similarity4. Homology modeling (or com-
parative protein structure modeling) techniques have been devel-
oped to build three-dimensional models of a protein (target) from
its amino-acid sequence on the basis of an alignment with a similar
protein with known structure (template)5–7. In cases where no
suitable template structure can be identified, de novo (a.k.a. ab
initio) structure prediction methods can be used to generate three-
dimensional protein models without relying on a homologus
template structure. However, despite recent progress in the field,
de novo predictions are limited to relatively small proteins and fall
short in terms of accuracy compared with comparative models8–12.
Therefore, homology modeling is the method of choice to build
reliable three-dimensional in silico models of a protein in all cases
where template structures can be identified.
Homology models are widely used in many applications, such as
virtual screening, designing site-directed mutagenesis experiments
or in rationalizing the effects of sequence variations13–17.Stable,
reliable and accurate systems for automated homology modeling
are therefore required, which are easy to use for both nonspecialists
and experts in structural bioinformatics.
Homology modeling
Homology modeling in general consists of four main steps: (i)
identifying evolutionarily related proteins with experimentally
solved structures that can be used as template(s) for modeling
the target protein of interest; (ii) mapping corresponding residues
of target sequence and template structure(s) by means of sequence
alignment methods and manual adjustment; (iii) building the
three-dimensional model on the basis of the alignment; and
(iv) evaluating the quality of the resulting model14,15.Thisproce-
dure can be iterated until a satisfactory model is obtained (Fig. 1).
Protein structure homology modeling relies on the evolutionary
relationship between the target and template proteins. Potential
structural templates are identified using a search for homologous
proteins in a library of experimentally determined protein struc-
tures. From the resulting list of possible candidate structures, a
template structure is chosen on the basis of its suitability according
to various criteria such as the level of similarity between the query
and template sequences, the experimental quality of the solved
structures, the presence of ligands or cofactors and so on. Ideally, a
large segment of the query sequence should be covered by a single
high-quality template, although in many cases, the available tem-
plate structures will correspond to only one or more distinct
structural domains of the protein.
p
uor
G
gn
i
h
s
i
lb
uP eru
t
a
N
800
2
©natureprotocol
s
/moc.erut
a
n.w
ww//:ptth
Known structures
(templates)
Target
sequence Template selection
Alignment
template–target
Structure modeling
Structure evaluation and
assessment
Homology
model(s)
Figure 1
|
The four main steps of comparative protein structure modeling:
template selection, target–template alignment, model building and model
quality evaluation.
NATURE PROTOCOLS |VOL.4 NO.1 |2009 |1
PROTOCOL
Estimating the accuracy of a protein structure model is a crucial
step in the whole process, as it is the quality of the model that
determines its possible applications18. The quality of the obtained
models will depend on the evolutionary distance between the target
and the template proteins. It has been shown that there is a direct
correlation between the sequence identity level of a pair of protein
structures and the deviation of the Caatoms of their common core.
The more similar two sequences are, the closer the corresponding
structures can be expected to be and the larger the fraction of the
model that can be directly inferred from the template4. As com-
parative models result from a structural extrapolation guided by a
sequence alignment, the percentage of sequence identity between
target and template is generally accepted as a reasonable first
estimate of the quality of the structurally conserved core of the
model. As a rule of thumb, the core Caatoms of protein models
sharing 50% sequence identity with their templates will deviate by
B1.0 A
˚root mean square deviation from experimentally eluci-
dated structures. Although the atomic coordinates of the three-
dimensional model, for regions of the target protein aligned to the
template, can be modeled on the basis of the information provided
by template structure5–7, regions that are not aligned with a tem-
plate (insertions/deletions) require specialized approaches19–22.
Unaligned regions of the target that are modeled using de novo
techniques, such as loops, will on average be less accurate than
structurally conserved regions of the model on the basis of
information derived directly from the template.
As the percentage identity falls below B30% (in the so-called
‘twilight zone’), model quality estimation on the basis of sequence
identity becomes unreliable, as the relationship between sequence
and structure similarity gets increasingly dispersed18,23. With
decreasing sequence identity, alignment errors and the incorrect
modeling of large insertions become the major source of inaccura-
cies. Correctly aligning the target sequence with the template is a
crucial step and one of the primary sources of errors in the whole
modeling procedure. The development of algorithms for sequence
comparison and alignment is a major topic of study in bioinfor-
matics, and the advancements in this field have been comprehen-
sively reviewed in ref. 24. For estimating the overall quality of
protein structure models and comparing predictions on the basis of
alternative alignments25–27, scoring functions such as statistical
potentials of mean force have been developed. Methods that
allow identifying local errors in models are currently an active
field of research28–32. The stereochemical plausibility of the gener-
ated models can be assessed using tools such as PROCHECK33 and
WHATCHECK34, which help to identify amino-acid conforma-
tions deviating from expected values for structural features such as
bond lengths and angles.
Accuracy and limitations of homology modeling
Comparative modeling relies on establishing an evolutionary
relationship between the sequence of the protein of interest and
other members of the protein family, whose structures have been
solved experimentally by X-ray or NMR. For this reason, the major
limitation of this technique is the availability of homologous
templates, i.e., only regions of the protein corresponding to an
identified template can be modeled accurately. As experimental
protein structures are often available only for individual structural
domains, it is often not possible to infer the correct relative domain
orientation in a model.
Modeling oligomeric proteins, i.e., complexes composed of more
than one polypeptide chain, may be straightforward in cases where
the complex of interest is similar to a homologous complex of
known structure. However, this situation is relatively rare, as most
experimental structures in the PDB consist of individual proteins
rather than complexes. Modeling complexes from individual com-
ponents is a daunting task35 and rarely successful without integrat-
ing additional information about the assembly36.
Comparative protein modeling techniques rely on structural
information from the template to derive the structure of the target.
Large structural changes, e.g., caused by mutations, insertions,
deletions and fusion proteins, are therefore, in general, not
expected to be modeled accurately by comparative techniques.
Nonetheless, homology models of a protein under investigation
can provide a valuable tool for the interpretation of sequence
variation and the design of mutagenesis experiment to elucidate the
biological function of proteins16,17,37.
The reliability of different protein modeling methods can be
objectively evaluated by examining the quality of predictions made
during blinded tests. For example, in CASP7, the ‘Community
Wide Experiment on the Critical Assessment of Techniques for
Protein Structure Prediction’ in 2006, predictions for 108 homo-
logy modeling targets were analyzed in detail to identify progress
and limitations of current protein structure prediction methods11.
Particular emphasis was also given to the analysis of the results of
automated prediction servers whose accuracy has significantly
increased over the last years. Details about the participating
servers and public accessibility are given in Table 1 of ref. 38.
Similarly, the EVA39 project provides a continuous assessment of
the stability and accuracy of automated modeling servers on
the basis of a large number of blind predictions. SWISS-MODEL
was the first comparative modeling server to join the EVA
project in May 2000. All results of this evaluation are available at
http://eva.compbio.ucsf.edu/~eva/.
Availability
SWISS-MODEL workspace40 can be freely accessed by the biologi-
cal community on the Web at http://swissmodel.expasy.org/
workspace/. SWISS-MODEL has been the first automated model-
ing server publicly available7. In the meantime, similar services have
been developed by other groups, e.g., ModPipe41,3D-JIGSAW
42 or
M4T43. For a more complete listing of other publicly available
comparative modeling servers, we refer the readers to the annual
Nucleic Acids Research Web server issue44.
SWISS-MODEL workspace
Each of the four steps in homology modeling requires specialized
software as well as access to up-to-date protein sequence and
structure databases. The SWISS-MODEL workspace40 integrates
the software required for homology modeling and databases in an
easy-to-use, Web-based modeling environment. The workspace
assists the user in building and evaluating protein homology
models at different levels of complexity—depending on the diffi-
culty of the individual modeling task. A highly automated model-
ing procedure with a minimum of user intervention is provided for
modeling scenarios where highly similar structural templates are
available7,14,45–47. For more complex modeling tasks where target
and template have lower sequence similarity, expert users are given
control over the several steps of model building to construct a
p
uor
G
gn
i
h
s
i
lb
uP eru
t
a
N
800
2
©natureprotocol
s
/moc.erut
a
n.w
ww//:ptth
2|VOL.4 NO.1 |2009 |NATURE PROTOCOLS
PROTOCOL
protein model that is optimally adapted to their scientific pro-
blem48. Modeling can be performed from within a Web browser
without the need for downloading, compiling and installing large
program packages or databases. The results of different modeling
tasks are presented in a graphical summary. As quality evaluation is
indispensable for a predictive method like homology modeling,
every model is accompanied by several quality checks. In the
following sections, we describe the main components of the
SWISS-MODEL workspace.
Tools for target sequence feature annotation. Functional and
structural domain annotation of the target sequence of interest is
the first step toward the identification of a suitable template for
building its three-dimensional model. Individual structural
domains of multidomain proteins often correspond to units of
distinct molecular function49–51. Furthermore, the sensitivity of
profile-based template detection methods can be enhanced when
the search is performed at the domain level rather than searching
the whole protein sequence. IprScan, a PERL-based InterProS-
can52,53 utility, has been integrated in the SWISS-MODEL work-
space for the analysis of the domain architecture of the target
protein and the annotation of its functional features. Prediction
tools for secondary structure54, disorder55 and transmembrane
(TM) regions56 complement the tools for protein sequence analysis
and aid the selection of suitable modeling templates for specific
regions of the target proteins. In the twilight zone of sequence
alignments, applying secondary structure prediction to the protein
of interest may help deciding whether a putative template shares
essential structural features. Intrinsically unstructured regions in
proteins have been associated with numerous important biological
cellular functions, from cell signalling to transcriptional regula-
tion57,58; several examples of such disordered regions undergoing
the transition to an ordered state upon binding their ligand
proteins have been reported59. Prediction of disordered and tras-
membrane regions therefore complement the analysis of protein
domain boundaries and functional annotation of the target pro-
tein.
Tools for template identification. The SWISS-MODEL work-
space provides a set of increasingly sensitive sequence-based search
methods for template detection that are applied depending on the
evolutionary divergence between the target protein and the closest
structurally characterized template protein. Close homologs of the
target can be identified using a gapped BLAST60 query against the
SWISS-MODEL Template Library (SMTL)40. When no closely
related templates are found, or can be identified only for some
segments of the target protein, more sensitive approaches for
detecting evolutionary relationships are provided. (i) In the itera-
tive profile Blast approach60, which has been initially introduced as
PDB-Blast by Godzik and coworkers, a profile for the target
sequence is compiled from homologous sequences by iterative
searches of the NR database61 and used subsequently to search
SMTL for homologous structures. (ii) Alternatively, to detect more
distantly related template structures, a Hidden Markov Model
(HMM) for the target sequence is built on the basis
of a multiple sequence alignment, similarly to the profile Blast
approach discussed above. The HMM for the target sequence is
subsequently used to search against the template library of
HMMs generated for a nonredundant set of the sequences of the
SMTL template library culled at 70% sequence identity.
HMM building, calibration and library searches are performed
using the HHSearch (v. 1.5.01) software package62. For the
selection of a suitable template, the following issues need to be
considered:
1. The selection of thebest template structure not only depends on
sequence similarity but should also take into account other
factors, such as experimental quality, bound substrate molecules
or different conformational states of the template. For example,
certain proteins undergo large conformational changes upon
substrate binding as observed, e.g., between the apostructure
and ATP- and ADP-bound forms of enzymes in the nucleotide
kinase family63. Depending on the planned model applications,
such as structure-based ligand design, it is necessary to choose a
structural template in the correct conformation.
2. For low-homology templates, the InterPro functional annota-
tion of the target sequence can be used to verify that putative
templates share essential functional features.
3. If the target–template alignment falls within the ‘twilight
zone’ of sequence alignments (i.e., below 30% sequence
identity), secondary structure prediction of the target protein
may help to decide whether a putative template shares structural
features with your protein and may therefore be used as
template.
4. Predicted disorder regions may indicate the boundaries of
protein domains and provide additional functional annotation
of the protein.
Modeling. Automated mode: If the alignment between the target
and the template sequences displays a sufficiently high similarity, a
fully automated homology modeling approach can be applied. As a
rule of thumb, automated sequence alignments are sufficiently
reliable when target and template share more than 50% of sequence
identity. Submissions in ‘Automated mode’ require only the amino-
acid sequence or the UniProt2accession code of the target protein
as input data. The hierarchical approach for template detection of
the modeling pipeline will automatically select suitable templates
on the basis of a Blast search or using an adapted sequence-to-
HMM comparison HHSearch protocol64. In cases where several
similar template structures are available, the automated template
selection will favor high-resolution template structures with good-
quality assessment. Optionally, a specific template from the SMTL
template library can be specified.
Alignment mode: For more distantly related target and template
sequences, the number of errors in automated sequence alignments
increases23. This poses a major problem for automated homology
modeling, as current methods are not capable of recovering from
an incorrect input alignment. In many molecular biology projects,
multiple sequence alignments are often the result of extensive
theoretical and experimental exploration of a family of proteins.
Such alignments can be used for comparative modeling using the
‘Alignment mode’ if at least one of the member sequences repre-
sents a protein for which the three-dimensional structure is known.
The ‘Alignment mode’ allows the user to test several alternative
alignments and evaluate the quality of the resulting models to
achieve an optimal result.
Project mode: In the so-called ‘twilight zone’ of sequence align-
ments, when the sequence identity between target and template is
below 30%, it is advisable to visually inspect and manually edit the
p
uor
G
gn
i
h
s
i
lb
uP eru
t
a
N
800
2
©natureprotocol
s
/moc.erut
a
n.w
ww//:ptth
NATURE PROTOCOLS |VOL.4 NO.1 |2009 |3
PROTOCOL
target–template alignment. This will lead to
a significant improvement of the quality of
the resulting model. The program Deep-
View (Swiss-PdbViewer)48 can be used to
display, analyze and manipulate modeling
projects. DeepView project files contain one
or more superposed template structures
and the alignment between the target and
template(s). Project files are also generated
by the workspace template selection tools
and are the default output format of the
modeling pipeline. Project files with mod-
ified alignments can then be saved to disk
and submitted as ‘Project mode’ to the
workspace for model building by the
SWISS-MODEL pipeline, thereby giving
the user full control over essential modeling
parameters: several template structures can
be compared simultaneously to identify
structurally conserved and variable regions
and select the most suitable template. The
placement of insertions and deletions in the
target–template alignment can be visualized
in their structural context and adjusted
accordingly.
Protein structure assessment and model
quality estimation. The percentage of
sequence identity between target and tem-
plate is generally accepted as a reasonable
first estimate of the quality of a model.
However, the accuracy of individual models
may vary significantly from the expected
average quality due to suboptimal target–
template alignments, low template quality,
structural flexibility or inaccuracies intro-
duced by the modeling program. Individual
assessment of each model is therefore essen-
tial. As a global indicator of the quality of a
given model, the results of QMEAN65,a
composite scoring function for model
quality estimation, and DFIRE30, an all-
atom distance-dependent statistical poten-
tial, are provided in the SWISS-MODEL workspace. However, a
good global score does not guarantee that important functional
sites of a protein have been modeled correctly. Therefore,
tools for local model quality estimates are included: graphical
plots of ANOLEA mean force potential28,GROMOSempirical
force field energy66 and the neural network-based approach
ProQres32 are provided as indicators for local model quality.
Finally, Whatcheck34 and Procheck33 reports enable the user
to assess the conformational quality of both models and
template structures.
In this protocol, we describe in a step-by-step procedure (Fig. 2)
how users can benefit from the integrated design of the
SWISS-MODEL workspace to build and assess the accuracy of
homology models.
MATERIALS
EQUIPMENT
.Amino-acid sequence of the protein to be modeled
.A computer with access to the Internet and a Web browser
.A multiple protein sequence alignment, including at least the
sequences of the target protein and the template structure
(optional; see Step 6B in PROCEDURE for information on sequence
alignment formats)
.DeepView for protein structure analysis and visualization (optional
software). DeepView can be freely downloaded from the ExPASy website
(http://www.expasy.org/spdbv/)
p
uor
G
gn
i
h
s
i
lb
uP eru
t
a
N
800
2
©natureprotocol
s
/moc.erut
a
n.w
ww//:ptth
Template(s)
Target–template
alignment (PF)
DeepView
Model
Alignment
Mode
Alignment
mode 6B Project
Mode
Project
mode 6C
Automated
Mode
Automated
mode 6A
Model building
Model
Quality Estimation
Quality estimation
Quality OK?
Annotation
Target feature
annotation
Selection
Template
selection
…
Target sequence
Model Model
report
Yes
No
Target–template
alignment (MSA)
MSA tools
SWISS-MODEL Workspace
Step 2
(optional)
Step 3
Step 4
Step 5A
Step 5B
Step 6
Step 7
Step 8
Return to
Steps 4, 5A, 5B
Template
Automated
target–template
alignment
Step 5
Template(s)
Figure 2
|
Workflow of comparative protein structure modeling using SWISS-MODEL workspace.
Starting from the amino-acid sequence of a ‘Target’ protein, three alternative routes for model building
are provided—depending on the difficulty of the modeling task. Individual steps are described in
detail in PROCEDURE.
4|VOL.4 NO.1 |2009 |NATURE PROTOCOLS
PROTOCOL
PROCEDURE
Access and personal user account
1| Access and create a personal
account for the SWISS-MODEL workspace
at http://swissmodel.expasy.org/
workspace/. The user data are stored in
a password-protected personal user
space, which is identified by the user’s
email address. It is also possible to
access the workspace system
anonymously without providing an email
address. However, it is then necessary to
bookmark the URLs of individual work
units in the Web browser to be able to
retrieve the results once the browser
session has been closed. Once logged in
your personal user account, individual
modeling tasks are organized in work
units under ‘Workspace’; their current
computational status is represented
graphically (Fig. 3).
Sequence feature annotation
2| Examine your target sequence. The
resultsofthisanalysiswillassistyouin
deciding which of the possible template(s) (obtained in Step 3) to use to build homology model(s). Submit your protein
sequence (as plain text, in FASTA format or its UniProt Accession Code) to one or more of the tools available in the ‘Sequence
Features Scan’ session of the server, you find under ‘Tools’: use option A for InterPro domain scan, option B for PsiPred, option C
for DISOPRED and option D for MEMSAT.
(A) InterPro domain scan
(i) InterPro domain scan52 identifies known protein domains and functional sites of the target sequence and possibly assigns
the protein to a specific family. The following databases, currently part of the InterPro scan method, can be selected:
HMMPfam—the target sequence is searched against the Pfam67 database, a large collection of multiple sequence align-
ments and hidden Markov models covering many common protein domains and families; ProfileScan—the target sequence
is searched against the profiles collection of PROSITE68, a database of protein families and domains, and it consists of bio-
logically significant sites, patterns and profiles that help in identifying to which known protein family a sequence belongs;
ScanRegExp—the target sequence is scanned for biologically significant patterns contained in the PROSITE database col-
lection, e.g., enzyme catalytic sites, phosphorylation sites and so on.
(ii) The occurrence of domains and functional sites are displayed on the target sequence. Domain boundaries and links to
InterPro database instruct about distinctive features of a given functional domain or provide documentation relative to a
specific protein family.
(B) PsiPred
(i) PsiPred54 predicts secondary structure elements of the target sequence. The graphical representation shows the probability
of a given residue of being part of an alpha helix (H), extended beta strand (E) or a coil region (C).
(C) DISOPRED
(i) DISOPRED255 predicts the occurrence of disordered regions in the target protein. The probability of being disordered
(ranging from 0 to 1) is plotted for each position in the sequence. The ‘output’ and ‘filter’ curves represent the raw and
filtered scores from the linear SVM classifier (DISOPREDsvm), respectively. Both outputs from DISOPREDsvm are included to
allow the user to identify shorter, low-confidence predictions of disorder. Asterisks (*) and dots (.) denote predicted
disorder and order, respectively. DISOPRED2 predictions are given at a default false-positive rate threshold of 2%, but this
value can be changed by the user.
?TROUBLESHOOTING
(D) MEMSAT
(i) MEMSAT56 predicts the occurrence of putative TM segment in the protein. Central TM helix segments are indicated with ‘X’
in the output sequence. Information about the predicted TM topology is also provided.
p
uor
G
gn
i
h
s
i
lb
uP eru
t
a
N
800
2
©natureprotocol
s
/moc.erut
a
n.w
ww//:ptth
Figure 3
|
Example of a personal user workspace. In SWISS-MODEL workspace, individual modeling tasks
are organized in work units; their current computational status is represented graphically.
NATURE PROTOCOLS |VOL.4 NO.1 |2009 |5
PROTOCOL
Template identification and target template alignment
3| Submit your target sequence (as plain text, in FASTA format or its UniProt Accession Code) to one or more of the template
identification tools of the ‘Template Identification’ session you find under ‘Tools’. The server provides access to a set of
increasingly complex and computationally demanding methods: use option A for BLAST, option B for PSI-BLAST and option C for
HMM-HMM-based searching.
(A) BLAST
(i) Closely related homologous templates are identified by running a gapped BLAST search60. Adjust standard BLAST parameter
like E-value cutoff or the choice of the substitution matrix to alter the sensitivity or specificity of your search.
?TROUBLESHOOTING
(B) PSI-BLAST
(i) More divergent template structures can be identified using iterative, profile-based BLAST60.
(ii) Profile generation: selectivity and sensitivity of the search can be adjusted in the profile generation step by altering the
number of iterations and the inclusion threshold for building the target profile. A more permissive E-value threshold and a
greater number of iterations will increase the sensitivity of your search. Note that the inclusion of false positives, i.e.,
proteins that do not belong to the family of interest, during profile building can cause a drift in the search and lead to an
increased false-positive rate among your hits.
(iii) Profile search: in the template library search step, the balance between selectivity and sensitivity can be adjusted by the
choice of the substitution matrix.
?TROUBLESHOOTING
(C) HMM-HMM-based searching
(i) Distantly related templates can be identified using HMM-based profile matching using HHSearch62. A profile of the query
sequence is generated and used for identifying matching HMM profiles in the template library. As this approach is compu-
tationally more intensive, compared with methods described in Steps 3A and B, the query is performed against a reduced
version of the PDB database (culled at the 70% sequence identity level).
4| Select one or more structures from the result hit list as template to build comparative models. Results of template selection
(Step 3) and domain identification (Step 2A) are displayed in a condensed graphical overview. This combined view allows you to
analyze template coverage with respect to the domain boundaries and to identify templates spanning one or more domains of
the target. Bars indicating matching regions will link you to the underlying target–template alignment and links to the SMTL
library are available to facilitate the choice of a suitable template.
?TROUBLESHOOTING
5| Once you have selected one or more suitable templates, the following options are possible to improve the initial target–
template alignment: Option A—DeepView Project or option B—alternative sequence alignment methods.
mCRITICAL STEP This is a particularly critical step, as homology modeling techniques cannot recover from an incorrect starting
target–template alignment.
(A) DeepView Project
(i) The target–template sequence alignments generated by the different template database search techniques can be used as
the basis for the subsequent model creation. The alignments can be downloaded as DeepView project file, which contains
the target sequence aligned to the template structure.
(ii) The program DeepView allows you to display and analyze the alignment in the structural context of the template to manu-
ally adjust misaligned regions.
?TROUBLESHOOTING
(iii) Once you have finished editing the alignment, save the project file on the local disk and submit it to the ‘Project Mode’ of
the Modeling session for model building (Step 6C).
(B) Alternative sequence alignment methods
(i) You might also want to apply alternative sequence alignment methods by using multiple sequence alignment programs to
align the target and the template sequences obtained in Step 3. For a list of the most widely used sequence alignments
tools, please refer to ref. 24, Table 1 therein.
(ii) The obtained sequence alignment between target and template (and additional homologous proteins) can be submitted to
the ‘Alignment mode’ of the modeling session for model building (Step 6B).
Modeling
6| To obtain an homology model of your target sequence, you can choose among three different approaches—accessible
through the ‘Modeling’ session of the server—whose applicability depends primarily on how distantly related your protein and
the homologous template are: option A—automated mode; option B—alignment mode; or option C—project mode.
p
uor
G
gn
i
h
s
i
lb
uP eru
t
a
N
800
2
©natureprotocol
s
/moc.erut
a
n.w
ww//:ptth
6|VOL.4 NO.1 |2009 |NATURE PROTOCOLS
PROTOCOL
(A) Automated mode
(i) In cases where the target–template similarity is
sufficiently high to allow for unambiguous sequence
alignment, homology modeling can be fully automated.
Submit your target sequence as plain text, FASTA format
or its UniProt accession code.
(ii) Optionally, a specific template, e.g., identified by a
Blast search in Step 3, can be specified by its PDB identi-
fier and chain ID. Make sure that the specified template
ID is present in the SWISS-MODEL template library.
?TROUBLESHOOTING
(B) Alignment mode
(i) With decreasing sequence similarity between target and template, the number of errors in automatically generated
sequence alignments increases. Therefore, you might choose to submit an alignment generated by alternative sequence
alignment tools (Step 5B). Provide a pairwise or multiple sequence alignment as input alignment in FASTA, MSF, ClustalW,
PFAM or SELEX format.
?TROUBLESHOOTING
(ii) After the alignment has been converted into a standard format, indicate which sequence corresponds to the target
protein and which corresponds to a protein with known structure in the template library.
?TROUBLESHOOTING
(iii) Submit your alignment for model calculation. Note that the SWISS-MODEL pipeline used for the modeling process might
introduce minor heuristic modifications to improve the placement of insertions and deletions during model building.
(C) Project mode
(i) In the twilight zone of sequence alignments, visual inspection and manual manipulation of the target–template alignment
can significantly improve the quality of the resulting model. Using Project mode, you can submit Project files that you
have obtained in Step 5A after adjusting the alignment in DeepView.
?TROUBLESHOOTING
(ii) In project mode, you can also submit projects generated directly inside DeepView. With this option, it is possible to
generate models using templates that are not part or not yet present in the SMTL library.
(iii) Oligomer modeling: template-based modeling of oligomeric assemblies (Fig. 4) is possible using DeepView and
subsequently submitting the file to the Project Mode (Box 1).
7| After completion of the modeling procedure, the results are stored in the workspace and, if specified in your personal
setting, you will be notified of the completion. Coordinates of the model, the underlying alignment, log files and quality
evaluations can be accessed and downloaded from the personal workspace (Fig. 5). The model coordinates are available in PDB
or DeepView project file format. The latter allows you to further inspect and manually modify the target–template alignment.
Modified project files can then be saved to disk and submitted as ‘Project mode’ to the workspace for a further iteration of the
model-building cycle (Step 6C). Energy profiles from the ANOLEA statistical potential28 as well as the GROMOS force field66 are
p
uor
G
gn
i
h
s
i
lb
uP eru
t
a
N
800
2
©natureprotocol
s
/moc.erut
a
n.w
ww//:ptth
>TARGET
QGQEPPPEPRITLTVGGQPVTFLVDTGAQH
SVLTQNPGPLSDRSAWVQGATGGKRYRWTT
DRKVHLATGKVTHSFLHVPDCPYPLLGRDL
LTKLKAQI;
QGQEPPPEPRITLTVGGQPVTFLVDTGAQH
SVLTQNPGPLSDRSAWVQGATGGKRYRWTT
DRKVHLATGKVTHSFLHVPDCPYPLLGRDL
LTKLKAQI
Figure 4
|
Example of the input format for an oligomeric target sequence
(Box 1).
BOX 1 |OLIGOMER MODELING
(i) Determine the correct quaternary state of the template. Asymmetric units of PDB files often do not correspond to the correct biological
assembly of a protein. Assembled coordinate files of the most likely biological assembly of the template can be retrieved from PQS77,PISA
78 or
the PDB.
!
CAUTION Homology between two proteins is not necessarily sufficient to signify that they share the same quaternary structure.
(ii) Download and save the oligomer template coordinates as PDB file to your local disk.
(iii) Open the file in DeepView and remove all nonamino-acid groups, such as ions, ligands, OXTand so on from the template (unless they are at the
very end of the file). You can do this by selecting the groups in the control panel of DeepView and remove the selected residues (‘Build’ menu).
(iv) Make sure each chain (protomer) hasa unique chain identifier, e.g., ‘A’,‘B’ and so on. Coloring the molecule by chain helps to check. You can
rename chains with DeepView (‘Edit -Rename’ menu).
(v) Create a FASTA file with the target sequences for each chain, i.e., ‘A’, then ‘B’ and so on, separated by semicolons. For hetero-oligomers, make
sure the order is the same as in the template (Fig. 4).
(vi) Adjust target–template alignment in DeepView. Load the FASTA file into DeepView (‘SwissModel’ Menu) and generate a preliminary target–
template alignment (‘Fit -Fit raw sequence’ Menu). Open the alignment window and adjust the alignment. Be sure not to align residues of
different chains or to align amino-acid residues to ligand (HETATM) groups or the C-terminal oxygen group (OXT) in the template. Make sure all
insertions and deletions are correctly positioned in the structural context.
(vii) Save the project to your local disk and submit the file to the Project mode of SWISS-MODEL workspace for model building (Step 6C).
NATURE PROTOCOLS |VOL.4 NO.1 |2009 |7
PROTOCOL
calculated in the course of the modeling
procedure, and the corresponding plots
are also accessible from the results
page. The percentage sequence identity
between target and template on the
basis of the alignment used to build the
model is also reported on the results
page.
?TROUBLESHOOTING
Quality estimation
8| To estimate the quality of your
model(s), submit it to the programs
provided in the ‘Structure Assessment’
section under ‘Tools,’ using the following
options: option A for sequence identity;
option B for stereochemistry check;
option C for global model quality
estimation; and option D for local model
quality estimation. Some of the tools
described below can help identify
incorrect regions in the predicted
protein structure. One possibility to
cope with the uncertainties in
comparative modeling (especially when
the sequence identity between target
and template is low) is to build multiple
models on the basis of alternative
templates and/or alignments (Step 3–6)
and to subsequently select the most
favorable model.
mCRITICAL STEP Protein structure
models generated by comparative mod-
eling may contain errors and thus need to
be treated with caution. Often, the qual-
ity varies between parts of the model.
?TROUBLESHOOTING
(A) Sequence identity
(i) The percentage sequence identity
between target and template is a
good predictor of the accuracy of a model. Model accuracy steadily increases with increasing sequence identity.
(B) Stereochemistry check
(i) The stereochemical plausibility of the model can be analyzed with the tools WHATCHECK34 and PROCHECK33. Deviations
from ideal stereochemical values are reported by these programs.
(C) Global model quality estimation
(i) The DFIRE statistical potential30 as well as the QMEAN composite scoring function65 both return a pseudo-energy for the
entire model.
(D) Local model quality estimation
(i) The following tools are available for analyzing the local (per-residue) model reliability that can help in identifying
potentially incorrect regions in the model: ProQres32—an artificial neural network trained to predict the local model
quality on the basis of the analysis of atom–atom contacts, residue–residue contacts, solvent accessibility surfaces and
secondary structure propensities. The plot shows the local reliability of the model ranging from 0 (unreliable) to 1
(reliable) for each residue in the sequence; ANOLEA28 is a statistical potential that can be used to analyze the packing
quality of the model on the basis of nonlocal atomic interactions. The plot shows the ANOLEA pseudo-energy for each
amino acid in the sequence. Negative values (colored in green) indicate that the amino acid is in a favorable environment,
whereas positive values (colored in red) suggest that this part of the model has been incorrectly built; the GROMOS66
p
uor
G
gn
i
h
s
i
lb
uP eru
t
a
N
800
2
©natureprotocol
s
/moc.erut
a
n.w
ww//:ptth
Figure 5
|
Typical view of a SWISS-MODEL workspace result. In this example, a model for methylpurine-
DNA glycosylase has been generated in automated mode. Upper panel: the green line represents the target
sequence (237 residues). Blue lines indicate for which segments of the target models have been
generated, in this case for residues 12–223. Middle panel: information on the template structure (2b6c,
chain A) and quality (sequence identity, E-value) of the target–template sequence alignment shown in the
lower panel. Model coordinates can be displayed within the Web browser window by clicking on the
preview image or downloaded for manipulation with external software.
8|VOL.4 NO.1 |2009 |NATURE PROTOCOLS
PROTOCOL
empirical force field is used to calculate the energy of each residue in the model. The graphical representation shows posi-
tion in the sequence against the empirical force field energy. Negative values (colored in green) represent energetically
favorable conformations, whereas positive values (colored in red) indicate unfavorable conformations.
?TROUBLESHOOTING
Troubleshooting advice can be found in Table 1.
p
uor
G
gn
i
h
s
i
lb
uP eru
t
a
N
800
2
©natureprotocol
s
/moc.erut
a
n.w
ww//:ptth
TABLE 1
|
Troubleshooting table.
Step Problem Solution
2B and C Secondary structure prediction predicts a
strand helix and the disorder prediction predicts
thesameregiontobedisordered
Examples of disordered regions undergoing the transition to an ordered
state upon binding their ligand proteins have been reported59.Incasethe
predicted regionaligns to a known template structure, check if the template
has been solved in complex with a binding partner or is otherwise known to
undergo structural rearrangement
3A BLAST reports too many matches Change the E-value cutoff for reporting hits to force BLAST to report only
hits with a low E-value
BLAST does not report any results When no suitable templates are identified, or only parts of the target
sequence are covered, two approaches for more sensitive detection of
distant relationships among protein families are provided (3B and 3C)
3A and B How will the choice of the substitution matrix
influence the output of Blast/Profile Blast?
Use a substitution matrix adapted to the expected divergence of the
searched sequences. For the BLOSUM family of matrices, the higher the
matrix index is (i.e., BLOSUM 80), the more selective your search will be: it
will exclude false positives but possibly miss true positives (closest to
PAM120). Vice versa, the lower the index is (i.e., BLOSUM 45), the more
sensitive the search will be: more true matches will be identified, but
eventually, more false positives will be included (closest to PAM250)
3B Profile Blast report too many matches Change the E-value cutoff for reporting hits, or in the template library
search step choose a library where the sequences of the templates are
clustered at a lower sequence identity, e.g., ExPDB 70
4 The template identification methods cover only
part of the sequence of my protein
The sensitivity of profile-based template detection methods (Step 3B and C)
can be increased when the search is performed at the domain level rather
than using the whole target sequence
The template identification method predicts two
templates with different structures
Similar structures in different conformation: protein structures can undergo
large conformational changes, e.g., upon binding of ligand or post-
translational modification. From the list of possible template structures,
select the one most suitable for your application on the basis of the
annotation provided, e.g., presence of ligands or cofactors and so on.
Template structures with different folds: ambiguous results in fold assign-
ment are expected if the evolutionary relationship between the target and
possible template structures is too weak to be reliably detected by
sequence-based methods, or no related template structure has been solved.
Template structures with unclear evolutionary relationship to the target
should not be used for homology modeling, unless supported by additional
(experimental) evidence
No templates are found for the protein or
domain of interest
In this case, it is not possible to produce a reliable three-dimensional model
for the protein by homology modeling. Alternatively, one can attempt to
apply de novo prediction methods. However, the results of these types of
prediction are currently far less accurate than comparative techniques and
often not sufficient for specific biological applications10
Can different templates, covering different
regions of my protein, be combined to obtain
a comparative model for the full length of
the protein?
Nonoverlapping templates cannot be combined, as the relative orientation
of the different structures is unknown. It is, however, in principle possible,
if the different templates are significantly overlapping (e.g., more than
20–40 amino acids). This feature is currently not supported by
SWISS-MODEL workspace; users are referred to other modeling programs
supporting this option, such as Modeller6
(continued)
NATURE PROTOCOLS |VOL.4 NO.1 |2009 |9
PROTOCOL
ANTICIPATED RESULTS
As an example, we apply the protocol described here to model the bacterial methylpurine-DNA glycosylase (AlkD, Uniprot AC:
Q2PAD8) and the Drosophila putative protein kinase C delta type (Pkcdelta, UniProt AC: P83099). Please note that the
results presented here illustrate a representative example at the time of writing. As sequence and structure databases are
continuously updated, new template structures may become available at a later point and may lead to different, in general,
better, modeling results.
AlkD is a DNA gylcosylase69 that functions as a DNA alkylation repair enzyme70. AlkD belongs to a newly characterized DNA
glycosylase superfamily71, for which no structures have yet been solved experimentally. Domain annotation (Step 2A) indicates
that the protein belongs to a multihelical fold called the armadillo-like fold72. This is in agreement with the results of the
secondary structure prediction analysis (Step 2B), predicting almost exclusively alpha-helices. A search for suitable structure
templates (Step 3) yields a highly significant match spanning almost the entire length of the target protein, a putative DNA
alkylation repair enzyme from Enterococcus faecalis (PDB: 2B6C) solved by the Midwest Center for Structural Genomics.
This template was used by Dalhus et al.73 to built a comparative model to elucidate the mechanism of AlkD.
The BLAST alignment between the target and the template displays only a single gap and a sufficiently high level of sequence
similarity for it to be modeled using the automated mode of SWISS-MODEL workspace (Step 6A, Fig. 5). The location of the
single gap in the target and sequence alignment can further be investigated in the structural context with the help of DeepView.
The project file resulting from the automated step containing the template and the modeled protein can be opened in DeepView
and the target–template alignment can be visualized with the help of the Alignment Window of DeepView. Secondary structure
elements of the template can be highlighted in different colors using the color menu of the software. In the alignment resulting
p
uor
G
gn
i
h
s
i
lb
uP eru
t
a
N
800
2
©natureprotocol
s
/moc.erut
a
n.w
ww//:ptth
TABLE 1
|
Troubleshooting table (continued).
Step Problem Solution
5A and 6C Where can I find information on how
to use the program DeepView?
Manuals for DeepView can be found on the program website: http://
www.expasy.org/spdbv/. It is highly recommended to start by following
the tutorial provided by Gale Rhodes (University of Maine): http://
www.usm.maine.edu/Brhodes/SPVTut/
6A and 6B The template of interest is not present
in the SMTL library
The SMTL library is updated biweekly, i.e., it might take a few days until
newly released PDB structures are included. You can check if a specific PDB
entry is available by querying the SMTL library in the ‘Tools’ section of the
server. Alternatively, you can use DeepView to create a model project file on
the basis of any template structure independently if this structure is part of
PDB or SMTL
6B Multiple sequence alignment is not correctly
recognized by the server
Please make sure the alignment is in one of the supported formats. Use short
unique names for sequences in the alignment; avoid nonalphanumerical
characters. Good examples: ‘THN_DENCL’,‘P01542’,‘1crnA’ and so on
7 Unable to obtain a model from the server In the majority of cases this is due to a poor alignment between the target
and template sequences. Take time to carefully edit the alignment as
suggested in Step 6A and B
Despite careful checking of the target–template
alignment, the server does not successfully
deliver a model
This is usually the case if the target–template alignment contains large gaps
(insertions and deletions) that were not successfully reconstructed. If this
is the case, consider using other modeling programs, e.g., Modeller6.
However, keep in mind that the results of de novo loop modeling techniques
are less reliable than the template-based part of the model when using the
model to answer the biological questions of interest
8 How to proceed when incorrect regions are
identified and how to interpret them
There are several possible explanations for regions predicted as potentially
incorrect (i.e., having low ProQres scores and/or high ANOLEA energies),
e.g., alignment errors, incorrect modeling of insertions, unfavorable side-
chain packing or false-positive assignment by the program itself. The
identified regions of low reliability should be further analyzed by visually
inspecting the alignment and the model. A model should always be
interpreted with respect to its future application: a model with local errors
outside the region of interest, such as the active site, can nevertheless be
valuable for certain experiments. On the contrary, if, e.g., surface loops
contain residues known to be involved in function, one needs to proceed
with great caution when using the model for refining functional hypothesis
10 |VOL.4 NO.1 |2009 |NATURE PROTOCOLS
PROTOCOL
from the automated mode, there is an inserted amino acid in
the target sequence in a position that corresponds to an inter-
nal alpha helix residue of the template. We assume that the
position of the gap could be improved by shifting it to the
loop region connecting two adjacent alpha helices. The
alignment can be edited directly within the alignment window
of DeepView (Step 5A). The resulting modified project file is
then saved locally and submitted to the Project Mode of the
server for model building (Step 6C).
Dalhus et al.73 predicted the location of a putative binding
pocket in the model by residues conservation analysis of homo-
logous proteins. The binding site of the obtained model was
then used to design site-specific mutations to characterize the
role of specific residues in the catalysis of DNA repair. Further
insights into the mechanism of activity of the enzyme were
gained by combining the obtained model with DNA coordinates
extracted from the homologous protein AlkA.
In the second example, we build a model for a putative pro-
tein kinase C delta from Drosophila (Pkcdelta d, UniProt AC:
P83099). Domain annotation (Step 2A, Fig. 6) confirms that
the target belongs to the protein kinase C (PKC) family74.
The PKC family members can be grouped into three classes:
(i) the conventional PKCs (a,gand b1andb2), which requires
diacylgycerol (DAG), phosphatydilserine (PS) and calcium for
activation; (ii) the novel PKCs (d,e,Z/l,y), which are activated by DAG and PS but are insensitive to calcium; and (iii) the
atypical PKCs (zand i/l), which require only PS for full activity75.
Several statistically significant matches for suitable templates are reported for the protein kinase domain (Step 4). Each
template in the list is linked to the SMTL, and from there to other external resources, to allow for verification and a plausibility
check. We have selected template 2JED (chain A) to build the model for our protein. Template 2JED corresponds to the crystal
structure of the human kinase domain of
the PKC y, which belongs to the same
class as our putative PKC d.Asinthe
previous example, the target–template
alignment (derived from the iterative
profile Blast search) is inspected with
the help of DeepView, e.g., to verify
that the residues corresponding to the
typical signatures for the serine/
threonine protein kinase active site
(Prosite Accession number PS00108)
and for the ATP-binding motif (Prosite
Accession number PS00107) are
correctly mapped in the target and
template alignment. This can be done
easily with the ‘search for PROSITE
pattern’ function in the ‘Edit’ menu of
DeepView. Subsequently, the target–
template alignment is saved as project
file and submitted to the Project Mode
to obtain a model for the protein kinase
domain of the protein (Fig. 7).
The two PE/DAG (Phorbol esters/dia-
cylglycerol)-binding domains of the PKC
dprotein can also be modeled on the
basis of templates identified by iterative
profile Blast and HMM-HMM search
p
uor
G
gn
i
h
s
i
lb
uP eru
t
a
N
800
2
©natureprotocol
s
/moc.erut
a
n.w
ww//:ptth
Figure 7
|
Model of the kinase domain of the putative PKC dfrom Drosophila shown as ribbon
representation in DeepView colored from blue (N terminus) to red (C terminus). Characteristic residues of
the Ser-Thr kinase active site and ATP-binding motif (identified by PROSITE) are shown as sticks. The
position of the inhibitor molecule Nvp-Xaa228 in the template structure is highlighted in green.
a
b
Figure 6
|
Target sequence annotation for the putative protein kinase C delta
from Drosophila (PKC d, UniProt AC: P83099). (a) Three functional domains are
indentified using InterPro scan: two PE/DAG (Phorbol esters/Diacylglycerol)-
binding domains and a PKC domain. (b) Secondary structure prediction for the
N-terminal 100 residues of the target sequence.
NATURE PROTOCOLS |VOL.4 NO.1 |2009 |11
PROTOCOL
methods. Few templates were detected by both methods and we decided to use the structure (PDB: 1PTQ76) of the second
activator-binding domain (PE/DAG) of an orthologous protein (the mouse PKC d, UniProt AC: P28867) to build the models for
the two PE/DAG domains. The alignment between the two PE/DAG domains of the PKC dand the template is largely unambigu-
ous, and the resulting model is of good quality according to the standard structure assessment tools (Step 8). Particular atten-
tion should be paid to the characteristic histidine and cysteine residues, which are assumed to be involved in the coordination
of zinc ions. Correctly mapping these residues in the target–template sequence alignment ensures that they have a chemically
plausible three-dimensional arrangement in the homology model.
Additional examples can be found in ref. 40 and in the tutorial provided on the SWISS-MODEL workspace website
http://swissmodel.expasy.org/workspace/tutorial/.
ACKNOWLEDGMENTS We are grateful to Dr Michael Podvinec for his enthusiastic
support and excellent coordination of the Scrum process for the SWISS-MODEL
team. We are thankful for financial support of our group by the Swiss Institute of
Bioinformatics (SIB).
Published online at http://www.natureprotocols.com/
Reprints and permissions information is available online at http://npg.nature.com/
reprintsandpermissions/
1. Berman, H.,Henrick, K., Nakamura, H.& Markley, J.L. The worldwideProtein Data
Bank (wwPDB): ensuring a single, uniform archive of PDB data. Nucleic Acids. Res.
35, D301–D303 (2007).
2. Wu, C.H. et al. The Universal Protein Resource (UniProt): an expanding universe
of protein information. Nucleic Acids. Res. 34, D187–D191 (2006).
3. Chothia, C. Proteins. One thousand families for the molecular biologist. Nature
357, 543–544 (1992).
4. Chothia, C. & Lesk, A.M. The relation between the divergence of sequence and
structure in proteins. EMBO J. 5, 823–826 (1986).
5. Topham, C.M. et al. An assessment of COMPOSER: a rule-based approach to
modelling protein structure. Biochem. Soc. Symp. 57, 1–9 (1990).
6. Sali, A. & Blundell, T.L. Comparative protein modelling by satisfaction of spatial
restraints. J. Mol. Biol. 234, 779–815 (1993).
7. Peitsch, M.C. Protein modelling by e-mail. BioTechnology 13, 658–660
(1995).
8. Tramontano, A. & Morea, V. Assessment of homology-based predictions in CASP5.
Proteins 53 (Suppl. 6): 352–368 (2003).
9. Tress, M., Ezkurdia, I., Grana, O., Lopez, G. & Valencia, A. Assessment of
predictions submitted for the CASP6 comparative modeling category. Proteins 61
(Suppl. 7): 27–45 (2005).
10. Jauch, R., Yeo, H.C., Kolatkar, P.R. & Clarke, N.D. Assessment of CASP7
structure predictions for template free targets. Proteins 69 (Suppl. 8): 57–67
(2007).
11. Kopp, J., Bordoli, L., Battey, J.N., Kiefer, F. & Schwede, T. Assessment of CASP7
predictions for template-based modeling targets. Proteins 69 (Suppl. 8): 38–56
(2007).
12. Kryshtafovych, A., Fidelis, K. & Moult, J. Progress from CASP6 to CASP7. Proteins
69 (Suppl. 8): 194–207 (2007).
13. Hillisch, A., Pineda, L.F. & Hilgenfeld, R. Utility of homology models in the drug
discovery process. Drug Discov. Today 9, 659–669 (2004).
14. Kopp, J. & Schwede, T. Automated protein structure homology modeling: a
progress report. Pharmacogenomics 5, 405–416 (2004).
15. Marti-Renom, M.A. et al. Comparative protein structure modeling of genes and
genomes. Annu. Rev. Biophys. Biomol. Struct. 29, 291–325 (2000).
16. Peitsch, M.C. About the use of protein models. Bioinformatics 18, 934–938
(2002).
17. Tramontano, A. In Computational Structural Biology (eds. Schwede T. & Peitsch
M.C.) (World Scientific Publishing, Singapore, 2008).
18. Baker, D. & Sali, A. Protein structure prediction and structural genomics. Science
294, 93–96 (2001).
19. Soto, C.S., Fasnacht, M., Zhu, J., Forrest,L. & Honig, B. Loop modeling: sampling,
filtering, and scoring. Proteins 70, 834–843 (2008).
20. Rohl, C.A., Strauss, C.E., Chivian, D. & Baker, D. Modeling structurally variable
regions in homologous proteins with rosetta. Proteins 55, 656–677 (2004).
21. Fiser, A., Do, R.K. & Sali, A. Modeling of loops in protein structures. Protein Sci. 9,
1753–1773 (2000).
22. Canutescu, A.A., Shelenkov, A.A. & Dunbrack, R.L. Jr. A graph-theory algorithm
for rapid protein side-chain prediction. Protein Sci. 12, 2001–2014 (2003).
23. Rost, B. Twilight zone of protein sequence alignments. Protein Eng. 12, 85–94
(1999).
24. Dunbrack, R.L. Jr. Sequence comparison and protein structure prediction. Curr.
Opin. Struct. Biol. 16, 374–384 (2006).
25. Sommer, I., Toppo, S., Sander, O., Lengauer, T. & Tosatto, S.C. Improving the
quality of protein structure models by selecting from alignment alternatives.
BMC Bioinformatics 7, 364 (2006).
26. Tress, M.L., Jones, D. & Valencia, A. Predicting reliable regions in protein
alignments from sequence profiles. J. Mol. Biol. 330, 705–718 (2003).
27. Vingron, M. Near-optimal sequence alignment. Curr. Opin. Struct. Biol. 6, 346–352
(1996).
28. Melo, F. & Feytmans, E. Assessing protein structures with a non-local atomic
interaction energy. J. Mol. Biol. 277, 1141–1152 (1998).
29. Sippl, M.J. Calculation of conformational ensembles from potentials of mean
force. An approach to the knowledge-based prediction of local structures in
globular proteins. J. Mol. Biol. 213, 859–883 (1990).
30. Zhou, H. & Zhou, Y. Distance-scaled, finite ideal-gas reference state improves
structure-derived potentials of mean force for structure selection and stability
prediction. Protein Sci. 11, 2714–2726 (2002).
31. Fasnacht, M., Zhu, J. & Honig, B. Local quality assessment in homology
models using statistical potentials and support vector machines. Protein Sci. 16,
1557–1568 (2007).
32. Wallner, B. & Elofsson, A.Identification ofcorrect regions in protein models using
structural, alignment, and consensus information. Protein Sci. 15, 900–913
(2006).
33. Laskowski, R.A., MacArthur, M.W., Moss, D.S. & Thornton, J.M. PROCHECK: a
program to check the stereochemical quality of protein structures. J. Appl. Cryst.
26, 283–291 (1993).
34. Hooft, R.W.,Vriend, G., Sander, C. & Abola, E.E. Errors in protein structures. Nature
381, 272 (1996).
35. Aloy, P., Pichaud, M. & Russell, R.B. Protein complexes: structure prediction
challenges for the 21st century. Curr. Opin. Struct. Biol. 15, 15–22 (2005).
36. Alber, F. et al. Determining the architectures of macromolecular assemblies.
Nature 450, 683–694 (2007).
37. Junne, T., Schwede, T., Goder, V. & Spiess, M. The plug domain of yeast Sec61p is
important for efficient protein translocation, but is not essential for cell viability.
Mol. Biol. Cell 17, 4063–4068 (2006).
38. Battey, J.N. et al. Automated server predictions in CASP7. Proteins 69 (Suppl. 8):
68–82 (2007).
39. Koh, I.Y. et al. EVA: evaluation of protein structure prediction servers. Nucleic
Acids. Res. 31, 3311–3315 (2003).
40. Arnold, K., Bordoli, L., Kopp, J. & Schwede, T. The SWISS-MODEL workspace: a
web-based environment for protein structure homology modelling.Bioinformatics
22, 195–201 (2006).
41. Eswar, N. et al. Tools for comparative protein structure modeling and analysis.
Nucleic Acids. Res. 31, 3375–3380 (2003).
42. Bates, P.A., Kelley, L.A., MacCallum, R.M. & Sternberg, M.J. Enhancement of
protein modeling by human intervention in applying the automatic programs
3D-JIGSAW and 3D-PSSM. Proteins (Suppl. 5): 39–46 (2001).
43. Fernandez-Fuentes, N., Madrid-Aliste, C.J., Rai, B.K., Fajardo, J.E. & Fiser, A. M4T:
a comparative protein structure modeling server. Nucleic Acids Res. 35,
W363–W368 (2007).
44. Fox, J.A., McMillan, S. & Ouellette, B.F. Conducting research on the web: 2007
update for the bioinformatics links directory. Nucleic Acids Res. 35,W3–W5
(2007).
45. Schwede, T., Diemand,A., Guex, N. & Peitsch, M.C. Protein structure computing in
the genomic era. Res. Microbiol. 151, 107–112 (2000).
46. Kopp, J. & Schwede, T. The SWISS-MODEL repository of annotated three-
dimensional protein structure homology models. Nucleic Acids Res. 32,
D230–D234 (2004).
47. Schwede, T., Kopp, J., Guex, N. & Peitsch, M.C. SWISS-MODEL: an automated
protein homology-modeling server. Nucleic Acids Res. 31, 3381–3385 (2003).
48. Guex, N. & Peitsch, M.C. SWISS-MODEL and the Swiss-PdbViewer: an environment
for comparative protein modeling. Electrophoresis 18, 2714–2723 (1997).
p
uor
G
gn
i
h
s
i
lb
uP eru
t
a
N
800
2
©natureprotocol
s
/moc.erut
a
n.w
ww//:ptth
12 |VOL.4 NO.1 |2009 |NATURE PROTOCOLS
PROTOCOL
49. Andreeva, A. et al. SCOP database in 2004: refinements integrate structure and
sequence family data. Nucleic Acids Res. 32, D226–D229 (2004).
50. Greene, L.H. et al. The CATH domain structure database: new protocols and
classification levels give a more comprehensive resource for exploring evolution.
Nucleic Acids Res. 35, D291–D297 (2007).
51. Finn, R.D. et al. The Pfam protein families database. Nucleic Acids Res. 36,
D281–D288 (2008).
52. Zdobnov, E.M. & Apweiler, R. InterProScan—an integration platform for the
signature-recognition methods in InterPro. Bioinformatics 17, 847–848
(2001).
53. Mulder, N.J. et al. New developments in the InterPro database. Nucleic Acids Res.
35, D224–228 (2007).
54. Jones, D.T. Protein secondary structure prediction based on position-specific
scoring matrices. J. Mol. Biol. 292, 195–202 (1999).
55. Jones, D.T. & Ward, J.J. Prediction ofdisordered regions in proteins from position
specific score matrices. Proteins 53 (Suppl. 6): 573–578 (2003).
56. Jones, D.T., Taylor, W.R. & Thornton, J.M. A model recognition approach to the
prediction of all-helical membrane protein structure and topology. Biochemistry
33, 3038–3049 (1994).
57. Fink, A.L. Natively unfolded proteins. Curr. Opin. Struct. Biol. 15, 35–41 (2005).
58. Radivojac, P. et al. Intrinsic disorder and functional proteomics. Biophys. J. 92,
1439–1456 (2007).
59. Dyson, H.J. & Wright, P.E. Intrinsically unstructured proteins and their functions.
Nat. Rev. Mol. Cell Biol. 6, 197–208 (2005).
60. Altschul, S.F. et al. Gapped BLAST and PSI-BLAST: a new generation of protein
database search programs. Nucleic Acids Res. 25, 3389–3402 (1997).
61. Wheeler, D.L. et al. Database resources of the National Center for Biotechnology
Information. Nucleic Acids Res. 33 Database Issue: D39–D45 (2005).
62. Soding, J. Protein homology detection by HMM-HMM comparison. Bioinformatics
21, 951–960 (2005).
63. Muller, C.W., Schlauderer, G.J., Reinstein, J. & Schulz, G.E. Adenylate kinase
motions during catalysis: an energetic counterweight balancing substrate
binding. Structure 4, 147–156 (1996).
64. So¨ding, J., Biegert, A. & Lupas, A.N. The HHpred interactive server for protein
homology detection and structure prediction. Nucleic Acids Res.33,W244–248
(2005).
65. Benkert, P., Tosatto, S.C. & Schomburg, D. QMEAN: a comprehensive scoring
function for model quality assessment. Proteins 71, 261–277 (2008).
66. van Gunsteren, W.F. et al. Biomolecular Simulations: the GROMOS96 Manual and
User Guide (VdF Hochschulverlag ETHZ, Zu¨rich, 1996).
67. Bateman, A. et al. The Pfam protein families database. Nucleic Acids Res. 32,
D138–D141 (2004).
68. Hulo, N. et al. The PROSITE database. Nucleic Acids Res. 34, D227–D230 (2006).
69. Stivers, J.T. & Jiang, Y.L. A mechanistic perspective on the chemistry of DNA
repair glycosylases. Chem. Rev. 103, 2729–2759 (2003).
70. Seeberg, E., Eide, L. & Bjoras, M. The base excision repair pathway. Trends
Biochem. Sci. 20, 391–397 (1995).
71. Alseth, I. et al. A new protein superfamily includes two novel 3-methyladenine
DNA glycosylases from Bacillus cereus, AlkC and AlkD. Mol. Microbiol. 59,
1602–1609 (2006).
72. Groves, M.R. & Barford, D. Topological characteristics of helical repeat proteins.
Curr. Opin. Struct. Biol. 9, 383–389 (1999).
73. Dalhus, B. et al. Structural insight into repair of alkylated DNA by a new
superfamily of DNA glycosylases comprising HEAT-like repeats. Nucleic Acids Res.
35, 2451–2459 (2007).
74. Nishizuka, Y. Membrane phospholipid degradation and protein kinase C for cell
signalling. Neurosci. Res. 15, 3–5 (1992).
75. Mellor, H. & Parker, P.J. The extended protein kinase C superfamily. Biochem. J.
332 (Part 2): 281–292 (1998).
76. Zhang, G., Kazanietz, M.G., Blumberg, P.M. & Hurley, J.H. Crystal structure of the
cys2 activator-binding domain of protein kinase C delta in complex with phorbol
ester. Cell 81, 917–924 (1995).
77. Henrick, K. & Thornton, J.M.PQS: a protein quaternarystructure file server. Trends
Biochem. Sci. 23, 358–361 (1998).
78. Krissinel, E. & Henrick, K. Inference of macromolecular assemblies from
crystalline state. J. Mol. Biol. 372, 774–797 (2007).
p
uor
G
gn
i
h
s
i
lb
uP eru
t
a
N
800
2
©natureprotocol
s
/moc.erut
a
n.w
ww//:ptth
NATURE PROTOCOLS |VOL.4 NO.1 |2009 |13
PROTOCOL
