Comparative modeling for protein structure prediction
Krzysztof Ginalski
With the progression of structural genomics projects, served regions; predicting structurally variable regions,
comparative modeling remains an increasingly important including insertions and missing N and C termini; mod-
method of choice. It helps to bridge the gap between the eling sidechains; and refining and evaluating the resulting
available sequence and structure information by providing model. Although each step can introduce errors that affect
reliable and accurate protein models. Comparative modeling the modeled structure, optimal use of structural informa-
based on more than 30% sequence identity is now
tion from available templates and correctness of
approaching its natural template-based limits and further
sequence-to-structure alignment are the most significant
improvements require the development of effective
determinants of final model quality.
refinement techniques capable of driving models toward
native structure. For difficult targets, for which the most
Traditionally, comparative modeling refers to cases in
significant progress in recent years has been observed,
which related proteins of known structure can be found
optimal template selection and alignment accuracy are
with PSI-BLAST [5]. The recent introduction of more
still the major problems.
sophisticated methods (reviewed in [6 ]) that derive
their power from profile-profile comparison [7 ,8,9] and
Addresses the effective use of structural information [10,11] has
Centre for Mathematical and Computational Modelling, Warsaw
significantly increased not only the resulting alignment
University, Pawińskiego 5a, 02-106 Warsaw, Poland
quality but also the remote homologue detection cap-
ability. Consequently, the boundary in template-based
Corresponding author: Ginalski, Krzysztof ([e-mail])
modeling between comparative modeling and fold recog-
nition is now quite blurred. Increased interest in devel-
Current Opinion in Structural Biology 2006, 16:172 177
oping new comparative modeling and fold recognition
algorithms has led to a variety of prediction services
This review comes from a themed issue on
available on the Internet [6 ], including structure pre-
Theory and simulation
Edited by Joel Janin and Michael Levitt
diction meta-servers [12]. The latter are of enormous
importance to biologists and modelers because they pro-
Available online 28th February 2006
vide convenient access to the results of various indepen-
0959-440X/$  see front matter dent two- and three-dimensional structure prediction
# 2005 Elsevier Ltd. All rights reserved.
methods. Also, they are frequently used as starting points
in sequence analysis and three-dimensional model build-
DOI 10.1016/j.sbi.2006.02.003
ing. This review summarizes recent progress, and dis-
cusses the current roles, limitations and challenges of
comparative modeling.
Introduction
Knowledge of three-dimensional protein structure is cru- Objective evaluation of methods  current
cial to answering many biological questions; however, the state of the art in comparative modeling
rapidly growing number of sequenced genes and gen- The launch of the biannual CASP (Critical Assessment of
omes is heavily outpacing the number of experimentally Techniques for Protein Structure Prediction) experiment
determined structures. Despite considerable progress in [13,14 ,15 ], established to detect the capabilities and
de novo structure prediction [1 ], comparative modeling limitations of current modeling methods, to determine
methods, when applicable, provide the most reliable and the progress made and to highlight specific bottlenecks,
accurate protein structure models [2]. Comparative mod- represented a crucial milestone in the protein structure
eling is based on the general observation that evolutio- prediction field. Results from the latest CASP experi-
narily related sequences have similar three-dimensional ments show that, in the comparative modeling category
structures [3]. As a consequence, a three-dimensional [16 ], the most successful approaches use consensus
model of a protein of interest (target) can be built from strategies [17,18 ] to build final models based on multi-
related protein(s) of known structure [template(s)] that ple templates or protein fragment recombination [19].
share statistically significant sequence similarity. The Consensus results from various fold recognition methods
traditional comparative modeling procedure consists of or multiple sequence searches [20] are frequently used for
several consecutive steps usually repeated iteratively template selection and detection of reliable alignment
until a satisfactory model is obtained [4]: finding suitable regions, whereas alternative alignment variants are eval-
template protein(s) related to the target; aligning target uated at the tertiary structure level using quality assess-
and template(s) sequences; identifying structurally con- ment methods [21,22] and/or visual inspection. Detailed
Current Opinion in Structural Biology 2006, 16:172 177 www.sciencedirect.com
Comparative modeling for protein structure prediction Ginalski 173
Figure 1
sequence analysis of the target and template families,
investigation of characteristic features of the fold and
extensive literature searches for any available biochem-
ical information (mutations, catalytic residues, etc.) are
usually mandatory, as even tiny details can serve as
alignment anchors and lead to the successful identifica-
tion of the correct sequence-structure mapping in ques-
tionable regions. Division of the target sequence into
single domains, removal of long insertions to the core
of the fold and iterative submission to prediction servers
are also strongly advised. Finally, model building for close
homologues of the target may enable the detection of
This example of difficult comparative modeling based on a distantly
significant alignment errors, which manifest themselves
related template illustrates the important role of human input in cases
in three-dimensional models only for some family mem-
of unexpected evolutionary changes in protein structure. (a)
bers [17].
Experimental structure of CASP6 target T0223, a putative
nitroreductase from Thermotoga maritima (PDB code 1vkw, green),
Modeling based on multiple templates is often advanta- and the best model (T0223TS450_1, blue). (b) The available template,
flavin reductase P from Vibrio harveyi (PDB code 1bkj, monomers in
geous, not least because it increases the chance that the
grey and orange), shares 18% sequence identity with the target. T0223
optimal template is among those used [18 ]. However, it
is a monomeric pseudo-dimer containing two duplicated reductase
is not easy to benefit from the large number of available
domains arranged exactly as within the dimeric template. Correct
templates, especially when their local structures differ
modeling of the complete protein chain required the use of a dimeric
template instead of a monomer.
significantly. Although existing methods can provide
reasonably accurate predictions for short loops, the mod-
eling of longer regions not present in available templates
remains a challenge and is frequently performed using and, more importantly, is not a blind prediction test. As
de novo methods [23 ], with anecdotal examples of rela- clearly demonstrated by the CAFASP and LiveBench
tive success [14 ,16 ]. Importantly, the quality of a experiments, significant progress in template-based auto-
modeled structurally variable region is greatly affected mated protein structure prediction has been achieved
by its length, correctness of the alignment and accuracy through the development of meta-servers [24,25 ],
of predicted neighboring regions [23 ]. Our ability to which detect common structural motifs (consensus) in
correctly predict sidechain conformations, which are the set of three-dimensional models generated by various
backbone conformation dependent, is, not surprisingly, independent structure prediction services. Meta-servers
rather limited [16 ]. Incorrect sidechain rotamers are either generate a new overall ranking and select the
mainly caused by misaligned residues and/or backbone potentially best model [28] or perform additional mod-
shifts, which must be either accurately modeled initially ifications (e.g. construct a hybrid from fragments of the
or refined simultaneously to improve sidechain predic- original models) [29]. A well-designed meta-predictor
tions. should perform at least as well as the best of its input
components; meta-servers do outperform individual ser-
Human expertise appears to be very valuable for model- vers and are already challenging most human expert
ing difficult targets (template detected by PSI-BLAST) predictions [25 ]. Nevertheless, the performance of sev-
and critical in cases of unexpected evolutionary changes eral newly developed autonomous servers appears to be
in protein structure (Figure 1). In contrast, for easy amongst the best in comparative modeling, suggesting
comparative modeling targets (related structure detected that further improvements of the individual methods
by simple BLAST), human improvements are often have recently been obtained [16 ,25 ]. These new
marginal, if not detrimental, as the performance of auto- autonomous methods base their strength on the compar-
matic methods on these targets has increased substan- ison of sequence profiles combined with predicted sec-
tially [14 ,16 ]. ondary structure [30] or, in addition, structure-based
profile energy scoring [31].
Evaluation of automatic structure prediction methods is
conducted by the CAFASP (Critical Assessment of Fully Quality and usefulness of comparative
Automated Structure Prediction) experiment [24], which models
runs in parallel with CASP on the same target set. A more Comparative models may be used to identify critical
continuous assessment of servers is provided by Live- residues involved in catalysis (or migration of catalytic
Bench [25 ] and EVA [26], which operate on a relatively residues), binding or structural stability, to examine pro-
large number of prediction targets compiled every week tein protein or protein ligand interactions (including
from newly released PDB [27] structures. However, drug design), to correlate genotypic and phenotypic
LiveBench excludes easy comparative modeling targets mutation data, and to guide experimental design. The
www.sciencedirect.com Current Opinion in Structural Biology 2006, 16:172 177
174 Theory and simulation
usefulness of comparative models for a specific applica- value) with respect to the template structure [34 ,35 ].
tion depends on their quality, which tends to decrease as Although the average accuracy of structure-derived prop-
evolutionary distance between target and template erties (residue exposure state, residue neighborhood,
increases [4,14 ,16 ]. accessible surface area, electrostatic potential, etc.)
decreases with lower target-template sequence identity,
Two important factors influence the ability to predict in general their added value (especially for sequence-
accurate models: the extent of structural conservation dependent properties) increases, making models rela-
between target and template, and the correctness of tively more informative in spite of their lower accuracy
alignment [4,14 ]. Models based on templates with more [34 ]. In addition, depending on the property, the accu-
than 50% sequence identity are generally very accurate racy of comparative models based on templates with 25
and can exhibit 1Å Ca atom rmsd from the experi- 40% sequence identity reaches the same value as differ-
mental structure. Proteins with 30 50% sequence iden- ences observed between NMR and X-ray structures
tity share at least 80% of their structures; the best CASP [35 ].
models within this range usually do not exceed 4 Å rmsd
(typically 2 3 Å) from the native structure, with errors Using comparative models instead of template structures
located mainly in loop regions. Structural conservation has also been shown to be invaluable in molecular repla-
can be as low as 55% for proteins that display 20 30% cement, whereby screening with a model or a diverse set
sequence identity or even lower when sequence identity of models can frequently be successful in cases in which
drops below 20%. Whereas alignments are most often the structural template used to build them failed [36 ].
near optimal for targets with more than 30% sequence Although target-template sequence identity is not a good
identity to template structures (easy targets), below this diagnostic for the success of this procedure, models based
threshold (mainly difficult targets), alignment quality on 30% sequence identity (and even sometimes less)
sharply decreases and even as many as half of all residues seem to be sufficiently accurate for molecular replace-
may be misaligned when sequence identity is less than ment [36 ].
20% [14 ].
Limitations and current challenges
Sequence identity between target and template is not, Despite steady but modest progress in difficult compara-
however, an effective parameter to estimate the difficulty tive modeling based on a distant evolutionary relationship
or the quality of a comparative model [14 ,16 ]. A much (template detected by PSI-BLAST), there is still room for
better estimator of the expected model quality is the further improvement in both optimal template selection
distribution of sequence identity in multiple sequence and the quality of sequence-to-structure alignments that
alignment, encompassing target, template and intermedi- are particularly error prone in cases of low sequence
ate homologous sequences [32 ]. Given the unprece- similarity [14 ,15 ]. Least progress has been made in
dented growth of both structural and sequence comparative modeling from relatively high sequence
databases, improvements in the quality of comparative identity templates, as measured in the last CASP experi-
models seem to be largely due to the increased avail- ment [14 ,16 ]. In general, predictions are not closer to
ability of sequences and structures homologous to the the experimental structure than the structure of the
protein of interest [32 ]. closest template [16 ]; however, for easier targets the
best models are now frequently as good as optimal multi-
Importantly, functional regions are not modeled with any templates, suggesting that the template-based methodol-
greater accuracy than the rest of the protein [16 ], unless ogy is approaching its natural limits for easy comparative
they are structurally better conserved than other regions, modeling [37 ]. Further progress in this area requires the
as is typically the case when the template structure has development of appropriate refinement techniques and
the same function and specificity. However, when func- potentials that are capable of making adjustments on an
tion or specificity differs, larger changes are usually atomic scale [14  16 ]. Typically, attempting the refine-
expected in functional regions [15 ]. Accurate modeling ment of coordinates derived from template structures
of the differences between similar structures (insertions leads to the deterioration of rather than improvement
and backbone shifts) is one of the most biologically in model quality. Nevertheless, there are some encoura-
relevant applications of comparative modeling, because ging signs in the field of final model improvement with
these structural changes usually add novel functions and/ respect to the initial template alignment [38]. High-
or specificities. Nevertheless, in many cases, relatively resolution refinement of comparative models remains a
good insight into the active site architecture and ligand formidable challenge, because of inaccuracies in current
binding can be obtained from comparative models, pro- force-fields and difficulties in sampling huge numbers of
viding there are no alignment errors [33 ]. alternatively packed conformations [39]. As recently
shown, some of these problems can be partially overcome
Recent systematic studies have proved that even simple by combining free energy optimizations with sampling
comparative models carry additional information (added along evolutionarily favored directions defined using
Current Opinion in Structural Biology 2006, 16:172 177 www.sciencedirect.com
Comparative modeling for protein structure prediction Ginalski 175
principal components of the backbone structure variation possibility of such structural changes in evolution has
within a homologous family [40 ]. important implications for protein design, but notably
impedes comparative modeling methods; the ability to
More effective energy-based methods, coupled with detect such cases from sequence is crucial.
relaxation techniques, should improve both best tem-
plate detection (or optimal local template fragment com- Conclusions
bination) and the selection of correct alignments for The number of unique folds in nature is expected to be
regions where the proper evaluation of alignment variants limited [43] and the principal goal of structural genomics
in the context of three-dimensional structure is not initiatives is to provide template structures for most
possible without defrosting (refining) the inherited tem- protein families [44,45]. Structure prediction approaches
plate backbone [18 ]. The development of such meth- are thus destined to become largely limited to compara-
ods may provide a stepping-stone toward new tive modeling, as evolutionarily related structures will be
comparative strategies that try to optimize all the mod- available for the majority of naturally occurring proteins
eling steps (template selection, alignment, modeling in the foreseeable future. As most modeling cases fall in
structurally variable regions and sidechain packing) in the 20 30% sequence identity range [46], where the
a more simultaneous way. For instance, one could apply a majority of new information is generated [34 ], further
refinement protocol to generate a diverse set of models progress is necessary to overcome the current major
based on different templates and alignment variants, and bottlenecks in comparative modeling: improving align-
use the energy function to discriminate near-native from ments and refining models. The development of effec-
erroneous models. tive all-atom structure refinement procedures should
tackle this problem, allowing the generation of high-
Additional challenges to the comparative modeling field, resolution models that can reproduce all key functional
which relies heavily on the assumption  similar sequences features.
 similar structures , are brought about by the existence
of evolutionarily related proteins that possess globally Acknowledgements
I would like to thank Lisa Kinch for critical reading of the manuscript.
distinct structures (Figure 2). The major mechanisms
by which proteins change their fold include insertions,
References and recommended reading
deletions and substitutions of structural elements, circular
Papers of particular interest, published within the annual period of
permutations, rearrangements of b-sheet topologies
review, have been highlighted as:
(strand invasions and withdrawals, b-hairpin flips and
of special interest
swaps) and fusion of duplicated domains [41,42]. The
of outstanding interest
Figure 2
1. Bradley P, Misura KM, Baker D: Toward high-resolution de novo
structure prediction for small proteins. Science 2005,
309:1868-1871.
A review of recent progress in de novo protein structure prediction.
2. Baker D, Sali A: Protein structure prediction and structural
genomics. Science 2001, 294:93-96.
3. Chothia C, Lesk AM: The relation between the divergence
of sequence and structure in proteins. EMBO J 1986,
5:823-826.
4. Marti-Renom MA, Stuart AC, Fiser A, Sanchez R, Melo F,
Sali A: Comparative protein structure modeling of genes
and genomes. Annu Rev Biophys Biomol Struct 2000,
29:291-325.
5. Altschul SF, Madden TL, Schaffer AA, Zhang J, Zhang Z, Miller W,
Lipman DJ: Gapped BLAST and PSI-BLAST: a new generation
 Same sequence adopts different structures  a highly challenging
of protein database search programs. Nucleic Acids Res 1997,
case for comparative modeling. (a) Experimental structure of CASP6
25:3389-3402.
target T0240, a 92-residue C-terminal fragment of TonB from
6. Ginalski K, Grishin NV, Godzik A, Rychlewski L: Practical lessons
Escherichia coli (TonB-92; PDB code 1u07, monomers in green and
from protein structure prediction. Nucleic Acids Res 2005,
olive), and the best model (T0240TS450_1, blue). (b) The available
33:1874-1891.
template, an 85-residue C-terminal fragment of TonB (TonB-85; PDB
This comprehensive review outlines currently available practical
code 1ihr, monomers in grey and orange). TonB-85 forms a swapped
approaches to protein structure prediction, including recent advances
dimer through the exchange of a b-hairpin and a C-terminal b-strand, in model quality assessment.
whereas TonB-92 dimerizes with considerably different structure
7. Ohlson T, Wallner B, Elofsson A: Profile-profile methods provide
without undergoing b-hairpin exchange. Although critical changes in
improved fold-recognition: a study of different profile-profile
the structure were modeled correctly, the model predicts TonB-92 to
alignment methods. Proteins 2004, 57:188-197.
be a monomer (as indeed it is expected to be in solution), with no An evaluation of different profile-profile alignment methods.
swapping of C-terminal b-strands seen in the experimental crystal
8. Wang G, Dunbrack RL Jr: Scoring profile-to-profile sequence
structure.
alignments. Protein Sci 2004, 13:1612-1626.
www.sciencedirect.com Current Opinion in Structural Biology 2006, 16:172 177
176 Theory and simulation
9. Sadreyev RI, Grishin NV: Quality of alignment comparison by Sali A et al.: EVA: evaluation of protein structure prediction
COMPASS improves with inclusion of diverse confident servers. Nucleic Acids Res 2003, 31:3311-3315.
homologs. Bioinformatics 2004, 20:818-828.
27. Deshpande N, Addess KJ, Bluhm WF, Merino-Ott JC,
10. Wrabl JO, Grishin NV: Gaps in structurally similar proteins: Townsend-Merino W, Zhang Q, Knezevich C, Xie L,
towards improvement of multiple sequence alignment. Chen L, Feng Z et al.: The RCSB Protein Data Bank: a
Proteins 2004, 54:71-87. redesigned query system and relational database
based on the mmCIF schema. Nucleic Acids Res 2005,
11. Przybylski D, Rost B: Improving fold recognition without folds.
33:D233-D237.
J Mol Biol 2004, 341:255-269.
28. Ginalski K, Elofsson A, Fischer D, Rychlewski L: 3D-Jury: a simple
12. Bujnicki JM, Elofsson A, Fischer D, Rychlewski L:
approach to improve protein structure predictions.
Structure prediction meta server. Bioinformatics 2001,
Bioinformatics 2003, 19:1015-1018.
17:750-751.
29. Fischer D: 3D-SHOTGUN: a novel, cooperative, fold-
13. Moult J, Fidelis K, Tramontano A, Rost B, Hubbard T: Critical
recognition meta-predictor. Proteins 2003, 51:434-441.
assessment of methods of protein structure prediction (CASP)
30. Ginalski K, von Grotthuss M, Grishin NV, Rychlewski L: Detecting
- round VI. Proteins 2005.
distant homology with Meta-BASIC. Nucleic Acids Res 2004,
14. Kryshtafovych A, Venclovas C, Fidelis K, Moult J: Progress 32:W576-W581.
over the first decade of CASP experiments. Proteins 2005,
31. Zhou H, Zhou Y: SPARKS 2 and SP(3) servers in CASP 6.
61(suppl 7):225-236.
Proteins 2005.
A description of the progress made in protein structure prediction during
the course of the CASP experiments.
32. Cozzetto D, Tramontano A: Relationship between multiple
sequence alignments and quality of protein comparative
15. Moult J: A decade of CASP: progress, bottlenecks and
models. Proteins 2005, 58:151-157.
prognosis in protein structure prediction. Curr Opin Struct Biol
The distribution of sequence identity in multiple sequence alignments is
2005, 15:285-289.
demonstrated to be a good estimator of the quality of comparative
This paper reviews the state of the art in protein structure prediction in the
models.
context of a decade of CASP experiments.
33. DeWeese-Scott C, Moult J: Molecular modeling of protein
16. Tress M, Ezkurdia I, Grana O, Lopez G, Valencia A: Assessment of
function regions. Proteins 2004, 55:942-961.
predictions submitted for the CASP6 comparative modelling
The authors explore the usefulness of comparative models in deducing
category. Proteins 2005, 61(suppl 7):27-45.
details of molecular function. They demonstrate that, in general, good
An assessment of the state of the art in comparative modeling from
insight into ligand binding can be obtained, providing there are no
CASP6.
alignment errors.
17. Ginalski K, Rychlewski L: Protein structure prediction
34. Chakravarty S, Sanchez R: Systematic analysis of added-value
of CASP5 comparative modeling and fold recognition
in simple comparative models of protein structure. Structure
targets using consensus alignment approach and 3D
2004, 12:1461-1470.
assessment. Proteins 2003, 53(suppl 6):410-417.
This study justifies the use of comparative models instead of templates to
estimate structure-derived properties of proteins, showing that, in gen-
18. Venclovas C, Margelevicius M: Comparative modeling in
eral, their added value increases with lower target-template sequence
CASP6 using consensus approach to template selection,
identity.
sequence-structure alignment and structure assessment.
Proteins 2005, 61(suppl 7):99-105.
35. Chakravarty S, Wang L, Sanchez R: Accuracy of
A report from one of the best performing groups in the comparative
structure-derived properties in simple comparative
modeling category of CASP6.
models of protein structures. Nucleic Acids Res 2005,
33:244-259.
19. Kolinski A, Bujnicki JM: Generalized protein structure
In an extension of their previous work [34 ], the authors show that the
prediction based on combination of fold-recognition with
average accuracy of structure-derived properties of comparative models
de novo folding and evaluation of models. Proteins 2005,
increases with higher target-template sequence identity. They also reveal
61(suppl 7):84-90.
that, for most properties, the differences observed between NMR and X-
20. Margelevicius M, Venclovas C: PSI-BLAST-ISS: an ray structures are similar to the errors in models based on templates with
intermediate sequence search tool for estimation of the 40% sequence identity.
position-specific alignment reliability. BMC Bioinformatics
36. Giorgetti A, Raimondo D, Miele AE, Tramontano A: Evaluating the
2005, 6:185.
usefulness of protein structure models for molecular
21. Luthy R, Bowie JU, Eisenberg D: Assessment of protein replacement. Bioinformatics 2005, 21:ii72-ii76.
models with three-dimensional profiles. Nature 1992, This study reveals that there is a clear relationship between the quality of
356:83-85. comparative models and their suitability for molecular replacement. It
also shows that target-template sequence identity is not a good diag-
22. Sippl MJ: Recognition of errors in three-dimensional
nostic for the success of the procedure.
structures of proteins. Proteins 1993, 17:355-362.
37. Contreras-Moreira B, Ezkurdia I, Tress ML, Valencia A:
23. Rohl CA, Strauss CE, Chivian D, Baker D: Modeling structurally
Empirical limits for template-based protein structure
variable regions in homologous proteins with Rosetta. Proteins
prediction: the CASP5 example. FEBS Lett 2005,
2004, 55:656-677.
579:1203-1207.
A de novo method for modeling structurally variable regions in compara- An analysis of the empirical limits of template-based modeling of protein
tive models based on the Rosetta structure prediction algorithm is
structure suggests that the methodology is approaching its limits for easy
described and evaluated.
comparative modeling and that additional improvements in quality require
information not available from template structures.
24. Fischer D, Rychlewski L, Dunbrack RL Jr, Ortiz AR, Elofsson A:
CAFASP3: the third critical assessment of fully automated
38. Zhang Y, Skolnick J: Automated structure prediction of weakly
structure prediction methods. Proteins 2003, 53(suppl 6):
homologous proteins on a genomic scale. Proc Natl Acad Sci
503-516.
USA 2004, 101:7594-7599.
25. Rychlewski L, Fischer D: LiveBench-8: the large-scale, 39. Misura KM, Baker D: Progress and challenges in high-
continuous assessment of automated protein structure resolution refinement of protein structure models.
prediction. Protein Sci 2005, 14:240-245. Proteins 2005, 59:15-29.
A report on the performance of protein structure prediction servers in the
40. Qian B, Ortiz AR, Baker D: Improvement of comparative
LiveBench-8 experiment.
model accuracy by free-energy optimization along
26. Koh IY, Eyrich VA, Marti-Renom MA, Przybylski D, principal components of natural structural variation.
Madhusudhan MS, Eswar N, Grana O, Pazos F, Valencia A, Proc Natl Acad Sci USA 2004, 101:15346-15351.
Current Opinion in Structural Biology 2006, 16:172 177 www.sciencedirect.com
Comparative modeling for protein structure prediction Ginalski 177
The authors present a novel approach to refining comparative models by 44. Yan Y, Moult J: Protein family clustering for structural
free energy optimization along evolutionarily favored sampling directions. genomics. J Mol Biol 2005, 353:744-759.
They show that improvement in model quality can be obtained.
45. Liu J, Hegyi H, Acton TB, Montelione GT, Rost B: Automatic
41. Grishin NV: Fold change in evolution of protein structures.
target selection for structural genomics on eukaryotes.
J Struct Biol 2001, 134:167-185.
Proteins 2004, 56:188-200.
42. Kinch LN, Grishin NV: Evolution of protein structures and
46. Sanchez R, Pieper U, Melo F, Eswar N, Marti-Renom MA,
functions. Curr Opin Struct Biol 2002, 12:400-408.
Madhusudhan MS, Mirkovic N, Sali A: Protein structure
modeling for structural genomics. Nat Struct Biol 2000,
43. Grant A, Lee D, Orengo C: Progress towards mapping the
7(suppl):986-990.
universe of protein folds. Genome Biol 2004, 5:107.
www.sciencedirect.com Current Opinion in Structural Biology 2006, 16:172 177