Protein complexes: structure prediction challenges for
the 21st century
Patrick Aloy, Matthieu Pichaud and Robert B Russell
Only a tiny fraction of the many hundreds of known protein model. As the coverage of the space of protein structure
complexes are also of known three-dimensional structure. grows, so does the average quality of such models.
The experimental difficulties surrounding structure
determination of complexes make methods that are able to At the same time, molecular biology and, in turn, struc-
predict structures paramount. The challenge of predicting tural biology have moved beyond studies of single mole-
complex structures is daunting and raises many issues that
cules. The great structural successes of the past five years
need to be addressed. To produce the best models, new
have almost invariably involved large macromolecular
prediction methods have to somehow combine partial
complexes (i.e. [5 8]). Unlike single proteins, whose
structures with a mixed bag of experimental data,
structures can now often be solved very rapidly, complex
including interactions and low-resolution electron
structures are difficult to obtain because of several tech-
microscopy images.
nical problems that will probably take years to overcome
[9]. However, many complexes have now been discov-
Addresses
ered by high-throughput interaction discovery techniques
EMBL, Meyerhofstrasse 1, 69117 Heidelberg, Germany
[10,11]. These complexes are thus now known in the
sense that their constituents have been determined, but it
Corresponding author: Russell, Robert B ([e-mail])
will probably be many years before high-resolution struc-
tures are available. Electron microscopy (EM) promises to
Current Opinion in Structural Biology 2005, 15:15 22
provide low-resolution structures of many of these com-
plexes, as this method can work with comparatively small
This review comes from a themed issue on
samples (a major stumbling block in X-ray crystallogra-
Folding and binding
Edited by Gideon Schreiber and Luis Serrano
phy). But, to be most useful, these images must be
complemented with models of the separate subunits
Available online 26th January 2005
and how they interact with each other. The new structure
0959-440X/$  see front matter prediction challenge is somehow to convert low-resolu-
# 2005 Elsevier Ltd. All rights reserved.
tion images, homology models of individual subunits and
a mixed bag of other information, including two-hybrid
DOI 10.1016/j.sbi.2005.01.012
interactions or expression data, to create the best possible
model. Here, we discuss the problems associated with
predicting complex structures, with a particular emphasis
Introduction
on parallels with prediction strategies for single proteins,
The challenge of predicting the structure adopted by a
and review recent work in this exciting new direction.
protein given only its amino acid sequence has been
taken up by many scientists over the past 50 years.
The difficulty of the problem has meant that, at best, The scope of the problem
limited progress was made until comparatively recently A typical complex modeling problem (Figure 1) involves
[1 4]. The availability of more sequence and structural a set of components known experimentally to co-exist in
data, faster computers and increasingly clever algorithms some biological context. Techniques such as the two-
and concepts now means that it is often possible to hybrid system or affinity purification (i.e. pull-down),
provide structural models with reasonable accuracy for whether applied in individual instances or to complete
a large fraction of the proteins present in an organism. genomes, have provided many hundreds of complexes.
This is a boon to those wishing to understand the mole- Standard bioinformatics will typically reveal the presence
cular details of particular molecules or those seeking to of known domains within the protein components and a
design new pharmaceuticals. No doubt the picture will growing fraction of these ( 60% of sequences in Swis-
continue to improve over time, but it is clear that much of sProt) either will be of known structure or will share
the field has matured to the point where progress is sufficient homology with others to permit models of
largely incremental. Indeed, the speed at which struc- varying degrees of quality to be constructed [12]. Some-
tures of individual proteins are now solved makes it likely times the complex will show complete or near complete
that the grand challenge of predicting the structure of a similarity to another complex of known structure, that is,
protein given only its amino acid sequence will probably homologues of most components have been seen pre-
become largely academic. Increasingly, structure predic- viously in a single structure. However, this situation is
tion involves simply finding the best homologous protein comparatively rare owing to the paucity of known com-
of known structure (template) on which to construct a plex structures. Most complexes cannot be modeled
www.sciencedirect.com Current Opinion in Structural Biology 2005, 15:15 22
16 Folding and binding
Figure 1
Ski2 Ski8 Ski3
1 1287 1 397 1 1492
Domain
assignment
Modelling components
Ski2a 300 500 Ski2b 600 800 Ski8 Ski3 700 800
*
Modelling of binary
interactions
*
*
Ski2a Ski8 Ski2b Ski3
Ski2a Ski2b
Complex assembly
*
Current Opinion in Structural Biology
Overview of the complex modeling procedure. The first step is to annotate each protein for domains and build three-dimensional models for as
many components as possible. Binary interactions can be constructed by identifying all suitable interaction templates of known three-dimensional
structure and using them to build all possible alternative interaction models. These can then be combined into one or more models of the
complete complex, of which the best can be selected based on measures of model quality (e.g. avoiding bad structural clashes, interface
preservation, etc.). An asterisk denotes the models chosen; several other possible models are not shown for clarity.
completely in one step and indeed even large complexes thus often need to be treated as if they were separate
with known structures are often found to have additional components.
components not seen during experimental structure
determination (e.g. transient components or components Model accuracy for single proteins
present only in specific cellular locations or conditions). In any attempt to predict the structure of a protein
For most complexes, there is a problem of how to take a complex, the first step is necessarily to model the separate
set of predictions for individual subunits or indeed sub- components. Homology modeling procedures have im-
complexes, and combine them to give the best and most proved during the past decade, although the quality of the
complete model (Figure 1). model is still tightly linked to the sequence identity
shared between the target sequence and template [2].
Many single proteins, particularly in eukaryotes, contain If the two proteins share over 60% sequence identity, the
more than one domain and this adds another complication resulting model can be of high quality, even comparable
to the modeling process. It is often the case that one to lower resolution X-ray ( 3 Å) or NMR structures.
can find suitable templates for modeling domains sepa- Below this value, accuracies are poorer, with insertions/
rately, but then lack any suitable single structure on deletions and most exposed sidechains being incorrect
which to model their relative orientation. These domains [13], and techniques exploiting fine atomic details
Current Opinion in Structural Biology 2005, 15:15 22 www.sciencedirect.com
Complex structure prediction Aloy, Pichaud and Russell 17
Figure 2
(a)
IL-1
FGF2
FGF2 receptor IL-1 receptor
(b)
CheY CheA-P2
CheY CheA-P2
(E. coli) (E. coli)
(T. maritima) (T. maritima)
Current Opinion in Structural Biology
Extreme cases of the relationship between sequence and interaction similarity. Molscript [41] figures showing (a) structurally similar interactions
in the absence of sequence similarity and (b) orthologous protein pairs interacting differently. Protein complexes are shown in similar
orientations [CheY in (b)] and colored from blue to red (N to C terminus). FGF, fibroblast growth factor.
(e.g. molecular dynamics force-fields) will not usually same is true of their receptors, which are distant relatives
give reliable results. within the immunoglobulin superfamily. Somewhat
surprisingly, the complex structures are also similar
Subcomplexes: modeling interactions (Figure 2a). On the other hand, proteins with similar
If one knows that a pair of proteins are in physical contact sequences can differ in the way they interact. For exam-
and has some idea about what the structures of the ple, the interaction between the clearly homologous
separated components look like (either experimental or signaling proteins CheY and CheA-P2 differs by a rotation
modeled), then there are several possibilities for model- of 1808 in different bacterial species [14 ] (Figure 2b).
ing how they interact. The best models come when one
can identify another experimental structure containing Anecdotes are important to remember, but general trends
similar proteins in complex with each other. As for model- are always needed. Here, much can be learned from
ing individual subunits, the accuracy or quality of the studies of individual proteins. The classic work of
model depends on the degree of sequence similarity Chothia and Lesk [15] showed how protein structure
between the proteins. A critical question is thus whether diverged as a function of sequence divergence: as
homologous proteins will interact in a similar way. In sequence similarity dropped, the root mean square devia-
other words, if one knows that A and B interact and there tion (rmsd) between Ca carbons increased up to a max-
is a crystal structure in which homologous proteins (A0 imum value that probably represented the upper limit
and B0) interact, can we use it to model the A B inter- possible between proteins sharing a similar fold (or
action? And, related to this, what degree of sequence indeed how the Ca equivalences were chosen). This
similarity is needed to infer that the interactions will also work was later extended by others to determine the
be similar? degree of sequence similarity needed to infer whether
proteins would adopt a similar three-dimensional struc-
A quick survey of the interactions of known structures ture [16]. It is now well established that pairs of sequences
suggests that extreme cases are possible. For example, it sharing more than about 25% sequence identity can be
is possible for interactions to be similar despite very weak confidently said to be similar in structure. Below this,
sequence similarity. Fibroblast growth factors are similar there lies a twilight zone, where dissimilar pairs of sequ-
in structure but not in sequence to interleukins-1 and the ences are mixed with those sharing structure similarity.
www.sciencedirect.com Current Opinion in Structural Biology 2005, 15:15 22
18 Folding and binding
Figure 3
140
120
100
Low Medium High
80
60
80th percentile
40
20
10 Å iRMSD
0
0 20 40 60 80 100
Sequence identity (%)
Current Opinion in Structural Biology
Relationship between sequence and interaction divergence. Plot showing interaction similarity (iRMSD) [17 ] versus percentage sequence
identity. Visual inspection reveals that two interactions are structurally similar if their iRMSD is below 10 Å. The curve shows the 80th percentile
(i.e. 80% of the data are below the curve). The plot is split into three bins to illustrate the level of detail expected to be correct at particular
degrees of sequence similarity. In the high-quality range (i.e. >60%), nearly atomic details of the interaction interface can be accurately predicted,
with those points with poor similarity generally corresponding to well-documented cases (e.g. different antibodies for the same lysozyme).
In the medium range (30 60%), the overall structural similarity will be conserved, although the molecular details of the interaction (i.e. interacting
residue pairs) are often different, due, for instance, to the reorientation of some secondary structure elements. Finally, in the low-quality range
(<30%), many structural elements will be distorted or missing (Ca trace in the figure) and only the rough relative orientation of the two
proteins can be predicted.
We performed a similar study using interacting pairs be predicted correctly. However, as for individual struc-
rather than single proteins and found similar trends tures, techniques related to threading (i.e. fold recogni-
[17 ]. Above about 25% sequence identity, pairs of pro- tion) can help recognize remote structural similarities and
teins tend to interact similarly (Figure 3), with a twilight thus detect interaction similarities with identities as low
zone below this, where interactions may or may not be as 15% [18].
similar. This provides a rough framework similar to that
often used in modeling [13] concerning the degree of How many thousands of interactions?
detail one can expect from predictions (Figure 3). At the The fact that many homologous proteins interact simi-
high end of identity (e.g. >60%), modeling of interactions larly raises the question as to how many different kinds of
can be expected to give fairly accurate details of the protein protein interactions there are in Nature. This is
interaction interface. In the medium range (30 60%), very similar to the situation for protein families or folds
one expects substantial differences in the precise details that existed in the early 1990s, when the then available
despite an overall similarity. At the lowest end (<30%), it genomic sequences, together with the Protein Data Bank
is difficult to say whether or not interactions will resemble (PDB), were used to suggest that there were probably a
each other at all and, even if they do, it is unlikely that limited number of folds in Nature that most protein
more than the rough orientation of the two proteins will domains could adopt [19,20]. We recently applied similar
Current Opinion in Structural Biology 2005, 15:15 22 www.sciencedirect.com
iRMSD
Complex structure prediction Aloy, Pichaud and Russell 19
thinking to interactions instead of single proteins. Using years through several technological developments and,
interaction data from genome-scale two-hybrid or affinity most importantly, by incorporating experimental informa-
purification screens, together with the current set of tion into the predictions [27,28]. There are now several
known structures of protein complexes, we estimated examples of docking applied successfully in close colla-
that the number of  interaction types (the interaction boration with experiment (e.g. [29,30]). It is likely that
equivalent of  fold ) is limited to about 10 000, of which the integration of these approaches with NMR, EM or
we currently know about 2000 [21 ]. other coarse information about shape or interaction sur-
faces will be of great use in complex modeling. The
Is the interface right? identification of putative binding patches on the surfaces
The fact that so many interaction types are now known of the two interacting proteins can be a helpful guide in
and that this number is growing at a rate of 300 400 per docking experiments [31 ], as it drastically restricts the
year [21 ] means that there will be an increasing number number of possibilities in a global space search. Recent
of possibilities for modeling interactions. Even in the work studying the structural characteristics of interaction
absence of evidence for an interaction between a pair interfaces across different complex types [32 ] is also of
of proteins, it will be increasingly possible to model the great value.
interaction based on previously determined structures
involving homologues (i.e. effectively using interactions There are clear parallels between docking and the most
of known structure to predict new interactions between difficult predictions of single protein structures, namely
homologous proteins). However, before doing so, it is those for which one has no template on which to model
critical to assess the fit of the putatively interacting the structure. These ab initio or de novo predictions
protein pairs on the interaction of known structure, as improve as the number of known structures increases
multiple interactions of the same type might exist in a [4] and it is likely that docking will be similar. More
particular species or even a single complex. To do this, structural information about how proteins interact will
two methods have emerged recently that use crude undoubtedly increase performance. However, this will
protein models to assess the interface complementarity come at a cost: just as more structures of single proteins
of any modeled interaction (i.e. the specificity). The first makes de novo prediction methods applicable to fewer
approach relies on empirical pair potentials, derived from proteins, ultimately, a complete repertoire of three-
a non-redundant database of interactions of known three- dimensional interaction types will reduce the number
dimensional structure [22], to score how well a pair of of interactions to which docking can be applied. Just like
proteins fits into a homologous complex [23,24]. The the symbiosis between homology modeling and de novo
second method, inspired by fold recognition, threads predictions for single proteins (for instance, with empiri-
the sequences of two putatively interacting proteins cal potentials or energy minimization), there will probably
onto a library of interactions of known structure and be a merger of ideas, with docking principles aiding the
scores the fits of the individual components and their modeling of interactions by homology and vice versa.
interface [18]. The two approaches are computationally
inexpensive (i.e. less than 5 s per protein pair) and From binary interactions into complexes:
are thus suited to genome-scale studies. The second the cellular puzzle
approach has already been used to construct a genome- Given a set of binary interaction models, the next logical
wide structure-based protein interaction network in yeast step in complex prediction is to assemble them into
[25 ]. higher order structures, that is, to put them together into
one or more multisubunit models. In the best cases,
When one can obtain high-quality models (i.e. sequence interactions can be combined based on a shared compo-
identity between the model and the template above nent, an interacting protein that is common to two or more
60%), methods using atomic potentials can also be binary interactions. Problems arise when clashes occur as
applied to discriminate binding partners among a collec- interactions are combined or if there are multiple ways in
tion of homologous proteins [26]. However, the high which particular subunits can interact with each other
computational cost means that they are not currently (Figure 1). Disentangling this problem will probably
applicable to genome-scale studies. involve selecting the best binary interactions based
on their quality, for example, the degree of sequence
Docking: new roles for an old technique similarity, the preservation of the interaction interface,
In the absence of a template on which to model an and whether or not the interaction within the complex
interaction, it is still sometimes possible to obtain accu- is supported by additional experiments, such as the
rate models of the interaction using theoretical docking two-hybrid system.
techniques. Methods have been around for many years,
but their utility has been hampered by poor accuracy, There are again certain parallels between complex pre-
probably related to a lack of structural data. Encoura- dictions and predictions for single proteins, and again it is
gingly, the methods have improved over the past few likely that the latter can greatly aid the former. The most
www.sciencedirect.com Current Opinion in Structural Biology 2005, 15:15 22
20 Folding and binding
Figure 4
(a) (b)
164 1063
SPT5
TFG2
1 42
? ?
Current Opinion in Structural Biology
Complex assembly. Molscript [41] figures illustrating (a) the problem of assembling several homologous, but non-identical, proteins into a
repetitive structure (the 20S proteasome) and (b) the identification of new potentially transient interactions involving RNA polymerase II and the
transcription initiator factors SPT5 and TFG2.
successful de novo structure predictions work by first particularly EM images. This makes the prediction pro-
identifying possible structures for fragments of a larger blem very difficult, as these subunits are often highly
polypeptide sequence (e.g. [33,34]). These fragments can similar in sequence and can thus stymie the interface
be overlapping and there may be several alternatives for a fitting approaches discussed above. In this case, addi-
particular stretch of sequence. Once these fragments have tional experiments can make a world of difference. For
been identified, the challenge is then to assemble them instance,  micro-complexes (individual pull-downs with
into the most accurate possible model. This is done by pairs of subunits) were used in conjunction with EM
generating many thousands of alternatives, and somehow data to suggest a pseudo-atomic model for the CCT
selecting the best one based on several theoretical con- chaperonin [35]. Similarly, the two-hybrid system was
siderations and on how well each model accounts for used to suggest a putative structure for the exosome [36],
experimental data. which revealed several errors in our predictions based on
interface preservation [37].
Complex prediction clearly presents a similar challenge.
Instead of fragments of a single protein, one can, in Space and time
principle, generate many alternatives for how pairs of The prediction problem is further complicated by the fact
components interact (Figure 1). One must then assemble that few complexes are static entities. They shuttle from
these into several alternative possible models, at which one cellular compartment to another, they adopt different
point one is faced with the same challenge of selecting subunits under different conditions and even change
the best using various criteria. These could include the shape while going about their daily business. A set of
interface preservation measures mentioned above, in known components might well be an average over several
addition to experimental data from the two-hybrid system  flavors of one complex and even the most accurate
or other interaction experiments. model will probably only reflect the state of a complex
at one particular time (Figure 4b).
Many complexes also present the quite separate chal-
lenge of having to assemble several similar molecules into To date, details about a complex s location in the cell and
a repetitive usually ring-like structure (Figure 4). This the time at which it is present have mostly been ignored,
means that there are often many hundreds of alterna- mainly because of a lack of precise data. However, even in
tive models (e.g. for a ring of 7 there are 7!/7 = 720) that the absence of very specific experiments, it is still possible
would all probably fit with many experimental details, to get some hints from high-throughput analyses.
Current Opinion in Structural Biology 2005, 15:15 22 www.sciencedirect.com
Complex structure prediction Aloy, Pichaud and Russell 21
10. Gavin AC, Bosche M, Krause R, Grandi P, Marzioch M, Bauer A,
Genome-scale cellular location assays in yeast provide
Schultz J, Rick JM, Michon AM, Cruciat CM et al.: Functional
insights into where proteins are in the cell [38 ] and
organization of the yeast proteome by systematic analysis of
protein complexes. Nature 2002, 415:141-147.
expression studies [39 ] give some indication of when
subunits are present at the same time. These data are 11. Ho Y, Gruhler A, Heilbut A, Bader GD, Moore L, Adams SL, Millar A,
Taylor P, Bennett K, Boutilier K et al.: Systematic identification of
particularly relevant for cell cycle proteins, for which
protein complexes in Saccharomyces cerevisiae by mass
expression is tightly coupled with their presence in the spectrometry. Nature 2002, 415:180-183.
cell [40]. Structural information can also provide clues as
12. Pieper U, Eswar N, Braberg H, Madhusudhan MS, Davis FP,
Stuart AC, Mirkovic N, Rossi A, Marti-Renom MA, Fiser A et al.:
to when proteins and complexes interact. For instance, if
MODBASE, a database of annotated comparative protein
one protein uses the same surface to interact with differ-
structure models, and associated resources. Nucleic Acids Res
ent partners, then it is unlikely that these interactions 2004, 32:D217-D222.
can occur simultaneously (i.e. the interaction of Ras
13. Marti-Renom MA, Stuart AC, Fiser A, Sanchez R, Melo F, Sali A:
Comparative protein structure modeling of genes and
with its GAP activator protein and the exchanging factor
genomes. Annu Rev Biophys Biomol Struct 2000, 29:291-325.
GEF).
14. Park SY, Beel BD, Simon MI, Bilwes AM, Crane BR: In different
organisms, the mode of interaction between two signaling
Concluding remarks
proteins is not necessarily conserved. Proc Natl Acad Sci USA
2004, 101:11646-11651.
Complexes represent the cornerstone of much of modern
Crystal structures show that the mode of interaction between the signal-
biology and nanotechnology. A deeper understanding
ing proteins CheY and CheA-P2 is different in Thermotoga maritima and
comes from knowledge of their three-dimensional struc- Escherichia coli. Although these orthologous proteins use roughly the
same residues for the interaction, the relative orientations differ by a 1808
tures. The current bottlenecks surrounding experimental
rotation.
structure determination make the need for new predic-
15. Chothia C, Lesk AM: The relation between the divergence of
tive methods acute. The next decade will undoubtedly
sequence and structure in proteins. EMBO J 1986, 5:823-826.
see many advances in the exciting new field of complex
16. Sander C, Schneider R: Database of homology-derived protein
structure prediction.
structures and the structural meaning of sequence alignment.
Proteins 1991, 9:56-68.
Acknowledgements
17. Aloy P, Ceulemans H, Stark A, Russell RB: The relationship
MP is supported by an E-STAR fellowship funded by the EC s FP6 Marie between sequence and interaction divergence in proteins.
J Mol Biol 2003, 332:989-998.
Curie Host fellowship for Early Stage Research Training under contract
An interaction equivalent of the study published by Chothia and Lesk for
number MEST-CT-2004-504640.
single proteins [15]. A plot of similarity of interaction, as measured by
studying known three-dimensional structures, reveals how interactions
References and recommended reading diverge as a function of sequence similarity. Binding modes of interacting
Papers of particular interest, published within the annual period of protein pairs that share above 30% sequence identity are generally
review, have been highlighted as: similar. Below this cutoff lies a twilight zone, where the modes of inter-
action may or may not be similar.
of special interest
18. Lu L, Lu H, Skolnick J: MULTIPROSPECTOR: an algorithm for
of outstanding interest
the prediction of protein-protein interactions by multimeric
threading. Proteins 2002, 49:350-364.
1. Tramontano A: Of men and machines. Nat Struct Biol 2003,
10:87-90. 19. Chothia C: One thousand families for the molecular biologist.
Nature 1992, 357:543-544.
2. Tramontano A, Morea V: Assessment of homology-based
predictions in CASP5. Proteins 2003, 53(suppl 6):352-368. 20. Vitkup D, Melamud E, Moult J, Sander C: Completeness in
structural genomics. Nat Struct Biol 2001, 8:559-566.
3. Kinch LN, Wrabl JO, Krishna SS, Majumdar I, Sadreyev RI,
Qi Y, Pei J, Cheng H, Grishin NV: CASP5 assessment of 21. Aloy P, Russell RB: Ten thousand interactions for the molecular
fold recognition target predictions. Proteins 2003, biologist. Nat Biotechnol 2004, 22:1317-1321.
53(suppl 6):395-409. This study combines interaction data and three-dimensional structures to
suggest that, as for protein folds, Nature is apparently restricted to a
4. Aloy P, Stark A, Hadley C, Russell RB: Predictions without
limited number of interaction types with which to perform a vast set of
templates: new folds, secondary structure, and contacts in
functions. The authors estimate that most interactions in Nature will
CASP5. Proteins 2003, 53(suppl 6):436-456.
conform to one of about 10 000 types, of which we currently know about
2000.
5. Zhang G, Campbell EA, Minakhin L, Richter C, Severinov K,
Darst SA: Crystal structure of Thermus aquaticus core RNA
22. Stein A, Russell RB, Aloy P: 3did: interacting domains of known
polymerase at 3.3 Å resolution. Cell 1999, 98:811-824.
three-dimensional structure. Nucleic Acids Res 2005, in press.
6. Cramer P, Bushnell DA, Fu J, Gnatt AL, Maier-Davis B,
23. Aloy P, Russell RB: Interrogating protein interaction networks
Thompson NE, Burgess RR, Edwards AM, David PR,
through structural biology. Proc Natl Acad Sci USA 2002,
Kornberg RD: Architecture of RNA polymerase II and
99:5896-5901.
implications for the transcription mechanism. Science 2000,
288:640-649.
24. Aloy P, Russell RB: InterPreTS: protein interaction prediction
through tertiary structure. Bioinformatics 2003, 19:161-162.
7. Ban N, Nissen P, Hansen J, Moore PB, Steitz TA: The complete
atomic structure of the large ribosomal subunit at 2.4 Å
25. Lu L, Arakaki AK, Lu H, Skolnick J: Multimeric threading-based
resolution. Science 2000, 289:905-920.
prediction of protein-protein interactions on a genomic scale:
application to the Saccharomyces cerevisiae proteome.
8. Yusupov MM, Yusupova GZ, Baucom A, Lieberman K, Earnest TN,
Genome Res 2003, 13:1146-1154.
Cate JH, Noller HF: Crystal structure of the ribosome at 5.5 Å
In this study, the authors predict the protein interaction network in
resolution. Science 2001, 292:883-896.
S. cerevisiae by applying their MULTIPROSPECTOR threading algorithm
9. Russell RB, Alber F, Aloy P, Davis FP, Korkin D, Pichaud M, Topf M, to the entire proteome. This reveals a network that the authors argue
Sali A: A structural perspective on protein-protein interactions. is similar in size and accuracy to those uncovered by genome-scale
Curr Opin Struct Biol 2004, 14:313-324. two-hybrid screens.
www.sciencedirect.com Current Opinion in Structural Biology 2005, 15:15 22
22 Folding and binding
26. Kiel C, Wohlgemuth S, Rousseau F, Schymkowitz J, Borg J, 34. Bradley P, Chivian D, Meiler J, Misura KM, Rohl CA, Schief WR,
Wittinghofer F, Serrano L: Recognizing and defining true Wedemeyer WJ, Schueler-Furman O, Murphy P, Schonbrun J
Ras binding domains II: in silico prediction based on homology et al.: Rosetta predictions in CASP5: successes, failures,
modelling and energy calculations. J Mol Biol 2005, and prospects for complete automation. Proteins 2003,
in press. 53(suppl 6):457-468.
27. Mendez R, Leplae R, De Maria L, Wodak SJ: Assessment of blind 35. Valpuesta JM, Martin-Benito J, Gomez-Puertas P, Carrascosa JL,
predictions of protein-protein interactions: current status of Willison KR: Structure and function of a protein folding
docking methods. Proteins 2003, 52:51-67. machine: the eukaryotic cytosolic chaperonin CCT.
FEBS Lett 2002, 529:11-16.
28. Wodak SJ, Mendez R: Prediction of protein-protein
interactions: the CAPRI experiment, its evaluation and 36. Raijmakers R, Noordman YE, van Venrooij WJ, Pruijn GJ:
implications. Curr Opin Struct Biol 2004, 14:242-249. Protein-protein interactions of hCsl4p with other human
exosome subunits. J Mol Biol 2002, 315:809-818.
29. Hothorn M, Wolf S, Aloy P, Greiner S, Scheffzek K: Structural
insights into the target specificity of plant invertase and 37. Aloy P, Ciccarelli FD, Leutwein C, Gavin AC, Superti-Furga G,
pectin methylesterase inhibitory proteins. Plant Cell 2004, Bork P, Boettcher B, Russell RB: A complex prediction:
16:3437-3447. three-dimensional model of the yeast exosome. EMBO Rep
2002, 3:628-635.
30. Morillas M, Gomez-Puertas P, Rubi B, Clotet J, Arino J, Valencia A,
Hegardt FG, Serra D, Asins G: Structural model of a malonyl- 38. Huh WK, Falvo JV, Gerke LC, Carroll AS, Howson RW,
CoA-binding site of carnitine octanoyltransferase and Weissman JS, O Shea EK: Global analysis of protein
carnitine palmitoyltransferase I: mutational analysis of a localization in budding yeast. Nature 2003, 425:686-691.
malonyl-CoA affinity domain. J Biol Chem 2002, The first genome-wide experiment to uncover protein localization. Using a
277:11473-11480. collection of full-length GFP-tagged proteins for the 6234 annotated open
reading frames in yeast, the authors could classify 75% into 22 subcellular
31. Halperin I, Wolfson H, Nussinov R: Protein-protein interactions;
localization categories, from the cell periphery to the nucleolus.
coupling of structurally conserved residues and of hot spots
across interfaces. Implications for docking. Structure 2004, 39. Ghaemmaghami S, Huh WK, Bower K, Howson RW, Belle A,
12:1027-1038. Dephoure N, O Shea EK, Weissman JS: Global analysis of
This study finds that clusters of structurally conserved residues (hot protein expression in yeast. Nature 2003, 425:737-741.
spots) in proteins tend to be coupled across the interface. This suggests A novel strategy to study protein expression on a large scale was applied
that the prediction of hot spots can be used as a filter to complement to the yeast S. cerevisiae. All 6234 open reading frames were tagged with
protein docking experiments. a high-affinity epitope for immuno-detection. About 80% of the proteins
were expressed under normal growth conditions, with abundances rang-
32. Nooren IM, Thornton JM: Diversity of protein-protein
ing from 50 to 106 molecules per cell. This technology provides the perfect
interactions. EMBO J 2003, 22:3486-3492.
framework for monitoring protein abundance dynamically during different
The authors use three-dimensional structures to study the functional
cellular events and under different conditions.
diversity of protein protein interactions and complexes. The authors
pay particular attention to the difference between transient interactions 40. de Lichtenberg U, Jensen L, Brunak S, Bork P: Dynamic complex
and those of a more permanent nature in the cell. formation during the yeast cell cycle. Science 2005, in press.
33. Bystroff C, Baker D: Prediction of local structure in proteins 41. Kraulis PJ: MOLSCRIPT: a program to produce both detailed
using a library of sequence-structure motifs. J Mol Biol 1998, and schematic plots of protein structures. J Appl Crystallogr
281:565-577. 1991, 24:946-950.
Current Opinion in Structural Biology 2005, 15:15 22 www.sciencedirect.com