Membrane Protein Structure: Prediction versus Reality Arne Elofsson and Gunnar von Heijne Center for Biomembrane Research, Stockholm Bioinformatics Center, Department of Biochemistry and Biophysics, Stockholm University, SE-106 91 Stockholm, Sweden; email: [email protected], [email protected]Annu. Rev. Biochem. 2007. 76:125–40 First published online as a Review in Advance on January 11, 2007 The Annual Review of Biochemistry is online at biochem.annualreviews.org This article’s doi: 10.1146/annurev.biochem.76.052705.163539 Copyright c 2007 by Annual Reviews. All rights reserved 0066-4154/07/0707-0125$20.00 Key Words bioinformatics, membrane protein structure prediction, topology Abstract Since high-resolution structural data are still scarce, different kinds of theoretical structure prediction algorithms are of major impor- tance in membrane protein biochemistry. But how well do the cur- rent prediction methods perform? Which structural features can be predicted and which cannot? And what can we expect in the next few years? 125 Annu. Rev. Biochem. 2007.76:125-140. Downloaded from www.annualreviews.org by NORTH CAROLINA STATE UNIVERSITY on 09/29/12. For personal use only.
Transcript
ANRV313-BI76-06 ARI 30 April 2007 17:26
Membrane ProteinStructure: Predictionversus RealityArne Elofsson and Gunnar von HeijneCenter for Biomembrane Research, Stockholm Bioinformatics Center, Department ofBiochemistry and Biophysics, Stockholm University, SE-106 91 Stockholm, Sweden;email: [email protected], [email protected]
Annu. Rev. Biochem. 2007. 76:125–40
First published online as a Review in Advance onJanuary 11, 2007
The Annual Review of Biochemistry is online atbiochem.annualreviews.org
This article’s doi:10.1146/annurev.biochem.76.052705.163539
bioinformatics, membrane protein structure prediction, topology
AbstractSince high-resolution structural data are still scarce, different kindsof theoretical structure prediction algorithms are of major impor-tance in membrane protein biochemistry. But how well do the cur-rent prediction methods perform? Which structural features can bepredicted and which cannot? And what can we expect in the next fewyears?
Membrane proteins are crucial players in thecell and take center stage in processes rang-ing from basic small-molecule transport tosophisticated signaling pathways. Many arealso prime contemporary or future drug tar-gets, and it has been estimated that more thanhalf of all drugs currently on the market aredirected against membrane proteins (1). Bycontrast, it is still frustratingly hard to ob-tain high-resolution three-dimensional (3D)
structures of membrane proteins, and theyrepresent less than 1% of the structures in theProtein Data Bank (2). Even if the numberof experimentally known membrane proteinstructures is on the rise (3, 4), methods to pre-dict their topology (i.e., the transmembranesegments and their in-out orientation acrossthe membrane) and fold type from the aminoacid sequence will be needed for many yearsto come.
In this review, we discuss current topol-ogy and structure prediction methods againsta background of knowledge that has beengleaned from membrane protein structuresand from studies of protein insertion and fold-ing in cellular membranes. We attempt to pro-vide a realistic picture of what one may andmay not expect from the various predictionschemes and to identify major issues yet to beresolved.
MEMBRANE PROTEINSTRUCTURES: THE BASICFACTS
Integral membrane proteins come in two ba-sic architectures: the α-helix bundle and theβ-barrel. Helix-bundle proteins are found inall cellular membranes and represent an es-timated 20% to 25% of all open readingframes (ORFs) in fully sequenced genomes(5). The number of β-barrel membrane pro-teins is more uncertain because they are moredifficult to identify by sequence gazing; forbacteria, a rough estimate, based on the factthat all known β-barrel proteins are in theouter membrane and hence are made withan (easily predicted) N-terminal signal pep-tide, suggests that they account for no morethan a few percent of all ORFs. The EcoCycdatabase (6) currently lists 58 outer membraneand 511 inner membrane proteins out of atotal of 4332 proteins; considering that thenumber of inner membrane proteins in Es-cherichia coli has been estimated to be closeto 1000 (5), one may guess at somewhere be-tween 100 and 150 outer membrane proteins
126 Elofsson · von Heijne
Ann
u. R
ev. B
ioch
em. 2
007.
76:1
25-1
40. D
ownl
oade
d fr
om w
ww
.ann
ualr
evie
ws.
org
by N
OR
TH
CA
RO
LIN
A S
TA
TE
UN
IVE
RSI
TY
on
09/2
9/12
. For
per
sona
l use
onl
y.
ANRV313-BI76-06 ARI 30 April 2007 17:26
in total. This number of ∼100 E. coli outermembrane proteins is consistent with theresults from recent attempts to identifybacterial outer membrane proteins computa-tionally (7–9).
Whether a helix bundle or β-barrel, all in-tegral membrane proteins share common sur-face characteristics with a belt of hydrophobic(mainly aliphatic) amino acids flanked by two“aromatic girdles” composed of Trp and Tyrresidues (10–12). This mirrors the structureof the surrounding lipid bilayer, with the lipidheadgroup regions corresponding to the aro-matic girdles and the hydrocarbon tail regionto the hydrophobic belt, and ensures a seam-less fit of the proteins to the membrane.
Even though their surface structures aresimilar, the two classes of proteins have com-pletely different secondary structures andfolds. As their names imply, helix-bundle pro-teins are built from long transmembrane α-helices that pack together into more or lesscomplicated bundles, whereas β-barrel pro-teins are large antiparallel β-sheets rolled upinto a barrel closed by the first and last strandsin the sheet. In both cases, all backbone hy-drogen bonds in the membrane-buried partsof the protein are internally satisfied withinthe helices or between the β-strands. Anotherfundamental difference between the helix-bundle and β-barrel proteins pertains to theirbiosynthesis and mechanism of membrane in-sertion; this is discussed in the next section.
Because all current membrane proteintopology and structure prediction schemesfirst seek to identify the transmembrane seg-ments, they are obviously quite different, andtheir variations depend on the class of pro-tein for which they are designed. Generallyspeaking, long hydrophobic transmembranehelices are easier to recognize in an aminoacid sequence than the much shorter and lesshydrophobic transmembrane β-strands, andpartly for this reason, much more bioinfor-matics work has been devoted to the helix-bundle proteins—another instance of thewell-known dictum “always go for the easyproblems.”
Translocon: aprotein complex thatassures thetranslocation ofproteins across acellular membrane
Endoplasmicreticulum (ER):organelle into whichsecretory andmembrane proteinsare delivered uponsynthesis on theribosome
As do all other proteins, a membrane pro-tein starts its life on the ribosome. But alreadyat this early stage, helix-bundle and β-barrelproteins are handled differently (13, 14): ri-bosomes making helix-bundle proteins typi-cally bind cotranslationally to translocons ina target membrane [the inner membrane inbacteria, the endoplasmic reticulum (ER) ineukaryotes], whereas bacterial β-barrel pro-teins are initially transferred from the ribo-some to the soluble cytoplasmic SecB chaper-one, Figure 1.
The cotranslational membrane insertionof helix-bundle proteins has been studied in-tensely for many years, and it now appearsthat the transmembrane helices move later-ally from the translocon channel into the sur-rounding lipid bilayer, either one at a timeor in pairs, depending on their hydropho-bicity and their ability to form stable helix-helix interactions. Furthermore, it appearsthat the molecular features that allow thetranslocon to recognize a stretch of a polypep-tide in transit as a transmembrane helix arethe same as those seen to mediate protein-lipid interactions in the known membraneprotein structures (15), strongly suggestingthat the translocon is designed such that it al-lows a translocating nascent chain to samplethe surrounding bilayer. At its simplest, trans-membrane helix insertion may thus be ap-proximated as a thermodynamic partitioningbetween the aqueous milieu in the transloconchannel and the lipid membrane.
The β-barrel proteins in the bacterialouter membrane are also translocated throughthe inner membrane translocon, but they doso posttranslationally with the aid of the SecAATPase, and their short transmembrane β-strands are not sufficiently hydrophobic to getstuck across the inner membrane (16). Instead,they are chaperoned through the periplas-mic space and finally insert into the outer
www.annualreviews.org • Membrane Protein Structure Prediction 127
Ann
u. R
ev. B
ioch
em. 2
007.
76:1
25-1
40. D
ownl
oade
d fr
om w
ww
.ann
ualr
evie
ws.
org
by N
OR
TH
CA
RO
LIN
A S
TA
TE
UN
IVE
RSI
TY
on
09/2
9/12
. For
per
sona
l use
onl
y.
ANRV313-BI76-06 ARI 24 May 2007 23:6
Figure 1Biogenesis of α-helix bundle (left) and β-barrel (right) membrane proteins in Escherichia coli.
membrane with the aid of the resident YaeThetero-oligomeric outer membrane integra-tion complex (14).
Folding and Stability
Once inserted into the membrane, transmem-brane helices pack into the typical helix-bundle folds, and many then go on toform homo- or hetero-oligomeric complexes.Membrane proteins form closely packedstructures, and it is believed that an im-portant driving force for folding is bettershape complementarity between the trans-membrane helices than between the helicesand the lipid (17). Other factors that come
into play are hydrogen bonding between po-lar side chains (18) and possibly even the for-mation of Cα−H−O hydrogen bonds (19, 20;but also see Reference 21).
Many membrane proteins, both helix bun-dle and β-barrel, form stable structures withlittle flexibility, whereas others undergo sub-stantial rearrangements of their transmem-brane domains as part of a reaction cycle. Pro-teins involved in proton and electron transfertypically coordinate a range of cofactors thatneed to be positioned relative to each otherwith A-level precision and hence must be quiterigidly packed (22), whereas small-moleculetransporters must flip between dramaticallydifferent conformations open either toward
128 Elofsson · von Heijne
Ann
u. R
ev. B
ioch
em. 2
007.
76:1
25-1
40. D
ownl
oade
d fr
om w
ww
.ann
ualr
evie
ws.
org
by N
OR
TH
CA
RO
LIN
A S
TA
TE
UN
IVE
RSI
TY
on
09/2
9/12
. For
per
sona
l use
onl
y.
ANRV313-BI76-06 ARI 30 April 2007 17:26
the external or the internal side of the cell(23, 24).
MEMBRANE PROTEINBIOINFORMATICS: WHAT THESEQUENCES TELL
For the helix-bundle membrane proteins,amino acid sequences told their story long be-fore the first high-resolution structures weredetermined: the typical transmembrane seg-ment is formed by a stretch of predominantlyhydrophobic residues long enough to spanthe lipid bilayer as an α-helix (25–29). Theearly topology prediction methods were con-sequently little more than plots of the seg-mental hydrophobicity (averaged over 10–20 residues) along the sequence (30–32). Withmore sequences came the realizations thataromatic Trp and Tyr residues tend to clus-ter near the ends of the transmembrane seg-ments (10, 33) and that the loops connectingthe helices differ in amino acid composition,depending on whether they face the inside oroutside of the cell (the “positive-inside” rule)(34–36). More recent analyses have focusedon the higher-than-random appearance of se-quence motifs, such as the GxxxG-motif intransmembrane segments (37, 38) as well asother periodic patterns within the membranehelices (39), with the aim of providing infor-mation that may help in predicting helix-helixpacking and 3D structure.
MEMBRANE PROTEINBIOINFORMATICS: WHAT THESTRUCTURES TELL
For a long time, the general view has beenthat membrane proteins form simple he-lix bundles, with their transmembrane he-lices crisscrossing the membrane in more orless perpendicular orientations. Indeed, manymembrane proteins abide by this principle.However, some more recently solved mem-brane protein structures show that reality isnot always this simple. This is illustrated bythe structure of the glutamate transporter ho-
Figure 2(a) The glutamate transporter homolog (1XFH) contains both disruptedtransmembrane helices and reentrant loops. Disrupted helices are shown(cyan and green), and reentrant loops are also shown. The mesh indicates theapproximate extent of the lipid tail region ( ± 15 A). (b) Topology (upperpart) and z-coordinate plot. The z-coordinate plot shows the distance fromthe center of the membrane for each residue. The coloring is the same as inpanel a. Modified with permission of Oxford University Press (79).
molog from Pyrococcus horikoshii (40), shown inFigure 2. This protein has six typical trans-membrane helices and two irregular heliceswith breaks inside the lipid bilayer. The struc-ture also contains two reentrant loops thatgo only halfway through the membrane andthen turn back to the side from which they
www.annualreviews.org • Membrane Protein Structure Prediction 129
Ann
u. R
ev. B
ioch
em. 2
007.
76:1
25-1
40. D
ownl
oade
d fr
om w
ww
.ann
ualr
evie
ws.
org
by N
OR
TH
CA
RO
LIN
A S
TA
TE
UN
IVE
RSI
TY
on
09/2
9/12
. For
per
sona
l use
onl
y.
ANRV313-BI76-06 ARI 30 April 2007 17:26
originate. The two reentrant loops meet inthe middle of the membrane, a feature alsoseen in the aquaporin structures (41).
Other structural elements, largely ignoreduntil recently in statistical studies of mem-brane protein structure, are found in thoseparts of the protein that are located in themembrane-water interface region. Here onefinds irregular structure and interfacial helicesrunning roughly parallel to the membranesurface, while β-strands are extremely rare(42–44). The average amino acid compositionis different between the interfacial helices, theparts of the transmembrane helices located inthe interface region, and the irregular struc-tures. Hydrophobic and aromatic residues inthis region tend to point toward the centerof the membrane, whereas charged and polarresidues tend to point away from the mem-brane. The interface region thus imposes dif-ferent constraints on protein structure thando the central hydrocarbon core of the mem-brane and the surrounding aqueous phase.
For β-barrel membrane proteins, a num-ber of structural rules have been deduced fromthe known structures (45): The number of β-strands is even; the N and C termini are atthe periplasmic barrel end; the β-strand tilt is∼45◦; all β-strands are antiparallel and con-nected locally to their next neighbors alongthe chain; and the β-barrel surface, contact-ing the nonpolar membrane interior, consistsof a belt of aliphatic side chains lined by twogirdles of aromatic side chains. These gener-alizations provide a framework for the devel-opment of topology prediction algorithms.
MEMBRANE PROTEINSTRUCTURE PREDICTION:FROM 2D TO 2.5D AND 3D
Over the years, many topology and structureprediction schemes have been developed forhelix-bundle membrane proteins. In general,there has been a logical progression from sim-ple to more complicated models. This hasbeen enabled by the increase in available train-ing data, a better understanding of membrane
protein structure, and advances in machine-learning methods.
2D Predictions
The earliest topology prediction methods re-lied only on the fact that transmembrane he-lices are on average more hydrophobic thanloop regions. Although these simple methodsworked surprisingly well, they left a lot to bedesired. Inclusion of the positive-inside ruleled to a significant improvement in the predic-tions (46), and a further step was taken whenhidden Markov models (HMMs), Figure 3,and other machine-learning techniques wereemployed to extract the relevant sequence fea-tures (5, 47–50). In addition, HMM-basedmethods that also include evolutionary and/orlimited experimental information to improvetopology predictions have been developed(51–53).
One particular problem faced by all topol-ogy predictors is to discriminate between sig-nal peptides and transmembrane helices—twokinds of topogenic elements that look quitesimilar. This problem has recently been ad-dressed with the development of Phobius (54),an HMM-based method that predicts bothsignal peptides and transmembrane segmentssimultaneously and thereby significantly de-creases the confusion between them.
Much less work has been devoted to the de-velopment of topology prediction schemes forβ-barrel membrane proteins, in part becausethe membrane-spanning β-strands are bothconsiderably shorter and much less conspicu-ous in terms of amino acid sequence than thelong, hydrophobic transmembrane helices inthe helix-bundle proteins. The simplest meth-ods attempt to identify bacterial outer mem-brane β-barrel proteins using only two cri-teria: the presence of an N-terminal signalpeptide [predicted using a program such asSignalP (55, 56)] and the overall amino acidcomposition of the protein (9, 57). The moreadvanced methods also predict the individualβ-strands and the topology of the protein, inmost cases using HMMs (58–60).
130 Elofsson · von Heijne
Ann
u. R
ev. B
ioch
em. 2
007.
76:1
25-1
40. D
ownl
oade
d fr
om w
ww
.ann
ualr
evie
ws.
org
by N
OR
TH
CA
RO
LIN
A S
TA
TE
UN
IVE
RSI
TY
on
09/2
9/12
. For
per
sona
l use
onl
y.
ANRV313-BI76-06 ARI 30 April 2007 17:26
Figure 3Typical hiddenMarkov models(HMMs) fortopology predictionsof (a) helix-bundlemembrane proteins(5) and (b) β-barrelmembrane proteins.Modified withpermission of OxfordUniversity Press(58).
Benchmarking
The topology of more than 400 helix-bundlemembrane proteins has been determined ex-perimentally by a variety of genetic, biochem-ical, and structural techniques, and such datahas been used both to train and to bench-mark the various topology prediction meth-ods. However, much of the experimental dataare of quite low resolution and not always cor-rect. In addition, when benchmarking differ-ent methods, many diverse quality measureshave been used. In principle, these can be di-vided into residue-based and topology-basedinstruments. In our experience, benchmark-ing with residue-based measures is not opti-mal as (a) all methods perform more or lessequally well by such measures, and (b) mostexperimental topology information is not ofhigh enough quality to exactly define the bor-ders between membrane and nonmembraneparts of the proteins.
The different data sets used in bench-marking studies can also impact the results.An especially important point is if single-spanning membrane proteins are included ornot. On average, transmembrane helices insingle-spanning proteins are more hydropho-bic than in multispanning (polytopic) mem-brane proteins (A. Bernsel & G. von Heijne,unpublished data). Including single-spanningproteins in training data may compromisethe performance on multispanning proteinsand vice versa. Ideally, single-spanning pro-teins should, therefore, be treated differentlythan multispanning proteins. However, to ourknowledge no method has successfully in-cluded this distinction in a prediction scheme,and it is rarely taken into account in bench-marking studies.
Given these caveats, it is nevertheless clearfrom several recent benchmarking studies(61–63) that HMM-based methods that also
www.annualreviews.org • Membrane Protein Structure Prediction 131
Ann
u. R
ev. B
ioch
em. 2
007.
76:1
25-1
40. D
ownl
oade
d fr
om w
ww
.ann
ualr
evie
ws.
org
by N
OR
TH
CA
RO
LIN
A S
TA
TE
UN
IVE
RSI
TY
on
09/2
9/12
. For
per
sona
l use
onl
y.
ANRV313-BI76-06 ARI 30 April 2007 17:26
include evolutionary information, e.g., poly-Phobius (64), HMMTOP (50), and prodiv-TMHMM (52), perform best. These methodspredict the correct topology (i.e., the correctnumber of transmembrane helices and thecorrect overall orientation of the protein inthe membrane) for close to 70% of all mem-brane proteins, and this number can be im-proved even further when additional experi-mental information is available to constrainthe predictions (51, 53, 65, 66).
A recent benchmarking of β-barrel mem-brane protein topology predictors concludedthat HMM-based methods also perform bet-ter than other methods for these proteins (67).The best HMM methods predicted the cor-rect topology for 14 out of 20 test proteins(70%), although the real accuracy is likelylower than this because many of the proteinsin the test set were used also to develop thedifferent methods.
It should be kept in mind that a correcttopology prediction does not mean that thepredicted starts and ends of the transmem-brane α-helices or β-strands can be trusted;only the number of transmembrane helicesand their approximate positions are correct. Infact, this part of the structure prediction prob-lem has not yet been satisfactorily resolved.With the availability of more exact structuralinformation, it should be possible to evalu-ate how well different methods can predictthe exact helix locations. Our experience hasshown that it is possible to predict the pointsof entrance and exit from the membrane en-vironment with acceptable accuracy, whereasthe prediction of helix start and terminationpoints is very hard.
The rapid increase in high-resolutionstructural data for membrane proteins meansthat in the future both benchmarking anddevelopment of novel prediction methodsshould be based on structural data only.Luckily, the general conclusions from usingstructure-based benchmarks are similarto those of earlier studies (52, 68). Oneremaining problem with the structural datais that the precise location of the protein in
the lipid membrane is often not immediatelyavailable because crystals are grown fromdetergent-solubilized proteins. Automaticmethods that optimize the fit between a 3Dprotein structure and a model membrane areavailable (69, 70), although it is difficult toassess their accuracy.
Genome Annotation
An important application of topology predic-tion algorithms is to annotate genome se-quencing data. It has been reported that al-gorithms such as TMHMM can discriminatehelix-bundle proteins from other proteinswith better than 95% sensitivity and speci-ficity (5), meaning that the helix-bundle mem-brane proteome of an organism can be quitereliably predicted from its genome sequence.
It was initially observed that the distri-bution of helix-bundle membrane proteintopologies in a genome seemed to follow apower law with respect to the number oftransmembrane helices, i.e., that proteins withfew transmembrane helices are more frequentthan proteins with many transmembrane he-lices (71–73). As the topology predictors im-proved, several exceptions to this generaltrend were noted (5, 65, 66, 74). In particular,bacterial genomes encode large numbers ofsmall-molecule transporters with 6 or (moreoften) 12 transmembrane helices, whereasmammalian genomes are strongly enrichedfor G protein–coupled receptors (GPCRs)with 7 transmembrane helices, as well asfor small-molecule transporters. With theavailability of more accurate predictors andgenome-wide experimental topology data, itwas also noted that there is a strong overrep-resentation of proteins with an even numberof transmembrane helices and with their Nand C termini located on the cytoplasmic sideof membrane, both in bacteria and eukaryotes(5, 65, 66).
2.5D Predictions
As noted above, the general view, until re-cently, has been that the basic structural
132 Elofsson · von Heijne
Ann
u. R
ev. B
ioch
em. 2
007.
76:1
25-1
40. D
ownl
oade
d fr
om w
ww
.ann
ualr
evie
ws.
org
by N
OR
TH
CA
RO
LIN
A S
TA
TE
UN
IVE
RSI
TY
on
09/2
9/12
. For
per
sona
l use
onl
y.
ANRV313-BI76-06 ARI 30 April 2007 17:26
feature of helix-bundle membrane proteinsis the perpendicularly penetrating transmem-brane helices. But as many 3D structures nowshow, membrane protein structures are of-ten too complex to fit completely into sucha simple topology model. To advance further,a more fine-grained definition of topology,“a two-and-a-half dimensional (2.5D) struc-ture,” is needed where structural elements,such as interfacial helices and reentrant loops,are taken into account. In addition, severalother limitations of the current generation ofpredictors exist, e.g., it has been noted thatsuch a trivial characteristic as the exact lengthof transmembrane helices is very difficult topredict using current methodologies.
Reentrant loops are a common feature inmany membrane proteins. They were firstseen in the aquaporin-1 water channel (75)and in the KcsA potassium channel (76). De-tailed analysis suggests that reentrant loopscan be divided into three distinct categoriesbased on secondary structure content: longloops with a helix-coil-helix structure, loops ofmedium length with a helix-coil or coil-helixstructure, and loops of short to medium lengthconsisting entirely of an irregular secondarystructure (77). Residues in reentrant loops aresignificantly smaller on average compared toother parts of the protein, and they can be de-tected in regions between the transmembranehelices with ∼70% accuracy based on theiramino acid composition. Reentrant loops of-ten contain particular functional motifs thatenable them to be detected (78). On the ba-sis of a novel predictor for reentrant loops(TOP-MOD), it appears that more than 10%of all multispanning membrane proteins con-tain such loops (77). Reentrant loops seem tobe most commonly found in ion and waterchannel proteins and least commonly in cellsurface receptors.
Although the division of a membrane pro-tein into different substructures is clearly use-ful, distinguishing different types of struc-tural elements is not always straighforward.Reentrant loops can vary quite dramatically intheir secondary structure and depth of pen-
etration into the membrane, and the lengthof transmembrane helices varies significantly.An alternative approach to membrane protein2.5D structure prediction is to directly predictthe distance from the center of the membrane(i.e., the z-coordinate) for each residue in aprotein, rather than the type of structural el-ement of which it is a part. One recent al-gorithm of this kind correctly classified 88%of all residues in the test set proteins to beinside or outside the membrane, with an aver-age error of 2.5 A in the predicted residue dis-tances from the center of the membrane (79).A similar z-coordinate predictor has also beendeveloped for β-barrel membrane proteins(80).
An important characteristic of residues intransmembrane helices is their degree of lipidexposure in the folded structure. In contrastto globular proteins, membrane proteins donot show a large difference in hydrophobicitybetween the lipid-exposed and buried residuesin the membrane-embedded region, and theprediction of surface accessibility becomesmuch harder. The major features distinguish-ing the lipid-exposed and buried residues arethe polarity of the side chain (more hydropho-bic residues tend to be more lipid exposed) andthe degree of sequence conservation (less con-served residues tend to be more lipid exposed).Many attempts to predict lipid exposure havebeen published; one of the most recent studiesreports the prediction of lipid-exposed surfacepatches in transmembrane helices that inter-face with lipid molecules with a per residueaccuracy of 88% (81).
A final 2.5D characteristic that can be pre-dicted with reasonable accuracy is the pres-ence of proline-induced kinks in the trans-membrane helices. Interestingly, such kinkscan be preserved even when the Pro residue ismutated (82, 83), and a kink can confidentlybe predicted if proline is conserved in a par-ticular position in a transmembrane helix inmore than 10% of the sequences in a multiplealignment (83).
The prediction of 2.5D features of mem-brane proteins should not only be useful as a
www.annualreviews.org • Membrane Protein Structure Prediction 133
Ann
u. R
ev. B
ioch
em. 2
007.
76:1
25-1
40. D
ownl
oade
d fr
om w
ww
.ann
ualr
evie
ws.
org
by N
OR
TH
CA
RO
LIN
A S
TA
TE
UN
IVE
RSI
TY
on
09/2
9/12
. For
per
sona
l use
onl
y.
ANRV313-BI76-06 ARI 30 April 2007 17:26
step toward 3D predictions. Such predictorswill also help in the classification of mem-brane protein families because the differentsubstructures provide unique sequence signa-tures separating different membrane proteinfamilies with the same topology. In addition,the identification of suitable peptide antibodyepitopes may be facilitated if z-coordinatescan be accurately predicted.
3D Predictions
Interestingly, 3D structure predictions ofmembrane proteins were attempted even be-fore the first high-resolution structure of anymembrane proteins was solved. Using infor-mation from low-resolution experiments, inparticular electron microscopy, quite accuratemodels of bacteriorhodopsin (84) as well asGPCRs (85, 86) were made.
Despite these early successes, the field ofnonhomology-based 3D structure predictionfor membrane proteins has followed a simi-lar trend as that seen in the globular proteinstructure prediction field, wherein the generalexperience is that most methods when testedin blind predictions show a much lower ac-curacy than first reported. However, throughrounds of iterative refinement, the best meth-ods can now predict the structure of smallglobular proteins quite accurately (87). As itturns out, one of the most important improve-ments to the methodology has been to base the3D prediction on short sequence fragmentsextracted from known protein structures (88–90).
It is, however, not straightforward toapply similar schemes to membrane pro-teins because the different environmentintroduced by the membrane has to be mod-eled in some way, and because most mem-brane proteins are significantly larger thanthe globular proteins successfully predicted sofar. Therefore the success to date has beenquite limited even using the most advancedmethods adapted from the globular proteinfield (91, 92). If experimentally derived dis-
tance constraints from techniques, such asFourier transform infrared spectroscopy, elec-tron paramagnetic resonance spectroscopy,and chemical cross-linking, or low-resolutionmodels based on electron microscopy, areavailable, more reliable models can be built(93).
An interesting attempt to model allGPCRs of the human proteome was recentlymade by Skolnick and coworkers (94) usingthe TASSER algorithm. Although the accu-racy of the predicted rhodopsin structure wasquite good, the correctness of the GPCRstructures can not be verified until more struc-tures are available.
Many membrane proteins, in particularchannels and transporters, undergo substan-tial structural changes during a reaction cy-cle. Often the structure of only one of thestates is available, and methodology to reli-ably model structural changes will be neededfor a long time to come. In a recent study, theROSETTA membrane folding algorithm wasused to model the closed and open states ofa voltage-dependent potassium channel (95),generating a number of testable hypothesesthat may guide further experimental work.For a more thorough review of ab initiostructure prediction methods, see Reference96.
While ab initio structure modeling can atbest predict the overall fold of a protein, struc-ture modeling based on a preexisting struc-ture of a close homologue promises atomic-level structural detail. Homology modelingof membrane proteins is still in its infancy,however, because so few structures are known.A recent benchmarking study suggests that,when a template is available, homology mod-els of membrane proteins are comparable inquality to those that can be made for globularproteins; i.e., when the sequence identity be-tween the template and the target is >30%,one can expect the root mean-square devia-tion between the modeled and correct struc-ture to be less than 2 A in the transmembraneregions (97).
Hierarchically organized databases of pro-tein structure have found many uses, bothin studies of protein evolution, as a meansto put together nonredundant sequence andstructure collections for statistical studies,and as test beds for benchmarking foldrecognition and structure prediction algo-rithms. For globular proteins, several well-established hierarchical, structure-based do-main classification schemes, such as SCOP(98) and CATH (99), exist. However, thenumber of membrane protein structuresis still too low for such classifications tocover an important part of the membraneprotein universe (4). All currently knownhigh-resolution membrane protein structuresare listed at http://blanco.biomol.uci.edu/MemPro resources.html and have alsobeen organized into a database at http://pdbtm.enzim.hu/ (100).
Sequence-based, nonhierarchical classi-fications of membrane protein domains areavailable in Pfam (101) and other similardatabases, and specialized databases col-lecting families of membrane transporters(http://www.tcdb.org/), GPCRs (http://www.gpcr.org/7tm/), and potassium channelproteins (http://www.receptors.org/KCN)have been set up. Because these databases arenot based on 3D structure information, theydo not contain complete information aboutvery distant evolutionary relationships (e.g.,the SCOP class, fold, and superfamily levels).
The detection of distantly related glob-ular proteins has seen great progress dur-ing the past decade through the use offold-recognition methods, improved use ofevolutionary relationships, and careful bench-marking. Because of the low incidence of po-lar and charged residues in transmembranehelices, the use of algorithms optimized forglobular proteins for the detection of distantlyrelated membrane proteins is problematic. Tocompound these difficulties, a major obstacle
for the development of improved methods toidentify distantly related membrane proteinsis the lack of a structure-based “gold standard”such as SCOP.
Nevertheless, large, divergent membraneprotein families, such as the GPCRs, canbe used to benchmark fold-recognition algo-rithms for membrane proteins. As for glob-ular proteins, it appears that HMM-basedsequence family models, profile-profile simi-larity searches, and the inclusion of secondarystructure information in the form of predictedtopology models all help in the detection ofdistant homologues (102, 103). It has alsobeen shown that the best alignment of re-lated membrane proteins is obtained usingprofile-profile methods in combination withpredicted secondary structures (97).
PROTEIN-PROTEININTERACTIONS
The final step on the structure predictionladder is the prediction of quaternary struc-ture, i.e., protein-protein interactions. Thisis especially pertinent for membrane proteinsbecause membrane-integral protein domainsin most cases seem to be encoded by sep-arate polypeptides rather than as multido-main polypeptides as often found in globu-lar proteins (104). Large-scale experimentalprotein-protein interaction studies tend to ig-nore membrane proteins, although some dataare now starting to appear in the literature(105–107).
The current state of predicting interac-tions between membrane proteins may besummarized in a few words: much remains tobe done. For example, a recent attempt to pre-dict interacting proteins in the Saccharomycescerevisiae membrane proteome by integratingdata, such as amino acid sequence, annotatedfunction, subcellular localization, mRNA andprotein abundance, transcriptional coregula-tion, and gene knock-out phenotype, resultedin a predictor that could identify ∼40% of304 experimentally well-documented gold
www.annualreviews.org • Membrane Protein Structure Prediction 135
Ann
u. R
ev. B
ioch
em. 2
007.
76:1
25-1
40. D
ownl
oade
d fr
om w
ww
.ann
ualr
evie
ws.
org
by N
OR
TH
CA
RO
LIN
A S
TA
TE
UN
IVE
RSI
TY
on
09/2
9/12
. For
per
sona
l use
onl
y.
ANRV313-BI76-06 ARI 30 April 2007 17:26
standard interactions while minimizingthe number of false-positive predictions(108).
CONCLUSIONS AND OUTLOOK
If bioinformatics methods are evaluated byhow they are received by the scientific com-munity at large, it is clear that membrane pro-tein structure prediction algorithms hold theirground; to give but one example, TMHMM(5, 47) has been cited well over 1200 times. Butprecisely what sort of information can one ex-pect to get from the various prediction meth-ods? And what sort of advances can we see onthe horizon?
First and foremost, do not expect the com-puter to tell you the truth! Topology predic-tions are just predictions. True, high-scoringpredictions are nearly always right (51, 63,109), but this only means that the really clear-cut cases (i.e., those that can equally wellbe done by hand) are easy to predict. Still,if taken with a grain of salt, topology mod-els, predicted lipid-exposed residues, poten-tial reentrant loops, and lists of possibly in-teracting partner proteins can be invaluableguides for planning experiments and inter-preting results. And large-scale computational
studies of entire genomes can provide tanta-lizing clues to everything from the basic pat-terns of membrane protein evolution (110,111) to differences in lifestyle between differ-ent organisms—show me your transporters,and I will tell you where you live.
Still, much remains to be done, both inperfecting the current arsenal of predictionmethods and devising entirely new algorithmsto do new things. Our current representationof membrane protein topology as a simplestring of membrane-spanning α-helices or β-strands does not fully capture the structuraldiversity seen in membrane proteins; defin-ing a fuzzy area between the 2D and 3Dstructure is in need of more exploration. Therapid growth in known membrane protein 3Dstructures improves the prospects for effec-tive fold-recognition and homology model-ing approaches, although the day when mostof membrane protein fold space has beenmapped experimentally seems desperately faroff (4). Computational means to map out themembrane interactome will become an im-portant complement to high-throughput (buterror-prone) experimental studies, and here,as in so many other areas, tight integrationbetween the “wet” and “dry” approaches iscertainly the best way forward.
SUMMARY POINTS
1. Integral membrane proteins come in two basic architectures: α-helix bundles andβ-barrels.
2. The lipid-facing surface of integral membrane proteins is composed of a central“hydrophobic belt” flanked by two “aromatic girdles.”
3. In the helix-bundle proteins, nontranslocated loops are enriched in Lys and Arg com-pared to translocated loops (the positive-inside rule).
4. Helix-bundle membrane proteins are built from transmembrane α-helices, interfacialhelices lying flat on the membrane, reentrant loops, and extramembraneous globulardomains.
5. For the β-barrel protein, the number of β-strands is even, the N and C termini are atthe periplasmic barrel end, the β-strand tilt is ∼45◦, and all β-strands are antiparalleland connected locally to their next neighbors along the chain.
136 Elofsson · von Heijne
Ann
u. R
ev. B
ioch
em. 2
007.
76:1
25-1
40. D
ownl
oade
d fr
om w
ww
.ann
ualr
evie
ws.
org
by N
OR
TH
CA
RO
LIN
A S
TA
TE
UN
IVE
RSI
TY
on
09/2
9/12
. For
per
sona
l use
onl
y.
ANRV313-BI76-06 ARI 30 April 2007 17:26
6. The best topology prediction algorithms forecast the correct topology for ≤70%of all proteins but cannot accurately predict the start and end of a transmembranesegment.
7. Only a few recent prediction algorithms attempt to identify surface helices and reen-trant loops.
8. Ab initio high-resolution 3D structure prediction is still not feasible for membraneproteins. Homology-based structure modeling of membrane proteins performs on apar with homology modeling of globular proteins.
ACKNOWLEDGMENTS
The authors’ laboratories are supported by grants from the Swedish Foundation for StrategicResearch, the Marianne and Marcus Wallenberg Foundation, the Swedish Cancer Founda-tion, the Swedish Research Council, and the European Commission (BioSapiens, Genefun,EMBRACE).
28:235–423. White SH. 2004. Protein Sci. 13:1948–494. Oberai A, Ihm Y, Kim S, Bowie JU. 2006. Protein Sci. 15:1723–345. Krogh A, Larsson B, von Heijne G, Sonnhammer EL. 2001. J. Mol. Biol. 305:567–806. Keseler IM, Collado-Vides J, Gama-Castro S, Ingraham J, Paley S, et al. 2005. Nucleic
Acids Res. 33:D334–377. Wimley WC. 2002. Protein Sci. 11:301–128. Casadio R, Fariselli P, Finocchiaro G, Martelli PL. 2003. Protein Sci. 12:1158–689. Garrow AG, Agnew A, Westhead DR. 2005. BMC Bioinformatics 6:56
10. Wallin E, Tsukihara T, Yoshikawa S, von Heijne G, Elofsson A. 1997. Protein Sci. 6:808–1511. Seshadri K, Garemyr R, Wallin E, von Heijne G, Elofsson A. 1998. Protein Sci. 7:2026–3212. Ulmschneider MB, Sansom MS, Di Nola A. 2005. Proteins 59:252–6513. Luirink J, von Heijne G, Houben E, de Gier JW. 2005. Annu. Rev. Microbiol. 59:329–5514. Ruiz N, Kahne D, Silhavy TJ. 2006. Nat. Rev. Microbiol. 4:57–6615. White SH, von Heijne G. 2004. Curr. Opin. Struct. Biol. 14:397–40416. MacIntyre S, Freudl R, Eschbach ML, Henning U. 1988. J. Biol. Chem. 263:19053–5917. Popot J-L, Engelman DM. 2000. Annu. Rev. Biochem. 69:881–92218. Curran AR, Engelman DM. 2003. Curr. Opin. Struct. Biol. 13:412–1719. Senes A, Ubarretxena-Belandia I, Engelman DM. 2001. Proc. Natl. Acad. Sci. USA
98:9056–6120. Arbely E, Arkin IT. 2004. J. Am. Chem. Soc. 126:5362–6321. Yohannan S, Faham S, Yang D, Grosfeld D, Chamberlain AK, Bowie JU. 2004. J. Am.
www.annualreviews.org • Membrane Protein Structure Prediction 137
Ann
u. R
ev. B
ioch
em. 2
007.
76:1
25-1
40. D
ownl
oade
d fr
om w
ww
.ann
ualr
evie
ws.
org
by N
OR
TH
CA
RO
LIN
A S
TA
TE
UN
IVE
RSI
TY
on
09/2
9/12
. For
per
sona
l use
onl
y.
ANRV313-BI76-06 ARI 30 April 2007 17:26
24. Toyoshima C, Nomura H, Tsuda T. 2004. Nature 432:361–6825. Henderson R, Unwin PNT. 1975. Nature 257:28–3226. Tomita M, Marchesi VT. 1975. Proc. Natl. Acad. Sci. USA 72:2964–6827. Ovchinnikov YA, Abdulaev NG, Feigina MY, Kiselev AV, Lobanov NA. 1977. FEBS Lett.
84:1–428. Khorana HG, Gerber GE, Herlihy WC, Gray CP, Anderegg RJ, et al. 1979. Proc. Natl.
Acad. Sci. USA 76:5046–5029. Ovchinnikov YA, Abdulaev NG, Feigina MY, Kiselev AV, Lobanov NA. 1979. FEBS Lett.
100:219–2430. von Heijne G, Blomberg C. 1979. Eur. J. Biochem. 97:175–8131. Engelman DM, Steitz TA. 1981. Cell 23:411–2232. Kyte J, Doolittle RF. 1982. J. Mol. Biol. 157:105–3233. Weiss MS, Kreusch A, Schiltz E, Nestel U, Welte W, et al. 1991. FEBS Lett. 280:379–8234. von Heijne G. 1986. J. Mol. Biol. 189:239–4235. von Heijne G. 1986. EMBO J. 5:3021–2736. von Heijne G. 1989. Nature 341:456–5837. Senes A, Gerstein M, Engelman DM. 2000. J. Mol. Biol. 296:921–3638. Kim S, Jeon TJ, Oberai A, Yang D, Schmidt JJ, Bowie JU. 2005. Proc. Natl. Acad. Sci.
USA 102:14278–8339. Samatey FA, Xu C, Popot J-L. 1995. Proc. Natl. Acad. Sci. USA 92:4577–8140. Yernool D, Boudker O, Jin Y, Gouaux E. 2004. Nature 431:811–1841. Hedfalk K, Tornroth-Horsefield S, Nyblom M, Johanson U, Kjellbom P, Neutze R.
2006. Curr. Opin. Struct. Biol. 16:447–5642. Orgel JPRO. 2004. J. Struct. Biol. 148:51–6543. Granseth E, von Heijne G, Elofsson A. 2005. J. Mol. Biol. 346:377–8544. Liang J, Adamian L, Jackups RJ. 2005. Trends Biochem. Sci. 30:355–5745. Schulz GE. 2000. Curr. Opin. Struct. Biol. 10:443–4746. von Heijne G. 1992. J. Mol. Biol. 225:487–9447. Sonnhammer ELL, von Heijne G, Krogh A. 1998. A hidden Markov model for predicting
transmembrane helices in protein sequences. Proc. Int. Conf. Intell. Syst. Mol. Biol. 6:175–8248. Jones DT, Taylor WR, Thornton JM. 1994. Biochemistry 33:3038–4949. Tusnady GE, Simon I. 1998. J. Mol. Biol. 283:489–50650. Tusnady GE, Simon I. 2001. Bioinformatics 17:849–5051. Melen K, Krogh A, von Heijne G. 2003. J. Mol. Biol. 327:735–4452. Viklund H, Elofsson A. 2004. Protein Sci. 13:1908–1753. Bernsel A, von Heijne G. 2005. Protein Sci. 14:1723–2854. Kall L, Krogh A, Sonnhammer ELL. 2004. J. Mol. Biol. 338:1027–3655. Nielsen H, Engelbrecht J, Brunak S, von Heijne G. 1997. Protein Eng. 10:1–656. Dyrløv-Bendtsen J, Nielsen H, von Heijne G, Brunak S. 2004. J. Mol. Biol. 340:783–9557. Nakai K, Horton P. 1999. Trends Biochem. Sci. 24:34–3558. Martelli PL, Fariselli P, Krogh A, Casadio R. 2002. Bioinformatics 18(Suppl. 1):S46–5359. Bagos PG, Liakopoulos TD, Spyropoulos IC, Hamodrakas SJ. 2004. Nucleic Acids Res.
32:W400–460. Bigelow HR, Petrey DS, Liu J, Przybylski D, Rost B. 2004. Nucleic Acids Res. 32:2566–7761. Chen CP, Kernytsky A, Rost B. 2002. Protein Sci. 11:2774–9162. Chen CP, Rost B. 2002. Protein Sci. 11:2766–7363. Kall L, Sonnhammer ELL. 2002. FEBS Lett. 532:415–1864. Kall L, Krogh A, Sonnhammer ELL. 2005. Bioinformatics 21(Suppl. 1):i251–57
138 Elofsson · von Heijne
Ann
u. R
ev. B
ioch
em. 2
007.
76:1
25-1
40. D
ownl
oade
d fr
om w
ww
.ann
ualr
evie
ws.
org
by N
OR
TH
CA
RO
LIN
A S
TA
TE
UN
IVE
RSI
TY
on
09/2
9/12
. For
per
sona
l use
onl
y.
ANRV313-BI76-06 ARI 30 April 2007 17:26
65. Daley DO, Rapp M, Granseth E, Melen K, Drew D, von Heijne G. 2005. Science308:1321–23
66. Kim H, Osterberg M, Melen K, von Heijne G. 2006. Proc. Natl. Acad. Sci. USA 103:11142–47
67. Bagos PG, Liakopoulos TD, Hamodrakas SJ. 2005. BMC Bioinformatics 6:768. Cuthbertson JM, Doyle DA, Sansom MS. 2005. Protein Eng. Des. Sel. 18:295–30869. Tusnady GE, Dosztanyi Z, Simon I. 2005. Bioinformatics 21:1276–7770. Lomize AL, Pogozheva ID, Mosberg HI. 2004. Protein Sci. 13:2600–1271. Wallin E, von Heijne G. 1998. Protein Sci. 7:1029–3872. Liu J, Rost B. 2001. Protein Sci. 10:1970–7973. Gerstein M. 1998. Proteins: Struct. Funct. Genet. 33:518–3474. Lehnert U, Xia Y, Royce TE, Goh CS, Liu Y, et al. 2004. Q. Rev. Biophys. 37:121–4675. Walz T, Hirai T, Murata K, Heymann J, Mitsuoka K, et al. 1997. Nature 387:624–2776. Doyle D, Cabral J, Pfuetzner R, Kuo A, Gulbis J, et al. 1998. Science 280:69–7777. Viklund H, Granseth E, Elofsson A. 2006. J. Mol. Biol. 361:591–60378. Lasso G, Antoniw JF, Mullins JGL. 2006. Bioinformatics 22:e290–9779. Granseth E, Viklund H, Elofsson A. 2006. Bioinformatics 22:e191–9680. Diederichs K, Freigang J, Umhau S, Zeth K, Breed J. 1998. Protein Sci. 7:2413–2081. Adamian L, Liang J. 2006. BMC Struct. Biol. 6:1382. von Heijne G. 1991. J. Mol. Biol. 218:499–50383. Yohannan S, Faham S, Yang D, Whitelegge JP, Bowie JU. 2004. Proc. Natl. Acad. Sci.
USA 101:959–6384. Baldwin JM. 1993. EMBO J. 12:1693–70385. Baldwin JM, Schertler GF, Unger VM. 1997. J. Mol. Biol. 272:144–6486. Unger VM, Hargrave PA, Baldwin JM, Schertler GF. 1997. Nature 389:203–887. Bradley P, Misura KM, Baker D. 2005. Science 309:1868–7188. Jones TA, Thirup S. 1986. EMBO J. 5:819–2289. Simons KT, Bonneau R, Ruczinski I, Baker D. 1999. Proteins 3(Suppl.):171–7690. Bowie JU, Eisenberg D. 1994. Proc. Natl. Acad. Sci. USA 91:4436–4091. Pellegrini-Calace M, Carotti A, Jones DT. 2003. Proteins 50:537–4592. Yarov-Yarovoy V, Schonbrun J, Baker D. 2006. Proteins 62:1010–2593. Fleishman SJ, Unger VM, Ben-Tal N. 2006. Trends Biochem. Sci. 31:106–1394. Zhang Y, Devries ME, Skolnick J. 2006. PLoS Comput. Biol. 2:e1395. Yarov-Yarovoy V, Baker D, Catterall WA. 2006. Proc. Natl. Acad. Sci. USA 103:7292–9796. Fleishman SJ, Ben-Tal N. 2006. Curr. Opin. Struct. Biol. 16:496–50497. Forrest LR, Tang CL, Honig B. 2006. Biophys. J. 91:508–1798. Murzin AG, Brenner SE, Hubbard T, Chothia C. 1995. J. Mol. Biol. 247:536–4099. Orengo CA, Michie AD, Jones S, Jones DT, Swindells MB, Thornton JM. 1997. Structure
5:1093–108100. Tusnady GE, Dosztanyi Z, Simon I. 2005. Nucleic Acids Res. 33:D275–78101. Bateman A, Coin L, Durbin R, Finn RD, Hollich V, et al. 2004. Nucleic Acids Res. 32:D138–
41102. Hedman M, DeLoof H, von Heijne G, Elofsson A. 2002. Protein Sci. 11:652–58103. Wistrand M, Kall L, Sonnhammer EL. 2006. Protein Sci. 15:509–21104. Liu Y, Gerstein M, Engelman DM. 2004. Proc. Natl. Acad. Sci. USA 101:3495–97105. Gavin AC, Bosche M, Krause R, Grandi P, Marzioch M, et al. 2002. Nature 415:141–47106. Stenberg F, Chovanec P, Maslen SL, Robinson CV, Ilag LL, et al. 2005. J. Biol. Chem.
280:34409–19
www.annualreviews.org • Membrane Protein Structure Prediction 139
Ann
u. R
ev. B
ioch
em. 2
007.
76:1
25-1
40. D
ownl
oade
d fr
om w
ww
.ann
ualr
evie
ws.
org
by N
OR
TH
CA
RO
LIN
A S
TA
TE
UN
IVE
RSI
TY
on
09/2
9/12
. For
per
sona
l use
onl
y.
ANRV313-BI76-06 ARI 30 April 2007 17:26
107. Lasserre JP, Beyne E, Pyndiah S, Lapaillerie D, Claverol S, Bonneu M. 2006. Electrophore-sis 27:3306–21
108. Xia Y, Lu LJ, Gerstein M. 2006. J. Mol. Biol. 357:339–49109. Nilsson J, Persson B, von Heijne G. 2000. FEBS Lett. 486:267–69110. Shimizu T, Mitsuke H, Noto K, Arai M. 2004. J. Mol. Biol. 339:1–15111. Rapp M, Seppala S, Granseth E, von Heijne G. 2006. Nat. Struct. Mol. Biol. 13:112–16
An online log of corrections to Annual Review of Biochemistry chapters (if any, 1997to the present) may be found at http://biochem.annualreviews.org/errata.shtml
