Canonical section: 02_One_Dimensional_Prediction_of_Protein_Structure
English title: One-Dimensional Prediction of Protein Structure
PDF range: PDF page 206-220；在 PDF page 221 的 Predicting Protein Function 真实标题前停止
Boundary note: 本节包含 Synopsis、Secondary Structure and Solvent Accessibility、Performance Assessment of Secondary Structure Prediction、Transmembrane Alpha Helices and Beta Strands、Disordered Regions 等内部子标题；不将其作为 progress/README 的 peer entries 暴露。
Extraction note: PDF 文本存在明显单词粘连、断字与 caption/Box 排版噪声；以下原文保留页标记，供人工核对。

186 Predictive Methods Using Protein Sequences

One-Dimensional Prediction of Protein Structure

Synopsis

The 1D structure of a protein can be written as a simple string of characters representing

the set of natural amino acids – that is, the information content is one dimensional. While

more details on protein structure can be found in Chapter 12, in this chapter, we will

focus specifically on 1D prediction methods. Predictions of 1D features are relevant for two

reasons. First, features such as the number of membrane helices, the disorder in a protein,

or the surface residues are often important for protein function. We could determine 1D

structure from 3D structures if such structures were experimentally available but, given the

sequence–structure gap discussed above, experimental 3D structures are available for fewer

than 1% of all known sequences, while 1D predictions can be obtained for all 180 million

proteinsequencesknowntoday.Second,predictionsof1Dstructureareusedasaninputfor

most of the methods that will be described in the section about functional prediction that

follows. All of the features that will be described here are available from the PredictProtein

server,illustratedinFigure7.1andprovidingpre-computeddataonover20millionproteins

(Rostetal.2004;Kajanetal.2013;Yachdavetal.2014).

Figure 7 .1 Dashboard of the PredictProtein web server. PredictProtein (Yachdav et al. 2014) provides a centralized interface to many

methods that predict aspects of protein structure and function from sequence. Shown here is a sample of the dashboard for the pro-

tein picturesquely named Mothers against decapentaplegic homolog 7 (UniProtKB identiﬁer smad7_human). The black, numbered line in the

upper middle indicates the input amino acid sequence of length 428. Below follow predictions of different sequence-based tools, along

with a synopsis of the protein family. Predictions include protein–protein binding, protein–DNA/RNA binding, residue exposure (solvent

accessibility), secondary structure, and residue ﬂexibility; if found, the predictions also include membrane, long disordered, and coiled-coil

regions. Additional information is shown through mouse-over events, here illustrated through the beta strand prediction from the method

ReProf. T abs on the left give access to more detailed views of various predictions and analyses.

One-Dimensional Prediction of Protein Structure 187

Secondary Structure and Solvent Accessibility

Background Secondary structures are local macrostructures formed from short stretches of

aminoacidresiduesthatorganizethemselvesinspecificwaystoformtheoverall3Dstructure

ofaprotein.Physically,thedrivingforcebehindtheformationofsecondarystructuresisacom-

plexcombinationoflocalandglobalforces.Secondarystructurepredictionmethodsseekto

predictthesecondarystructureofashortproteinstretchbasedonthesumoftheseforces.For

instance,alphahelices arestabilizedbyhydrogenbondsbetweentheCOgroupofoneamino

acid and the NH group of the amino acid that is four positions C-terminal to it. Strands are

structuresinwhichthebackbonezigzagstocreateanextendedstructure.Themostcommon

amongtheseiscalledthe betastrand.Twoormorestretchesofbetastrandsofteninteractwith

eachother,throughhydrogenbondsformedbetweenthedifferentstrands,tocreateaplanar

structureknownasa betasheet.Structuresthatareneitherhelicesnorstrandsarereferredto

as“coils,”“others,”or“loops”(Figure7.2).

Predicting secondary structure is an important step in the experimental study of proteins

and toward inferring their function and evolution. Dozens of ideas, approaches, and meth-

ods for secondary structure prediction have been suggested over the last decades. They are

basedonvariousbiochemicalinsightsandawiderangeofcomputationaltechniques.Attheir

core, all methods identify patterns of amino acid frequencies that correlate with the forma-

tionofsecondarystructureelements.Thefirstapproachwasbasedjustontheoccurrenceof

prolines,whosestructuretendstointerrupthelices(Szent-GyörgyiandCohen1957).Further

work was based on the clustered occurrence of residues statistically identified to be either

helix- or strand-“former” residues (Chou and Fasman 1974). However, single amino acids

do not inherently contain sufficient information for reliable secondary structure prediction.

Therefore,thenextsetofmethodsthatweredevelopedtypicallyemployedaslidingwindow

approach in which the secondary structure prediction for a central residue is based on the

informationofthesurroundingresidues.Forexample,inaslidingwindowofsevenresidues,

theinputconsistsofthecentralresiduetogetherwiththethreeprecedingandthreefollowing

residues.Theinformationabouttheresiduesinthiswindowcanthenbeextractedinseveral

ways; for example, one could count how often a specific combination of seven residues was

found in known 3D structures, then use that information to predict the secondary structure

classthatwasmostoftenobservedintheknowndata.Thisistheapproachofoneoftheold-

estsecondarystructurepredictionmethods,calledGOR(Garnieretal.1978,1996).Acrucial

extensionofthisconceptwastheincorporationofevolutionaryinformation(RostandSander

Figure 7 .2 Protein secondary structure. Experimentally determined three-dimensional structure of alco-

hol dehydrogenase (UniProtKB identiﬁer ADHX_HUMAN) rendered in PyMOL (PDB structure 1m6h, chain

A; Schrodinger 2015). The protein is shown in a “ cartoon” view that highlights its secondary structure

elements: alpha helices in are shown in red, beta strands are depicted as yellow arrows, and all other

(“loops”) are colored in green.

188 Predictive Methods Using Protein Sequences

1993).Typically,thisisdonebyfindinghomologsofthequerysequenceinalargedatabase,

using position-specific iterated (PSI)-BLAST (Altschul and Gish 1996) or the more sensitive

HHblits, which uses hidden Markov models (HMMs; Box 7.1) to perform sequence profile

comparisons(Remmertetal.2012).Theresultinghitsarethenalignedinamultiplesequence

alignment (MSA; Chapter 8) that, in turn, is often represented as a position-specific scoring

matrix (PSSM; Chapter 3) or similar construct. As homologous proteins typically have simi-

larstructures,thesubstitutionpatternsfoundwithintheMSAcontainvaluableinformation

abouttheseproteins’secondarystructure.Conservedregionslikelymeanthatsecondarystruc-

tureispresentatthosepositions,thusaminoacidfrequenciescanbeweightedbytheoverall

conservationofresiduesateachpositionwithintheMSA.

B o x7 . 1 H i d d e nM a r k o vM o d e l s

Hidden Markov models (HMMs) provide a statistical representation of real biological pro-

cesses. In the context of gene prediction, HMMs describe information about the controlled

syntax of the structure of a gene. In the context of protein analysis, different syntactical

rules or “ grammars” are used for different applications. The ﬁrst and arguably the most

important application of HMMs for proteins is the creation of a reliable multiple sequence

alignment (MSA) for a given protein family.

Consider a simple MSA of length six:

Q-WKPG

Q-WKPG

Q-WRPG

QIWK-G

Q-WRPG

Q-WRPG

Some of the positions, such as the ﬁrst glutamine (Q), are absolutely conserved, while

position 4 is occupied only by positively charged residues. Position 2 shows a gap intro-

duced by an insertion in one sequence; position 5 shows a gap introduced by a deletion

in the same sequence.

Each of these observations can be represented by different states in the HMM. The

match state represents the most probable amino acid(s) found at each position of the

alignment; match is a bit of a misnomer, because the match state considers the proba-

bility of ﬁnding a given amino acid at that position of the alignment. If the position is

not absolutely conserved, the probabilities are adjusted accordingly, given the residues

found in that column. The insert state represents the case when an additional amino acid

needs to be accommodated, as it is here with the isoleucine (I) in sequence 4. The delete

state represents the case when an amino acid is missing and when the sequence (here,

sequence 4) must jump over a position to continue the alignment.

Each of these states, as well as the relationship between the states, can be illustrated

graphically using a commonly used representation originally introduced by Anders Krogh:

D1 D2 D3 D4 D5

I0

BM 1 M 2 M 3 M 4 M 5E

I1 I2 I3 I4 I5

Here, the lower row represents the match states, with B representing the beginning of

the alignment and E the end of the alignment. Each of the diamonds represents the insert

states, and each of the circles represents the delete states. Arrows represent a movement

from state to state, and each is associated with the probability of moving from one state

to another.

One-Dimensional Prediction of Protein Structure 189

Returning to the alignment, M1 would require a glutamine, M2 a tryptophan, M3 either

a lysine or an arginine (50% of the time for either), M4 a proline, and M5 a glycine. Given

this, to represent the sequence EWRPG, the movement through the model would take

the form B → M1 → M2 → M3 → M4 → M5 → E, because the sequence does not require an

insertion or a deletion to align with the rest of the sequences in the group, with the

exception of sequence 4. The path for sequence 4 obviously would be different than for

most of the sequences; it would take the form B → M1 → I1 → M2 → M3 → D4 → M5 → E.

The movement from M1 to I1 accounts for the insertion at position 2, and the movement

from M3 to D4 accounts for the deletion at position 6.

The usefulness and elegance of this model comes from the ability to train the model

based on a set of sequences. Returning to the original sequences, imagine that they were

fed into the model one by one, without knowing the alignment in advance. For each

sequence, the most probable path through the model is determined. As soon as this is

carried out for all the sequences, the transition probabilities and probabilities for each

match state are determined. This, in turn, can be used to generate the best alignment

of the sequences. Knowledge of these probabilities also allows for new sequences to be

aligned to the original set.

Using the collective characteristics of the input sequences either allows these proﬁle

HMMs to be used to scan databases for new sequences belonging to the set or individual

sequences can be scanned against a series of HMMs to see whether a new sequence of

interest belongs to a previously characterized family.

Another important feature of a single residue is its solvent accessibility or accessible sur-

facearea(ASA);thisistheareaofaresidue’ssurfacethatisexposedtothesolventandthat

couldinteractwithotherproteinsorsmallermolecules(Figure7.3).Aftertheelucidationof

thefirstproteinstructure,LeeandRichardsstartedexaminingtheASAsofproteinsbydevel-

opingaprogramtodrawvanderWaalsradiiontheprotein’sstructure,essentiallyby“rolling

aball”overtheentiresurfaceoftheprotein(LeeandRichards1971).Theaccessibilityisthen

definedbytheprotein’ssurfacearea,typicallygiveninangstromssquared,thatcanbetouched

byawatermolecule’sradius.Biologically,theconceptofaccessibilityisofinterest,asresidues

deeply buried inside a protein may be crucial in stabilizing its structure even though they

cannot partake directly in binding other molecules. Many methods that perform secondary

structurepredictionalsopredictaccessibilitytoaidtheidentificationofactiveresidues,which

isimportantforfunctionalstudies.

Thesecondarystructureannotationsandsolventaccessibilitydataneededfortrainingthe

predictionmethodscanbeobtainedfromknown3Dproteinstructures.Themostpopulartool

for this task, the Dictionary of Secondary Structure for Proteins (DSSP; Kabsch and Sander

1983),analyzeshydrogenbondsin3Dstructuresandassignseighttypesofsecondarystructure

elementsthatcanbegroupedintothepreviouslymentionedclasses:helices(alpha,310,andpi

helices),strands(extendedandbridge),andother(usuallyreferredtoas“turn,”“bend,”and

“other”withintheDSSP).Solventaccessibilityisgivenforeachresidueinangstromssquared.

OthersimilarprogramsincludeSTRIDE,whichincorporatesbackbonegeometryinformation

andaimstoprovideannotationsthatareconsistentwiththoseprovidedbyexperimentalists

(FrishmanandArgos1995;HeinigandFrishman2004),andNACCESS,whichcalculatessol-

ventaccessibilityfromstructures(HubbardandThornton1993).

Wenowdescribethemostpopularsecondarystructureandsolventaccessibilityprediction

servers.Nearlyalloftoday’smethodspredictjustthreetypesofsecondarystructure(helix,beta

strand,andother)andnumerical(0–9)orcategorical(e.g.buried,partiallyburied,orexposed)

measuresofaccessibility.

Methods PHDsec is among the earliest methods applying the idea of machine learning to

the prediction of secondary structure (Rost and Sander 1993). I

190 Predictive Methods Using Protein Sequences

Figure 7 .3 Accessible surface area (ASA). The ASA describes the surface that is in direct contact with

the solvent. Practically, the solvent is usually water, and the ASA can be deﬁned by rolling a probe

over the surface of the protein. The probe is a sphere representing the size of a water molecule (blue),

while the protein’s surface is deﬁned by the van der Waals volume of every amino acid’s atoms (gray).

The ASA is then described by the center of the probe as it moves along the surface (red). Software such

as DSSP that calculate ASA employ more efﬁcient algorithms; however, the underlying principle of what

is being measured remains the same.

obtained through the Homology-derived Secondary Structure of Proteins (HSSP) database

(Touwetal.2015),whichcontainsalignmentsofproteinsequencestothoseofproteinswith

known structures contained in the Protein Data Bank (PDB) (Rose et al. 2017). PHDsec

uses two consecutive neural networks to make its predictions (Box 7.2): the first layer uses

conservation values derived from an MSA within the HSSP database, using a window of 13

residues to predict one of three secondary structure states. The first layer’s output is then

used as input to the second layer, which smooths out biologically implausible predictions,

such as an alphahelix that is interruptedby a single beta sheet residue in the middleof the

helix.Methodimprovementssincethefirstversionincludetheabilitytofactorinadditional

weighting information derived from MSAs or global amino acid composition features (Rost

andSander1994a,b).ThisprocessalsoledtotherenamingofPHDsecasPROFsec.Themost

recentiteration,ReProf,isaccessibleaspartofthePredictProteinwebserver(Yachdavetal.

2014). ReProf also includes the most recent version of the solvent accessibility prediction

algorithm,initiallynamedPHDaccandPROFacc(RostandSander1994a,b).

Box 7 .2 Neural Networks

With the development of deep learning, neural networks have regained their popularity

as machine learning models for biology (Punta and Rost 2008; Jensen and Bateman 2011).

Neural networks attempt to mimic the way the human brain processes information when

trying to make a meaningful conclusion based on previously seen patterns. The type of

network design that is typically used for machine learning has several layers: an input

layer, zero or more hidden layers, and one output layer. The ﬁgure shows an example with

one hidden layer.

Here, the input layer is a protein sequence (PKRPSSAY), and the output layer is one of

the possible outcomes: whether a particular amino acid lies within an alpha helix, a beta

strand, or other (loop) region. The neural network receives its signals from the input layer

and passes information to the hidden layer through a neuron, similar to how a neuron

would ﬁre across a synapse. A representative subset of the neurons is shown here by

arrows. In most applications all input units are connected to all hidden, and all hidden to

all output units.

One-Dimensional Prediction of Protein Structure 191

Output layer

Hidden layer

Input layer

αβ O

PKR P SSAY

In the ﬁgure, a proline residue is highlighted in the input layer; it inﬂuences several

nodes in the hidden layer, the strength or weight of the input being controlled by the

neurons. In the hidden layer, one of the nodes is shown in orange; many of the positions

in the sequence may inﬂuence that node, with the neurons again controlling the degree

to which those inputs inﬂuence the node. The ﬁgure illustrates a feed-forward neural

network, where the ﬂow of information is in one direction. Recurrent neural networks,

where neurons connect back to neurons in earlier layers, are also possible.

Why use a neural network? If the direct relationship between the input and output layer

was perfectly understood, such a complicated approach would not be necessary, because

one could write cause-and-effect rules that would make the intermediate, hidden layer(s)

unnecessary. In the absence of this direct relationship, neural networks can be used to

deduce the probable relationship between the input and output layers; the only require-

ment is the knowledge that the input and output layers are, indeed, related. Here, there is

an obvious relationship between the input and output layers, because the individual amino

acids found within a protein must belong to one of the three given secondary structure

classes modeled by the output layer.

T o start deducing the relation between input and output, a supervised learning approach

is used, i.e. the connections are optimized to ﬁt a set of training examples for which the

input→output mapping is known. For example, in the realm of secondary structure pre-

diction, one would construct datasets based on known 3D structures, noting not only the

secondary structure in which a particular residue is found but also other factors inﬂuenc-

ing structural conformation. Based on these training data, the network attempts to learn

the relationship between the input and output layers, adjusting the strength of each of

the interconnected neurons to ﬁne-tune the predictive power of the network.

PSIPREDisanotherneuralnetwork-basedsecondarystructurepredictorusinganapproach

similartoPHDsec(Jones1999).Aprofilebasedontheinitialquerysequenceiscreatedusing

PSI-BLAST and then fed to the first neural network using a window size of 15. The output

is then further processed using a second network and the same window size. The network

architecturecurrentlyusedforperformingthesepredictionshasbeensignificantlyimproved

overtime(Buchanetal.2013).

Proteus directly transfers secondary structure annotation information obtained from

homologs of the protein being analyzed into the prediction pipeline (Montgomerie et al.

2006). A BLAST search is performed in order to find homologs of known structure within

PDB that match the query protein. If successful, the secondary structure annotation of

the homolog found in PDB is copied onto all aligned residues of the query protein. Any

non-alignedresiduesarethenannotatedusingpredictionsbasedonthequerysequenceitself.

Once this process is complete, three methods (including PSIPRED) are executed and their

outputisfedintoaneuralnetworkthatyieldsaconsensuspredictionthatisfinallyreported.

192 Predictive Methods Using Protein Sequences

SANNisasolventaccessibilitypredictor(Jooetal.2012).Usingasetofproteinswithknown

accessibility,theauthorsusedPSI-BLAST-basedPSSMswithaslidingwindowofsize15asfea-

turevectors.Thevectorofaresidueinthequeryproteinisthencomparedwithallvectorsinthe

database,withthecentralresiduehavingthehighestweightanddecreasingoutwards.Based

onthis,the64nearestneighborsarethenusedtoperformapredictionbasedona z-scorethat

weightsneighborsbytheirclosenesstothequeryresidues.SANNpredictssolventaccessibility

intwo(buried,exposed)orthree(buried,intermediate,exposed)discretestates.Italsogivesa

fractionalpredictionforrelativesolventaccessibility(RSA)between0and1.

SSpro5predictssecondarystructureinthreeoreight(“SSpro8”)states.Itcanmakepredic-

tions directly, based on the query sequence, or based on homology of the query sequence to

proteins of known structure, similar to the approach used by the Proteus method (Magnan

and Baldi 2014). First, BLAST is used to identify query protein sequence matches to known

structuresinPDB.Oncefound,themostcommonsecondarystructureclassassignedbyDSSP

ateachknownstructureresidueisoutputastheprediction.Ifthisfails(e.g.whenBLASTfinds

nomatchestoPDBorthereisnomost-commonsecondarystructureclass),asequence-based

predictionisused.Forthis,asetof100bidirectionalrecurrentneuralnetworks(BRNNs)has

beentrained,usingPSI-BLASTPSSMsasitsinput(Box7.2).Thebidirectionalneuralnetwork

architectureenablesthepredictortoconsiderinformationfromacentralsequencewindow,as

wellassequenceregionsoccurringbeforeandafterthecentralwindow(Pollastrietal.2002).

ACCpro5 uses the same methodology to perform predictions of solvent accessibility in two

states(RSAlargerorsmallerthan25%)or20states(RSA0–95%in5%increments).

RaptorX Property is part of a larger structure prediction service and predicts secondary

structure, solvent accessibility, and disorder (Wang et al. 2016a,b) (Figure 7.4). The method

Figure 7 .4 Protein secondary structure. Prediction of secondary structure, solvent accessibility, and disordered regions of the same protein

by the web server RaptorX Property. Summary statistics about the prediction results are shown at the top left. Below, the ﬁrst row contains

the input amino acid sequence, followed by the predicted secondary structure in the three DSSP classes (SS3) followed by the eight

DSSP classes (SS8). The last two rows contain solvent accessibility prediction in three states (exposed, medium, and buried) and a binary

prediction of disordered residues (none detected here).

One-Dimensional Prediction of Protein Structure 193

also uses a deep learning approach called deep convolutional neural fields, using sequence

profilesasinput.Thisdeeplearningapproachisexpectedtocapturebothrelevantaspectsof

global protein structure and observed correlations with neighboring residues. The network

achieveditsbestperformanceusingawindowsizeof11andsevenhiddenlayers.Disordered

residues (discussed in more detail in Disordered Regions) and solvent accessibility (either

buried,medium,orexposed)arepredictedusingthesameapproach.

SPIDER3isadeeplearning-basedmethodthatpredictssecondarystructure,solventacces-

sibility,andbackbonetorsionangles,aswellastwootheranglesinvolvingtheC 𝛼atomsand

neighboring residues (Heffernan et al. 2017). The correlation between these three elements

is exploited by an iterative prediction approach. In the first iteration, the various structure

propertiesarepredictedseparatelyusingtwoBRNNs.Inthefollowingiterations,theoutput

fromoneoftheseBRNNsisusedasinputfortheothertoiterativelypredicteachoutputstruc-

turepropertyinturnbasedonallotherstructuralproperties.Theinputinthefirstiteration,

to which intermediate predictions are then added in later iterations, consists of just seven

physicochemicalaminoacidproperties,togetherwithsequenceprofilesfromPSI-BLASTand

HHblits. Owing to the design of the network, the method is not window based but uses the

wholesequenceasinput.

Performance Assessment of Secondary Structure Prediction

Themostcommonmeasurefortheassessmentofsecondarystructurepredictioniscalledthe

Q3 score(Box7.3).Additionalmeasuresshouldalsobeconsideredbecauseoftheimbalance

of the three possible states (alpha helix, beta strand, or random coil). Most methods predict

strandsleastaccuratelyowingtoinherentbiasintheexperimentaltrainingdata(Rost1996,

2001); only about 20% of all residues are in beta strands, over 30% are in helices, and about

50%inotherstructures.Manymethodsthatreachrelativelyhigh Q3 scoresperformasbadly

asrandomwhenpredictingbetastrands.Anotheraspect,consideredbyper-segmentscores,is

howwellanentiresecondarystructuresegmentispredicted.Themostprominentexampleof

suchascoreisthesegmentoverlapscore(SOV),whichmeasurescorrectlypredictedresidues

andsegments(seeBox7.4)(Fidelisetal.1999).

Box 7 .3 Secondary Structure Prediction Scoring Schemes and Receiver Operating Char-

acteristic Curves

A classiﬁcation task can discern between two or more classes. For example, a secondary

structure prediction typically yields a per-residue classiﬁcation of either an alpha helix,

beta strand, or non-structured region. Measures of classiﬁcation performance, such as the

number of true positives, can differ in meaning between publications; thus, one must care-

fully identify the relevant deﬁnition within the publication. Most commonly, multi-class

prediction performance is reported using accuracy (see Box 5.4), which intuitively trans-

lates to multiple classes when its deﬁnition is regarded as “the number of correct pre-

dictions among all predictions.” For some prediction tasks such as secondary structure

prediction, this score has also been termed Qn,with n denoting the number of classes. As

secondary structure predictions ultimately result in the detection of one of three classes,

this metric is often called the Q3 score.

Binary classiﬁcation usually treats prediction results as discrete. However, machine

learning models usually output a continuous number, representing a likelihood for that

class. During development of the method, a threshold is chosen that yields the best

194 Predictive Methods Using Protein Sequences

Box 7 .3 (Continued)

output of 0). Given a query residue, the network will now predict a value between 0 and

1; the closer that value is to 1, the more likely the residue is found inside the membrane.

The receiver operating characteristic (ROC) curve is used in these cases to describe how

the choice of threshold for the predictor output affects the true-positive rate (TPR) and

the false-positive rate (FPR; see Box 5.4).

Receiver operating characteristic curves

False positive rate

0.0

0.0

0.2

0.4

0.6

0.8

0

0.2 0.4 0.6 0.8 1 .0

True positive rate

AUC_ROC = 0.91

AUC_ROC = 0.80

This curve is derived by making predictions for every possible decision threshold, then

plotting the resulting TPR and FPR values. The plot forms a curve that describes the dis-

crimination performance of the prediction method. A perfect predictor would have a TPR

of 1 and FPR of 0 across all relevant thresholds, while a random prediction is represented

by the dashed black line. Instead of displaying the ROC curve itself, one can represent

the curve as a single numerical score, called “the area under the curve” (AUC). T ypically,

AUC denotes the area under the ROC curve (AUC_ROC); however, this term is also used to

describe other areas, such as the area under the precision vs. recall curve (AUC_PR).

Recent methods estimateQ3 and SOV values around 80–85% for their prediction perfor-

mance (Mirabelloand Pollastri 2013;Heffernan et al. 2015;Wang et al. 2016a);those values

havebeenconfirmedinarecentreviewforasmallsetofmethods(Yangetal.2016a).These

measuresalsoseemrealisticgiventheincreasinglyslowadvancesinthedevelopmentofnew

methodologiessincethelastindependentevaluationpublishedseveralyearsbeforethiswrit-

ing, where only small performance increases were reported (Zhang et al. 2011). Because of

knownerrorsinstructuredeterminationandresultinginconsistenciesinsecondarystructure

assignments from experimentally determined protein 3D structures, among other issues, a

perfectQ3 of100%isnotattainable(Kihara2005;Heffernanetal.2017).Arguably,themost

importantchallengeinthisfieldhasshiftedfrompredictingsecondarystructuretothehigher

goalofpredictingtertiarystructure(Chapter12).

One-Dimensional Prediction of Protein Structure 195

For solvent accessibility, scoring schemes differ as widely as the values that are predicted.

TocalculateRSA,agoodnormalizationschemeisnecessary,butthisisnottrivial(Tienetal.

2013).Giventhatcategoricalvaluessuchasapredictionintwostates(buriedorexposed)are

used,Q2 orQ3 measurescanbeusedinafashionanalogoustohowthesemeasuresareapplied

tosecondarystructureprediction.ForpredictionsofRSAthatfallintoarangesuchas0–100,

accuracy measures are not meaningful, and scores such as Pearson’s correlation coefficient

aremoreappropriate.Fromthemethodsdescribedabove,SANNandRaptorXPropertyboth

reportQ3 valuesof66%,whileSANNreportsa Q2 valueof81%andACCproreportsa Q2 value

of88%.Pearson’scorrelationcoefficientsforRSApredictionare0.68forSANNand0.8forSPI-

DER3.Unfortunately,thesevaluescannotbedirectlycompared,astheywerecomputedusing

differenttrainingdataorusingdifferentpredictedoutputcategorydefinitions.Suchdecisions

can impact performancesince residues with 0% solvent exposure are much easier to predict

thanthosewith,forexample,36–49%exposure.

Inthepast,thewebserversEVAandLivebenchprovidedautomatedpredictorperformance

assessments by comparing prediction results with about-to-be-released experimentally

determinedproteinstructures(Eyrichetal.2001;RychlewskiandFischer2005),but,unfortu-

nately,thisserviceisnolongeractiveandnoindependentlarge-scaleassessmentofsecondary

structurepredictioncurrentlyexists.TheCriticalAssessmentofProteinStructurePrediction

(or CASP) is a biannual community-led challenge in which developers submit predictions

for structures that have been solved but not yet publicly disclosed (Moult et al. 1995). At

the time of submission, none of the structures are known to either the organizers or any

of the participating groups, making CASP a double-blind experiment. Once the structures

are released, assessors evaluate the previously submitted predictions to chart progress in

the field in an unbiased manner. The focus of the challenge is on developing new, more

accuratealgorithmsforproteintertiarystructureprediction.InCASP5,thelastiterationthat

considered secondary structure predictions (in 2002), method performance had approached

asaturationpoint(Aloyetal.2003).Giventhislikelysaturationofmethodperformanceand

becauseofthegenerallyveryhighperformanceavailable,thechoiceofpredictoris,tosome

degree, a matter of personal preference. Any recent method (such as RaptorX Property or

SPIDER3)willprovidehigh-performancepredictionsthrougheasy-to-usewebservers.Other

services,suchasReProfandPSIPRED,arepartoflargerweb-basedsuitesthatallowtheuser

torunadditionalpredictiontoolsonthequeryproteinwiththeclickofabuttonandarethus

wellsuitedtogiveanoverviewofawiderangeofproteinfeatures.

Transmembrane Alpha Helices and Beta Strands

Background The communication between a cell and its surroundings takes place almost

exclusivelythroughproteinsthatareembeddedinthecellmembrane,withthese transmem-

brane proteins interacting with molecules on both the intracellular and the extracellular

sidesofthemembrane.Transmembraneproteinshavebeenestimatedtomakeup20–30%of

all proteins found in any organism (Stevens and Arkin 2000; Liu and Rost 2001; Fagerberg

et al. 2010). These include well-known and highly studied protein classes such as the

G-protein-coupledreceptors,proteinsthatareoftenmajortargetsofdrugdevelopmentefforts

(Jacoby et al. 2006; Overington et al. 2006). In fact, almost two-thirds of all drug targets are

transmembrane proteins. The ability to identify transmembrane proteins and to decipher

their molecular mechanisms is, therefore, of high interest in many fields of biomedicine.

Unfortunately, experimental structural determination of membrane-bound proteins is sig-

nificantlymoredifficultthanforsolubleproteins,andtransmembraneproteinstructuresare

strongly under-represented in PDB (Kloppmann et al. 2012). Therefore, computational pre-

dictionsareessentialforunderstandingthestructuresofthisclassofproteins.Typically

196 Predictive Methods Using Protein Sequences

(a) (b)

Figure 7 .5 T ypes of transmembrane proteins. Experimentally determined three-dimensional structures

of two transmembrane proteins rendered in PyMOL (Schrodinger 2015). Protein segments that are anno-

tated as being inside the membrane are highlighted in yellow. (a) Alpha helical transmembrane protein

aquaporin (PDB structure 3llq, chain A). (b) Beta barrel transmembrane protein OmpA (PDB structure

1bxw, chain A).

structures: helices or strands. Usually, the proteins consist only of these two secondary

structureelements;however,bothfulfillthesametaskofmaskingthepolarproteinbackbone

fromthehydrophobicmembraneenvironment.Proteinsusingalphahelicaltransmembrane

segmentsforthispurposearemuchmorecommon,whilethoseconsistingofbetastrandsare

typicallyporinsfoundonlyintheoutermembraneofGram-negativebacteria,mycobacteria,

chloroplasts,andmitochondria(KesselandBen-Tal2011)(Figure7.5).Asmembraneproteins

consist of the same types of secondary structure discussed in the previous section, one may

wonder why specialized predictors were developed for use with transmembrane proteins

rather than just using already available predictors. The underlying reason is that trans-

membrane proteins have evolved unique structural properties that allow them to be firmly

embeddedinthecellmembrane.Thesephysicochemicalpropertiesaredifferentenoughfrom

thoseofsolubleproteinsthatspecializedpredictorsarerequired.Fortunately,theseproperties

are easier to identify, making it easier to predict transmembrane segments compared with

predicting “general” secondary structure. The basic biophysical property that is responsible

for a residue to be buried within a membrane ishydrophobicity, the property that enables

mostofthetransmembranesegmentstoremainwithinthemembraneandavoidexposureto

thesolventoneitherofitssides.Hence,thefirsttransmembraneregionpredictionmethods

focusedonasearchoflonghydrophobicstretchesofsequence(KyteandDoolittle1982)and

hydrophobicity metrics remain a crucial input feature for today’s most advanced methods

(Reebetal.2014).Additionally,manymethodspredictthetransmembranesegment topology,

providinginformationregardingtheorientationofhelicesorbetastrandswithrespecttothe

cytoplasmic and non-cytoplasmic sides of the membrane. An important concept underlying

the determination of topology is the “positive-inside rule,” which describes the observation

that the loops on the cytoplasmic side of the membrane typically contain more positively

chargedresiduesthanthoseonthenon-cytoplasmicside(VonHeijneandGavel1988).Early

transmembrane region prediction methods focused on the combination of window-based

searches for hydrophobic stretches and the analysis of electrical charges, leading to the first

widely used prediction method (Von Heijne 1992; Claros and Von Heijne 1994). Since then,

andwiththeadventofmachinelearningapproaches,methodstodeterminethestructureof

membraneproteinshavemadeimmenseadvances,asdescribedbelow.

One-Dimensional Prediction of Protein Structure 197

Methods Phobius, like its predecessor TMHMM, predicts transmembrane segments using

an HMM that is based on representations of globular domains, loops, helix ends, and the

transmembrane helix core (Krogh et al. 2001). With respect to TMHMM, Phobius adds

another set of states that model signal peptides (Käll et al. 2004) because signal peptides

contain a stretch of hydrophobic residues that can easily be mistaken for transmembrane

segments. By combining the use of known transmembrane segments and signal peptides

in one model, Phobius is capable of accurately distinguishing between these two classes.

PolyPhobius is an extension of Phobius that keeps the HMM unchanged but achieves a

better performance by using a decoding algorithm that harnesses evolutionary information

fromMSAsofthequerysequence(Källetal.2005).

Proteus-2 is an extension of the original Proteus method described above for secondary

structureprediction(Montgomerieetal.2008).Themethodfirstpredictssignalpeptides,then

checksthesepredictionsagainstadatabaseofexperimentallyannotatedsignalpeptides.Next,

ifaqueryproteinishomologoustoonewithknowntransmembranesegments,thesearetrans-

ferredtothequery.Ifthisfails,transmembranesegmentsarepredictedusingacombinationof

TMHMMandTMB-HUNT(Garrowetal.2005).Finally,anyunassignedresiduesareassigned

a secondary structure class by homology-based inference or, if not possible, as predicted by

Proteus.

MEMSAT-SVMconsists of four separate support vector machines (SVMs) for the predic-

tion of transmembrane helices, signal peptides, loops, and re-entrant helices (Nugent and

Jones 2009). This last class is a special case of helical transmembrane segments in which a

helixentersandexitsthemembraneatthesameside(Viklundetal.2006;vonHeijne2006).

MEMSAT-SVMisoneofthefewmethodsthataccuratelymodelsthiscase.

TMSEGis a recent method that combines machine learning with empirical filters (Bern-

hofer et al. 2016). First, a random forest model predicts scores for each input residue to be

eitherinthemembrane,asignalpeptide,oraloop.Scoresarethensmoothedandcontinuous

proteinsequencepositionsthatareconsistentlyhighscoringforagivenclass(e.g.transmem-

brane)areidentifiedasaproteinsequencesegmentofthatclass(e.g.atransmembraneregion).

Next,aneuralnetworkrefinesthepreviouslypredictedsegmentsand,finally,anotherrandom

forestmodelpredictsthetopologyoftheprotein(Figure7.6).Astrengthofthismethodisthat

itcanaccuratelydistinguishbetweenproteinswithandwithoutmembranehelices.

BETA W AREfocusesonpredictingwhetheraqueryproteinisatransmembranebetabarrel

(TMBB)protein(Savojardoetal.2013),anunderstudiedclassoftransmembraneproteins.This

is achieved using a neural network with an “N-to-1” encoding that supports the input of a

0 50 100 150

Protein sequence

0.0

0.2

0.4

0.6

0.8

0

Predictor output

TMSEG prediction for PTGES_HUMAN

Non-membrane Transmembrane Signal peptide TM segment

Figure 7 .6 T ransmembrane helix prediction by TMSEG. TMSEG (Bernhofer et al. 2016) predictions of

transmembrane helices for prostaglandin E synthase (UniProtKB identiﬁer PTGES_HUMAN). The graph

distinguishes non-membrane residues (green), helical transmembrane residues (raw prediction in blue),

and signal peptides (red: here all close to 0). The purple line marks the ﬁnal prediction of transmembrane

helices after smoothing and reﬁnement by a segment-based neural network classiﬁer. In this case, all

four helices are correctly identiﬁed.

198 Predictive Methods Using Protein Sequences

variablelengthquerysequenceintothenetwork,whichpredictswhetherthequerysequenceis

aTMBBprotein.ForallputativeTMBBproteins,agrammaticalrestrainedhiddenconditional

randomfield,similartoanHMM,isusedtopredictthestateofeachresidueintheproteinand

theproteinmembranetopology(Farisellietal.2009).

BOCTOPUS2(Hayat et al. 2016) also predicts TMBB strands and topology. First, an SVM

predicts the preference for each residue to be either pore facing, lipid facing, outer loop, or

innerloop.Next,theregionofthebetabarrelisdeterminedbyfindingtheregionmostcon-

sistentwiththepredictedpore-andlipid-facingresidues.Onceset,thepredictedclassesare

adjusted such that no membrane residues are predicted outside the barrel region. Lastly, an

HMMthatmodelsthesamefourstatesastheSVMisusedtopredictthefinaltopology.

Performance Predictionperformanceoftransmembranesegmentscanbemeasuredonthree

differentlevels:perresidue,persegment,orperprotein(Chenetal.2002)(Boxes7.3and7.4;

seealsoBox5.4).Mostmethodsreach Q2 scoresof80%ormore(Reebetal.2014).However,

owing to disadvantages of per-residue scores (Box 7.4) per-segment scores are important to

consider when interpreting membrane prediction results. If we consider a transmembrane

helix to be correctly predicted if the endpoints deviate by no more than five residues from

knowntrue-positiveexamples,thenTMSEGandMEMSAT-SVMreachsegmentrecallandpre-

cisionvaluesofaround85%andPolyPhobiusaround75%(Bernhoferetal.2016).Onamore

stringentper-proteinscore, Qok,TMSEGreaches66%,MEMSAT-SVM61%,andPolyPhobius

54%.Thesescoreshavehighvariancebecauseofthesmallsizeoftheindependentbenchmark

usedtocomputethem( n =44),sothesescoresshouldbeusedasageneralguidelineinstead

of a perfect measure of the expected performance of each method. Similarly, performance

assessment for TMBB protein prediction methods is challenging owing to a small indepen-

dentbenchmarksize.TheBOCTOPUS2publicationreportsa Q3(membrane,inside,oroutside

loop)scoreof89%andSOVscore(Box7.4)of94%fortheirownmethod,and74% Q3,75%SOV

forBETAWARE.Thesenumbersareroughlyconfirmedinanotherrecentpublication(Tsirigos

etal.2016).GiventhissuggestionofhigherperformancealongwiththefactthatBETAWARE

isonlyavailableasacommandlinetool,werecommendBOCOTPUS2asagoodstartingpoint

forthepredictionoftransmembranebetastrandsandTMBBproteins.

Box 7 .4 Scoring Schemes for Structural Protein Segments

A disadvantage of per-residue scores such as Q2 or Q3 is that they do not punish biologi-

cally meaningless predictions. For example, predicting every transmembrane helix with a

single non-membrane residue in its middle would not have a large effect on Q2, but this

outcome does not make biological sense, so should therefore be penalized numerically.

Per-segment scores account for this possibility, and both precision and recall (see Box 5.4)

can also be measured on a per-segment basis for every protein.

Names Formula

Segment recall, Qtmhobs Number of correctly predicted transmembrane helices

Number of observed transmembrane helices

Segment precision, Qtmhpred Number of correctly predicted transmembrane helices

Number of predicted transmembrane helices

T o calculate these scores, one needs to deﬁne what constitutes a correctly predicted

structural segment. For example, the observed and predicted positions for a helix should

roughly encompass the same region of a protein. It is crucial to predict a continuous seg-

ment, while exact matches to the start and end position are not as important, as these

start and end points can differ even between almost identical homologs. The problem

One-Dimensional Prediction of Protein Structure 199

becomes much more complex when dealing with transmembrane proteins, as it is not at

all trivial to exactly determine the beginning and end of a membrane segment even when

a high-resolution 3D protein structure is available. The ﬁgure below shows the C-terminal

residues of aquaporin (see Figure 7 .5a), including its ﬁrst transmembrane helix, experi-

mentally determined to range from residue positions 5 to 25. Black arrows indicate the

range within which a predicted helix would be considered correct if deviations of up to

ﬁve residues are allowed. The same concept is employed by the segment overlap score

(SOV), which measures the average overlap between the predicted and observed structural

segment (Fidelis et al. 1999).

51 0 1 52 02 5 3 0

–5 +5 –5 +5

Protein sequence

Observed transmembrane helix

MGRKLLAEFFGTFWLVFGGCGSAVFAAAFPE...

Allowed variation

Correctly predicted helix

Incorrectly predicted helix

Given the per-segment scores, one can deﬁne an additional (per protein) score that mea-

sures the fraction of proteins where every transmembrane helix was correctly predicted:

Qok = 1

N

N∑

i=0

𝛿i;𝛿i =

⎧

⎪

⎨

⎪⎩

1, ifQobs

tmh(i)= Qpred

tmh(i)= 1

0, else

Here, N is the number of proteins in the respective dataset. Clearly, this provides a much

stricter cut-off than just predicting the majority of residues correctly. However, it is crucial

to apply such stringency, as missing a single membrane segment can reverse the overall

membrane segment topology of the protein, the orientation of which provides important

clues about protein function. For example, all G-protein-coupled receptors have seven

membrane helices, with the N-terminus typically residing outside of the cell (Coleman

et al. 2017).

Disordered Regions

Background Disorderedregionsinproteinsdonotadoptawell-definedstructureinisolation

but fluctuate in a large conformational space (Habchi et al. 2014; Wright and Dyson 2014).

Disordered regions are experimentally determined, for example using nuclear magnetic res-

onancespectroscopy,whichallowsthedynamicobservationofaproteininsolution(Habchi

et al. 2014). The phenomenon ofproteindisorder ordisorderedregions is described by many

differenttermsintheliterature;theseinclude intrinsicallyunstructured,nativelyunstructured,

nativelydisordered,and loopy.Proteinswithsuchregionsareoftenreferredtoasintrinsically

disorderedproteins(IDPs)orintrinsicallyunstructuredproteins.Typically,theseintrinsically

disorderedregions(IDRs)havelowsequencecomplexityandanaminoacidcompositionbias

towardhydrophilicresidues.

Disordered regions have many functions. One feature is to cover a larger set of potential

bindingmodes,i.e.amoreflexibleregioncanbeinducedtofittomanydifferentshapes.Con-

sequently,manyIDPsareinvolvedinregulationorcellsignalingandappearashubswithmany

bindingpartnersintheinteractionnetwork.ThisisespeciallytrueforeukaryoticIDPs,while

thoseinprokaryotesareofteninvolvedinlongerlastinginteractionsthatformcomplexes(van

200 Predictive Methods Using Protein Sequences

derLeeetal.2014).PredictingIDPsanddisorderedregionsfromsequenceisofvaluesinceit

canhelpinstructuredetermination,drugdesign,anddiseaseriskevaluation(Dengetal.2015).

ThreedifferentapproacheshavebeenappliedtopredictIDRs(vanderLeeetal.2014).The

firstusessequencepatternssuchastheaminoacidcompositionandphysicochemicalproper-

ties.ThesecondusessequenceandMSAinformationwithmachinelearning.Thethirduses

aconsensuspredictorstrategytocombinethepredictionsofothertools.Protein3Dstructure

mayalsobeusedwithanyoftheseapproaches.

Alllarge-scaleestimatesfortheabundanceofdisorderingenomesarebasedonprediction

methods. On average, eukaryotes have more IDRs longer than 30 residues (33% of proteins)

thanprokaryotes(4%ofproteins)(Schlessingeretal.2011;vanderLeeetal.2014).However,

individualorganismsdifferalot,asdoesthedisordercontentofproteinswithinanorganism

(Habchi et al. 2014; van der Lee et al. 2014; Lobanov and Galzitskaya 2015). Furthermore,

disorder correlates with the extremity of an organism’s habitat (Habchi et al. 2014; Vicedo

etal.2015).

Several databases store disorder protein region information. DisProt contains manually

curated,experimentallyvalidatedIDPannotations(Piovesanetal.2016).IDEALstoresIDRs

thatbecomeordereduponproteinbindingandincludesinteractionpartnersofthedisordered

proteins(Fukuchietal.2014),whileD2P2 providesprecompileddisorderpredictionsofnine

differentmethodsonover1700genomes(Oatesetal.2013).

Methods PrDOSpredictsIDRsusingmachinelearningintwoparts.First,anSVMissupplied

withinformationfromaPSI-BLAST-generatedPSSMina27-residueslidingwindow(Ishida

and Kinoshita 2007). Second, any query homologs with known 3D structures are identified.

Themorealignedresiduesinstructuresaredisordered,thehighertheassignedlikelihoodto

bedisorderedforagivenresidue.Predictionsfromeachapproacharecombinedinaweighted

average and smoothed to provide a final output. The more recent incarnationPrDOS-CNF

appliesthesamemethodologyusingconditionalneuralfieldsinsteadofSVMs(Monastyrskyy

etal.2014).Thesamegroupalsomaintainsameta-predictor, metaprdos2,whichcombines

thepredictionsoffivepredictionmethodsincludingDISOPRED2andPOODLE-S(seebelow)

inanSVMclassifier(IshidaandKinoshita2008;Monastyrskyyetal.2014).

DISOPRED3 employs three prediction models: a neural network tuned to capture long

disorderedregions,anSVM,andanearestneighborpredictorthatcanbeadaptedtonewexper-

imentaldatasinceitdoesnotrequireanytraining(JonesandCozzetto2015).Theoutputofthe

threepredictorsiscombinedintothefinalpredictionusingasimplewindow-basedneuralnet-

work.AnewfeatureofDISOPRED3isthepredictionofputativeproteinbindingsiteswithin

theidentifiedIDRs.ThisisperformedbythreeseparateSVMswithincreasinglycomplexinput

signals,allofwhicharebasedonsequenceorevolutionaryinformation.

POODLEisanSVM-basedfamilyofdisorderpredictors.Threepredictorsarecombinedin

themeta-predictorPOODLE-Isuchthataspecificmethodistrainedforaspecificlengthofdis-

orderedsegment(Hiroseetal.2010).POODLE-SpredictsshortIDRsbasedontheirdistance

from the N-terminus using information from PSI-BLAST PSSMs, as well as using physico-

chemicalfeaturesasinput(Shimizuetal.2007).POODLE-LpredictslongIDRs(Hiroseetal.

2007). While predictions are refined in a second level, the first level of POODLE-L predicts

IDRs in a 40-residue window using the amino acid composition of the input sequence, as

wellasphysicochemicalfeatures.POODLE-Wpredictsproteinsas(mostly)orderedorasIDPs

basedonasemi-supervisedlearningschemethatwastrainedon608proteinswithknowndis-

order status and on random samples of 30000 proteins without labels (Shimizu et al. 2007).

Finally, POODLE-I combines these predictors by first running POODLE-L and POODLE-W

todeterminelongdisorderedregions.ThepredictedIDRsaretruncatedbasedontheoutput

of three secondary structure prediction methods. POODLE-S is then applied to the remain-

ing ordered regions and its predictions are compared with 3D structures of similar proteins, if

available. If a region is predicted as ordered by POODLE-S but not resolved in the structure,

the prediction for this segment is changed to “disorder.”

导出整本书

导出设置