Predictive Methods Using Protein Sequences

Predictive Methods Using Protein Sequences

Jonas Reeb, T atyana Goldberg, Yanay Ofran, and Burkhard Rost

Introduction

Simply put, DNA encodes the instructions for life, while proteins constitute the machinery

oflife.DNAistranscribedintoRNAandfromthereinformationisdeliveredintotheamino

acidsequenceofaprotein.Thissimplifiedversionofthe“centraldogmaofmolecularbiology”

formulatedbyFrancisCrick(1958)essentiallyremainsvalid,althoughnewdiscoverieshave

extendedourview(Elbarbaryetal.2016).Furthermore,epigeneticstudieshavedemonstrated

that chromatincontains more complex informationthan just a one-dimensional(1D)string

ofletters,withtheheritabilityofepigenetictraitshavingaprofoundeffectongeneexpression

(Allis and Jenuwein 2016). Nonetheless, the 1D protein sequence ultimately determines the

three-dimensional (3D) structure into which the protein will fold (Anfinsen 1973), where it

willresideinthecell,withwhichothermoleculesitwillinteract,itsbiochemicalandphysio-

logicalfunction,andwhenandhowitwilleventuallybebrokendownandreducedbackinto

itsbuildingblocks.Insum,thefunction(or,inthecaseofadisease,themalfunction)ofevery

proteinisencodedinthesequenceofaminoacids.

The central dogma suggests that everything about a protein can be inferred from its DNA

sequence – so, why then analyze protein sequences? It turns out that computationally con-

verting DNA to protein sequence is challenging and we still do not understand exactly how

toidentifythestructureofaproteinbasedontheDNAthatencodesit.Itisevenmorediffi-

culttopredicttranscriptsfromDNA.Fortunately,manyexperimentalapproaches,including

proteomicsmethods,canbeusedtodeduceproteinsequences,asdiscussedinChapter11.

Theadventof“next-generation”DNAsequencingtechnologiesisgeneratingawealthofraw

sequence data about which very little is known (Martinez and Nelson 2010; Goodwin et al.

2016).Thepaceatwhichsequencesareaccumulatingfarexceedstheabilityofexperimental

biologiststodeciphertheirbiochemicaltraitsandbiologicalfunctions.Thegapbetweenthe

numberofproteinsofknownsequenceandknownfunction–the“sequence–functiongap”–is

everincreasing,requiringimprovedcomputationalapproachestopredictaspectsofaprotein’s

functionfromitsaminoacidsequence.Asimilarsequence–structuregapexistsforproteins,in

thatthereare180millionproteinsequencesavailablebutonlyabout150000differentknown

protein 3D structures have been determined as of this writing (Berman et al. 2000; UniProt

Consortium2016).

Determining a protein’s function begins with an analysis of what is already known. This

means that every protein must be compared with all others, which implies that the compu-

tational time needed to study protein function grows as the square of sequence growth – a

tremendouschallengeforcomputationalbiologyandbioinformatics.Inthefollowingsections,

we survey some of the research approaches and computational tools that have been shown

tosuccessfullypredictaspectsofthestructureandfunctionofaproteinfromitsaminoacid

sequence.

Bioinformatics,FourthEdition.EditedbyAndreasD.Baxevanis,GaryD.Bader,andDavidS.Wishart.

©2020JohnWiley&Sons,Inc.Published2020byJohnWiley&Sons,Inc.

CompanionWebsite:www.wiley.com/go/baxevanis/Bioinformatics_4e

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.

• 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

1. 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

1. 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.”

Predicting Protein Function

The primary reason for our interest in protein structure and its prediction is to learn about

proteinfunction.Beingabletopredictaspectsofproteinstructureandfunctiondirectlyfrom

theproteinsequenceenablesfastfunctionalannotationofthewealthofsequencesavailable

andcanguidefurtherexperimentalstudiesandvalidation.

Synopsis

Mostproteinsequencesarepredictedfromgenomicsdataandtheonlywaytostudytheirfunc-

tionisthroughcomputationalproteinfunctionprediction.Themostwidelyusedapproachfor

predictingfunctionisbytransferringknownfunctionalannotationfromhomologs.However,

large-scale assessments of homology transfer show that this approach has caveats and must

beappliedcarefully(Rost2002;Ashkenazietal.2012).Moreover,manyproteinsdonothave

annotatedhomologs;thus,manycomputationalmethodsforproteinfunctionpredictionhave

been developed to use other information, such as sequence features with known functional

properties.

Therearemanyaspectsofproteinfunction,includingtheprotein’scellularlocation,interact-

ingmolecules,andbiologicalprocessesthatitparticipatesin.Here,wefocusmostlyonaspects

ofmolecularfunction,definedasphysicochemicalprocessesthattheproteinmoleculecanpar-

ticipatein,suchasanenzymaticactivity,bindingotherproteinsormolecules,orfunctioning

asatransporter.Webeginwithproteinmotifsanddomains,consideredtobethefunctional

unitsofaprotein.Then,wediscussthepredictionofGeneOntology(GO)terms,adictionary

capturingthevariousaspectsofproteinfunction(GeneOntologyConsortium2000,2015).A

202 Predictive Methods Using Protein Sequences

relatedendeavoristheCriticalAssessmentofFunctionAnnotation(CAFA),which,overthe

last6years,hasestablisheditselfasamajorresourceforgauginghowwellproteinfunction

predictionworks.Wealsoprovideseveralexamplesandtoolsformethodsthatpredictsubcel-

lular localization, protein interaction binding sites, and the prediction of the effect of single

aminoacidsequencevariants(SAVs)onproteinfunction.

Motifs and Domains

Background Proteinscontainmanytypesofshortandlongersequenceregionsthatarerespon-

sibleforcarryingourparticularfunctions.Motifsaretypicallyshort;forexample,thenuclear

localization signals responsible for import into the nucleus, typically ranging from 6 to 20

residues,orthesix-residue-longmotifcharacterizingtheactivesiteofserineproteases.Struc-

tural domains, on the other hand, are defined as protein fragments that fold independently

intoacharacteristicstructure,existinmanyotherwisedifferentproteins,andcarryoutsome

functions. Most such structural domains are 100–500 residues long (Liu and Rost 2003) and

arethebasisforthemodulararchitectureofproteins.Forexample,thePleckstrinhomology

domainplaysaroleintargetingofproteinstomembranes;itisfoundinbothdynamin,which

is involved in vesicle processing, as well as many kinases such as those of the Akt/Rac fam-

ily that play a role in signal transduction. Finding a motif or domain in a newly discovered

sequencecanofferinsightsintoitsstructureandfunction.

Motifsanddomainsaretypicallyoriginallydiscoveredbyidentifyingconservedregionsin

a set of proteins, assuming that more conserved protein sequence regions are more impor-

tantforfunctionthanlessconservedregions.Thesearethenfunctionallycharacterizedusing

experiments,suchasmutatingthemotifandobservingwhatfunctiontheproteinloses.Many

databasesarededicatedtocataloguingsuchinformationfromavarietyofsources.Motifsand

domains are described by sequence patterns learned from known examples. For example, a

simple pattern consisting of four amino acids, where the first is always an alanine,followed

byanyaminoacid,theneitherglutamicorasparticacid,followedbyafinalhistidine,canbe

capturedasaregularexpression.However,mostsequencepatternsarelargerandmorecom-

plex–forinstance,someaminoacidsmayappearmoreoftenthanothersinasetofmotifs.Such

patternsarebetterexpressedusingPSSMsorHMMs(Box7.1),constructedbasedonanMSA

(Chapters3and8)usingtoolssuchasHHsearchandHMMER(Söding2005;Eddy2011).These

methods are limited to analyzingknown patterns, and prediction methods have been devel-

opedthatpredictdomains,domainboundaries,ordomainlinkerregionsfromsequencealone,

using amino acid propensities or homology information from sequence alignments. Other

methods employ structural templates, that is, they compare the query protein with known

proteinstructuresandtheirdomainannotations.

Databases InterProistheprimaryresourcethatcatalogsinformationaboutsequencemotifs

anddomains(Finnetal.2016)(Figure7.7a).Itcollectsdatafrom14memberdatabases,each

withitsownspecialization,andmakestheannotationsavailablethroughaunifiedsearchtool,

InterProScan, which exists online as well as a standalone tool (Jones et al. 2014). InterPro

comprisesthefollowingdatabases.

•PROSITE,whichcontainsregularexpression-likepatterns,aswellasprofileswhichiden-

tifyproteinfamilies,domains,andfunctionalsites(Sigristetal.2013).

•Pfam,adatabaseofprofileHMMs,whicharebuiltfrommanuallycuratedseedalignments

thatarespecifictoaproteinfamily(Finnetal.2014a,2016).

•SMART,whichcontainsmanuallycuratedHMMsthatidentifydomainsandcanprovide

informationonorthologsbasedonitsannotationprocess(Letunicetal.2015).

Predicting Protein Function 203

•TIGRFAM, containing families represented by curated HMMs that have been collected

with an emphasis on longer annotations that describe more specific functions (Haft et al.

2013).

•SUPERFAMILY,acollectionofHMMsderivedbySCOPfrom3Dstructures,modelingall

regionssimilartoknown3Dstructures(Oatesetal.2015).

•PRINTS,which,asthenamesuggests,isasetofproteinfamily“fingerprints”thatisman-

uallycuratedfromMSA(Attwoodetal.2012).

•PRODOM,a set ofautomaticallygenerateddomainfamiliesbasedmostly onPSI-BLAST

searchesfromSCOPdomains(Bruetal.2005).

•CATH-Gene3D,whichcontainsprofileHMMsbasedonsuperfamilyannotationsinCATH

(seefollowingparagraph)(Lametal.2016).

•PIRSF,whichclusterssequencesintohomeomorphicfamilieswhosemembershavesimilar

sequencesovertheirentirelengthandwhoseproteindomainsareinthesameorder(Wu

etal.2004).

•PANTHER, containing families of homologous genes, each with a phylogenetic tree that

displaystheevolutionaryrelationshipbetweenthefamilymembers(Mietal.2016).

•HAMAP, providing general protein family profiles built from manually curated MSAs

based on sequences that are synchronized with those in UniProtKB (Pedruzzi et al. 2013,

2015).

•SFLD,oneofthemostrecentadditionstoInterPro(Akivaetal.2014)thatcontainsman-

uallycuratedhierarchicalclustersofenzymesandaimstoprovideanalternativetopurely

sequence-orhomology-basedclusteringthatcannotaccuratelyaccountfortheevolutionof

theseproteins.

•CDD,ameta-databasethatincorporatesdomaindefinitionsfromPfam,SMART,clustersof

orthologousgroup(COG),PRK(Klimkeetal.2009),andTIGRFAM(Finnetal.2016);also

containsdomainboundariesdeterminedfromknown3Dstructures.

•MobiDB contains consensus predictions of disordered regions longer than 20 residues

(Neccietal.2017).

Thesedescriptionsrevealtheinherent(andintended)redundancyofInterPro.Thereason

forthisredundancyliesinthehopethatalternativeapproachestothesameproblemcanleadto

amorecompleteandreliableannotationofmotifsanddomains.Furthermore,somedatabases

canprovidemorespecificannotationsthanothers,allowingtheusertochoosethedesiredlevel

ofgranularityfromthecombinedresults.

Twootherimportantrelateddatabasesaretheabove-mentionedSCOPandCATH. SCOP2

containsmanuallycurateddomainsfromPDBstructuresorganizedinahierarchicalformat,

defined as a directed acyclic graph (Andreeva et al. 2014).CATHcontains semi-automated

structure-based domain annotations at four hierarchical levels (Sillitoe et al. 2015). To offer

morespecificdescriptions,anextensionofCATH,calledSOLID,clusterssuperfamilymem-

bers at increasingly higher sequence identity cut-offs into another five levels (Greene et al.

2007).CATHrecentlyaddedFunFams,whicharerepresentedbyHMMsthathavebeencre-

ated by clustering sequences based on specificity-determining positions (SDPs) (Das et al.

2015).SDPsarepositionsthatareuniquetoaproteinsubfamilyascomparedwiththeover-

allfamily,andthuspresumablyareimportantforfunctionspecifictothesubfamily.SDPsare

a more sensitive pattern determinant than just conserved sequence positions. Finally,COG

is a database containing COGs which are formed from full-length proteins from microbial

genomes(Galperinetal.2015).

Nosingledatabasecaptureseveryaspectoffunctionannotation.Therefore,meta-databases

such as InterPro are crucial resources for proteins without annotations since they allow

researchersastraightforwardcomparisonandtojudgereliabilityoftheinformationpresented

while maintaining transparency of their origin. Another approach has been pursued by

Swiss-Prot (Bairoch and Boeckmann 1994; Boutet et al. 2016), a database that has largely

204 Predictive Methods Using Protein Sequences

(a)

(b)

Figure 7 .7 Annotations of human tumor suppressor P53 (P53_HUMAN). (a) InterPro (Finn et al. 2016) shows the entry as belonging

to one single family, namely IPR002117 . Families are sets of proteins assumed to share the evolutionary origin and inferred to be

similar in function and structure based on homology. Domains (distinct functional or structural units) are shown, followed by a list

of the specific hits against all InterPro databases (see text for details). Here, the identifiers of each segment in the source database

are shown on the right and as a mouse-over. For example, PF08563 is a reference to the Pfam family “P53 transactivation motif” and

cd08367 links to the CDD entry for “P53 DNA-binding domain.” (b) neXtProt expertly curates the current knowledge about protein

function (Gaudet et al. 2017). While many annotations originate from UniProtKB/Swiss-Prot, neXtProt carefully adds information

based on experimental data.

Predicting Protein Function 205

beenassembledandcuratedbyexpertswhoreadtheliteratureandusemanyofthedatabases

and methods described in this chapter. Over recent years, Swiss-Prot, which is part of

UniProtKB (UniProt Consortium 2016; see Chapter 1), has moved its objective from “many

annotations” to “deep annotations,” i.e. from the effort to annotate as many proteins as

possibletothatoffocusingonprovidingasdetailedannotationsaspossibleforfewerproteins.

neXtProt is a related effort that focuses solely on collecting the most detailed possible

annotationofallhumanproteins(Gaudetetal.2017)(Figure7.7b).

Methods Motif- and domain-finding methods used by the databases and tools mentioned

above essentially are derivatives of sequence alignment methods, such as those discussed

in Chapter 3. In the following section, we cover methods for the identification of structural

domainsusingonlyinformationaboutproteinsequence.

DomCutisoneoftheearliestandsimplestapproachesforthepredictionofdomainlinker

regions (Suyama and Ohara 2003). At every position over a sliding window of size 15, it

comparestheaminoacidfrequencybetweendomainandlinkerregions,ascalibratedfroma

high-resolutionsetofstructuresfromPDB.

Scooby-Domainidentifiesdomainsbytheircharacteristichydrophobicaminoacidcompo-

sition. The method calculates the fraction of hydrophobic residues in sliding windows of all

possiblesizes,fromthesmallesttothelargestobserveddomain(Pangetal.2008).Thisleads

to a matrix that contains the average hydrophobicity for every combination of window size

andcentralresidueinthewindow.Startingfromthe10best-scoringvaluesinthematrix,the

A* heuristic search algorithm is used to filter through a set of alternative domain architec-

tures. The method was further improved by integration of domain boundary predictions by

DomCutandPDLI(Dongetal.2006),aswellashomologyinformation.Asaspecialfeature,

Scooby-Domaincanalsopredictadomaincomposedofsegmentsthatarenotcontinuousin

sequence.Theoutputofthenetworkisfirstsmoothedanddomainsarethenassignedusinga

pattern-matchingalgorithmthatmergestworegionswithhighdomainscores.

DOMpropredictsproteindomainswitha1Drecursiveneuralnetworkthatisfedthewhole

sequenceasinput,with25inputsperproteinresidue(Chengetal.2006).Twentyoftheseinput

units code for amino acid probabilities from a PSSM, capturing evolutionary information,

whiletheotherthreecodeforpredictedsecondarystructureandtwoforsolventaccessibility

featuresatthatposition.

Dom-Predisadomainpredictionwebserverconsistingoftwocomponents(Brysonetal.

2007):DomSSEArecognizesknowndomainsbypredictingthesecondarystructureofaquery

sequence,thencomparingthesecondarystructureprofileagainstthatofdomainsdeposited

in PDB. DPS, on the other hand, tries to predict domain boundaries of yet-uncategorized

domains.ThisexploitsthefactthatlocalalignmentsresultingfromaPSI-BLASTsearchhave

been observed to create consistent N- and C-terminal signals. These boundaries likely mark

domainsthatareexpectedtobeconserved.

Performance Predictions of domains can be assessed on two levels. The first is to correctly

predict the number of domains in a protein. For this measure, Dom-Pred reports a perfor-

mance of 85% correctly identified domains using homology inference (DomSSEA) and 68%

from sequence alone (DPS). DOMpro reports 69% correctly identified domains. The second

and more difficult level involves the correct identification of the domain boundaries – or,

alternatively,thelinkerregionsbetweendomains.Here,DomCutreportsasensitivityof53%,

similar to Scooby-Domain and DPS both with 50%. DomSSEA improves the performance to

55%.Scooby-Domainalsoclaimsaprecisionof29%,comparedwith85%(DomSSEA)and72%

(DPS).Finally,DOMproreachesanaccuracyof25%forproteinswithtwodomainsand20%for

morethantwo.Unfortunately,theseperformancemeasurescanonlybecomparedtoalimited

degree,astheywerecalculatedondifferentsetsofdataandusingvaryingdefinitionsofhow

206 Predictive Methods Using Protein Sequences

muchapredictedboundarycandeviatefromtheexperimentalannotationwhilestillcounting

ascorrect.

CASP8providedthelastindependentevaluationofdomainboundarypredictions(Ezkurdia

et al. 2009). Eighteen predictors were assessed on up to 122 targets with manually curated

domaindefinitions.Thebestmethodspredictedthecorrectnumberofdomainsinupto90%of

thecasesandalsoperformedwellindeterminingexactdomainboundaries(asassessedbyhow

close predictions are to the experimental annotation). All of the top-scoring methods made

use of structure templates where available, and the CASP assessors noted that performance

significantlydecreasedfortargetswhichhadtobepredictedabinitio,withoutpriorstructural

knowledge. Among the methods mentioned above, only Dom-Pred was evaluated during

CASP8, predicting the correct number of domains for less than 70% of the targets. It should

alsobekeptinmindthatsingle-domainproteinsareover-representedamongallproteinsof

known3Dstructure(i.e.allCASPtargets)andthatpredictionsinvolvingthesesingle-domain

proteinsyieldsignificantlyhigherperformance(LiuandRost2004;Ezkurdiaetal.2009).

Gene Function Prediction Based on the Gene Ontology

Background To systematically predict and evaluate protein function requires a well-defined

dictionaryoftermsdescribingsuchfunctions.GO(GeneOntologyConsortium2000,2015)is

thestandardandmostcomprehensivesuchdictionary.GOcapturesmultipleaspectsofprotein

function, described using three different hierarchies/ontologies: one for molecular function

(MFO; e.g. “alcohol dehydrogenase activity”), one for biological process (BPO; negative reg-

ulation of eye pigmentation’), and one for cellular component (CCO, “extracellular matrix

component”).Eachoftheseconsistsoftermsorderedhierarchicallyinadirectedacyclicgraph,

withthemostspecificdescriptionsoffunctionatthegraph’sleaves.Forexample,parentsof

“negativeregulationineyepigmentation”inBPOinclude“regulationofpigmentationduring

development” and, simply, “pigmentation.” GO, in collaboration with genome and protein

databasegroups,alsomaintainsannotationsofthesetermstogenesformultiplespecies,with

eachannotationassociatedwithanevidencecodethatprovidesinformationaboutwherethe

annotationinformationisderivedfrom(e.g.“inferredfromsequencesimilarity”or“traceable

authorstatement”).

Methods Metastudent(Hampetal.2013)predictsGOtermsforaproteinbasedonhomol-

ogy. BLAST (Chapter 3) is used to find similar sequences with known GO annotations, and

these annotations are then transferred to the query protein. If no homolog is found, no pre-

dictionismade.Ingeneral,thismethodhasrelativelylowreliability,butcertaintermscanbe

highlyreliablypredicted.

COGICuses annotation transfer by homology and machine learning to predict GO terms

(Cozzetto et al. 2013). Specifically it uses: (i) annotation transfer from PSI-BLAST hits

againstUniRef90withatleast85%alignmentcoverage;(ii)anaiveBayesclassifierbasedon

text-miningdatafromUniProtKB/Swiss-Prot;(iii)anaiveBayesclassifierassessingtheasso-

ciationofthefrequencyofaminoacid3-merswithspecificGOterms;(iv)predictionsbythe

SVM-based method FFPred for eukaryotic targets (described below); (v) further annotation

transfer based on distant relationships described by the orthologous groups found within

the EggNOG database (Huerta-Cepas et al. 2016); (vi) a simple neural network, based on a

similarityscorethatcomparesthePSSMsofquerysequenceswiththosefromtheannotated

dataset;and(vii)GOtermpredictionsbasedonhigh-throughputdataprovidedbythemethod

FunctionSpaceforhumanproteins(Lobley2010).Predictionsbythesesevencomponentsare

thencombinedandpropagatedtoproteinannotationsbasedontheGOgraphstructure.

FFPred 3.0 is an SVM-based de novo function predictor focusing on human proteins

(Cozzetto et al. 2016). Therefore, it is a particularly helpful tool when annotation transfer

Predicting Protein Function 207

cannotbeappliedorwhenannotationtransfermethodsprovidefewresults.Astheprediction

of GO terms is a large multi-label problem, a single protein can be associated with many

GO terms, but SVMs are binary classifiers; FFPred trains a total of 868 separate SVMs, each

predictingasingleGOterm.TheinputfeaturesetfortheSVMsconsistsofsequencefeatures

includingaminoacidcomposition,predictedsecondarystructure,transmembranesegments,

disorderedregions,signalpeptides,subcellularlocalization,andmore.

FunFams,shortforFunctionalFamilies,areclustersofproteindomainswhicharelikelyto

sharethesamefunction(Dasetal.2015).ToconstructFunFams,allsequencesfromaCATH

superfamily with a sequence identity of 90% are clustered (Sillitoe et al. 2013, 2015). These

initialclustersarethenfurthermergedusingprofilescreatedfromtheMSAsofeachclusterin

aniterativefashionuntilasingleclusterremains.Intheprocess,ahierarchicalclusteringtree

isconstructed.Next,theoptimalcutofthetreeisdeterminedsuchthattheremainingparent

nodesrepresentFunFams,consideringSDPs(describedinMotifsandDomains,Databases).

TheresultingFunFamscanthenbeusedtopredictGOterms.Allsuperfamiliesidentifiedin

aproteinarelinkedtoallcorrespondingFunFamsandalltheassociatedGOtermsandtheir

GODAGancestorsaretransferredtothequeryprotein.

Performance BecauseofthehierarchicalstructureofGO,evaluationmetricsconsidernotjust

perfectmatchesofpredictedGOtermsbutalsotheirparents.Furthermore,anevaluationcan

beeitherproteincentric,measuringhowwellallGOtermsofasingleproteinhavebeenpre-

dictedandaveragingoverallproteins,ortermcentric,evaluatinghowwellaspecifictermhas

beenpredictedoverallproteinsintheset.

Themostcommonlyreportedscorebymethoddevelopersistheprotein-centricF

max,corre-

spondingtothemaximalF1score(seeBox5.4)achievedalongtheprecision-recallortheROC

curve(Box7.3).ThisisalsothemainscoreintheCAFAalgorithms(Radivojacetal.2013),the

defactostandardinevaluatingGOtermpredictionmethods.

CAFA is the equivalent of CASP for the protein function prediction community. The idea

istodefineasetoftargetproteinsforwhichallparticipantsmustsubmittheirpredictionsby

agivendeadline.Thisisfollowedbyanannotationgrowthphaseofaround8monthsduring

whichtheorganizerswaitfortheaccumulationofexperimentaldatatobeannotatedbyGO

curators;thesedatacanthenbeusedtoevaluatethepreviouslysubmittedpredictionresults.

In thefirstCAFAexperimentconductedfrom2010to2011,onlytwoGOaspectswereeval-

uated:MFOandBPO.CAFA2(2013–2014)extendedthisevaluationbyalsoaddingthethird

GOontology–CCO–andfurtheraddedtermsdescribinghumanphenotypicabnormalitiesas

catalogedinthehumanphenotypeontology(HPO;Köhleretal.2016).CAFA3startedin2016

andfinishedevaluationsin2019.Itexpandedthescopebyincludingpredictionsofmetal-and

nucleotide-bindingresidues,aswellasamanuallycuratedsetofmoonlightingproteinsthat

performmultipledistinctfunctions,onlyone(ornone)ofwhichismadeknowntoparticipants

inthechallenge.

Resultsfortheprotein-centricF

max fromCAFA2areshowninTable7.2(Jiangetal.2016).

ComparedwiththePSI-BLASTbaseline,overallperformanceisbestforMFO,worseforBPO,

andaslowasanaivepredictionusingCCO.However,thelatterlowperformancewaslikelydue

toaproblemwiththeassessmentdesign,asafewverycommontermssuchas“organelle”seem

tohavedominatedtheevaluation.Whenadaptingthealternativescoringscheme(S

min)that

de-emphasizesgeneraltermsovermoredescriptiveonesfurtherdowntheGOgraph,thenaive

predictionperformancedroppedsignificantly–belowthatofthe10bestpredictionmethods.

UsingS

min forscoring,FunFams(MFOandBPO)andjfpred(CCO,basedonCOGIC/FFPred)

jointheranksofthebestmethods.Independentofweighting,nomethodwasabletoperform

better than the naive baseline on human phenotypes described by HPO. The generally bad

performance could be attributed to the fact that target proteins on average have many more

HPOtermsannotatedthantheydoforthethreeGOontologies.Furthermore,HPOislimited

tohuman,whileGOallowsthetransferofannotationsfromhomologsinotherorganisms.

208 Predictive Methods Using Protein Sequences

Ta b l e 7. 2Performance of selected gene ontology term prediction methods in CAFA2.

Fmax

Prediction

method

Molecular

function (MFO)

Biological

process (BPO)

Cellular

component (CCO)

jfpred(FFPred+COGIC) 0.544 0.446 0.35

FunFams 0.563 0.43 0.352

Metastudent 0.523 0.441 0.288

Bestmethod 0.595 0.468 0.372

PSI-BLASTBaseline 0.451 0.347 0.251

ShownistheperformanceoftheGeneOntology(GO)termpredictionmethodsevaluatedinthe

mostrecentindependentassessmentperformedbyCAFA2(Jiangetal.2016),basedontheprimary

score,F

max,forthethreeGOontologies.Forreference, Bestmethod showsthebestperformance

achievedintherespectiveontologyforanymethod.Allvaluesshouldonlybeseenasarough

performanceestimate,asthecoverageofmethodsshownherediffersanderrorestimationshave

beenomitted.Theoriginalpublicationcontainsadditionalperformancescoresthatare

complementarytoF

max.

Overall,performancehasimprovedfromCAFA1toCAFA2(JiaandLiu2006).Itisfurther-

moreencouragingthatthebestmethodsdonotperformsignificantlyworsefordifficulttargets

thathavelowsequencesimilaritytothetrainingdata.Finally,whileseveralmethodsstandout

foroneortwomeasures,thereisnosinglemethodthatconsistentlyperformsbestacrossall

ontologiesandscoringschemes.

Subcellular Localization

Background Predictingthesubcellularlocalizationofproteinscomputationallyisanimpor-

tant challenge in bioinformatics, as the compartment in which a protein functions natively

is informative about its function. Experimental studies have shown that proteins may travel

between different subcellular compartments, yet most of them are functional within a sin-

glecompartmentformostoftheirlifetime(Huhetal.2003;Fosteretal.2006).Furthermore,

thecellularsortingmechanismthatdirectsproteinstotheirmainlocationoffunctionisrel-

ativelywellunderstood,providingusefulfeaturesforcomputationalpredictionmethods,and

experimental localization data are available in public databases for many proteins, provid-

ing useful training and test data. The manually annotated database UniProtKB/Swiss-Prot

(O’Donovanetal.2002)containsexperimentallocalizationinformationformorethan30000

proteins(release2018_08).However,theseconstituteonly0.03%ofallknownproteinsinthe

currentUniProtrelease(UniProtConsortium2016).Thus,thevastmajorityofproteinsdonot

have any experimentally determined cell location information, so computational prediction

methodscanhelpfillthisgap.

The best computational methods achieve impressive levels of prediction performance

(GardyandBrinkman2006;Hortonetal.2007;Huetal.2009;Goldbergetal.2014)andare

routinelyusedforlarge-scaleproteinannotation(Graesseletal.2015;Ramilowskietal.2015).

However, most of these methods focus on one or a few cellular compartments or specific

organisms. The most reliable localization predictions result from careful homology-based

inference,i.e.wherelocalizationinformationistransferredfromanexperimentallyannotated

proteintoitsunannotatedsequencehomolog.Unfortunately,thismethodcannotbeapplied

tomostproteinsbecauseoflimitedexistingcellularlocationinformation.Forthese,denovo

machinelearningmethodsprovidefairlyreliableresults.Otherautomaticmethodsannotate

proteinsbyminingbiologicalliteratureandmolecularbiologydatabases(NairandRost2002).

Forexample,asimpleapproachcouldbetocompareannotatedUniProtKBkeywordsbetween

the query protein and proteins of known subcellular localization. Location annotations are

Predicting Protein Function 209

transferredfromaproteinwithsimilarkeywords.Finally,meta-predictorsintegratedifferent

methods,usingthemostreliablepredictionforthefinalprediction.

Methods LocTree3is the latest generation of the LocTree method, employing a hierarchi-

cal structure of SVMs to mimic the protein-trafficking pathway in a cell and predict protein

localization(Goldbergetal.2014)(Figure7.8).LocTree3predicts18subcellularlocalization

classesineukaryotes,sixinbacteria,andthreeinarchaea.LocTree3isahybridapproachthat

firstusesaPSI-BLASTsearchtotransferanyknownlocalizationinformationtoaquerypro-

tein from its sequence homolog. If no homology-based information is available, it uses a de

novo-basedmethodcalledLocTree2thatemploysaprofile-kernelSVM,usinganevolutionary

profile-basedconservationscoresofshortstretchesof kconsecutiveresidues( k-mers)asinput

(Goldbergetal.2012).

MultiLoc2predictsproteinlocalizationbyintegratingtheoverallaminoacidcomposition

with known cellular sorting signals, phylogenetic profiles, and GO (Gene Ontology Consor-

tium 2015) terms for the prediction of protein localization in eukaryotes (Blum et al. 2009).

MultiLoc2isavailableintwoversions:MultiLoc2-LowRescanbeappliedtoeukaryoticglobu-

larproteinsandpredictsfivelocalizationclasses,whileMultiLoc2-HighResadditionallycovers

transmembraneproteinsandpredicts11eukaryoticsubcellularlocalizationclasses.MultiLoc2

uses a two-level prediction approach, where the first layer consists of several SVMs process-

ing different encodings of the query sequence, forwarding their output to the second-level

SVM-basedclassifiersthatproduceprobabilityestimatesforeachlocalizationclass.

DeepLoc predicts 10 different eukaryotic localization classes (Almagro Armenteros et al.

2017).Unlikethemethodsmentionedabove,DeepLocperformsitspredictionabinitiousing

only the sequence and no ancillary homology information. The method first uses a convo-

lutional neural network to extract sequence motifs of varying lengths that are then used as

Figure 7 .8 Prediction of subcellular localization. Visual output from LocT ree3 (Goldberg et al. 2014), a web server predicting subcellular

localization. Upon submission of a sequence (and specification of its kingdom: eukaryota, bacteria, or archaea), the predicted localization,

along with the corresponding Gene Ontology terms (if available), are shown. The last column (Annotation T ype) denotes whether the result

was obtained through a homology search or, as is the case here, predicted by the machine learning model. The tree on the right is a

visualization of the hierarchical network of support vector machines (SVMs). Highlighted in orange is the path that was used to predict the

example, the alpha subunit of human ATPase, as being localized to the mitochondrion.

210 Predictive Methods Using Protein Sequences

inputforarecurrentneuralnetworkandfedthroughadditionalfilteringlayers.Thefinalstep

performspredictionsusingahierarchicalapproachsimilartothatofLocTree3.

Performance Generally, the performance of subcellular localization prediction methods can

beassessedusingabinaryaccuracyscore(Box7.3;seealsoBox5.4),eventhoughmanymeth-

odspredict10ormoreclasses.Ingeneral,newermethodsthatintegratemoresourcesofinfor-

mationoutperformolderandsimplermethods(Huetal.2009;Mooneyetal.2013;Goldberg

et al. 2014). The reported performance levels suggest accuracies>85% for extracellular pro-

teins,∼80%forthosefoundwithintheplasmamembraneandnucleus, >65%forcytoplasmic

proteins,and >70%formitochondrialproteins.DeepLoc,themostrecentmethod,reportsan

overallaccuracyof78%basedontheirnewlyassembleddataset.Usingthesamedataset,Loc-

Tree2reaches61%accuracy,whileMultiLoc2reaches56%accuracy.

Protein Interaction Sites

Background Most proteins do not function alone. Instead, they work with other molecules

within molecular complexes (Rao et al. 2014; Keskin et al. 2016) to perform specific func-

tions. Protein interactions and resulting interaction networks are covered in greater depth

in Chapter 13. Here, we focus on describing methods that predict physical, non-covalently

bondedprotein–proteininteractionsitesforanindividualprotein1Dsequence.Manymethods

arealso availableto predictprotein–proteininteractionsites based on known3D structures,

and the reader is referred to two comprehensive references for more information on these

methods(Esmaielbeikietal.2016;Keskinetal.2016).

Therearemanytypesofproteininteractionbindingsites.Forinstance,asitecouldbedefined

byalargesurfaceofaprotein’s3Dstructure,correspondingtoaminoacidpositionsscattered

acrossa1Dsequencethatarecloseinthe3Dfold,whichisusedtobindtootherlargeproteins.

Smallsitesthatbindsmallmoleculesareprevalentinenzymes,andusuallydefinetheenzyme

activesite.Finally,shortlinearsitesarecontinuousaminoacidstretches,oftenfoundindisor-

deredregions,thatbindproteins(Tompaetal.2014).Proteininteractionsareusuallymediated

bydomainsandmotifs,eachwiththeirowncharacteristicbindingsitesandpreferredbinding

patternsintheirpartnermolecules(PawsonandNash2003).Proteinsmayhavemultipleinter-

actingpartnersandbindingsitesandsomesitesmaybindtomorethanonepartner,whichmay

leadtocompetitionamongthepartners.Interestingly,evenbetweentwoidenticalproteinsA

andB,wesurprisinglyobservealternativeinterfacesthatarebiologicallymeaningful(Hamp

andRost2012).

Typically,bindinginterfaces,whichcaninvolve20–50residues,aredefinedbyhavingalarge

difference in ASA between the monomer and the complexor by the distance between two

residueswithinaspecificprotein3Dstructure(Figure7.3).However,thecharacterizationof

bindinginterfacesisnottrivial,soaspecificsetoftoolshasbeendevelopedthatenabletheuser

to discern artifacts from true interactions (Tuncbag et al. 2009; Keskin et al. 2016). Binding

sitescanalsobeexperimentallyidentifiedusingarangeofmethods,suchasalaninescanning,

wherebyeachpositioninaproteinsequenceismutatedtoalanineonebyone,andtheresulting

changeinbindingstrengthisthenmeasured.Positionsthataffectthebindingstrengthwhen

mutatedarepartofthebindingsite.Someinteractingresiduescontributemoretobindingthan

others.Thosecontributingmostareoftenreferredtoas hotspots (Tuncbagetal.2009;Morrow

andZhang2012),withabout10%ofallbindingresiduesbeinghotspots.Theseresiduesare

often buried and differ by their native amino acid composition, with tryptophan, arginine,

and tyrosine being particularly common. As hot spots are prime drug targets, several tools

have been developed to predict them (Keskin et al. 2016). Newer methods for binding site

predictioninvolvesearchingforpatternsofcorrelatedmutationspresentinMSAs,basedonthe

conceptthatevolutionarymutationsinonesitemustbecompensatedforinapartnersitefor

Predicting Protein Function 211

theinteractiontobemaintained(Marksetal.2012).Severaltoolshavebeendevelopedrecently

basedontheseadvances,suchasFILM3,EVfold,andEVcomplex(Marksetal.2011;Nugent

andJones2012;Hopfetal.2014).EvaluatedaspartoftheCASP11andCASP12assessments,

therehavebeentremendousadvancesinthedevelopmentofthesemethods,anditisexpected

thattheywillbefurtherimprovedinthenearfuture(Kinchetal.2016;Yangetal.2016b;Wang

etal.2017;Schaarschmidtetal.2018).

About5%ofalleukaryoticproteinsareassumedtobindnucleotides;theirfunctionsinclude

processes such as the regulation of gene expression, DNA replication, and DNA repair (Yan

etal.2016).Topredictbindingresiduesforthesecases,aseparatesetofspecializedpredictors

has been developed that are based on either the sequence or template 3D structures where

available(Zhaoetal.2013).

Many databases collect protein interaction sites (Tuncbag et al. 2009; de Las Rivas and

Fontanillo2010;Keskinetal.2016).Proteinbindinginterfacesandtheirinteractingresidues

are stored in PISA (Krissinel and Henrick 2007), Inferred Biomolecular Interactions Server

(IBIS) (Shoemaker et al. 2012), IPfam (Finn et al. 2014b), PIFACE (Cukuroglu et al. 2014),

and 3did (Mosca et al. 2014). Results from experimental alanine scanning mutagenesis are

available within BID (Fischer et al. 2003), as well as from a legacy database called ASEdb

(ThornandBogan2001).

Methods Predicting PPI interface residues. A popular approach for predicting binding

sites in proteins is to learn their physicochemical patterns from known examples and then

search for these patterns in other protein sequences (Ofran and Rost 2003a,b; Esmaielbeiki

etal.2016).OfranandRost(2003a,b)developedapredictionmethodbasedonaneuralnetwork

withaslidingwindowofsizenineandusingonlythesequenceasinput.Šiki ´cetal.(2009)more

recently showed that this remains a viable approach. Subsequent methods improved perfor-

mance by including evolutionary information, as well as predicted secondary structure and

solvent accessibility data (Res et al. 2005; Wang et al. 2006; Ofran and Rost 2007; Chen and

Jeong2009;ChenandLi2010).Developersofthe PSIVERmethodfoundthatpredictedacces-

sibilityinformationisalsohighlyinformative,evenifusedalone(MurakamiandMizuguchi

2010). HomPPIisapurelyhomology-basedpredictorthattransfersbindingsiteinformation

from known sites in homologous proteins that are part of molecular complex 3D structures

inPDB(Xueetal.2011).ELMisadatabaseofshortlinearproteinbindingsitepatterns,typi-

callydescribedbyregularexpressions,withassociatedtoolsforidentifyingknownpatternsin

aquerysequence(Gouwetal.2018).

Predicting protein–DNA and –RNA binding. DBS-PSSM extends an earlier approach

forthepredictionofDNA-bindingresiduesbytrainingarelativelysmallneuralnetwork(103

nodesintotal)withaslidingwindowoffiveandconservationinformationfromPSI-BLAST

PSSMs(AhmadandSarai2005). DP-Bindisaconsensusapproachcombiningtheprediction

results from an SVM and two logistic regression models (Hwang et al. 2007; Rezácová et al.

2008). Both of these methods found that performance is significantly increased when using

homologyinformation. SomeNA(Hönigschmid2012;Yachdavetal.2014)usesahierarchical

set of neural networks that predict whether a protein binds DNA or RNA and, if so, where

itbinds.RNA-bindingresiduesarealsopredictedby PprintusinganSVMwithPSSMinput

(Kumaretal.2008). RNABindRPlustacklesthesametaskusingthesameinputtoanSVM

andawindowsizeof21(Waliaetal.2014).Alogisticregressionclassifierisusedtocombinethe

predictionswiththeresultsfromthehomology-basedpredictor HomPRIP,whichgenerally

achieveshigherperformancebutcannotbeappliedtoqueryproteinswithouthomologs.

Performance Predicting PPI interface residues. Predicted protein–protein interaction

(PPI)sitesaretypicallyevaluatedusingthesamesetofstandardscoresasoutlinedforother

binary classification tasks (see Box 5.4). Without a recent independent evaluation, the only

212 Predictive Methods Using Protein Sequences

values available are those provided by method authors. These values have been determined

basedondifferenttestdataandarethereforenotcomparable.Theyarealsohighlyvariable.

Independent of these difficulties, performance has been reported to have reached a plateau

(Esmaielbeikietal.2016).

Predicting protein–DNA and –RNA binding. A recent review by Yan et al. (2016) has

comprehensivelystudied30predictionmethodsofDNA-andRNA-bindingresidues,evaluat-

ingnineonanindependentdatasetof3Dstructures.Theauthorsfoundthatmostpredictors

reachasimilarperformanceforthepredictionofDNA-bindingresidueswithanAUC_ROC

around0.79(Box7.3).Theseresultsareencouraging,butworkremainstobedoneandpredic-

torsdodifferintermsofwhethertheyemphasizehigherspecificity(most)orsensitivity.Infact,

theareaundertheROCcurvemaynotbethemostappropriatewaytoevaluatesuchprediction

methods. For example, DBS-PSSM has the best sensitivity, at 72%, but a specificity of only

75%,significantlylowerthanothermethodsthatreachmorethan90%.Aconsensuspredictor

developed by Yan et al. (2016) that combined prediction results using a logistic regression

approach outperformed all individual predictors. Similar results describe the prediction of

RNA-binding residues, albeit with lower AUC_ROCs (0.724 for RNABindR and 0.681 for

Pprint).Interestingly,thesemethodsseemtocapturegeneralpropertiesofnucleicacidbind-

ing. For example, an RNA-binding site prediction method also predicts many DNA-binding

residuesaspositives.Tocounterthis,newmethods,suchasSomeNA,thattrainonbothdata

classes (DNA and RNA) in an attempt to distinguish them have been proposed (Yachdav

etal.2014).Oneimportantproblemthathasbeenbenchmarkedpoorlysofaristhedegreeto

whichamethodconfusesproteinsthatbindnucleotidesandthosethatdonot.Alltheabove

performanceestimatesapplyifandonlyiftheproteinisalreadyknowntobindDNAorRNA.

Effect of Sequence Variants

Background Anytwounrelatedindividualsdifferintheirgenomesataround5millionsites

(Autonetal.2015)andbyabout20000SAVs.Animportantquestioniswhetherthesemuta-

tionsaffectproteinfunction.Becauseofthelargenumberofpossibleproteinmutations,itis

currently impossible to evaluate all of them using experimental methods, so computational

prediction methods are needed to fill the gap. Using such methods, surprising findings are

revealed.Forinstance,manySAVsinhealthyindividualsappeartohavestrongimpactupon

function,commonSAVsappeartoaffectfunctionmorethanrareSAVs,andSAVsbetweenindi-

vidualshave,onaverage,moreimpactthanthoseseenbetweenhumanandrelatedorganisms

(suchasapesormice;Mahlichetal.2017).Mostmethods,includingtheonespresentedbelow,

are either limited or strongly biased toward human sequence variants. Chapter 17 provides

moreinformationabouthowthesemethodsareusedinpractice.

Methods SIFT was one of the first methods to predict SAV effects on protein function (Ng

and Henikoff 2003; Kumar et al. 2009). The central information used in this method (and

most other methods) is to evaluate the SAV in terms of evolutionary sequence conservation

identifiedinanMSA.IftheSAVchangesthenativeaminoacidtoonethatisobservedinthe

familyofthevariedprotein,thevariantispredictedtobeneutral.Incontrast,iftheSAValters

aconservedposition,achangeinfunctionispredicted.SIFTusesPSI-BLASTtoidentifysim-

ilarproteinsandbuildanMSA,thenusesthisMSAtopredictwhethertheSAVislikelytobe

deleterious(i.e.bymutatingarelativelyconservedposition).

PROVEAN is based on the same idea as SIFT but contains some important extensions,

includingtheabilitytopredicttheeffectofinsertionsanddeletionsandmultiplesimultaneous

variantsonproteinfunction(Choietal.2012).PSI-BLASTisusedtoidentifysequencefamilies

thatareclusteredat80%sequenceidentity.The45clustersmostsimilartothequerysequence

areselected.Foreachcluster,thequerysequenceiscomparedwiththeclustersequencesto

Predicting Protein Function 213

computeadeltascoreusingthesubstitutionscoresintheBLOSUM62matrix(seeChapter3).

Thearithmeticaveragesoverallscoreswithineachclusterareagainaveragedtodeterminethe

finalpredictionscore.Ifthescoreisbelowanempiricallydefinedthreshold,theSAVisconsid-

ereddeleterious.Calculatingaveragesforeveryclusterbeforemergingresultsinsteadofjust

averagingallsequencescoresisimportant,aslargegroupsofsimilarproteinsmightotherwise

biastheresults;theclusteringstepreducesredundancyinthedata,therebyavoidingthisbias.

Thedefaultscorethresholdwasestimatedbasedonasetofhumandisease-causingSAVsand

commonvariants(assumedtobeneutral)fromtheUniProtKB(humsavar)databasethatwas

thenextendedtocoverinsertionsanddeletions.

PolyPhen-2predictsSAVeffectsbasedonconservationandexperimentalorpredictedinfor-

mation about protein structure (Adzhubei et al. 2010). From a larger set of features, eight

sequence-basedandthreestructure-basedfeatureswerecombinedwithinanaiveBayesclassi-

fiertodetermineafinalpredictionscore.PolyPhen-2wastrainedondisease-causingmutations

annotatedintheUniProtKBdatabase.Themethodaimstopredicttheeffectuponthe“system”

(a disease) rather than upon the protein (a molecular effect). PolyPhen-2 offers two models:

onefocusesonSAVswithstrongeffectsthatcauseaMendeliandiseaseordisorderinvolving

asinglegene,whiletheotheristrainedtoidentifylessdeleteriousallelesthatmaycontribute

tocomplexdiseaseinvolvingmorethanonegene.

SNAP2predicts the effect of SAVs on protein function using a system of neural networks

(Hecht et al. 2015). As for other methods, the most important signal for prediction comes

from evolutionary conservation. To further improve performance, SNAP2 includes protein

sequence-based predictions as input, including secondary structure, binding site, and dis-

ordered region information. Global features such as the protein length and the amino acid

compositionarealsoconsidered.Themethodwastrainedonadatasetthatincludeshuman

disease-causing SAVs, mostly from the Protein Mutant Database that catalogs how variants

affect an experimentally measured level of molecular function. Thus, while SNAP2 predicts

howanSAVaffectsmolecularfunction,thismaynotcorrespondtoadeleteriouseffectonthe

phenotypebeingstudied.SNAP2wasdevelopedtosupporttheideaof“comprehensiveinsilico

mutagenesis,”whichattemptstolearnthefunctionaleffectsofallpossibleproteinmutations;

thisinformation,inturn,canbeusedtocomputeaheatmapillustratingthefunctionaleffects

ofeachmutation(Figure7.9;Hechtetal.2013).Althoughconservationprovidesthedominant

signal,theadditionalfeaturesunderconsiderationarealsoimportant,particularlyforproteins

withfewknownsimilarsequences.RecentanalysesshowthatSNAP2resultsareverydifferent

formanyvariantswhencomparedwiththepredictionsfromtheothermethodsdescribedhere

(Reebetal.2016).

CADD(Kircheretal.2014)istheonlymethodpresentedherethatpredictstheeffectofa

variant on genomic sequences rather than for proteins; however, it also can handle SAVs. It

can score any SNV or small indel in either coding or non-coding regions. CADD uses a set

of63featuresthatarecombinedwithinanSVMclassifier.Thesefeaturesincludeevolution-

ary,regulatory,andtranscriptinformation(suchastranscriptionfactorbindingsitesorintron

andexonboundaries),aswellaspredictedeffectscoresfromPolyPhenandSIFT.Thetraining

was performedon a set of variants that capture differences between human and an inferred

human–chimpcommonancestor.Allofthesevariants,whichalsoappearinlargenumbersin

thehumanpopulation,areconsideredneutralvariantssincetheyhavebeensubjectedtonat-

uralselection.Anothersetofsimulatedinsilicovariantsareconsidereddeleteriousbecause

theyarenotunderevolutionaryconstraint.CADDprovidesasinglescoremeasuringthedele-

teriousnessofavariant,independentofwhattypethevariantisandwhereinthegenomeit

lies.Thismakesthemethodmoregenerallyapplicablethantheothersdescribedabove.

Performance Generally, variant effect prediction is evaluated with the same measures used

forotherclassificationtaskssuchasAUC_ROCoraccuracy(Box7.3).Forexample,inarecent

independent evaluation SIFT, PolyPhen, and CADD reached AUC_ROC values of 0.59–0.63

214 Predictive Methods Using Protein Sequences

L

A

R

N

D

C

Q

E

G

H

I

L

K

M

F

P

S

T

W

Y

V

KD I NFKI ERG QL LA VAG STG AG KTSL LM VI M G ELEPSEG KI KHS

Figure 7 .9 From predicting single amino acid sequence variant (SAV) effects to landscapes of suscepti-

bility to change. Shown are the SNAP2 (Hecht et al. 2015) predictions resulting from a complete in silico

mutagenesis study for the cystic fibrosis transmembrane conductance regulator, as shown by Predict-

Protein (Yachdav et al. 2014). Prediction results are represented as a heatmap in which every column

corresponds to one residue in the sequence. Rows represent mutations to all other non-native amino

acids. Note that not all those SAVs are reachable by a single nucleotide variant (SNV). SAVs predicted

as neutral are highlighted in green, while those predicted to affect molecular function are in red. Syn-

onymous mutations are black. While traditionally focusing on a few select variants that are of particular

interest, some modern tools are computationally efficient and accurate enough to predict the effect of

every possible variant in a protein. While predictors are not sensitive enough to regard every high-scoring

variant as potentially interesting, such an approach allows for the identification of sites that may be func-

tionally important, as the effect of every possible kind of variation for that residue is predicted. Here,

the clustering of high-effect scores falls exactly into the known nucleotide binding region from residues

458–465. This approach represents one very effective way in which the application of residue-based

prediction tools can lead to knowledge at the level of the whole protein.

on a dataset containing variants known to affect protein function. Using a different set of

variants that affect the function of transcription factor TP53, performance was significantly

higher(0.83–0.87).Performancewasevenhigheronasetofvariantsimplicatedinhumandis-

ease (0.83–0.94). Other published rankings of methods often disagree widely, depending on

thedatasetsused(Thusbergetal.2011;Dongetal.2015;Grimmetal.2015).Thishighvaria-

tionexemplifiesthedifficultiesinevaluatingtheeffectofaSAV.Onereasonforthevariation

isthatthetoolsdescribedabovepredictdifferenttypesofeffects,suchastheeffectonmolec-

ular function, a pathway, or the organism in general. One extreme example of the problem

is that SIFT, PolyPhen-2, and SNAP2 predictions seem to generally agree on the base set of

SAVswith knownexperimentalinformationthatshouldbeusedasthebasisforpredictions,

butdiffersubstantiallyonhowtoapplynever-before-seendatasuchasthenaturalsequence

variationobservedbetween60000individuals(Mahlichetal.2017)orbyrandomSAVs(Reeb

etal.2016).Anotherissueatplayisascertainmentbias,wherewell-studiedproteinsaremore

likelytobeincludedintrainingdata,leadingtooverestimationofgeneralizationperformance

(Grimmetal.2015).Giventheseissuesandtheoverallestimatedperformanceofthesetools,

theyarebestusedtogeneratehypothesesforfurthertesting(Miosgeetal.2015).

---

Source extraction for internal review only — for educational use only, not for public distribution.

版权声明：本提取文本

Extracted%20from%3A%20

Page range: PDF page 234-246 (printed pages 214-226)

---

214 Predictive Methods Using Protein Sequences

L

A

R

N

D

C

Q

E

G

H

I

L

K

M

F

P

S

T

W

Y

V

KD I NFKI ERG QL LA VAG STG AG KTSL LM VI M G ELEPSEG KI KHS

Figure 7 .9 From predicting single amino acid sequence variant (SAV) effects to landscapes of suscepti-

bility to change. Shown are the SNAP2 (Hecht et al. 2015) predictions resulting from a complete in silico

mutagenesis study for the cystic fibrosis transmembrane conductance regulator, as shown by Predict-

Protein (Yachdav et al. 2014). Prediction results are represented as a heatmap in which every column

corresponds to one residue in the sequence. Rows represent mutations to all other non-native amino

acids. Note that not all those SAVs are reachable by a single nucleotide variant (SNV). SAVs predicted

as neutral are highlighted in green, while those predicted to affect molecular function are in red. Syn-

onymous mutations are black. While traditionally focusing on a few select variants that are of particular

interest, some modern tools are computationally efficient and accurate enough to predict the effect of

every possible variant in a protein. While predictors are not sensitive enough to regard every high-scoring

variant as potentially interesting, such an approach allows for the identification of sites that may be func-

tionally important, as the effect of every possible kind of variation for that residue is predicted. Here,

the clustering of high-effect scores falls exactly into the known nucleotide binding region from residues

458–465. This approach represents one very effective way in which the application of residue-based

prediction tools can lead to knowledge at the level of the whole protein.

on a dataset containing variants known to affect protein function. Using a different set of

variants that affect the function of transcription factor TP53, performance was significantly

higher(0.83–0.87).Performancewasevenhigheronasetofvariantsimplicatedinhumandis-

ease (0.83–0.94). Other published rankings of methods often disagree widely, depending on

thedatasetsused(Thusbergetal.2011;Dongetal.2015;Grimmetal.2015).Thishighvaria-

tionexemplifiesthedifficultiesinevaluatingtheeffectofaSAV.Onereasonforthevariation

isthatthetoolsdescribedabovepredictdifferenttypesofeffects,suchastheeffectonmolec-

ular function, a pathway, or the organism in general. One extreme example of the problem

is that SIFT, PolyPhen-2, and SNAP2 predictions seem to generally agree on the base set of

SAVswith knownexperimentalinformationthatshouldbeusedasthebasisforpredictions,

butdiffersubstantiallyonhowtoapplynever-before-seendatasuchasthenaturalsequence

variationobservedbetween60000individuals(Mahlichetal.2017)orbyrandomSAVs(Reeb

etal.2016).Anotherissueatplayisascertainmentbias,wherewell-studiedproteinsaremore

likelytobeincludedintrainingdata,leadingtooverestimationofgeneralizationperformance

(Grimmetal.2015).Giventheseissuesandtheoverallestimatedperformanceofthesetools,

theyarebestusedtogeneratehypothesesforfurthertesting(Miosgeetal.2015).

Summary

Seminal discoveries made in the 1960s by Anfinsen and others have clearly established that

the sequence of a protein determines its structure and, ultimately, its function. Owing to

therelativesimplicitywithwhichproteinsequencescanbeobtainedexperimentally,alarge

enterprisedevotedtopredictingstructureandfunctionfromsequencehasemerged.Structure

predictionhastremendouslymaturedtothepointthatsomeaspectsmaybeconsideredsolved,

atleasttotheextentthatcurrentexperimentaldataallow(Hopfetal.2012).Despitethesesub-

stantialadvances,thegeneralproblemofpredictingproteinfunctionfromsequencehasnot

Summary 215

beensolved.The1Dpredictionmethodspresentedinthischapter(secondarystructure,trans-

membrane,solventaccessibility,anddisorder)areimportantasinputtohigherlevelprediction

methods. Fortunately, given the wide range of prediction methods available, many of which

arediscussedinthischapter,itispossibletoannotateproteinsequenceswithamultitudeof

information, even without any prior knowledge. These predictions will still clearly contain

errors–but,oncetheuserunderstandsthestrengthsandweaknessesofeachmethod,these

toolscanbeincrediblyusefultowardallowingtheusertofilterthedelugeofsequencedatagen-

eratedtodayand,hopefully,generatehypothesesthatcanbeexperimentallytested.Giventhat

errorsmayexistinthedatathatthesemethodsdependupon,itisimportanttoidentifythepri-

maryevidenceusedforproteinfunctionprediction,whetherusingbest-in-classtools,mapped

by high-throughput experiments, or those carefully gathered by experts based on detailed

experiments. No existing resource informs users about these situations. Thus, the analysis

oftheproteinsthatyouareinterestedinistypicallydonebestbyusingthebest,appropriate

predictiontoolsforanyparticularquestion,alongwiththemostreliabledatabaseannotations.

Internet Resources

Essentialdatabasesandpredictionevaluations

CAFA biofunctionprediction.org/cafa

CAGI genomeinterpretation.org

CASP predictioncenter.org

CATH www.cathdb.info

InterPro www.ebi.ac.uk/interpro

neXtProt www.nextprot.org

PDB www.wwpdb.org

Pfam pfam.xfam.org

SCOP2 scop2.mrc-lmb.cam.ac.uk

UniProtKB www.uniprot.org

Predictionofproteinstructure

BETAWARE biocomp.unibo.it/savojard/betawarecl

BOCTOPUS2 boctopus.bioinfo.se

PolyPhobius%20phobius.sbc.su.se/poly.html

POODLE cblab.my-pharm.ac.jp/poodle

PrDOS prdos.hgc.jp/cgi-bin/top.cgi

Proteus wks80920.ccis.ualberta.ca/proteus

Proteus2 www.proteus2.ca/proteus2

PSIPRED,MEMSAT-SVM,and

DISOPRED3

bioinf.cs.ucl.ac.uk/psipred

RaptorX raptorx.uchicago.edu/StructurePropertyPred/predict

ReProf,TMSEG,andMeta-Disorder predictprotein.org

SPIDER3 sparks-lab.org/server/SPIDER3

SSpro5,ACCpro5 scratch.proteomics.ics.uci.edu

Predictionofproteinfunction

CADD cadd.gs.washington.edu

DeepLoc www.cbs.dtu.dk/services/DeepLoc

DomCut www.bork.embl-heidelberg.de/~suyama/domcut

DomPred,FFPred3.0,COGIC bioinf.cs.ucl.ac.uk/psipred

DOMpro scratch.proteomics.ics.uci.edu

DP-Bind lcg.rit.albany.edu/dp-bind

216 Predictive Methods Using Protein Sequences

FunFams www.cathdb.info/search/by_sequence

HomPPI ailab1.ist.psu.edu/PSHOMPPIv1.3

HomPRIP-NB%20ailab1.ist.psu.edu/HomPRIP-NB/index.html

LocTree3 rostlab.org/services/loctree3

MultiLoc2 abi-services.informatik.uni-tuebingen.de/multiloc2/webloc.cgi

PolyPhen-2 genetics.bwh.harvard.edu/pph2

Pprint crdd.osdd.net/raghava/pprint

PROVEAN provean.jcvi.org/index.php

PSIVER mizuguchilab.org/PSIVER

RNABindRPlus ailab1.ist.psu.edu/RNABindRPlus

ScoobyDomain www.ibi.vu.nl/programs/scoobywww

SIFT sift.bii.a-star.edu.sg

SNAP2 rostlab.org/services/snap2web

SomeNA,Metastudent,Ofran,and

RostPPIpredictor

www.predictprotein.org

Further Reading

Keskin,O.,Tuncbag,N.,andGursoy,A.(2016).Predictingprotein-proteininteractionsfromthe

moleculartotheproteomelevel. Chem.Rev. 116:4884–4909.Keskinetal.giveanexpansive

overviewofproteinbindinginallitsfacets,cove ringprotein–proteinandprotein–nucleicacid

bindingontheproteinandresiduelevel,aswellasadditionaltopicsnotcoveredinthischapter,

suchasdockingandotherpredictionalgorithmsbasedonproteinstructureinsteadofsequence.

Moult,J.,Fidelis,K.,Kryshtafovych,A.etal.(2016).Criticalassessmentofmethodsofprotein

structureprediction:progressandnewdirectionsinroundXI. Proteins84(Suppl1):4–14.This

isthemostrecentevaluationoftheCASPexperiment,whichisanindependentassessmentof

allmajoraspectsofproteinstructureprediction.Similarly,readersinterestedinfunction

predictionshouldinvestigateCAFA(Jiangetal.2016)andCAGI(seeInternetResources)for

varianteffectprediction.

References

Adzhubei,I.A.,Schmidt,S.,Peshkin,L.etal.(2010).Amethodandserverforpredictingdamaging

missensemutations. Nat.Methods. 7:248–249.

Ahmad,S.andSarai,A.(2005).PSSM-basedpredictionofDNAbindingsitesinproteins. BMC

Bioinf.6:33.

Akiva,E.,Brown,S.,Almonacid,D.E.etal.(2014).Thestructure-functionlinkagedatabase.

NucleicAcidsRes. 42:521–530.

Allis,C.D.andJenuwein,T.(2016).Themolecularhallmarksofepigeneticcontrol. Nat.Rev.

Genet.17:487–500.

AlmagroArmenteros,J.J.,Sønderby,C.K.,Sønderby,S.K.etal.(2017).DeepLoc:predictionof

proteinsubcellularlocalizationusingdeeplearning. Bioinformatics.33:3387–3395.

Aloy,P.,Stark,A.,Hadley,C.,andRussell,R.B.(2003).Predictionswithouttemplates:newfolds,

secondarystructure,andcontactsinCASP5. ProteinsStruct.Funct.Genet. 53

(Suppl6):436–456.

Altschul,S.F.andGish,W.(1996).Localalignmentstatistics. MethodsEnzymol. 266:460–480.

Andreeva,A.,Howorth,D.,Chothia,C.etal.(2014).SCOP2prototype:anewapproachtoprotein

structuremining. NucleicAcidsRes. 42:310–314.

References 217

Anfinsen,C.B.(1973).Principlesthatgovernthefoldingofproteinchains. Science.181:223–230.

Ashkenazi,S.,Snir,R.,andOfran,Y.(2012).Assessingtherelationshipbetweenconservationof

functionandconservationofsequenceusingphotosyntheticproteins. Bioinformatics.28:

3203–3210.

Attwood,T.K.,Coletta,A.,Muirhead,G.etal.(2012).ThePRINTSdatabase:afine-grainedprotein

sequenceannotationandanalysisresource-itsstatusin2012. Database.2012:1–9.

Auton,A.,Abecasis,G.R.,Altshuler,D.M.etal.(2015).Aglobalreferenceforhumangenetic

variation.Nature.526:68–74.

Bairoch,A.andBoeckmann,B.(1994).TheSWISS-PROTproteinsequencedatabank:current

status.NucleicAcidsRes. 22:3578–3580.

Berman,H.M.,Westbrook,J.,Feng,Z.etal.(2000).Theproteindatabank. NucleicAcidsRes. 28:

235–242.

Bernhofer,M.,Kloppmann,E.,Reeb,J.,andRost,B.(2016).TMSEG:novelpredictionof

transmembranehelices. Proteins84:1706–1716.

Blum,T.,Briesemeister,S.,andKohlbacher,O.(2009).MultiLoc2:integratingphylogenyandgene

ontologytermsimprovessubcellularproteinlocalizationprediction. BMCBioinf. 10:274.

Boutet,E.,Lieberherr,D.,Tognolli,M.etal.(2016).UniProtKB/Swiss-Prot,themanually

annotatedsectionoftheUniProtknowledgeBase:howtousetheentryview. MethodsMol.Biol.

1374:23–54.

Bru,C.,Courcelle,E.,Carrère,S.etal.(2005).TheProDomdatabaseofproteindomainfamilies:

moreemphasison3D. NucleicAcidsRes. 33:212–215.

Bryson,K.,Cozzetto,D.,andJones,D.T.(2007).Computer-assistedproteindomainboundary

predictionusingtheDomPredserver. Curr.ProteinPept.Sci. 8:181–188.

Buchan,D.W.A.,Minneci,F.,Nugent,T.C.O.etal.(2013).ScalablewebservicesforthePSIPRED

proteinanalysisworkbench. NucleicAcidsRes. 41:349–357.

Chen,X.W.andJeong,J.C.(2009).Sequence-basedpredictionofproteininteractionsiteswithan

integrativemethod. Bioinformatics.25:585–591.

Chen,P.andLi,J.(2010).Sequence-basedidentificationofinterfaceresiduesbyanintegrative

profilecombininghydrophobicandevolutionaryinformation. BMCBioinf. 11:402.

Chen,C.P.,Kernytsky,A.,andRost,B.(2002).Transmembranehelixpredictionsrevisited. Protein

Sci.11:2774–2791.

Cheng,J.,Sweredoski,M.J.,andBaldi,P.(2006).DOMpro:proteindomainpredictionusing

profiles,secondarystructure,relativesolventaccessibility,andrecursiveneuralnetworks. Data

Min.Knowl.Discovery 13:1–10.

Choi,Y.,Sims,G.E.,Murphy,S.etal.(2012).Predictingthefunctionaleffectofaminoacid

substitutionsandindels. PLoSOne 7(10):e46688.

Chou,P.Y.andFasman,G.D.(1974).Predictionofproteinconformation. Biochemistry.13(2):

222–245.

Claros,M.G.andVonHeijne,G.(1994).TopPredII:animprovedsoftwareformembraneprotein

structurepredictions. Comput.Appl.Biosci. 10:685–686.

Coleman,J.L.J.,Ngo,T.,andSmith,N.J.(2017).TheGprotein-coupledreceptorN-terminusand

receptorsignalling:N-teringanewera. Cell.Signalling. 33:1–9.

Cozzetto,D.,Buchan,D.W.A.,Bryson,K.,andJones,D.T.(2013).Proteinfunctionpredictionby

massiveintegrationofevolutionaryanalysesandmultipledatasources. BMCBioinf. 14:S1.

Cozzetto,D.,Minneci,F.,Currant,H.,andJones,D.T.(2016).FFPred3:feature-basedfunction

predictionforallgeneontologydomains. Sci.Rep. 6:31865.

Crick,F.H.(1958).Onproteinsynthesis. Symp.Soc.Exp.Biol. 12:138–163.

Cukuroglu,E.,Gursoy,A.,Nussinov,R.,andKeskin,O.(2014).Non-redundantuniqueinterface

structuresastemplatesformodelingproteininteractions. PLoSOne. 9:e86738.

Das,S.,Lee,D.,Sillitoe,I.etal.(2015).FunctionalclassificationofCATHsuperfamilies:a

domain-basedapproachforproteinfunctionannotation. Bioinformatics.31:3460–3467.

Deng,X.,Gumm,J.,Karki,S.etal.(2015).Anoverviewofpracticalapplicationsofproteindisorder

predictionanddriveforfaster,moreaccuratepredictions. Int.J.Mol.Sci. 16:15384–15404.

218 Predictive Methods Using Protein Sequences

Dong,Q.,Wang,X.,Lin,L.,andXu,Z.(2006).Domainboundarypredictionbasedonprofile

domainlinkerpropensityindex. Comput.Biol.Chem. 30:127–133.

Dong,C.,Wei,P.,Jian,X.etal.(2015).Comparisonandintegrationofdeleteriousnessprediction

methodsfornonsynonymousSNVsinwholeexomesequencingstudies. Hum.Mol.Genet. 24:

2125–2137.

Eddy,S.R.(2011).AcceleratedprofileHMMsearches. PLoSComput.Biol. 7:e1002195.

Elbarbary,R.A.,Lucas,B.A.,andMaquat,L.E.(2016).Retrotransposonsasregulatorsofgene

expression.Science.351:aac7247.

Esmaielbeiki,R.,Krawczyk,K.,Knapp,B.etal.(2016).Progressandchallengesinpredicting

proteininterfaces. BriefingsBioinf. 17:117–131.

Eyrich,V.,Martí-Renom,M.A.,Przybylski,D.etal.(2001).EVA:continuousautomaticevaluation

ofproteinstructurepredictionservers. Bioinformatics.17:1242–1243.

Ezkurdia,L.,Grana,O.,Izarzugaza,J.M.G.,andTress,M.L.(2009).Assessmentofdomain

boundarypredictionsandthepredictionofintramolecularcontactsinCASP8. ProteinsStruct.

Funct.Bioinf. 77:196–209.

Fagerberg,L.,Jonasson,K.,andHeijne,G.V.(2010).Predictionofthehumanmembrane

proteome.Proteomics.10:1141–1149.

Fariselli,P.,Savojardo,C.,Martelli,P.L.,andCasadio,R.(2009).Grammatical-restrainedhidden

conditionalrandomfieldsforbioinformaticsapplications. AlgorithmsMol.Biol. 4:13.

Fidelis,K.,Rost,B.,andZemla,A.(1999).AmodifieddefinitionofSov,asegment-basedmeasure

forproteinsecondarystructurepredictionassessment. Proteins.223:220–223.

Finn,R.D.,Bateman,A.,Clements,J.etal.(2014a).Pfam:theproteinfamiliesdatabase. Nucleic

AcidsRes. 42:222–230.

Finn,R.D.,Miller,B.L.,Clements,J.,andBateman,A.(2014b).IPfam:adatabaseofproteinfamily

anddomaininteractionsfoundintheproteindataBank. NucleicAcidsRes. 42:364–373.

Finn,R.D.,Attwood,T.K.,Babbitt,P.C.etal.(2016).InterProin2017-beyondproteinfamilyand

domainannotations. NucleicAcidsRes. 45:gkw1107.

Fischer,T.B.,Arunachalam,K.V.,Bailey,D.etal.(2003).Thebindinginterferencedatabase(BID):

acompilationofaminoacidhotspotsinproteininterfaces. Bioinformatics.19:1453–1454.

Foster,L.J.,deHoog,C.L.,Zhang,Y.etal.(2006).Amammalianorganellemapbyprotein

correlationprofiling. Cell.125(1):187–199.

Frishman,D.andArgos,P.(1995).Knowledge-basedproteinsecondarystructureassignment.

ProteinsStruct.Funct.Genet. 23(4):566–579.

Fukuchi,S.,Amemiya,T.,Sakamoto,S.etal.(2014).IDEALin2014illustratesinteraction

networkscomposedofintrinsicallydisorderedproteinsandtheirbindingpartners. Nucleic

AcidsRes. 42:320–325.

Galperin,M.Y.,Makarova,K.S.,Wolf,Y.I.,andKoonin,E.V.(2015).Expandedmicrobialgenome

coverageandimprovedproteinfamilyannotationintheCOGdatabase. NucleicAcidsRes. 43:

D261–D269.

Gardy,J.L.andBrinkman,F.S.(2006).Methodsforpredictingbacterialproteinsubcellular

localization.Nat.Rev.Microbiol. 4(10):741–751.

Garnier,J.,Osguthorpe,D.J.,andRobson,B.(1978).Analysisoftheaccuracyandimplicationsof

simplemethodsforpredictingthesecondarystructureofglobularproteins. J.Mol.Biol. 120:

97–120.

Garnier,J.,Gibrat,J.-F.,andRobson,B.(1996).GORmethodforpredictingproteinsecondary

structurefromaminoacidsequence. MethodsEnzymol. 266:540–553.

Garrow,A.G.,Agnew,A.,andWesthead,D.R.(2005).TMB-Hunt:awebservertoscreensequence

setsfortransmembranebeta-barrelproteins. NucleicAcidsRes. 33(Suppl2):188–192.

Gaudet,P.,Michel,P.A.,Zahn-Zabal,M.etal.(2017).TheneXtProtknowledgebaseonhuman

proteins:2017update. NucleicAcidsRes. 45(D1):D177–D182.

GeneOntologyConsortium(2000).Geneontology:toolfortheunificationofbiology. Nat.Genet.

25:25–29.

References 219

GeneOntologyConsortium(2015).GeneOntologyConsortium:goingforward. NucleicAcidsRes.

43:D1049–D1056.

Goldberg,T.,Hamp,T.,andRost,B.(2012).LocTree2predictslocalizationforalldomainsoflife.

Bioinformatics.28:i458–i465.

Goldberg,T.,Hecht,M.,Hamp,T.etal.(2014).LocTree3predictionoflocalization. NucleicAcids

Res.42(WebServerissue):1–6.

Goodwin,S.,McPherson,J.D.,andMcCombie,W.R.(2016).Comingofage:tenyearsof

next-generationsequencingtechnologies. Nat.Rev.Genet. 17:333–351.

Gouw,M.,Michael,S.,Samano-Sanchez,H.etal.(2018).Theeukaryoticlinearmotif

resource–2018update. NucleicAcidsRes. 46(D1):D428–D434.

Graessel,A.,Hauck,S.M.,vonToerne,C.etal.(2015).Acombinedomicsapproachtogeneratethe

surfaceatlasofhumannaiveCD4 +TcellsduringearlyT-cellreceptoractivation. Mol.Cell.

Proteomics.14(8):2085–2102.

Greene,L.H.,Lewis,T.E.,Addou,S.etal.(2007).TheCATHdomainstructuredatabase:new

protocolsandclassificationlevelsgiveamorecomprehensiveresourceforexploringevolution.

NucleicAcidsRes. 35:291–297.

Grimm,D.G.,Azencott,C.A.,Aicheler,F.etal.(2015).Theevaluationoftoolsusedtopredictthe

impactofmissensevariantsishinderedbytwotypesofcircularity. Hum.Mutat. 36:513–523.

Habchi,J.,Tompa,P.,Longhi,S.,andUversky,V.N.(2014).Introducingproteinintrinsicdisorder.

Chem.Rev. 114:6561–6588.

Haft,D.H.,Selengut,J.D.,Richter,R.A.etal.(2013).TIGRFAMsandgenomepropertiesin2013.

NucleicAcidsRes. 41:387–395.

Hamp,T.andRost,B.(2012).Alternativeprotein-proteininterfacesarefrequentexceptions. PLoS

Comput.Biol. 8(8):e1002623.

Hamp,T.,Kassner,R.,Seemayer,S.etal.(2013).Homology-basedinferencesetsthebarhighfor

proteinfunctionprediction. BMCBioinf. 14(Suppl3):S7.

Hayat,S.,Peters,C.,Shu,N.etal.(2016).Inclusionofdyad-repeatpatternimprovestopology

predictionoftransmembrane β-barrelproteins. Bioinformatics.32:1571–1573.

Hecht,M.,Bromberg,Y.,andRost,B.(2013).Newsfromtheproteinmutabilitylandscape.

J.Mol.Biol. 425(21):3937–3948.

Hecht,M.,Bromberg,Y.,andRost,B.(2015).Betterpredictionoffunctionaleffectsforsequence

variants.BMCGenomics. 16(Suppl8):S1.

Heffernan,R.,Paliwal,K.,Lyons,J.etal.(2015).Improvingpredictionofsecondarystructure,

localbackboneangles,andsolventaccessiblesurfaceareaofproteinsbyiterativedeeplearning.

Sci.Rep. 5:11476.

Heffernan,R.,Yang,Y.,Paliwal,K.,andZhou,Y.(2017).Capturingnon-localinteractionsbylong

short-termmemorybidirectionalrecurrentneuralnetworksforimprovingpredictionofprotein

secondarystructure,backboneangles,contactnumbersandsolventaccessibility.

Bioinformatics.33(18):2842–2849.

vonHeijne,G.(2006).Membrane-proteintopology. Nat.Rev.Mol.CellBiol. 7:909–918.

Heinig,M.andFrishman,D.(2004).STRIDE:awebserverforsecondarystructureassignment

fromknownatomiccoordinatesofproteins. NucleicAcidsRes. 32(WebServerissue):500–502.

Hirose,S.,Shimizu,K.,Kanai,S.etal.(2007).StructuralbioinformaticsPOODLE-L:atwo-level

SVMpredictionsystemforreliablypredictinglongdisorderedregions. Struct.Bioinf. 23:

2046–2053.

Hirose,S.,Shimizu,K.,andNoguchi,T.(2010).POODLE-I:disorderedregionpredictionby

integratingPOODLEseriesandstructuralinformationpredictorsbasedonaworkflow

approach.InSilicoBiol. 10:185–191.

Hönigschmid,P.(2012). ImprovementofDNA-andRNA-proteinbindingprediction .Diploma

thesis.TUM–TechnicalUniversityofMunich.

Hopf,T.A.,Colwell,L.J.,Sheridan,R.etal.(2012).Three-dimensionalstructuresofmembrane

proteinsfromgenomicsequencing. Cell.149:1607–1621.

220 Predictive Methods Using Protein Sequences

Hopf,T.A.,Schärfe,C.P.I.,Rodrigues,J.P.G.L.M.etal.(2014).Sequenceco-evolutiongives3D

contactsandstructuresofproteincomplexes. eLife.3:e03430.

Horton,P.,Park,K.J.,Obayashi,T.etal.(2007).WoLFPSORT:proteinlocalizationpredictor.

NucleicAcidsRes. 35(WebServerissue):W585–W587.

Hu,Y.,Lehrach,H.,andJanitz,M.(2009).Comparativeanalysisofanexperimentalsubcellular

proteinlocalizationassayandinsilicopredictionmethods. J.Mol.Histol. 40(5–6):343–352.

Hubbard,S.J.andThornton,J.M.(1993). NACCESS.DepartmentofBiochemistryandMolecular

Biology.UniversityCollegeLondon.

Huerta-Cepas,J.,Szklarczyk,D.,Forslund,K.etal.(2016).EGGNOG4.5:ahierarchicalorthology

frameworkwithimprovedfunctionalannotationsforeukaryotic,prokaryoticandviral

sequences.NucleicAcidsRes. 44:D286–D293.

Huh,W.K.,Falvo,J.V.,Gerke,L.C.etal.(2003).Globalanalysisofproteinlocalizationinbudding

yeast.Nature.425(6959):686–691.

Hwang,S.,Guo,Z.,andKuznetsov,I.B.(2007).DP-bind:awebserverforsequence-based

predictionofDNA-bindingresiduesinDNA-bindingproteins. Bioinformatics.23:

634–636.

Ishida,T.andKinoshita,K.(2007).PrDOS:predictionofdisorderedproteinregionsfromamino

acidsequence. NucleicAcidsRes. 35:460–464.

Ishida,T.andKinoshita,K.(2008).Predictionofdisorderedregionsinproteinsbasedonthemeta

approach.Bioinformatics.24(11):1344–1348.

Jacoby,E.,Bouhelal,R.,Gerspacher,M.,andSeuwen,K.(2006).The7TMG-protein-coupled

receptortargetfamily. ChemMedChem.1:761–782.

Jensen,L.J.andBateman,A.(2011).Theriseandfallofsupervisedmachinelearningtechniques.

Bioinformatics.27:3331–3332.

Jia,Y.andLiu,X.-Y.(2006).Fromsurfaceself-assemblytocrystallization:predictionofprotein

crystallizationconditions. J.Phys.Chem.B. 110:6949–6955.

Jiang,Y.,Oron,T.R.,Clark,W.T.etal.(2016).Anexpandedevaluationofproteinfunction

predictionmethodsshowsanimprovementinaccuracy. GenomeBiol 17(1):184.

Jones,D.T.(1999).Proteinsecondarystructurepredictionbasedonposition-specificscoring

matrices.J.Mol.Biol. 292:195–202.

Jones,D.T.andCozzetto,D.(2015).DISOPRED3:precisedisorderedregionpredictionswith

annotatedprotein-bindingactivity. Bioinformatics.31:857–863.

Jones,P.,Binns,D.,Chang,H.Y.etal.(2014).InterProScan5:genome-scaleproteinfunction

classification.Bioinformatics.30:1236–1240.

Joo,K.,Lee,S.J.,andLee,J.(2012).SANN:solventaccessibilitypredictionofproteinsbynearest

neighbormethod. Proteins80(7):1791–1797.

Kabsch,W.andSander,C.(1983).Dictionaryofproteinsecondarystructure:patternrecognition

ofhydrogen-bondedandgeometricalfeatures. Biopolymers.22:2577–2637.

Kajan,L.,Yachdav,G.,Vicedo,E.etal.(2013).Cloudpredictionofproteinstructureandfunction

withPredictProteinforDebian. Biomed.Res.Int. 2013:398968.

Käll,L.,Krogh,A.,andSonnhammer,E.L.L.(2004).Acombinedtransmembranetopologyand

signalpeptidepredictionmethod. J.Mol.Biol. 338:1027–1036.

Käll,L.,Krogh,A.,andSonnhammer,E.L.L.(2005).AnHMMposteriordecoderforsequence

featurepredictionthatincludeshomologyinformation. Bioinformatics.21:i251.

Keskin,O.,Tuncbag,N.,andGursoy,A.(2016).Predictingprotein-proteininteractionsfromthe

moleculartotheproteomelevel. Chem.Rev. 116:4884–4909.

Kessel,A.andBen-Tal,N.(2011). IntroductiontoProteins ,438–440.London,UK:CRCPress.

Kihara,D.(2005).Theeffectoflong-rangeinteractionsonthesecondarystructureformationof

proteins.ProteinSci. 14:1955–1963.

Kinch,L.N.,Li,W.,Monastyrskyy,B.etal.(2016).EvaluationoffreemodelingtargetsinCASP11

andROLL. Proteins84(Suppl1):51–66.

Kircher,M.,Witten,D.M.,Jain,P.etal.(2014).Ageneralframeworkforestimatingtherelative

pathogenicityofhumangeneticvariants. Nat.Genet.

46:310–315.

References 221

Klimke,W.,Agarwala,R.,Badretdin,A.etal.(2009).TheNationalCenterforBiotechnology

Information’sproteinclustersdatabase. NucleicAcidsRes. 37:216–223.

Kloppmann,E.,Punta,M.,andRost,B.(2012).Structuralgenomicspluckshigh-hanging

membraneproteins. Curr.Opin.Struct.Biol. 22:326–332.

Köhler,S.,Vasilevsky,N.A.,Engelstad,M.etal.(2016).Thehumanphenotypeontologyin2017.

NucleicAcidsRes. 45:gkw1039.

Krissinel,E.andHenrick,K.(2007).Inferenceofmacromolecularassembliesfromcrystalline

state.J.Mol.Biol. 372:774–797.

Krogh,A.,Larsson,B.,vonHeijne,G.,andSonnhammer,E.L.(2001).Predictingtransmembrane

proteintopologywithahiddenMarkovmodel:applicationtocompletegenomes. J.Mol.Biol.

305:567–580.

Kumar,M.,Gromiha,M.M.,andRaghava,G.P.S.(2008).PredictionofRNAbindingsitesina

proteinusingSVMandPSSMprofile. Proteins.71:189–194.

Kumar,P.,Henikoff,S.,andNg,P.C.(2009).Predictingtheeffectsofcodingnon-synonymous

variantsonproteinfunctionusingtheSIFTalgorithm. Nat.Protoc. 4:1073–1081.

Kyte,J.andDoolittle,R.F.(1982).Asimplemethodfordisplayingthehydropathiccharacterofa

protein.J.Mol.Biol. 157:105–132.

Lam,S.D.,Dawson,N.L.,Das,S.etal.(2016).Gene3D:expandingtheutilityofdomain

assignments.NucleicAcidsRes. 44:D404–D409.

deLasRivas,J.andFontanillo,C.(2010).Protein-proteininteractionsessentials:keyconceptsto

buildingandanalyzinginteractomenetworks. PLoSComput.Biol. 6:1–8.

Lee,B.andRichards,F.M.(1971).Theinterpretationofproteinstructures:estimationofstatic

accessibility.J.Mol.Biol. 55(3):379–400.

vanderLee,R.,Buljan,M.,Lang,B.etal.(2014).Classificationofintrinsicallydisorderedregions

andproteins. Chem.Rev. 114:6589–6631.

Letunic,I.,Doerks,T.,andBork,P.(2015).SMART:recentupdates,newdevelopmentsandstatus

in2015. NucleicAcidsRes. 43:D257–D260.

Liu,J.andRost,B.(2001).Comparingfunctionandstructurebetweenentireproteomes. Protein

Sci.10:1970–1979.

Liu,J.andRost,B.(2003).Domains,motifsandclustersintheproteinuniverse. Curr.Opin.Chem.

Biol.7:5–11.

Liu,J.andRost,B.(2004).CHOPproteinsintostructuraldomain-likefragments. ProteinsStruct.

Funct.Genet. 55:678–688.

Lobanov,M.Y.andGalzitskaya,O.V.(2015).Howcommonisdisorder?Occurrenceofdisordered

residuesinfourdomainsoflife. Int.J.Mol.Sci. 16:19490–19507.

Lobley,A.(2010).HumanProteinFunctionPrediction:applicationofmachinelearningfor

integrationofheterogeneousdatasources.PhDthesis.UniversityCollegeLondon,Lo0ndon,

UK.

Magnan,C.N.andBaldi,P.(2014).SSpro/ACCpro5:almostperfectpredictionofprotein

secondarystructureandrelativesolventaccessibilityusingprofiles,machinelearningand

structuralsimilarity. Bioinformatics.30:2592–2597.

Mahlich,Y.,Reeb,J.,Schelling,M.etal.(2017).Commonsequencevariantsaffectmolecular

functionmorethanrarevariants. Sci.Rep. 7:1608.

Marks,D.S.,Colwell,L.J.,Sheridan,R.etal.(2011).Protein3Dstructurecomputedfrom

evolutionarysequencevariation. PLoSOne 6:e28766.

Marks,D.S.,Hopf,T.A.,Chris,S.,andSander,C.(2012).Proteinstructurepredictionfrom

sequencevariation. Nat.Biotechnol. 30:1072–1080.

Martinez,D.A.andNelson,M.A.(2010).Thenextgenerationbecomesthenowgeneration. PLoS

Genet.6:e1000906.

Mi,H.,Poudel,S.,Muruganujan,A.etal.(2016).PANTHERversion10:expandedproteinfamilies

andfunctions,andanalysistools. NucleicAcidsRes. 44:D336–D342.

Miosge,L.A.,Field,M.A.,Sontani,Y.etal.(2015).Comparisonofpredictedandactual

consequencesofmissensemutations. Proc.Natl.Acad.Sci.USA. 112:E5189–E5198.

222 Predictive Methods Using Protein Sequences

Mirabello,C.andPollastri,G.(2013).Porter,PaleAle4.0:high-accuracypredictionofprotein

secondarystructureandrelativesolventaccessibility. Bioinformatics.29(16):2056–2058.

Monastyrskyy,B.,Kryshtafovych,A.,Moult,J.etal.(2014).Assessmentofproteindisorderregion

predictionsinCASP10. ProteinsStruct.Funct.Bioinf. 82:127–137.

Montgomerie,S.,Sundararaj,S.,Gallin,W.J.,andWishart,D.S.(2006).Improvingtheaccuracyof

proteinsecondarystructurepredictionusingstructuralalignment. BMCBioinf. 7:301–301.

Montgomerie,S.,Cruz,J.A.,Shrivastava,S.etal.(2008).PROTEUS2:awebserverfor

comprehensiveproteinstructurepredictionandstructure-basedannotation. NucleicAcidsRes.

36(WebServerissue):202–209.

Mooney,C.,Cessieux,A.,Shields,D.C.,andPollastri,G.(2013).SCL-Epred:ageneraliseddenovo

eukaryoticproteinsubcellularlocalisationpredictor. AminoAcids. 45(2):291–299.

Morrow,J.K.andZhang,S.(2012).Computationalpredictionofproteinhotspotresidues. Curr.

Pharm.Des. 18:1255–1265.

Mosca,R.,Céol,A.,Stein,A.etal.(2014).3did:acatalogofdomain-basedinteractionsofknown

three-dimensionalstructure. NucleicAcidsRes. 42:374–379.

Moult,J.,Pedersen,J.T.,Judson,R.,andFidelis,K.(1995).Alarge-scaleexperimenttoassess

proteinstructurepredictionmethods. ProteinsStruct.Funct.Genet. 23:ii–iv.

Murakami,Y.andMizuguchi,K.(2010).ApplyingtheNaiveBayesclassifierwithkerneldensity

estimationtothepredictionofprotein-proteininteractionsites. Bioinformatics.26:1841–1848.

Nair,R.andRost,B.(2002).Inferringsub-cellularlocalisationthroughautomatedlexicalanalysis.

Bioinformatics.18(Suppl1):S78–S86.

Necci,M.,Piovesan,D.,Dosztányi,Z.,andTosatto,S.C.E.(2017).MobiDB-lite:fastandhighly

specificconsensuspredictionofintrinsicdisorderinproteins. Bioinformatics.33:btx015.

Ng,P.C.andHenikoff,S.(2003).SIFT:predictingaminoacidchangesthataffectproteinfunction.

NucleicAcidsRes. 31:3812–3814.

Nugent,T.andJones,D.T.(2009).Transmembraneproteintopologypredictionusingsupport

vectormachines. BMCBioinf. 10:159.

Nugent,T.andJones,D.T.(2012).Accuratedenovostructurepredictionoflargetransmembrane

proteindomainsusingfragment-assemblyandcorrelatedmutationanalysis. Proc.Natl.Acad.

Sci.USA 109:E1540–E1547.

Oates,M.E.,Romero,P.,Ishida,T.etal.(2013).D2P2:databaseofdisorderedproteinpredictions.

NucleicAcidsRes. 41:508–516.

Oates,M.E.,Stahlhacke,J.,Vavoulis,D.V.etal.(2015).TheSUPERFAMILY1.75databasein2014:

adoublingofdata. NucleicAcidsRes. 43:D227–D233.

O’Donovan,C.,Martin,M.J.,Gattiker,A.etal.(2002).High-qualityproteinknowledgeresource:

SWISS-PROTandTrEMBL. BriefingsBioinf. 3:275–284.

Ofran,Y.andRost,B.(2003a).Analysingsixtypesofprotein-proteininterfaces. J.Mol.Biol. 325:

377–387.

Ofran,Y.andRost,B.(2003b).Predictedprotein-proteininteractionsitesfromlocalsequence

information.FEBSLett. 544:236–239.

Ofran,Y.andRost,B.(2007).ISIS:interactionsitesidentifiedfromsequence. Bioinformatics.23

(2):e13–e16.

Overington,J.,Al-Lazikani,B.,andHopkins,A.L.(2006).Howmanydrugtargetsarethere? Nat.

Rev.DrugDiscov. 5:993–996.

Pang,C.I.,Lin,K.,Wouters,M.A.etal.(2008).Identifyingfoldableregionsinproteinsequence

fromthehydrophobicsignal. NucleicAcidsRes. 36:578–588.

Pawson,T.andNash,P.(2003).Assemblyofcellregulatorysystemsthroughproteininteraction

domains.Science.300(5618):445–452.

Pedruzzi,I.,Rivoire,C.,Auchincloss,A.H.etal.(2013).HAMAPin2013,newdevelopmentsin

theproteinfamilyclassificationandannotationsystem. NucleicAcidsRes. 41:584–589.

Pedruzzi,I.,Rivoire,C.,Auchincloss,A.H.etal.(2015).HAMAPin2015:updatestotheprotein

familyclassificationandannotationsystem. NucleicAcidsRes. 43:D1064–D1070.

References 223

Piovesan,D.,Tabaro,F.,Mi ˇceti´c,I.etal.(2016).DisProt7.0:amajorupdateofthedatabaseof

disorderedproteins. NucleicAcidsRes. 45:gkw1056.

Pollastri,G.,Przybylski,D.,Rost,B.,andBaldi,P.(2002).Improvingthepredictionofprotein

secondarystructureinthreeandeightclassesusingrecurrentneuralnetworksandprofiles.

ProteinsStruct.Funct.Bioinf. 47:228–235.

Punta,M.andRost,B.(2008).Neuralnetworkspredictproteinstructureandfunction. Methods

Mol.Biol. 458:203–230.

Radivojac,P.,Clark,W.T.,Oron,T.R.etal.(2013).Alarge-scaleevaluationofcomputational

proteinfunctionprediction. Nat.Methods. 10:221–227.

Ramilowski,J.A.,Goldberg,T.,Harshbarger,J.etal.(2015).Adraftnetworkof

ligand-receptor-mediatedmulticellularsignallinginhuman. Nat.Commun. 6:7866.

Rao,V.S.,Srinivas,K.,Sujini,G.N.,andKumar,G.N.S.(2014).Protein-proteininteraction

detection:methodsandanalysis. Int.J.Proteomics. 2014:1–12.

Reeb,J.,Kloppmann,E.,Bernhofer,M.,andRost,B.(2014).Evaluationoftransmembranehelix

predictionsin2014. ProteinsStruct.Funct.Bioinf. 83:473–484.

Reeb,J.,Hecht,M.,Mahlich,Y.etal.(2016).Predictedmoleculareffectsofsequencevariantslink

tosystemlevelofdisease. PLoSComput.Biol. 12(8):e1005047.

Remmert,M.,Biegert,A.,Hauser,A.,andSöding,J.(2012).HHblits:lightning-fastiterative

proteinsequencesearchingbyHMM-HMMalignment. Nat.Methods. 9(2):173–175.

Res,I.,Mihalek,I.,andLichtarge,O.(2005).Anevolutionbasedclassifierforpredictionofprotein

interfaceswithoutusingproteinstructures. Bioinformatics.21(10):2496–2501.

Rezácová,P.,Borek,D.,Moy,S.F.etal.(2008).Crystalstructureandputativefunctionofsmall

Toprimdomain-containingproteinfromBacillusstearothermophilus. Proteins.70:311–319.

Rose,P.W.,Prli ´c,A.,Altunkaya,A.etal.(2017).TheRCSBproteindatabank:integrativeviewof

protein,geneand3Dstructuralinformation. NucleicAcidsRes. 45(D1):D271–D281.

Rost,B.(1996).PHD:predictingone-dimensionalproteinstructurebyprofilebasedneural

networks.MethodsEnzymol. 266:525–539.

Rost,B.(2001).Proteinsecondarystructurepredictioncontinuestorise. J.Struct.Biol. 134:

204–218.

Rost,B.(2002).Enzymefunctionlessconservedthananticipated. J.Mol.Biol. 318:595–608.

Rost,B.andSander,C.(1993).Improvedpredictionofproteinsecondarystructurebyuseof

sequenceprofilesandneuralnetworks. Proc.Natl.Acad.Sci.USA 90:7558–7562.

Rost,B.andSander,C.(1994a).Combiningevolutionaryinformationandneuralnetworksto

predictproteinsecondarystructure. ProteinsStruct.Funct.Bioinf. 19:55–72.

Rost,B.andSander,C.(1994b).Conservationandpredictionofsolventaccessibilityinprotein

families.ProteinsStruct.Funct.Genet. 20(3):216–226.

Rost,B.,Yachdav,G.,andLiu,J.(2004).ThePredictProteinserver. NucleicAcidsRes. 32(Suppl2):

W321–W326.

Rychlewski,L.andFischer,D.(2005).LiveBench-8:thelarge-scale,continuousassessmentof

automatedproteinstructureprediction. ProteinSci. 14(1):240–245.

Savojardo,C.,Fariselli,P.,andCasadio,R.(2013).BETAWARE:amachine-learningtooltodetect

andpredicttransmembranebeta-barrelproteinsinprokaryotes. Bioinformatics.29:

504–505.

Schaarschmidt,J.,Monastyrskyy,B.,Kryshtafovych,A.,andBonvin,A.M.J.J.(2018).Assessment

ofcontactpredictionsinCASP12:co-evolutionanddeeplearningcomingofage. Proteins86

(Suppl1):51–66.

Schlessinger,A.,Schaefer,C.,Vicedo,E.etal.(2011).Proteindisorder–abreakthroughinvention

ofevolution? Curr.Opin.Struct.Biol. 21:412–418.

SchrodingerLLC.(2015).ThePyMOLMolecularGraphicsSystem,Version1.9.

Shimizu,K.,Hirose,S.,andNoguchi,T.(2007).StructuralbioinformaticsPOODLE-S:web

applicationforpredictingproteindisorderbyusingphysicochemicalfeaturesandreduced

aminoacidsetofaposition-specificscoringmatrix. Struct.Bioinf. 23:2337–2338.

224 Predictive Methods Using Protein Sequences

Shoemaker,B.A.,Zhang,D.,Tyagi,M.etal.(2012).IBIS(inferredbiomolecularinteractionserver)

reports,predictsandintegratesmultipletypesofconservedinteractionsforproteins. Nucleic

AcidsRes. 40:834–840.

Sigrist,C.J.A.,DeCastro,E.,Cerutti,L.etal.(2013).Newandcontinuingdevelopmentsat

PROSITE.NucleicAcidsRes. 41:344–347.

Šiki´c,M.,Tomi ´c,S.,andVlahovi ˇcek,K.(2009).Predictionofprotein-proteininteractionsitesin

sequencesand3Dstructuresbyrandomforests. PLoSComput.Biol. 5(1):e1000278.

Sillitoe,I.,Cuff,A.L.,Dessailly,B.H.etal.(2013).Newfunctionalfamilies(FunFams)inCATHto

improvethemappingofconservedfunctionalsitesto3Dstructures. NucleicAcidsRes. 41:

490–498.

Sillitoe,I.,Lewis,T.E.,Cuff,A.etal.(2015).CATH:comprehensivestructuralandfunctional

annotationsforgenomesequences. NucleicAcidsRes. 43:D376–D381.

Söding,J.(2005).ProteinhomologydetectionbyHMM-HMMcomparison. Bioinformatics.21:

951–960.

Stevens,T.J.andArkin,I.T.(2000).Domorecomplexorganismshaveagreaterproportionof

membraneproteinsintheirgenomes? Proteins39:417–420.

Suyama,M.andOhara,O.(2003).DomCut:predictionofinter-domainlinkerregionsinamino

acidsequences. Bioinformatics.19:673–674.

Szent-Györgyi,A.G.andCohen,C.(1957).Roleofprolineinpolypeptidechainconfigurationof

proteins.Science.126:697.

Thorn,K.S.andBogan,A.A.(2001).ASEdb:adatabaseofalaninemutationsandtheireffectson

thefreeenergyofbindinginproteininteractions. Bioinformatics.17:284–285.

Thusberg,J.,Olatubosun,A.,andVihinen,M.(2011).Performanceofmutationpathogenicity

predictionmethodsonmissensevariants. Hum.Mutat. 32:358–368.

Tien,M.Z.,Meyer,A.G.,Sydykova,D.K.etal.(2013).Maximumallowedsolventaccessibilitiesof

residuesinproteins. PLoSOne. 8(11):e80635.

Tompa,P.,Davey,N.E.,Gibson,T.J.,andBabu,M.M.(2014).Amillionpeptidemotifsforthe

molecularbiologist. Mol.Cell. 55(2):161–169.

Touw,W.G.,Baakman,C.,Black,J.etal.(2015).AseriesofPDB-relateddatabanksforeveryday

needs.NucleicAcidsRes. 43(D1):D364–D368.

Tsirigos,K.D.,Elofsson,A.,andBagos,P.G.(2016).PRED-TMBB2:improvedtopologyprediction

anddetectionofbeta-barreloutermembraneproteins. Bioinformatics.32(17):i665–i671.

Tuncbag,N.,Kar,G.,Keskin,O.etal.(2009).Asurveyofavailabletoolsandwebserversfor

analysisofprotein-proteininteractionsandinterfaces. BriefingsBioinf. 10:217–232.

UniProtConsortium(2016).UniProt:theuniversalproteinknowledgebase. NucleicAcidsRes. 45:

1–12.

Vicedo,E.,Schlessinger,A.,andRost,B.(2015).Environmentalpressuremaychangethe

compositionproteindisorderinprokaryotes. PLoSOne. 10:1–21.

Viklund,H.,Granseth,E.,andElofsson,A.(2006).Structuralclassificationandpredictionof

reentrantregionsinalpha-helicaltransmembraneproteins:applicationtocompletegenomes. J.

Mol.Biol. 361:591–603.

VonHeijne,G.(1992).Membraneproteinstructureprediction.Hydrophobicityanalysisandthe

positive-insiderule. J.Mol.Biol. 225:487–494.

VonHeijne,G.andGavel,Y.(1988).Topogenicsignalsinintegralmembraneproteins. Eur.J.

Biochem.174:671–678.

Walia,R.R.,Xue,L.C.,Wilkins,K.etal.(2014).RNABindRPlus:apredictorthatcombines

machinelearningandsequencehomology-basedmethodstoimprovethereliabilityof

predictedRNA-bindingresiduesinproteins. PLoSOne 9(5):e97725.

Wang,B.,Chen,P.,Huang,D.S.etal.(2006).Predictingproteininteractionsitesfromresidue

spatialsequenceprofileandevolutionrate. FEBSLett. 580:380–384.

Wang,S.,Li,W.,Liu,S.,andXu,J.(2016a).RaptorX-property:awebserverforproteinstructure

propertyprediction. NucleicAcidsRes. 44(W1):W430–W435.

References 225

Wang,S.,Peng,J.,Ma,J.,andXu,J.(2016b).Proteinsecondarystructurepredictionusingdeep

convolutionalneuralfields. Sci.Rep. 6:18962.

Wang,S.,Sun,S.,Li,Z.etal.(2017).Accuratedenovopredictionofproteincontactmapby

ultra-deeplearningmodel. PLoSComput.Biol. 13(1):e1005324.

Wright,P.E.andDyson,H.J.(2014).Intrinsicallydisorderedproteinsincellularsignallingand

regulation.Nat.Rev.Mol.CellBiol. 16:18–29.

Wu,C.H.,Nikolskaya,A.,Huang,H.etal.(2004).PIRSF:familyclassificationsystematthe

proteininformationresource. NucleicAcidsRes. 32:D112–D114.

Xue,L.C.,Dobbs,D.,andHonavar,V.(2011).HomPPI:aclassofsequencehomologybased

protein-proteininterfacepredictionmethods. BMCBioinf. 12:244.

Yachdav,G.,Kloppmann,E.,Kajan,L.etal.(2014).PredictProtein–anopenresourceforonline

predictionofproteinstructuralandfunctionalfeatures. NucleicAcidsRes. 42:W337–W343.

Yan,J.,Friedrich,S.,andKurgan,L.(2016).Acomprehensivecomparativereviewofsequence

basedpredictorsofDNA-andRNA-bindingresidues. BriefingsBioinf. 17:88–105.

Yang,Y.,Gao,J.,Wang,J.etal.(2016a).Sixty-fiveyearsofthelongmarchinproteinsecondary

structureprediction:thefinalstretch. BriefingsBioinf. 19(3):482–494.

Yang,J.,Jin,Q.Y.,Zhang,B.,andShen,H.B.(2016b).R2C:improvingabinitioresiduecontact

mappredictionusingdynamicfusionstrategyandGaussiannoisefilter. Bioinformatics.32:

2435–2443.

Zhang,H.,Zhang,T.,Chen,K.etal.(2011).Criticalassessmentofhigh-throughputstandalone

methodsforsecondarystructureprediction. BriefingsBioinf. 12(6):672–688.

Zhao,H.,Yang,Y.,andZhou,Y.(2013).PredictionofRNAbindingproteinscomesofagefromlow

resolutiontohighresolution. Mol.Biosyst. 9:2417–2425.