AutoQSAR:AnAutomatedToolforBest-Prac8cesQSARModeling

Dr.Ma'RepaskyDr.SteveDixonOct.19,2016

Overview•  Introduc@ontoAutoQSAR•  Valida@onofAutoQSARmethodologyagainstQSARmodelsfromliterature

•  Predic@ngproteinsolubilitybyQSAR•  Exampleofintegra@onofQSARmodelsintoacheminforma@csplaLorm(LiveDesign)

Newmoleculepropertypredic@ons

QSAR–Quan8ta8veStructureAc8vityRela8onships

Newmolecules

Propertytomodelfortrainingsetcompounds

Descriptorsfortraining

setcompounds

Machinelearning

techniques

•  AutoQSAR– AccessedthroughMaestroorcommandline

– Automa@callycreatesQSARmodelsfollowingabest-prac@cesapproach

•  Canvas– Acompletesetofexpert-capabilitytoolstocreate,validateandemployQSARmodels

– AccessedthroughCanvasGUIorcommandline

•  Strike– AccessedthroughMaestroorcommandline

–  CreatebasicnumericQSARmodelswithproperty-baseddescriptors

Crea8ngQSARModelsintheSchrödingerSuite

AutoQSARisa“QSARexpertinabox”•  HighdemandforourQSARexpert’s@memo@vatedcrea@onofAutoQSAR

•  Automa@callygeneratepredic@veQSARmodelsfollowingbest-prac@cesmethodology

•  Easilycreate,validate,distribute,andemployQSARmodelsdirectlyfromMaestro– AutoQSARknowshowtogeneratenecessarydescriptorsandapplymodelsforpredic@ons

•  QuicklyandeasilydeployQSARmodelsviaLiveDesignorotherinforma@csplaLorms

CorwinHansch–FatherofQSAR

KeyBenefitsofUsingAutoQSARforQSARModeling

•  GenerateandemployQSARmodelswithconfidence•  Save@meiden@fyingapproachesmost/leastlikelytobesuccessfulbeforemanualQSARmodeling– Evidencesuggeststhereisnosingleidealdescriptorsetormachinelearningmethodapplicabletoallendpoints

•  EasilydeployQSARmodelsacrossprojectgroups/labsandarangeofinforma@csplaLorms

•  Takeadvantageoflargeamountsofpublicandprivatedata– ChEMBL,PubChem,ChemSpider

•  ImproveQSARmodelsasdataareaddedduringaproject11Mol.Inform.30(2-3),256-266(2011)

AutoQSARWorkflowEncodesQSARBestPrac8ces

Structures+Propertytofit

Structures

QSARModels

Build

Test

Report

!

Descriptors+

Fingerprints

Informa@veDescriptors

+Fingerprints

Calculate FeatureSelec@on

SingleModelMior

ConsensusΣiMi

Con@nuousY

CategoricalY

RPBayes

MLRKPLSetc.

Train/TestSplits

Bayes1

RP2

RP1

Bayes2­­­

­­­

KPLS1

MLR2

MLR1

KPLS2­­­

­­­

Rank,Sort,Cap

LearningMethods

ŷ1ŷ2...

InputsforQSARModelingwithAutoQSAR•  LigandsprovidedinSMILES,2D,or3Dformats•  Datacura@onrecommenda@ons:– Consistencyinligandprepara@ontonormalizespecificchemotypesandtreatmentoftautomericforms

– Removeduplicates– Removeundesirables–mixtures,salts,etc.– Minimizedatainadequacy

•  Predictedpropertyshouldn’tspanmul@pleordersofmagnitude–  Converttologarithmicscaleifnecessary,i.e.IC50topIC50

•  Endpointshouldbeinmolarandnotweightunits•  Dataheterogeneityshouldbeavoidedifpossible,i.e.datafromdifferentspecies/protocols

DescriptorSetsUsedbyAutoQSARFingerprintType Descrip8on

Dendri@c Linearandbranchedfragments

Linear Linearfragments+ringclosures

MOLPRINT2D Radial-likeFPthatencodesatomenvironmentsusinglistsofatomtypeslocatedatdifferenttopologicaldistances

Radial Fragmentsthatgrowradiallyfromeachatom.Alsoknownasextendedconnec@vityfingerprints

497topology-baseddescriptors•  Estatecounts•  2Dtopologicaldescriptors•  Func@onalgroupcounts

Employalterna@vedescriptorsifdesired

FeatureSelec8on•  Descriptors:– Eliminatedescriptorswhere>90%oflearningsethasthesamevalue– EnsurenopairofdescriptorshasanabsolutePearsoncorrela@oncoefficientgreaterthan0.8

•  Fingerprints:– Ofallbitssetbyatleastonecompound,onlythemostsignificant10,000withthegreatestvarianceoverthelearningsetareemployed

MachineLearningMethodsEmployed

SupportedVariableTypes

Descrip8on Dependent(y) Independent(x)

Bestsubsetsmul@plelinearregression(MLR)

Con@nuous Descriptors

Par@alleastsquaresregression(PLS)

Con@nuous Descriptors

Principalcomponentsregression(PCR)

Con@nuous Descriptors

Kernel-basedPLS(kPLS) Con@nuous Descriptors,fingerprints

NaïveBayesclassifica@on Categorical Descriptors,fingerprints

Ensemblerecursivepar@@oning(RP)

Categorical Descriptors

NeuralnetworkmodelsarenotcreatedbyAutoQSARastheywerefoundtobetooeasilyoverfit

CreateandvalidateQSAR

model

Splitdataset

SpliXngLearningSetintoTest/TrainingSets•  Intradi@onalsingleQSARmodelcrea@onbiasingtraining/testsetisooendesireddueto“allornothing”outcome

•  Toincreaseprobabilityoffindingapredic@vemodelwithgoodpredic@veabilityAutoQSARrandomlysplitslearningsetinto75%training/25%testsetN@mescrea@ngmanyQSARmodels

•  With

CreateandvalidateQSAR

model

LearningSet TrainingSet TestSetQSARModelLearningSet TrainingSet TestSetQSARModelLearningSet TrainingSet TestSetQSARModelLearningSet TrainingSet TestSetQSARModelLearningSet TrainingSet TestSetQSARModelLearningSet TrainingSet TestSetQSARModelLearningSet TrainingSet TestSetQSARModelLearningSet TrainingSet TestSetQSARModelLearningSet TrainingSet TestSetQSARModel

Splitdataset

ScoringFunc8ontoRank-OrderQSARModels

score = accuracytest • 1.0− accuracytrain − accuracytest( )accuracytrain,continuous = R

2 =1.0−σ 2err σ

2y

accuracytest,continuous =1.0−σ2err σ

2Ly

accuracycategorical =1n

fkk=1

n

∑

•  RankorderQSARmodelsbytheirpredic@vepower•  Retaintop10modelsbydefault– Useofaddi@onalmodelsmaydegradeaccuracyifthosemodelsarenotsufficientlypredic@ve

•  Defaultistocreateconsensusmodelfromtop10-rankedmodels

•  Forcon@nuousmodels

•  Forcategoricalmodels–  Eachofthemmodelsvotesforacategoryandcompoundisassignedtocategorywithmostvotes

–  Incaseofa@e,categorywithhighestaverageprobabilityofpredic@onisassigned

ApplyConsensusorSingleModelPredic8onsConsensusmodelsooenincreasepredic@onaccuracyandexpandapplicabilitydomainfordiverseQSARmodels

yi =1m

yijj=1

m

∑“Thisobserva@onsuggestthatconsensusmodelsaffordbothhighspacecoverageandhighaccuracyofpredic@on”

AutoQSARApplicabilityDomainEs8mate•  Domain_Alertindicateswhetheracompoundisoutsidethesimilarityrangeof95%oftrainingsetcompounds–  ForsingleQSARmodelsDomain_Alertsare1or0

•  Domain_Scoreisaz-scorethatindicateshowstructurally‘similar’acompoundistotrainingsetcompounds–  Tanimoto‘similarity’calculatedtoamodaldendri@cfingerprintfromtrainingsetstructures(logicalORofallonbits):

Domain_Scorei=(simi-µsim)/σsim

Domain_Alertissetto1whenDomain_Score<-2–  ForconsensusmodelstheDomain_Scoreisaveragedoverallmodelsintheconsensus

•  Applicabilitydomaines@matesmostusefulforlocalmodels

0.0

Domain_Alert1.0

AutoQSARPredic8onErrorRela8onshiptoDomain_ScoreforLocalModels

0

0.5

1

1.5

2

2.5

3

MeanAb

soluteError(log)

0.0-0.3

0.3-0.6

0.6-0.8

0.8-1.0

>1.0

Domain_ScoreRange

AutoQSARDomain_AlertsRecognizeCompoundsDissimilartotheTrainingSet

fxa_1f0r chk1_2e9u lxrb_3kfc throm_1mue cdk2_2bkz metap2_2adu cdk2_1fvt metap2_1r5g hsp90_2ccu fxa_1z6efxa_1f0r 4 100 100 100 100 100 100 100 100 100chk1_2e9u 100 29 100 100 100 100 100 100 100 100lxrb_3kfc 100 100 11 100 100 100 100 100 100 100throm_1mue 100 100 100 4 100 100 100 100 100 100cdk2_2bkz 100 100 100 100 11 100 100 100 100 100metap2_2adu 96 90 75 100 100 7 87 98 90 14cdk2_1fvt 100 100 100 100 100 100 4 100 100 100metap2_1r5g 0 57 36 83 94 100 26 2 50 14hsp90_2ccu 100 100 100 100 100 100 100 100 10 100fxa_1z6e 100 100 100 100 100 100 100 100 100 14

PercentofTestSetCompoundswithDomain_AlertsDataset/Model

SelectCompoundsFromfxa_1f0rtrainingset SelectCompoundsFrommetap2_1r5gtestset

AutoQSARDomain_AlertsRecognizeCompoundsDissimilartotheTrainingSet

fxa_1f0r chk1_2e9u lxrb_3kfc throm_1mue cdk2_2bkz metap2_2adu cdk2_1fvt metap2_1r5g hsp90_2ccu fxa_1z6efxa_1f0r 4 100 100 100 100 100 100 100 100 100chk1_2e9u 100 29 100 100 100 100 100 100 100 100lxrb_3kfc 100 100 11 100 100 100 100 100 100 100throm_1mue 100 100 100 4 100 100 100 100 100 100cdk2_2bkz 100 100 100 100 11 100 100 100 100 100metap2_2adu 96 90 75 100 100 7 87 98 90 14cdk2_1fvt 100 100 100 100 100 100 4 100 100 100metap2_1r5g 0 57 36 83 94 100 26 2 50 14hsp90_2ccu 100 100 100 100 100 100 100 100 10 100fxa_1z6e 100 100 100 100 100 100 100 100 100 14

PercentofTestSetCompoundswithDomain_AlertsDataset/Model

SelectCompoundsFromfxa_1f0rtrainingset SelectCompoundsFrommetap2_1r5gtrainingset

AutoQSARModelGenera8onTimings•  CPU@mescaleslinearlywiththenumberofmolecules– Minutesforlocalmodels,hoursforglobalmodelson1CPU

•  Builtincapof5ktraining/testsetligands•  Licensingfavorabletorunningonmul@pleCPUs(1token/subjob)

0

2

4

6

8

10

12

800 1000 1200 1400 1600 1800

HoursT

oMod

el

NumberofTraining/TestSetLigands

SingleCPU

50CPUs

Crea8ngAutoQSARModelsinMaestro

ApplyingAutoQSARModelinMaestro

Valida8onofAutoQSARMethodologyObjec8ve:Examinepredic@veperformanceofAutoQSARmodelscreatedforavarietyofproper@esandcomparetopublished,humanexpert-createdQSARmodels– CanAutoQSARgeneratemodelswithsimilarperformancetothosepublishedbyQSARexperts?

– ResearchpublishedinFutureMed.Chem.,2016,8(15),1825-1839.

ClassifierQSARModelsFishbioconcentra@onfactor

CarcinogenicityMutagenicity

NumericQSARModelsCongenericSeriesBindingaffinity

SolubilityBloodbrainbarrierpermeability

CongenericSeriesBindingAffinity

J.Chem.Inf.Model.,2013,53(9),2312-21.

•  PrimarilykPLSmodelswithafewPLSmodels

•  kPLSmodelsprimarilyfingerprint-based

•  CHK12e9upoorQ2forexternalvalida@onsetduetopredic@ngonecompoundwithIC50of10μMtobe10nM.RMSEis0.82.

54to150compoundsinlearningset18to50compoundsinexternalvalidset

Solubility

LearningsetR2=0.88Externalvalida@onsetQ2=0.89

LearningsetR2=0.88

J.Chem.Inf.Model.,2007,47(4),1395-1404

1708compoundsinlearningset427compoundsinexternalvalid

set

BloodBrainBarrierPermeability

!!BBB"############

predicted#BBB+################

predicted#BBB"# 25! 18!BBB+# 6! 50!!

144compoundsinlearningset99compoundsinexternalvalidset

Accuracy=(25+50)/(99)*100=75.7%Sensi8vity=50/(50+18)*100=73.5%Specificity=25/(25+6)*100=80.5%

Pharm.Res.25(8),1902-1914(2008)

Carcinogenicity

644compoundsinlearningset161compoundsinexternalvalid

set

Chem.Cent.J.,4Suppl1,S3(2010)

Mutagenicity

ChemistryCentralJournal4Suppl1S2(2010)

3367compoundsinlearningset837compoundsinexternalvalid

set

FishBioconcentra8onFactor

0.0

0.1

0.2

0.3

0.4

0.5

0.6

CAESARPredic@on AutoQSARconsensusmodelalldescriptors*

AutoQSARconsensusmodelonlyfingerprints

AutoQSARbestmodelalldescriptors

AutoQSARbestmodelonlyfingerprints

ExternalValida8onSetQ2

473compoundsinlearningset119compoundsinexternalvalidset

Chem.Cent.J.,4Suppl1,S1(2010)

UsingAutoQSARtoPredictProteinSolubility

Trainoretal.,JMB,Vol428,(6),Mar2016,pp1365–1374

31FG-Loopvariants

0%10%20%30%

Percen

tofFG

Loop

Seq

uences

PercentInsoluble

Goal:UseproteindescriptorstocreateanAutoQSARmodeltopredictproteinsolubilityforAdnec@ns

•  Adnec@nsareengineeredtargetbindingproteinsfromthefibnectrintypeIIIdomain

•  Muta@onsinFGLoopgreatlyinfluenceforma@onofinclusionbodies(IB)thatnega@velyimpactsolubility

•  Conforma@onallyverymobile,facilitatesbindingtotarget

•  IB-forma@oncannotbeexplainedbyvaria@oninthermalstabilityamongtheloopmutants(DSCdata)

•  IB-forma@oniscorrelatedwiththeaveragelocalstabilityofFG-loop

ProteinDescriptors•  Scalardescriptors(25):calculatedontheen@restructure

–  pI,netcharge,electrophore@cmobility,hydrodynamicradius,radiusofgyra@on,dipolemoment,hydrophobicmoment,aggrega@onpropensi@es,etc…

•  Region-specificdescriptors(25-residueaware):–  Charge,zetapoten@al,surfaceexposure,hydrogenbond

donors/acceptors,hydrophobicity,energe@ccontribu@onstoposi@ve,nega@veandhydrophobicpatches,AggrescanandZyggregatorprofilescores,etc…

–  e.g.full-lengthimmunoglobulin(25regions->625datapoints):

•  VL,VL_Fc,FR,FR1,FR2,FR3,FR4,L1,L2,L3,CL

•  VH,VH_Fc,FR,FR1,FR2,FR3,FR4,H1,H2,H3,CH1,CH2,CH3,Hinge

•  Patch-specificdescriptors(27): –  Basedonpatchsize,energyscores–  Notresidue-awareness

Workflowtocomputeproteindescriptors

ProteinDescriptors–Adnec8nBenchmarkSet•  Adnec8ns:engineeredtargetbindingproteinsfromthefibnectrintypeIIIdomain–  Residuemuta@onsinthreelociofasolvent-exposedbindingloop(FG-Loop)greatlyalterforma@onofinclusionbodies(IB)

–  FG-LoopisanalogoustoCDRH3domaininimmunoglobulins•  conforma@onallyverymobile,facilitatesbindingtotarget

–  IB-forma@oncannotbeexplainedbyvaria@oninthermalstabilityamongtheloopmutants(DSCdata)

–  IB-forma@oniscorrelatedwiththeaveragelocalstabilityofFG-loop•  Modelledwithcomputa@onallyexpensiveloopsamplingmethod(R2=0.61)

–  Zyggregatorpredic@on(sequence-basedaggrega@onpredic@onmethod)correlateswithIB-forma@on(R2=0.63)

–  Linearcombina@onofthemethodsrevealedoverallR2of0.8

ProteinDescriptors–Aggrega8onPredic8on•  AGGRESCAN Vendrelletal.BMC(2007)8:65,1-17

–  Intracellularaggrega@onpropensityformutantsofAβ42pep@de–  Algorithmwasparameterizedonmutants–  Validatedwithexperimentaldataof24fibril-formingpep@desa3v:amino-acidaggrega@onpropensityvalue(basedonexperiment)a4v:smoothedamino-acid(smoothingwindowsize:5,7,9,11)HS:aggrega@onhotspots

•  ZYGGREGATOR Pawaretal.J.Mol.Biol.(2005)350,379–392

–  Algorithmcalculatestherela@vefoldingpropensityandaggrega@on–  Combina@onof4factorsintrinsictoasequence:

•  charge,hydrophobicity,SSEandpa'erning

–  Parameterizedonshortpep@des–  Validatedonknownamyloidogenicpep@des

chains, such as charge,26–29 hydrophobicity,30–32

patterns of polar and non-polar residues,33 andthe propensities to adopt diverse secondary struc-ture elements.27,32,34,35 In the case of globularproteins, the propensities to form amyloid struc-tures are generally inversely related to the stabilityof their native states.24,36–42 Many of the proteinsassociated with amyloid diseases are, however, atleast partially unfolded under physiological con-ditions. In addition, it is thought that many globularproteins unfold, at least partially, before aggre-gating. The present study is therefore focused onthe conversion of unfolded or partially unfoldedstates into amyloid aggregates.

One of the most intriguing recent observations instudies of the kinetics of amyloid formation is thatpolypeptide sequences appear to contain localregions that are “sensitive” for aggregation.32 Singleamino acid mutations in these regions can changethe aggregation rates dramatically, while similarchanges in other regions may have relatively littleeffect.32,43 In addition, it has been shown that it ispossible to describe with considerable accuracy thein vitro amyloid aggregation propensities of poly-peptides using algorithms that take into account thephysico-chemical properties of their sequences andof their environment.44–48 Here, our purpose is toapply this type of analysis to the rationalisation andthe prediction of the sensitive regions of poly-peptide sequences in general and of proteinsassociated with neurodegenerative diseases inparticular.

Results

Definition of intrinsic aggregation propensities

We define the intrinsic propensity of an unfoldedpolypeptide chain to form amyloid aggregates, Pagg,by considering just the intrinsic factors in theformula that we have recently introduced to definethe absolute aggregation rates of unstructuredpolypeptide chains:45

Pagg Z ahydrIhydr CaaI

a CabIb CapatI

pat

CachIch (1)

where Ihydr represents the hydrophobicity of thesequence,49,50 Ia is the a-helical propensity,51 Ib isthe b-sheet propensity,51 Ipat is the hydrophobicpatterning,52 and Ich is the absolute value of the netcharge of the sequence; the coefficients a, whichweight the individual factors, were determined asdescribed and are listed in Table 1 (see Materialsand Methods).45 Since pH influences some of theseterms, such as Ihydr, Ia and Ich, it must be specifiedfor equation (1), and the hydrophobicity andsecondary structure propensities are normalisedfor the length of the polypeptide chain. The valuesthat we found for the coefficients indicate thathydrophobicity, the presence of specific

hydrophobic patterns and the propensity to formb-sheets favour aggregation; in contrast, the netcharge and the propensity to form a-helices reducethe tendency to aggregate. It has been suggestedthat aromatic residues may play an important rolein promoting aggregation,48,53 although otherstudies have indicated that their importance maybe limited.54 The systematic collection of exper-imental data relevant to this factor will allow its roleto be clarified.

Definition of intrinsic Z-scores for aggregation

The intrinsic Z-score for aggregation, Zagg,enables comparisons to be made between theaggregation propensities of different polypeptidesequences. It is calculated as:

Zagg ZPagg Kmagg

sagg(2)

where magg is the average value of Pagg over a set ofrandom polypeptides having the same length as thesequence of interest, and sagg is the correspondingstandard deviation from the average (see Materialsand Methods). If ZaggO0, the sequence is moreprone to aggregation than a randomly generatedsequence at that particular pH, while it is less proneif Zagg!0.

Intrinsic aggregation propensities of individualamino acids

The intrinsic propensities pagg of individualamino acids to promote the conversion of apolypeptide chain into amyloid aggregates can becalculated from equation (1) if one considers apolypeptide sequence of length 1. In this case, theterm describing the patterning of polar and non-polar residues, Ipat, does not contribute to theresulting pagg value. The aggregation propensityscale obtained in this way should be useful for aqualitative estimate of the effect of mutations on theaggregation behaviour of a given polypeptidechain, at least when hydrophobic patterns are notinvolved. The scales at three different values of pHare shown in Table 2, where the various amino acidsare listed in decreasing order of amyloid formationpropensity at neutral pH. At this pH, tryptophan,phenylalanine, cysteine, tyrosine, and isoleucinehave the highest amyloid propensities, while

Table 1. Coefficients of equation (1)

Parameter a p-value

Hydrophobicity K1.99G0.31 !0.001a K5.7G2.3 !0.001b 5.0G1.7 !0.001Charge K0.08G0.03 0.17Patterns 0.39G0.02 !0.001

The coefficients a used in equation (1). These values and thecorresponding statistical errors were derived as described.45

380 Sensitive Regions for Protein Aggregation

chains, such as charge,26–29 hydrophobicity,30–32

patterns of polar and non-polar residues,33 andthe propensities to adopt diverse secondary struc-ture elements.27,32,34,35 In the case of globularproteins, the propensities to form amyloid struc-tures are generally inversely related to the stabilityof their native states.24,36–42 Many of the proteinsassociated with amyloid diseases are, however, atleast partially unfolded under physiological con-ditions. In addition, it is thought that many globularproteins unfold, at least partially, before aggre-gating. The present study is therefore focused onthe conversion of unfolded or partially unfoldedstates into amyloid aggregates.

One of the most intriguing recent observations instudies of the kinetics of amyloid formation is thatpolypeptide sequences appear to contain localregions that are “sensitive” for aggregation.32 Singleamino acid mutations in these regions can changethe aggregation rates dramatically, while similarchanges in other regions may have relatively littleeffect.32,43 In addition, it has been shown that it ispossible to describe with considerable accuracy thein vitro amyloid aggregation propensities of poly-peptides using algorithms that take into account thephysico-chemical properties of their sequences andof their environment.44–48 Here, our purpose is toapply this type of analysis to the rationalisation andthe prediction of the sensitive regions of poly-peptide sequences in general and of proteinsassociated with neurodegenerative diseases inparticular.

Results

Definition of intrinsic aggregation propensities

We define the intrinsic propensity of an unfoldedpolypeptide chain to form amyloid aggregates, Pagg,by considering just the intrinsic factors in theformula that we have recently introduced to definethe absolute aggregation rates of unstructuredpolypeptide chains:45

Pagg Z ahydrIhydr CaaI

a CabIb CapatI

pat

CachIch (1)

where Ihydr represents the hydrophobicity of thesequence,49,50 Ia is the a-helical propensity,51 Ib isthe b-sheet propensity,51 Ipat is the hydrophobicpatterning,52 and Ich is the absolute value of the netcharge of the sequence; the coefficients a, whichweight the individual factors, were determined asdescribed and are listed in Table 1 (see Materialsand Methods).45 Since pH influences some of theseterms, such as Ihydr, Ia and Ich, it must be specifiedfor equation (1), and the hydrophobicity andsecondary structure propensities are normalisedfor the length of the polypeptide chain. The valuesthat we found for the coefficients indicate thathydrophobicity, the presence of specific

hydrophobic patterns and the propensity to formb-sheets favour aggregation; in contrast, the netcharge and the propensity to form a-helices reducethe tendency to aggregate. It has been suggestedthat aromatic residues may play an important rolein promoting aggregation,48,53 although otherstudies have indicated that their importance maybe limited.54 The systematic collection of exper-imental data relevant to this factor will allow its roleto be clarified.

Definition of intrinsic Z-scores for aggregation

The intrinsic Z-score for aggregation, Zagg,enables comparisons to be made between theaggregation propensities of different polypeptidesequences. It is calculated as:

Zagg ZPagg Kmagg

sagg(2)

where magg is the average value of Pagg over a set ofrandom polypeptides having the same length as thesequence of interest, and sagg is the correspondingstandard deviation from the average (see Materialsand Methods). If ZaggO0, the sequence is moreprone to aggregation than a randomly generatedsequence at that particular pH, while it is less proneif Zagg!0.

Intrinsic aggregation propensities of individualamino acids

The intrinsic propensities pagg of individualamino acids to promote the conversion of apolypeptide chain into amyloid aggregates can becalculated from equation (1) if one considers apolypeptide sequence of length 1. In this case, theterm describing the patterning of polar and non-polar residues, Ipat, does not contribute to theresulting pagg value. The aggregation propensityscale obtained in this way should be useful for aqualitative estimate of the effect of mutations on theaggregation behaviour of a given polypeptidechain, at least when hydrophobic patterns are notinvolved. The scales at three different values of pHare shown in Table 2, where the various amino acidsare listed in decreasing order of amyloid formationpropensity at neutral pH. At this pH, tryptophan,phenylalanine, cysteine, tyrosine, and isoleucinehave the highest amyloid propensities, while

Table 1. Coefficients of equation (1)

Parameter a p-value

Hydrophobicity K1.99G0.31 !0.001a K5.7G2.3 !0.001b 5.0G1.7 !0.001Charge K0.08G0.03 0.17Patterns 0.39G0.02 !0.001

The coefficients a used in equation (1). These values and thecorresponding statistical errors were derived as described.45

380 Sensitive Regions for Protein Aggregation

μagg=averagePaggonrandomsequenceσagg=stdevofPaggonrandomsequence

MethodologytoPredictAdnec8nSolubility1.  FG-LoopmodelingusingPrimeandBioLuminate2.  Calculateproteindescriptorswithlowestenergyloopstructureofeachmutant3.  CreateQSARmodelwithAutoQSAR

•  Learningsetwasall31compounds•  Usedproteindescriptorsonly•  Otherwiseemployeddefaultse�ngs

4.  Create4-foldcrossvalidatedmodeltoaccountforsmalldatasetsize•  Extractfourexternalvalida@onsetsfromnon-overlappingrandomsamples

of25%ofthedata•  BuildAutoQSARmodelsfromeachoftheremaining75%ofthedata•  UseeachofthefourAutoQSARmodelstomakepredic@onsonthe

correspondingexternalvalida@onset

ProteinDescriptors–Adnec8nBenchmark

0

10

20

30

40

50

60

70

80

0 10 20 30 40 50 60 70 80

Pred

icted

Observed(Y)

Adnec8n4-FoldCrossValida8on

Q2=0.66RMSE=10.76

Predic8onsfromotherapproaches•  Computa@onallyexpensive

loopsamplingmethodR2=0.61

•  Zyggregatorsequence-basedaggrega@onpredic@onmethodR2=0.63

•  Linearcombina@onofZyggregatorandloopsamplingR2=0.8

Integra8ngAutoQSARintoExis8ngITInfrastructure

•  ToapplyanAutoQSARmodelonlya1D,2D,or3DstructureandanAutoQSARmodelarerequired– Noknowledgeofthemodelordescriptorsusedrequiredwithdefaultdescriptors

•  Benefitfromimprovedpredic@vepoweroverthecourseofaprojectbyautoma@callyupda@ngQSARmodelsperiodically

EasilyIntegrateAutoQSARintoAutomatedWorkflows

AutoQSAR

AutoQSARinLiveDesign

RunningAutoQSARtotestnewcompoundsisaone-clickopera@on

AutoQSARinLiveDesign

Onceclicked,thecalcula@onsrunonallmoleculesinthereport,andresults

arereturnedasanewcolumn

AutoQSARinLiveDesign

Whennewcompoundsaresketched,AutoQSARisruninreal-@metooffer

instantfeedbackonideas

AutoQSARinLiveDesign

Whenthenewideaisaddedtothereport,thecalculatedresultisstored.

FutureDirec8ons–Interpre8ngAutoQSARModels•  InterpretabilityishighlydesiredinanyQSARmodel– Atom/group-basedprojec@on–  Interpretcoefficientvalues,RPtrees,etc.

•  KPLSmodelsemployingfingerprintsareinterpretableinCanvasbutnoteasytovisualizeAutoQSARmodelsinCanvas

•  Inves@ga@ngamodetogenerateandvisualizeinterpretableQSARmodelsinAutoQSAR

Conclusions•  AutoQSARgenerateshigh-quality,predic@veQSARmodelsonparwithpublishedmodelsforawidevarietyofproper@esofinterest

•  CreateandvalidateQSARmodelsfarmorequicklythenmanualcrea@on

•  GenerateandemployQSARmodelswithconfidence•  Save@meiden@fyingapproachesmost/leastlikelytobesuccessfulbeforemanualQSARmodeling