AbstractPolycystic ovary syndrome (PCOS) is a common endocrine disorder thataffects up to 20% of women, however diagnosis is commonly unreliable andun-quantitative. Here we use supervised machine learning and measurementsof 51 cytokines from a large cohort of patients to identify a low-dimensionalset of potential biomarkers for diagnosis of PCOS. Both whole blood andindividual follicular fluid (FF) aspirates were collected women during pre-intracytoplasmic sperm injection with in vitro fertilization (ICSI/IVF) oocyteretrieval and linked with patients’ PCOS status as diagnosed by the Rotterdamcriteria (n = 69 PCOS, n = 222 non-PCOS). We trained a binary support vectormachine (SVM) using a random subset of patient data to determine cytokineprofile associated with PCOS. Our resultant model includes 3 variables and is76% accurate. This provides insight into the immunological basis of PCOS andmay define a potential non-invasivequantitative strategy for diagnosis.

Introduction

Definitions&Equations

StatisticalAnalysis UsingSupportVectorMachines OtherAnalyses

FutureDirectionsThe dataset described in this study is a small subset of a much largercollection of patient data. In the future we plan to incorporate more ofthese data to determine what more can be said about the classification ofPCOS and prediction of fertility treatment success. Another question wemight be interested in is if Follicular Fluids data could be better used topredict pregnancy results while blood plasma data could be better atpredicting the presence of PCOS. Applying different methods of analysis(i.e. different machine learning classifiers) may have different strengthsthan our current models.

References

1.http://www.med.nyu.edu/chibi/sites/default/files/chibi/Final.pdf2.http://www.ncbi.nlm.nih.gov/pmc/articles/PMC3872139/3.http://www.csie.ntu.edu.tw/~cjlin/papers/guide/guide.pdf4.Latchman,DavidS.GeneRegulation- FifthEdition.Taylor&FrancisGroup2005.Print.5.http://www.ncbi.nlm.nih.gov/pmc/articles/PMC2785020/

Thank you tothe members ofthe Gunawardena lab includingbut not limitedtoJohn,Sieu,Dan,David,Deepesh,Mohan,Felix,Javi.ThisworkwassupportedbytheGwill YorkandPaulMaeder ResearchAwardforSystemsBiologyandtheFASCenterforSystemsBiology. This work wassupported bythe NationalScience Foundation,Award id.1462629.

Acknowledgements

IdentificationofNon-invasiveCytokineBiomarkersforPolycysticOvarySyndromeUsingSupervisedMachineLearning

Sensitivity = 𝑇𝑃

𝑇𝑃 + 𝐹𝑁

Specificity = 𝑇𝑁

𝑇𝑁 + 𝐹𝑃

Accuracy = 𝑇𝑁

𝑇𝑁 + 𝐹𝑃

PCOS is an endocrine disorder that affects up to 20% of women. It is diagnosedusing the Rotterdam criteria, which are as follows:At least two of the followingmust be present: Hyperandrogenism, which could include clinical hirsutism(excessive hairiness) or biochemically raised testosterone levels; anovulation,or difficulties getting pregnant; the presence of cysts in ovaries analyzed by anultrasound.

With these diagnosis criteria in mind, one thing my projects aims to accomplishis to provide a different measurement of PCOS using concrete levels ofcytokines. This study involved taking samples from 291 women by firstexposing them to long-protocol ovarian hyperstimulation, which is a techniqueused to induce ovulation by multiple ovarian follicles. Then samples are takenfrom each woman’s blood plasma and follicular fluid. To note, the reason whythere is approximately double the amount of follicular fluid samples is that forthis set, each ovary is sampled, yielding roughly double the number ofsamples. Then, fifty-one whole blood and FF cytokines were measured by fluid-phase multiplex cytometric immunoassay (the resultant dataset is picturedbelow). The different cytokines were detected using different antibodies,which can be quite an expensive and lengthy test. So, another goal of thisprojec is to reduce the number of cytokines, or features, needed to predictPCOS in patients.

Cytokines:Cytokinesaresmallsecretedproteinsreleasedbycellshaveaspecificeffectontheinteractionsandcommunicationsbetweencells.

Lowlevels

Highlevels

P Dataset FFDataset

Datasetvisualization.Thesetwofiguresrepresenttheplasma(P)andfollicularfluid(FF)datasetswhichare291x51and530x51respectively.

Acomprehensivelistofthe48cytokinesmeasuredis:IL.1a,IL.1b,IL.1ra,IL.2,IL.2ra,IL.3,IL.4,IL.5,IL.6,IL.7,IL.8,IL.9,IL.10,IL.12..p40.,IL.12..p70., IL.13,IL.15,IL.16,IL.17,IL.18,CTACK,Eotaxin,FGF,G.CSF,GM.CSF,GRO.a,IFN.a,IFN.g,IP.10,LIF,MCP.1,MCP.3,M.CSF,MIF,MIG,MIP.1a,MIP.1b,b.NGF,PDGF,RANTES,SCF,SDF.1a,TGF.b,TGF.b,TNF.a,TNF.b,TRAIL,VEGF,CRP

ln 89:;89 = 𝛽= + 𝛽:XLogisticRegression:

LinearKernel: 𝐾 𝑥:, 𝑥B = (𝑥:·𝑥B)

f (x) = sign(w⋅ x + b)

SVMOptimizationProblem:

DanielaPerry1,2,Tathagata Dasgupta2,JosephDexter2,SarahField3, MicheleCummings3, VinaySharma3,NadiaGopichandran3,Ellis Baskind3,NicholasOrsi3,JeremyGunawardena21CornellUniversity,DepartmentofBiologicalStatisticsandComputationalBiology2HarvardMedicalSchool,DepartmentofSystemsBiology3UniversityofLeedsInstituteofCancerandPathology

First,weputall48variablesinalogisticregressionmodelandpredictedtheprobabilityofeachsamplebeingaPCOSorControl(ALL). Bothdatasetshadcomparablypoorperformance.Thenweconductedstepwiseregression,whichusestheAICcriteriatoeliminatetheleastsignificantvariable(STEP).Oncethisprocesswascomplete,weanalyzedtheperformanceofthemodelusingtheFFandPdatasetsonceagain.Finally,togetasenseofhowthemodelswouldperformwithamoreclinicallyfeasiblenumberofvariables,weremovedvariablesbasedonthehighestp-valueuntilwewereleftwithjustfourcytokines(FOUR). ROCcurveshighlightingtheperformanceofallsixmodelsaredisplayedbelow.Inaddition,avisualofaccuracy,specificity,andsensitivityarealsodisplayed.

PrincipalComponentAnalysis(PCA):ff.pca

Variances

05

1015

1 2 3 4 5 6 7 8 9

IL.1a

IL.1b

IL.1ra

IL.2

IL.2ra

IL.3

IL.4

IL.5

IL.6

IL.7

IL.8

IL.9

IL.10

IL.12..p40.

IL.12..p70.

IL.13

IL.15

IL.16

IL.17

IL.18

CTACK

Eotaxin

FGF

G.CSF

GM.CSF

GRO.a

HGF

ICAM.1

IFN.a

IFN.g

IP.10

LIF

MCP.1

MCP.3

M.CSF

MIF

MIG

MIP.1a

MIP.1b

b.NGF

PDGF

RANTES

SCF

SCGF.b

SDF.1aTGF.b

TNF.a

TNF.b

TRAIL

VCAM.1

VEGF

CRP

-30

-20

-10

0

10

20

0 20 40PC1 (36.2% explained var.)

PC

2 (1

9.1%

exp

lain

ed v

ar.)

Control PCOS

K-MeansClustering:K-meansclusteringisastochasticprocessthatgroupsdatapointsbasedontheirdistancefromeachother.Pointsarerandomlyassignedtoacluster,thenclusterplacementisoptimizedbyfindingthecenterofeachcluster.Ourclusteringanalysisresultedinthegraphtotheright,whichishighlycondensedbecauseofthepresenceofsomanyuniqueoutliers,whichisconsistentwithone-classclassification.

PCAusesanorthogonaltransformationtomakeasetofvariableslinearlyindependentvariablescalledprincipalcomponents.Theabovelinegraphrepresentshowmuchvarianceinthedataisaccountedforbyeachprincipalcomponent.Thegraphtotheleftrepresentsthegroupingsbasedonthetwomostsignificantprincipalcomponents.

FullModel

“StepwiseModel”

FourVariableModel

StepwiseRegression

P-valueelimination

Model #Variables TrainingSet TestingSet Specificity Sensitivity Accuracy

FF All 48 75%offullFF

25%offullFF

0.9207921 0.21875 0.7518797

FFStepwise 21 75%offullFF

25%offullFF

0.950495 0.25 0.7819549

FFFour 4 75%offullFF

25%offullFF

0.990099 0.0625 0.7669173

P All 48 75%offullP 25%offullP 0.8571429 0.2222222 0.7027027

PStepwise 12 75%offullP 25%offullP 0.8392857 0.1111111 0.6621622

PFour 4 75%offullP 25%offullP 0.9821429 0.05555556 0.7567568

0.0

0.1

0.2

0.3

Control PCOSGroup

IL.10

Group vs. IL.10 (P)

-0.1

0.0

0.1

Control PCOSGroup

IL.10

Group vs. IL.10 (FF)

0.00

0.05

0.10

0.15

Control PCOSGroup

IL.10

Group vs. IL.10 (FF All)

0.00

0.05

0.10

0.15

Control PCOSGroup

IL.10

Group vs. IL.10 (FF Step)

0.00

0.05

0.10

0.15

Control PCOSGroup

IL.10

Group vs. IL.10 (FF Four)

0.0

0.1

0.2

0.3

Control PCOSGroup

IL.10

Group vs. IL.10 (P All)

0.0

0.1

0.2

0.3

0.4

Control PCOSGroup

IL.10

Group vs. IL.10 (P Four)

0.0

0.1

0.2

0.3

Control PCOSGroup

IL.10

Group vs. IL.10 (P Step)

FirstweperformedgridsearchinordertooptimizeparametersforourSVMmodelusingalinearkernel.Then,wetrainedabinaryclassifierandtesteditsperformanceusing5-foldcrossvalidation(resultsinthetablebelow).Theresultsledustoconductone-classclassificationforoutlierdetection.Thenweretestedourmodelexcludingthepointswefoundtobeoutliers.

GridSearch

TrainabinaryclassificationSVM

Testmodelperformance

Applybestparameters

5-foldcrossvalidation

One-classclassification

Testmodelwithoutoutliers

5-foldcrossvalidation

Lowmodelstrength

HowSVMWorks1

Group2

Group1 Group1

Group2

Model #Variables TrainingSet TestingSet Specificity Sensitivity Accuracy

FFSVM 7 75%offullFF

25%offullFF

1 0.0625 0.7744361

5-foldCV 5-foldCV 0.9876543 0.03846154 0.7570093

5-foldCV(nooutliers)

5-foldCV(nooutliers)

1 0 0.7583333

PSVM 3 75%offullFF

25%offullFF

0.9821429 0.05555556 0.7567568

5-foldCV 5-foldCV 1 0 0.7627119

5-foldCV(nooutliers)

5-foldCV(nooutliers)

1 0.06666667 0.7878788