noninvasive prenatal detection and selective analysis of ... prenatal detectio… · all 167...
TRANSCRIPT
eta
Reports of Major Impact www.AJOG.org
Noninvasive prenatal detection and selective analysisof cell-free DNA obtained from maternal blood:evaluation for trisomy 21 and trisomy 18Andrew B. Sparks, PhD; Craig A. Struble, PhD; Eric T. Wang, PhD; Ken Song, MD; Arnold Oliphant, PhD
OBJECTIVE: We sought to develop a novel biochemical assay and algo-rithm for the prenatal evaluation of risk for fetal trisomy 21 (T21) andtrisomy 18 (T18) using cell-free DNA obtained from maternal blood.
STUDY DESIGN: We assayed cell-free DNA from a training set and ablinded validation set of pregnant women, comprising 250 disomy, 72T21, and 16 T18 pregnancies. We used digital analysis of selected re-gions in combination with a novel algorithm, fetal-fraction optimizedrisk of trisomy evaluation (FORTE), to determine trisomy risk for eachsubject.
RESULTS: In all, 163/171 subjects in the training set passed quality
control criteria. Using a Z statistic, 35/35 T21 cases and 7/7 T18 casesevaluation for trisomy 21 and trisomy 18. Am J Obstet Gynecol 2012;206:319.e1-9.
thicvargs
ltttf
4% DNA from a T21
had Z statistic �3 and 120/121 disomic cases had Z statistic �3.FORTE produced an individualized trisomy risk score for each subject,and correctly discriminated all T21 and T18 cases from disomic cases.All 167 subjects in the blinded validation set passed quality control andFORTE performance matched that observed in the training set correctlydiscriminating 36/36 T21 cases and 8/8 T18 cases from 123/123 dis-omic cases.
CONCLUSION: Digital analysis of selected regions and FORTE enableaccurate, scalable noninvasive fetal aneuploidy detection.
Key words: aneuploidy detection, Down syndrome, noninvasive
prenatal diagnostics, trisomyCite this article as: Sparks AB, Struble CA, Wang ET, et al. Noninvasive prenatal detection and selective analysis of cell-free DNA obtained from maternal blood:
cschmcmtstuFt
The American Congress of Obstetri-cians and Gynecologists (ACOG)
recommends that pregnant women beoffered noninvasive screening for fetalchromosomal abnormalities.1 However,xisting screening methods exhibit de-ection rates in the range of 90-95%nd false-positive rates in the range of
From Aria Diagnostics, San Jose, CA.
Received Nov. 14, 2011; revised Jan. 18,2012; accepted Jan. 23, 2012.
The first 2 authors contributed equally to thiswork.
All authors are employees of Aria Diagnostics.K.S. is a member of the board of the company.
Reprints: Ken Song, MD, Aria Diagnostics,5945 Optical Ct., San Jose CA [email protected].
0002-9378/free© 2012 Mosby, Inc. All rights reserved.doi: 10.1016/j.ajog.2012.01.030
For Editors’ Commentary, seeContents
See related editorial, page 269
VIDEOClick Supplementary Content underthe title of this article in the onlineTable of Contents
3-5%.2-6 Thus, ACOG also recommendshat patients categorized by screening asigh risk for fetal aneuploidy be offered
nvasive testing such as amniocentesis orhorionicvillussampling.Althoughthesein-asive procedures are highly accurate, theyre expensive and entail a risk of miscar-iage.7,8 To address these limitations, severalroups have pursued methods for noninva-ive fetal aneuploidy detection.
Initial efforts, which were focused on iso-ation and analysis of circulating fetal cells,urned out to be challenging.9,10 The realiza-ionthat fetalnucleicacidsarepresent inma-ernal blood spawned efforts to analyze cell-ree DNA (cfDNA) for fetal conditions.11-15
In the last few years, massively parallel shot-gun sequencing (MPSS) has been used toquantify precisely cfDNA fragments for fetaltrisomy detection.16-19 Several groups haverecently used this approach to identify fetaltrisomy 21 (T21), and with less success, tri-somy 18 (T18) and trisomy 13.20-24
The chromosomal dosage resultingfrom fetal aneuploidy is directly re-lated to the fraction of fetal cfDNA. Forexample, a cfDNA sample containing
fetus should ex-
APRIL 2012 Americ
hibit a 2% increase in the proportion ofreads from chromosome 21 (chr21) ascompared to a normal fetus. Distin-guishing these 2 scenarios with highconfidence requires a large number(�93,000) of chr21 observations.24 Be-ause MPSS is indiscriminate with re-pect to chromosomal origin, and be-ause chr21 represents �1.5% of theuman genome, �6.3 million uniquelyapped reads are required to ensure suffi-
ient chr21 counts. Given typical MPSSapping yields of �25%, this translates
o 25 million raw sequencing reads perample. This requirement constrains thehroughput, cost efficiency, and clinicaltility of MPSS for aneuploidy detection.or example, a recently launched producthat detects T21 via MPSS22 has a list price
of approximately $2700 per test.Selective sequencing of relevant chro-
mosomes can address these constraints.We recently described a novel assay, digitalanalysis of selected regions (DANSR),which enables highly multiplexed se-quencing of selected loci from specificchromosomes of interest (Appendix).25
We used DANSR to evaluate loci on chro-mosome 18 (chr18) and chr21 in a set of
subject samples whose aneuploidy statusan Journal of Obstetrics & Gynecology 319.e1
pna3otaaf
apmdea(Ft(
es
Reports of Major Impact www.AJOG.org
was known at the time of analysis and dem-onstrated accurate aneuploidy detection.
In this study, we extend DANSR to as-say simultaneously polymorphic andnonpolymorphic loci in a single reac-tion, enabling estimation of chromo-some proportion and fetal fraction. Wedescribe a novel analysis algorithm, thefetal-fraction optimized risk of trisomyevaluation (FORTE), which uses this in-formation to compute the likelihood offetal trisomy in each subject. We demon-strate the power of this approach in ablinded set of 167 pregnant women, in-cluding 36 T21 and 8 T18 pregnancies.
MATERIALS AND METHODSSubjectsSubjects were prospectively enrolled uponproviding informed consent, under proto-cols approved by institutional reviewboards. Subjects were required to be atleast 18 years of age, to be at least 10 weeks’gestational age, and to have singleton preg-nancies. A subset of enrolled subjects, con-sisting of 250 women with disomic preg-nancies, 72 with T21 pregnancies, and 16with T18 pregnancies, was selected for in-clusion in this study. The subjects wererandomized into a training set consistingof 127 disomic pregnancies, 36 T21 preg-nancies, and 8 T18 pregnancies, and a val-idation set consisting of 123 disomic preg-nancies, 36 T21 pregnancies, and 8 T18pregnancies. The trisomy status of eachpregnancy was confirmed by invasive test-ing (fluorescent in situ hybridizationand/or karyotype analysis). The trisomystatus of the training set was known at thetime of analysis; in the validation set, thetrisomy status was kept blinded until afterFORTE analysis.
DANSR assayWe designed DANSR assays against lociin the human genome as previously de-scribed.25 To assess chromosome pro-
ortion, we designed assays against 576onpolymorphic loci on each of chr18nd chr21, where each assay consisted oflocus-specific oligonucleotides: a left
ligo with a 5’ universal amplificationail, a 5’ phosphorylated middle oligo,nd a 5’ phosphorylated right oligo with3’ universal amplification tail. To assess
etal fraction, we designed assays against a
319.e2 American Journal of Obstetrics & Gynecolo
set of 192 single nucleotide polymor-hism (SNP)-containing loci on chro-osomes 1-12, where 2 middle oligos,
iffering by 1 base, were used to queryach SNP. SNPs were optimized for minorllele frequency in the HapMap 3 datasethttp://hapmap.ncbi.nlm.nih.gov/.Accessedeb. 14, 2012). Oligonucleotides were syn-hesized by Integrated DNA TechnologiesCoralville, IA)andpooledtogether tocreate
FIGURE 1Schematic of digital analysis of sel
ANSR process applied to unselected (left) vs’phosphate and 3’hydroxyl moieties, respectiith biotin (B) moiety and bound to streptavpecific DANSR oligos (orange) are annealedognate locus sequences in cfDNA, their termireation of ligation product capable of supportineaction (UPCR) primers (purple). Elution of thrimers containing 96 distinct 7-base sampleus sequencing of 96 different UPCR productontain universal tail sequences that supporample-specific bases, respectively. In additiohat support HiSeq (Illumina, San Diego, CA) cparks. Noninvasive prenatal detection and selective analysisynecol 2012.
single multiplexed DANSR assay pool.
gy APRIL 2012
DANSR product was generated fromach subject sample as previously de-cribed (Figure 1).25 Briefly, 8 mL of
blood per subject was collected into acfDNA tube (Streck, Omaha, NE) andstored at room temperature for up to 3days. Plasma was isolated from blood viadouble centrifugation and stored at�20°C for up to a year. cfDNA was iso-lated from plasma using viral nucleic acid
ted regions (DANSR) assay
ected (right) loci. Circles and arrows indicate. Cell-free DNA (cfDNA) (black) is first labeled-coated magnetic beads (SA). Next, locus-fDNA. When DANSR oligos hybridize to theirorm 2 nicks. Ligation of these nicks results inmplification using universal polymerase chainigation product followed by UPCR with UPCRs (purple box) enables pooling and simultane-n a single lane. Left and right UPCR primersquencing of locus-specific 56 bases and 7PCR primers contain universal tail sequencester amplification.ll-free DNA obtained from maternal blood. Am J Obstet
ec
D sel5 velyw idins to cc ni fc g ar is lp tago s oc t ses n, Ut lusS of ceG
DNA purification beads (Dynal, Grand
l8
totccsdp
wgsfat
flss
iustS
wtdssfpdcp
lathtlmrwC
dsp
w
nsCttxsj
w
tsrbtPstcttmm
a
www.AJOG.org Reports of Major Impact
Island, NY), biotinylated, immobilized onMyOne C1 streptavidin beads (Dynal),and annealed with the multiplexed DANSRoligonucleotide pool. Appropriately hy-bridized oligonucleotides were catenatedwith Taq ligase, eluted from the cfDNA,and amplified using universal polymerasechain reaction primers. Polymerase chainreaction product from 96 independentsamples was pooled and used as a templatefor cluster amplification on a single lane ofa TruSeq v2 SR flow slide (Illumina, SanDiego, CA). The slide was processed on anIllumina HiSeq 2000 to produce a 56-baselocus-specific sequence and a 7-base sam-ple tag sequence from an average of 1.18million clusters/sample. Locus-specificreads were compared to expected locus se-quences. An average of 1.15 million (97%)reads had �3 mismatches with expectedocus sequences, resulting in an average of54 reads/locus/sample.
Analysis of nonpolymorphic locifor chromosome proportionSequence counts were normalized by sys-tematically removing sample and assaybiases. Sequence counts follow a log nor-mal distribution, so biases were estimatedusing median polish on log-transformedcounts.25-27 A chr21 proportion metric washen computed for each sample as the meanf counts for selected chr21 loci divided byhe sum of the mean of counts for selectedhr21 loci and the mean of counts for all 576hr18 loci. A chr18 proportion metric wasimilarly calculated for each sample. A stan-ard Z test of proportions was used to com-ute Z statistics:
Zj �pj � po
�po�1 � p0�nj
(1)
here pj is the observed proportion for aiven chromosome of interest in a givenample j, p0 is the expected proportionor the given test chromosome calculateds the median pj, and nj is the denomina-or of the proportion metric.
Z statistic standardization was per-ormed using iterative censoring on eachane of 96 samples. At each iteration, theamples falling outside of 3 median ab-
olute deviations were removed. After 10 dterations, mean and SD were calculatedsing only the uncensored samples. Allamples were then standardized againsthis mean and SD. The Kolmogorov-mirnov test28 and Shapiro-Wilk test29
were used to establish the normality ofthe uncensored samples’ Z statistics.
Locus selection usingtraining samplesSequence count data from the trainingsamples were first normalized as describedabove and previously.25-27 These samples
ere subsequently analyzed to select 384 ofhe 576 loci on chr21 and chr18 best able toiscriminate T21 and T18 from normalamples. The 384 loci on each chromo-ome exhibiting the greatest residual dif-erence between normal and trisomy sam-les were identified using Z statisticserived from individual loci for the testhromosome and all 576 loci for the com-arison chromosome.
Analysis of polymorphicloci for fetal fractionInformative polymorphic loci were de-fined as loci where fetal alleles differ frommaternal alleles. Because DANSR exhib-its allele specificities �99%, informativeoci were readily identified when the fetalllele proportion of a locus was measuredo be between 1-20%. A maximum likeli-ood estimate using the binomial distribu-ion was employed to determine the mostikely fetal fraction based upon measure-
ents from several informative loci. Theesults correlate well (R2 �0.99) with theeighted average approach presented byhu and colleagues.30
Aneuploidy detection using FORTEThe FORTE algorithm estimates the risk ofaneuploidy using an odds ratio comparinga model assuming a disomic fetal chromo-some and a model assuming a trisomic fe-tal chromosome. Let xj � pj � p0 be the
ifference of the observed proportion pj forample j and the estimated reference pro-ortion p0. FORTE computes:
P�xj�T�P�xj�D�
, (2)
here T is the trisomic model and D is the
isomic model. The disomic model D is aAPRIL 2012 Americ
ormal distribution with mean 0 and aample specific SD estimated by Montearlo simulations described below. The
risomic model T is also a normal distribu-ion with mean 0, by transforming xj to
j � pj � p̂j, the difference between the ob-erved proportion and a fetal fraction ad-usted reference proportion defined by p̂j
p̂j ��1 � 0.5fj�p0
��1 � 0.5fj�p0� � �1 � p0�(3)
here fj is the fetal fraction for sample j.This adjustment accounts for the ex-pected increased representation of a tri-somic fetal chromosome.
Monte Carlo simulations are used to es-timate sample-specific SD for disomic andtrisomic models of proportion differences.Observed proportions for each sample canbe simulated by nonparametric bootstrapsampling of loci and calculating means, orparametric sampling from a normal distri-bution using the mean and SE estimatesfor each chromosome from the observednonpolymorphic locus counts. Similarly,the reference proportion p0 and fetal frac-ion fj can be simulated by nonparametricampling of samples and polymorphic lociespectively, or chosen from normal distri-utions using their mean and SE estimateso account for measurement variances.arametric sampling was used in thistudy. Simulations were executed 100,000imes, and proportion differences wereomputed for each execution to constructhe distributions. Based on the results ofhese simulations in the training set, nor-
al distributions were found to be goododels of disomy and trisomy.The final FORTE risk score is defined
s
P�xj�T�P�T�P�xj�D�P�D�
, (4)
where P�T�⁄P�D� is the prior risk of tri-somy vs disomy. The prior risk was takenfrom well-established tables capturingthe risk of trisomy associated with the
subject’s maternal and gestational age.31an Journal of Obstetrics & Gynecology 319.e3
Stni
wt
sniwewtfptf
s(l
o
ysis o Gyne
Reports of Major Impact www.AJOG.org
RESULTSSample characteristicsThe Table describes the clinical charac-teristics and demographics of the pa-tients whose samples were analyzedin this study. The mean maternal age ofthe disomic, T21, and T18 subjects was34, 34, and 37 years, respectively. Themean gestational age of the disomic, T21,and T18 subjects was 17.7, 19.6, and 17.0weeks. The mean maternal ages of thedisomic, T21, and T18 subjects were notsignificantly different between trainingand validation sets (all t test P � .05).
imilarly, the mean gestational ages ofhe disomic, T21, and T18 subjects wereot significantly different between train-
ng and validation sets (all t test P � .05).
Chromosome proportion Zstatistics in training setTo select loci to be used for aneuploidydetection, we evaluated a set of subjectswhose aneuploidy status was known.This training set consisted of 127 nor-mal, 36 T21, and 8 T18 pregnancies. Sixnormal, 1 T18, and 1 T21 samples (8/171, or 5%) did not meet quality control(QC) criteria (low count, fetal fraction�3%, and/or evidence from SNPs of anonsingleton pregnancy) and were re-moved from the data set. We computedchromosome proportion Z statistics inthe remaining samples for chr18 andchr21 (Figure 2). In all, 120/121 (99.2%)disomic samples had Z statistics �3; 1
TABLESample characteristics
Cohort Status Subjects, n
M
A
Training Normal 127 3.............................................................................
T18 8 3.............................................................................
T21 36 3.............................................................................
Total 171 3...................................................................................................................
Validation Normal 123 3.............................................................................
T18 8 3.............................................................................
T21 36 3.............................................................................
Total 167 3...................................................................................................................
T18, trisomy 18; T21, trisomy 21.
Sparks. Noninvasive prenatal detection and selective anal
disomic sample had a chr21 Z statistic of p
319.e4 American Journal of Obstetrics & Gynecolo
3.5. In all, 35/35 (100%) T21 and 7/7(100%) T18 samples had chromosomeproportion Z statistics �3. Thus, using Zstatistic analysis, DANSR exhibited99.2% specificity and 100% sensitivityfor T21, and 100% specificity and 100%sensitivity for T18.
Fetal fraction in training setA principal determinant of the chromo-some proportion response to aneuploidy isthe fraction of fetal DNA in the sample. Tomeasure fetal fraction reliably, we incorpo-rated 192 DANSR assays targeting SNPsinto our multiplex assay pool. By measur-ing fetal fraction and chromosome pro-portion in the same reaction we ensure es-timates of fetal fraction from polymorphicassays closely represent fetal fraction in thenonpolymorphic assays used to assesschromosome proportion. Fetal fractionexhibited a strong correlation (R2 �0.90)
ith the chromosome proportion Z statis-ic in trisomic pregnancies (Figure 2).
Importantly, the Z statistic was not re-ponsive to fetal fraction in normal preg-ancies (Figure 2), reflecting a major lim-
tation of the Z statistic metric: samplesith low Z statistic values arise from both
uploid samples and aneuploid samplesith modest fetal fraction. We reasoned
hat a metric that was responsive to fetalraction in euploid as well as aneuploidregnancies would be preferable. Weherefore developed FORTE, which usesetal fraction information to: (1) define ex-
rnal age, y
age SD Minimum Maximum
6.3 18 44..........................................................................................................................
6.1 27 44..........................................................................................................................
6.6 18 44..........................................................................................................................
6.3 18 44.........................................................................................................................
7.1 18 51..........................................................................................................................
6.5 25 43..........................................................................................................................
7.1 18 46..........................................................................................................................
7.1 18 51.........................................................................................................................
f cell-free DNA obtained from maternal blood. Am J Obstet
ected chromosome proportions for tri- i
gy APRIL 2012
omic vs disomic test chromosomes, and2) compute the odds that a sample be-ongs to one or the other group.
Analysis of training set using FORTEWe used FORTE to compute the odds oftrisomy vs disomy of chr18 and chr21 ineach sample within the training set (Fig-ure 3). As expected, the FORTE oddsdemonstrated a response to fetal fractionin both trisomic and disomic samples,and the response magnitude was ap-proximately equivalent in the 2 groups.FORTE correctly discriminated all eup-loid from aneuploid samples, and thedifference between the lowest aneuploidodds and the euploid odds was �1012.All aneuploidy samples had odds �1010.
Validation of FORTE aneuploidyanalysis on blinded setTo test the performance of the DANSR/FORTE assay in an independent set ofsubjects, we assayed a blinded validationset consisting of 123 normal, 36 T21, and8 T18 pregnancies. All samples passedQC criteria and were assigned FORTEodds scores for chr18 and chr21 (Figure4). As above, FORTE correctly discrimi-nated all trisomy from disomy subjects.The difference between the lowest aneu-ploid odds and the highest euploid oddswas 103.9. All 36 T21 and 8 T18 sampleshad trisomy odds �102.67 (�99.8% risk
f trisomy).Current prenatal aneuploidy screen-
tational age, wk
rage SD Minimum Maximum
1 4.3 10.3 32.4..................................................................................................................
4 4.4 13.0 25.9..................................................................................................................
9 4.4 11.0 33.0..................................................................................................................
6 4.4 10.3 33.0..................................................................................................................
3 3.5 11.0 30.4..................................................................................................................
7 3.3 11.7 21.0..................................................................................................................
1 5.1 12.3 36.1..................................................................................................................
6 4.0 11.0 36.1..................................................................................................................
col 2012.
ate Ges
ver Ave
4.4 17.......... .........
7.7 18.......... .........
4.2 18.......... .........
4.5 17.......... .........
3.1 18.......... .........
6.6 15.......... .........
4.1 20.......... .........
3.5 18.......... .........
ng tests employ varying risk thresholds,
it
www.AJOG.org Reports of Major Impact
APRIL 2012 Americ
but generally risks of 1 in 100 (10–2) to 1n 300 (10–2.5) are referred to invasiveesting.3,32 If this threshold range were
applied to the FORTE odds for theblinded set, it would yield 99.2% speci-ficity and 100% sensitivity for each chro-mosome. This compares favorably withcurrent screening methods, which canentail 5% false-positive and 10% false-negative rates.3,5 Moreover, because theminimum difference between the eup-loid and aneuploid subjects’ FORTEodds was almost 4 orders of magnitudefor T21 and 14 orders of magnitude forT18, a variety of thresholds produce per-fect sensitivity and specificity.
COMMENTPrincipal findings of this studyThis study demonstrates the analyticalperformance of DANSR and FORTE indetecting fetal T21 and T18 in pregnantwomen of at least 10 weeks’ gestationalage. DANSR refers to the biochemical as-say that involves directed analysis ofcfDNA. FORTE refers to the algorithmthat provides an individualized risk scorefor T21 and T18 taking into account age-related risks and fetal fraction of the sam-ple. The combination of DANSR andFORTE correctly identified all 36 cases ofT21 and 8 cases of T18 as having �99%risk for each trisomy in a blinded analy-sis. There was at least 1000-fold magni-tude separation in the risk score betweentrisomic and disomic samples.
Difference between DANSRand MPSS assayBy generating sequencing template fromchromosome-specific assays and by pro-ducing high mapping rates, DANSR per-mits aneuploidy detection using �1 mil-lion raw reads per subject, enablinganalysis of 96 subjects per sequencinglane. By contrast, MPSS evaluates the en-tire genome, and requires �25 millionraw reads per subject, which limits se-quencing throughput to 4-6 samples perlane. Thus, DANSR enjoys a substantiveadvantage over MPSS in sequencing costand throughput.
DANSR enables genotyping of individ-ual polymorphic loci that is not possibleusing current MPSS approaches. DANSR
FIGURE 2Training set Z statistics vs fetal fraction
Chromosome proportion Z statistic is plotted for A, chromosome 18 (chr18) or B, chromosome 21 (chr21)vs fraction of fetal DNA for each training set subject. Disomic subjects are represented as blue diamonds,trisomic subjects as red. Z statistic of trisomic subjects increases with increasing fetal fraction. By contrast,Z statistic of disomic subjects is not responsive to fetal fraction. Thus, when evaluating disomic subjects, Zstatistic does not reflect increase in certainty expected to result from increased fetal fraction.Sparks. Noninvasive prenatal detection and selective analysis of cell-free DNA obtained from maternal blood. Am J ObstetGynecol 2012.
allowed us to develop an integrated assay
an Journal of Obstetrics & Gynecology 319.e5
i
rbDat
caacl
Reports of Major Impact www.AJOG.org
to assess polymorphic as well as nonpoly-morphic loci, thereby permitting simulta-neous determination of fetal fraction andchromosome proportion. We used fetalfraction information by imposing a QC re-quirement that each sample have at least3% fetal DNA, thereby avoiding low con-fidence calls arising from low proportionsof fetal DNA. In addition, we developedthe FORTE algorithm to produce a fetalfraction-dependent risk score indicatingthe odds of a sample being trisomic vsdisomic.
Purpose of the FORTE algorithmFORTE is a novel algorithm that incor-porates multiple risk factors to generatean individualized odds score for trisomy.While the result format of FORTE is sim-ilar to that of current prenatal screeningresults, the combination of DANSRand FORTE yields greatly improvedperformance. FORTE analysis differsfrom chromosome proportion Z statis-tic analysis in several important re-spects. First, because we process 96samples in a single batch/lane, FORTEuses the observed variances within andbetween samples in a lane, rather thanestimating variance based upon infor-mation obtained from a previously an-alyzed reference data set. Thus, FORTEis less susceptible to process drift anddoes not require external referencesamples or normalizing adjustmentsbased upon historical information.
Second, FORTE is responsive to fetalfraction in both the trisomic and disomicstate, whereas Z statistic is only respon-sive to fetal fraction in the trisomic state.As a consequence, FORTE producesoverall better separation of trisomic vsdisomic samples. Moreover, becausesamples with low fetal fraction yield oddswith lower magnitudes in both disomicand trisomic samples, FORTE commu-nicates a more accurate understandingof the confidence with which a call is be-ing made in disomic samples as well astrisomic samples.
Third, because the risk of aneuploidyvaries significantly with maternal andgestational age, and because incorpo-rating these risks is standard practice inreporting screening results,5,33 FORTE
s designed to accommodate incorpo- b319.e6 American Journal of Obstetrics & Gynecolo
ation of age-related risks. Specifically,ecause both the risk computed fromANSR and the age-related risk reflectsubject’s odds of trisomy vs disomy,
hese risk components are readily com-
FIGURE 3Training set FORTE odds vs fetal fr
ORTE-computed odds of trisomy vs disomy forchr21) is plotted vs fraction of fetal DNA for eacented as blue diamonds, trisomic subjects as rencreasing fetal fraction among trisomic subjectreases with increasing fetal fraction among disrisomic subjects, FORTE metric reflects increaseORTE, fetal-fraction optimized risk of trisomy evaluation.
parks. Noninvasive prenatal detection and selective analysisynecol 2012.
ined. It would also be possible to in- t
gy APRIL 2012
orporate other risk components suchs prior pregnancy with trisomy fetusnd other prenatal testing results. Byontrast, the Z statistic reflects the like-ihood that a sample is disomic, and
ion
chromosome 18 (chr18) or B, chromosome 21aining set subject. Disomic subjects are repre-ORTE-computed odds of trisomy increases withimilarly, FORTE-computed odds of trisomy de-c subjects. Thus, when evaluating disomic andcertainty resulting from increased fetal fraction.
ll-free DNA obtained from maternal blood. Am J Obstet
act
F A,( h trs d. Fi s. Sc omit inF
S of ceG
herefore is not readily combined with
cat
www.AJOG.org Reports of Major Impact
age-related risks of trisomy vs disomy.One consequence of this deficiency isthat the Z statistic will exhibit differentperformance depending upon a sub-
FIGURE 4Validation set FORTE risk vs fetal f
FORTE-computed odds of trisomy vs disomy for(chr21) is plotted vs fraction of fetal DNA for eachrepresented as blue diamonds, trisomic subjectstrisomy increases with increasing fetal fraction aming fetal fraction among disomic samples. Thus,from increased fetal fraction in both disomic andFORTE, fetal-fraction optimized risk of trisomy evaluation.
Sparks. Noninvasive prenatal detection and selective analysisGynecol 2012.
ject’s age. For example, an 18-year-old
subject at 12 weeks’ gestation and witha Z statistic of 3 is �38 times morelikely to be a false positive than a 44-year-old subject at 12 weeks’ gestation
tion
chromosome 18 (chr18) or B, chromosome 21ded validation set subject. Disomic subjects are
red. As in training set, FORTE-computed odds ofg trisomic subjects and decreases with increas-TE metric reflects increase in certainty resultingomic samples.
ll-free DNA obtained from maternal blood. Am J Obstet
and with the same score.31
APRIL 2012 Americ
Importance of fetal fractionA principal determinant of the accuracy ofany cfDNA analysis method is the fractionof fetal cfDNA in the sample. The higherthe fraction of fetal cfDNA, the greater thedifference in the number of cfDNA frag-ments originating from trisomic vs dis-omic chromosomes and hence the easier itis to detect trisomy. The FORTE algorithmexplicitly accounts for fetal fraction in cal-culating trisomy risk.
Limitations of the studyWhereas this study involved blinded anal-ysis of a number of subjects, the total num-ber of T21 and T18 subjects was modestand therefore larger studies are warranted.This study was also limited to women con-sidered high risk for fetal trisomy, as allwomen went on to invasive testing. Somehave questioned whether the performanceof trisomy detection using cfDNA wouldvary between clinically high-risk and low-risk pregnancies.34 Additional studiesomparing fetal fraction between low- andverage-risk pregnancies would addresshis matter.
Future developmentsBecause DANSR enables directed anal-ysis of specific genomic regions, DANSRcould possibly be used to evaluate ge-netic conditions besides trisomy, suchas subchromosomal conditions (eg,microdeletions). Because the FORTEalgorithm can incorporate multipleclinical risk factors, incorporation ofadditional risk information, includingthat from ultrasound, warrants inves-tigation. Technical improvements inDNA sequencing continue to be madeat a rapid pace. Current cfDNA analy-sis turnaround times of 7-10 days arethus likely to be reduced as furthertechnical advances are made.22
Clinical implicationsThe clinical application of noninvasivetesting via cfDNA warrants consider-ation. Whereas the term “noninvasiveprenatal diagnosis” has been used exten-sively in the literature, it is premature toregard these tests as diagnostic. A re-cently completed large-scale validationstudy using MPSS demonstrated a 98.6%detection rate for T21, raising the ques-
rac
A,blinason
FORtris
of ce
tion of whether current cfDNA testing is
an Journal of Obstetrics & Gynecology 319.e7
ee51a
Reports of Major Impact www.AJOG.org
more appropriate as a higher perfor-mance screen or as a diagnostic test.22
To date, studies of cfDNA analysis havebeen focused on detection of common fetaltrisomies. Other prenatal screening modali-ties such as ultrasound are likely to continuetoplayanimportantroleinconjunctionwithcfDNA analysis since T21 and T18 representa subset of fetal anomalies.35 Should nonin-vasive testing via cfDNA become affordableand widely accessible, the myriad of prenatalscreening and testing options today mayevolve to a simpler model in which cfDNAanalysis and ultrasound become a new stan-dard for all pregnant women. f
ACKNOWLEDGMENTSWe would like to acknowledge the following individ-uals for their material contribution in recruiting sub-jects for this study: Herb Brar, MD, Prenatal Diag-nostic and Perinatal Center; Jonathan Weiss, MD,East Bay Perinatal Medical Associates; Louise Lau-rent, MD, PhD, University of California, San Diego;Hanmin Lee, MD, University of California, San Fran-cisco; Aaron Caughey, MD, PhD, Leonardo Pereira,MD, Oregon Health and Science University.
REFERENCES1. American College of Obstetricians and Gyne-cologists. ACOG practice bulletin no. 77:screening for fetal chromosomal abnormalities.Obstet Gynecol 2007;109:217-27.2. Nicolaides KH. Screening for fetal aneup-loidies at 11 to 13 weeks. Prenat Diagn 2011;31:7-15.3. Malone FD, Canick JA, Ball RH, et al. First- andsecond-trimester evaluation of risk (FASTER) re-search consortium. N Engl J Med 2005;353:2001-11.4. Wald NJ. Prenatal screening for open neuraltube defects and Down syndrome: three de-cades of progress. Prenat Diagn 2010;30:619-21.5. Nicolaides KH. Nuchal translucency andother first-trimester sonographic markers ofchromosomal abnormalities. Am J Obstet Gy-necol 2004;191:45-67.6. Rozenberg P, Bussières L, Chevret S, et al.Screening for Down syndrome using first-tri-mester combined screening followed by sec-ond-trimester ultrasound examination in an un-selected population. Am J Obstet Gynecol2006;195:1379-87.7. Caughey AB, Hopkins LM, Norton ME. Cho-rionic villus sampling compared with amniocen-tesis and the difference in the rate of pregnancyloss. Obstet Gynecol 2006;108:612-6.8. American College of Obstetricians and Gyne-cologists. ACOG practice bulletin no. 88: inva-sive prenatal testing for aneuploidy. Obstet Gy-necol 2007;110:1459-67.9. Bischoff FZ, Lewis DE, Nguyen DD, et al. Pre-
natal diagnosis with use of fetal cells isolated319.e8 American Journal of Obstetrics & Gynecolo
from maternal blood: five-color fluorescent insitu hybridization analysis on flow-sorted cellsfor chromosomes X, Y, 13, 18, and 21. Am JObstet Gynecol 1998;179:203-9.10. Bianchi DW, Hanson J. Sharpening thetools: a summary of a National Institutes ofHealth workshop on new technologies for de-tection of fetal cells in maternal blood for earlyprenatal diagnosis. J Matern Fetal NeonatalMed 2006;19:199-207.11. Al-Mufti R, Howard C, Overton T, et al. De-tection of fetal messenger ribonucleic acid inmaternal blood to determine fetal RhD status asa strategy for noninvasive prenatal diagnosis.Am J Obstet Gynecol 1998;179:210-4.12. Harper TC, Finning KM, Martin P, Moise KJJr. Use of maternal plasma for noninvasive de-termination of fetal RhD status. Am J ObstetGynecol 2004;191:1730-2.13. Geifman-Holtzman O, Grotegut CA,Gaughan JP. Diagnostic accuracy of noninvasivefetal Rh genotyping from maternal blood–a metaanalysis. Am J Obstet Gynecol 2006;195:1163-73.14. Fan HC, Blumenfeld YJ, El-Sayed YY,Chueh J, Quake SR. Microfluidic digital PCRnables rapid prenatal diagnosis of fetal an-uploidy. Am J Obstet Gynecol 2009;200:43.e1-7.5. Tynan JA, Angkachatchai V, Ehrich M, Pal-dino T, van den Boom D, Oeth P. Multiplexed
analysis of circulating cell-free fetal nucleicacids for noninvasive prenatal diagnosticRHD testing. Am J Obstet Gynecol 2011;204:251.e1-6.16. Chiu RW, Chan KC, Gao Y, et al. Noninva-sive prenatal diagnosis of fetal chromosomalaneuploidy by massively parallel genomic se-quencing of DNA in maternal plasma. Proc NatlAcad Sci U S A 2008;105:20458-63.17. Fan HC, Blumenfeld YJ, Chitkara U, et al.Noninvasive diagnosis of fetal aneuploidy byshotgun sequencing DNA from maternal blood.Proc Natl Acad Sci U S A 2008;105:266-71.18. Chen EZ, Chiu RWK, Sun H, et al. Noninva-sive prenatal diagnosis of fetal trisomy 18 andtrisomy 13 by maternal plasma DNA sequenc-ing. PLoS One 2011;6:e21791.19. Chiu RW, Akolekar R, Zheng YWL, et al.Non-invasive prenatal assessment of trisomy21 by multiplexed maternal plasma DNA se-quencing: large scale validity study. BMJ 2011;342:c7401.20. Ehrich M, Deciu C, Zwiefelhofer T, et al.Noninvasive detection of fetal trisomy 21 by se-quencing of DNA in maternal blood: a study in aclinical setting. Am J Obstet Gynecol 2011;204:205.e1-11.21. Shulman LP. One small step and one giantleap for noninvasive prenatal screening: an ed-itorial. Am J Obstet Gynecol 2011;205:S9-13.22. Palomaki GE, Kloza EM, Lambert-Messer-lian GM, et al. DNA sequencing of maternalplasma to detect Down syndrome: an interna-tional clinical validation study. Genet Med
2011;13:913-20.gy APRIL 2012
23. Sehnert AJ, Rhees B, Comstock D, et al.Optimal detection of fetal chromosomal abnor-malities by massively parallel DNA sequencingof cell-free fetal DNA from maternal blood. ClinChem 2011;57:1042-9.24. Fan HC, Quake SR. Sensitivity of noninva-sive prenatal detection of fetal aneuploidy frommaternal plasma using shotgun sequencing islimited only by counting statistics. PLoS One2010;5:e10439.25. Sparks AB, Wang ET, Struble CA, et al. Se-lective analysis of cell-free DNA in maternalblood for evaluation of fetal trisomy. Prenat Di-agn 2012 Jan 6 [Epub ahead of print]. doi:10.1002/pd.2922.26. Tukey JW. Exploratory data analysis. Read-ing, MA: Addison-Wesley; 1977.27. Irizarry RA, Bolstad BM, Collin F, Cope LM,Hobbs B, Speed TP. Summaries of AffymetrixGeneChip probe level data. Nucleic Acids Res2003;31:e15.28. Conover WJ. Practical nonparametric sta-tistics. New York: John Wiley & Sons; 1971:295-301.29. Royston P. An extension of Shapiro andWilk’s W test for normality to large samples.Appl Statistics 1982;31:115-24.30. Chu T, Bunce K, Hogge WA, Peters DG. Anovel approach toward the challenge of accu-rately quantifying fetal DNA in maternal plasma.Prenat Diagn 2010;30:1226-9.31. Nicolaides KH. Screening for chromosomaldefects. Ultrasound Obstet Gynecol 2003;21:313-21.32. Fang YM, Benn P, Campbell W, Bolnick J,Prabulos AM, Egan JF. Down syndromescreening in the United States in 2001 and2007: a survey of maternal-fetal medicine spe-cialists. Am J Obstet Gynecol 2009;201:97.e1-5.33. Cuckle H. Integrating antenatal Down’ssyndrome screening. Curr Opin Obstet Gynecol2001;13:175-81.34. Alberry MS, Maddocks DG, Hadi MA, et al.Quantification of cell free fetal DNA in maternalplasma in normal pregnancies and in pregnan-cies with placental dysfunction. Am J ObstetGynecol 2009;200:98.e1-6.35. Souka AP, Von Kaisenberg CS, Hyett JA,Sonek JD, Nicolaides KH. Increased nuchaltranslucency with normal karyotype. Am J Ob-stet Gynecol 2005;192:1005-21.
APPENDIXGlossary of technical terms● Cell-free DNA (cfDNA), DNA that
has been isolated from the plasma por-tion of maternal blood.
● Digital analysis of selected regions(DANSR), a process of analyzing thecounts from assays targeted against se-lected genomic regions.
● HapMap, an international project to
catalog human genome single nucleo-
www.AJOG.org Reports of Major Impact
tide polymorphisms, http://hapmap.ncbi.nlm.nih.gov/.
● Locus, a targeted region of the genome.● Massively parallel shot-gun sequenc-
ing (MPSS), a method that randomlyanalyzes cfDNA fragments for fetal an-euploidy detection.
● Nonpolymorphic locus, a locus with-out known variation.
● Oligo, oligonucleotides, short syn-thesized sequences of DNA that areused in combination to create assays.
● Polymorphic locus, a locus that in-cludes a known single nucleotide poly-morphism.
● (Sequencing) count, the number ofreads identified as coming from (mappedto) a specific locus or assay.
● (Sequencing) read, a single DNA se-quence as determined by a sequencer.
Overview of sequencingThe Illumina HiSeq 2000 DNA sequenc-ing system has been used to generate datafor both the MPSS approaches presentedin other papers and for the approachused in this paper. The basic unit of se-quencing on the system is a lane that iscapable of producing approximately 100million clusters. Each of these clusters issequenced and the resulting raw se-quences are called a “read.” Using a sys-tem of encoding with sequence sampletags these reads can be apportioned toseveral or even 100 different samples. InMPSS, random fragments from thecfDNA of several samples are used tomake the raw sequence reads on a lane.
These sequences are then compared tothe expected sample tags and the humangenome and mapped to a chromosomeof origin. In digital analysis of selectedregions, loci to be sequenced are selectedby a set of oligonucleotide assays specificfor the chromosomes and regions of in-terest. The products of these assays from96 samples are then pooled on a se-quencing lane, converted to raw reads,and compared to the expected 96 sampletags and sequences of the loci that havebeen selected by the assays. Errors in se-quencing can be observed as mismatchesbetween the observed raw sequences andthe expected sequences during compari-son. Typically 97% of the digital analysisof selected regions raw reads map to ex-pected sequences with �3 mismatches.
Statistical terminology● Chromosome proportion, the esti-
mated fraction of reads originatingon a specific chromosome. Fetal-frac-tion optimized risk of trisomy evalua-tion estimates this fraction as the meancount on the specified chromosomedivided by the sum of mean counts onall measured chromosomes.
● Disomic model, a statistical model ofchromosome proportion assumingthe fetus is disomic (has 2 copies of thechromosome).
● Fetal fraction, the estimated fractionof cfDNA in a sample that originatedin the fetus.
● Fetal-fraction optimized risk of tri-somy evaluation (FORTE), a process
of using fetal fraction of cfDNA andAPRIL 2012 Americ
results from sequencing cfDNA togenerating a risk score.
● Informative polymorphic locus, apolymorphic locus where the motherhas a homozygous genotype and thefetus has a heterozygous genotype.
● Log normal, a statistical distributionof data. The data, after log transforma-tion, follow a normal distribution.
● Mean count, the average count of locilocated on a specific chromosome.
● Median polish, an iterative algorithmfor estimating the parameters of a lin-ear model. Similar to estimating pa-rameters in an analysis of variance, re-placing mean estimates with mediansfor robustness to outlying measure-ments.
● Monte Carlo simulation, a computa-tional technique used to numerically es-timate values when mathematical anal-ysis is intractable. The term “MonteCarlo” implies that appropriately com-puted random values are used by simu-lations to generate estimates.
● Risk score, the relative likelihood ofthe disomic model compared to thetrisomic model, as is used in currentprenatal screening.
● Trisomic model, a statistical model ofchromosome proportion assumingthe fetus is trisomic (has 3 copies of thechromosome) and the sample has theestimated fetal fraction of cfDNA.
● Z statistic, a number indicating howfar an observation deviates from theaverage in a population. The unit of aZ statistic is the number of SD.
an Journal of Obstetrics & Gynecology 319.e9