Noninvasive detection of fetal subchromosomalabnormalities by semiconductor sequencing ofmaternal plasma DNAAi-hua Yina,b,c,1,2, Chun-fang Pengd,1, Xin Zhaoa,b,1, Bennett A. Caugheye, Jie-xia Yanga,b, Jian Liua,b, Wei-wei Huanga,b,Chang Liua,b, Dong-hong Luod, Hai-liang Liud, Yang-yi Chend, Jing Wua,b, Rui Houf, Mindy Zhangg, Michael Aie,Lianghong Zhengf, Rachel Q. Xuee, Ming-qin Maia,b, Fang-fang Guoa,b, Yi-ming Qia,b, Dong-mei Wanga,b,Michal Krawczyke, Daniel Zhange, Yu-nan Wanga,b, Quan-fei Huangd,2, Michael Karinh,2, and Kang Zhange,g,i,2

aPrenatal Diagnosis Centre, Guangdong Women and Children Hospital, Guangzhou 510010, China; bMaternal and Children Metabolic-Genetic KeyLaboratory, Guangdong Women and Children Hospital, Guangzhou 510010, China; cGuangdong Thalassemia Diagnostic Centre, Guangzhou 510010, China;dCapitalBio Genomics Co., Ltd., Dongguan 523808, China; eInstitute for Genomic Medicine and Shiley Eye Institute, University of California, San Diego, LaJolla, CA 92328; fKangrui Biological Pharmaceutical Technology Co., Ltd., Guangzhou 510005, China; gMolecular Medicine Research Center, State KeyLaboratory of Biotherapy, West China Hospital and Sichuan University, Chengdu 610041, China; hDepartment of Pharmacology, University of California, SanDiego, La Jolla, CA 92328; and iVeterans Administration Healthcare System, San Diego, CA 92161

Contributed by Michael Karin, September 28, 2015 (sent for review June 3, 2015)

Noninvasive prenatal testing (NIPT) using sequencing of fetal cell-free DNA from maternal plasma has enabled accurate prenataldiagnosis of aneuploidy and become increasingly accepted in clinicalpractice. We investigated whether NIPT using semiconductor se-quencing platform (SSP) could reliably detect subchromosomaldeletions/duplications in women carrying high-risk fetuses. We firstshowed that increasing concentration of abnormal DNA and se-quencing depth improved detection. Subsequently, we analyzedplasma from 1,456 pregnant women to develop a method forestimating fetal DNA concentration based on the size distribution ofDNA fragments. Finally, we collected plasma from 1,476 pregnantwomen with fetal structural abnormalities detected on ultrasoundwho also underwent an invasive diagnostic procedure. We used SSPof maternal plasma DNA to detect subchromosomal abnormalitiesand validated our results with array comparative genomic hybrid-ization (aCGH). With 3.5 million reads, SSP detected 56 of 78 (71.8%)subchromosomal abnormalities detected by aCGH. With increasedsequencing depth up to 10 million reads and restriction of the sizeof abnormalities to more than 1Mb, sensitivity improved to 69 of 73(94.5%). Of 55 false-positive samples, 35 were caused by deletions/duplications present in maternal DNA, indicating the necessity of avalidation test to exclude maternal karyotype abnormalities. Thisstudy shows that detection of fetal subchromosomal abnormalitiesis a viable extension of NIPT based on SSP. Although we focused onthe application of cell-free DNA sequencing for NIPT, we believethat this method has broader applications for genetic diagnosis,such as analysis of circulating tumor DNA for detection of cancer.

noninvasive prenatal testing | NIPT | maternal plasma DNA | cell-freeDNA | semiconductor sequencing

Genomic disorders are defined by loss, gain, or translocationof chromosomal material. Deletion/duplication syndromes

are known to be associated with a wide range of structural andfunctional abnormalities (1), such as Cri du Chat Syndrome (5pdeletion) (2) and DiGeorge Syndrome (22q11.2 deletion) (3).Such deletion/duplication syndromes can be reliably diagnosedprenatally from the DNA of fetal cells; fetal DNAmay be assessedfor chromosomal abnormalities by karyotyping, FISH, compara-tive genomic hybridization (CGH), and array-based technologies(4). G-banded karyotyping is the predominant technique for di-agnosis of chromosomal abnormalities, but it is limited to reso-lution of 5–10 Mb (5, 6). Genomic disorders of a smaller size aremore reliably detected by chromosomal microarray analysis (CMA),of which array CGH (aCGH) is an example.According to The American Society of HumanGenetics, CMA has

replaced the standard metaphase karyotype in postnatal assessment

of individuals with developmental delay, intellectual disability, con-genital anomalies, and autism (7). In December of 2013, The Amer-ican Congress of Obstetricians and Gynecologists and the Society ofMaternal Fetal-Medicine recommended prenatal CMA as a first-linetest in the case of abnormal ultrasound findings (8). This methodidentifies structural anomalies of chromosomes with great accuracy inprenatal and postnatal stages alike and increases detection of clinicallysignificant copy number changes, even when the conventional meta-phase karyotype appears normal. However, CMA or any of the otherabove-mentioned conventional techniques requires that an invasiveprocedure, such as amniocentesis, chorionic villus sampling (CVS), orumbilical cord blood collection, be performed to obtain a source offetal genomic DNA. These procedures carry small but well-validatedrates of pregnancy loss that may approach 1% (9) and can be per-formed only by specially trained obstetricians. These invasive di-agnostic tests are most often performed after a positive screening testor abnormal ultrasound. Widely used trisomy 21 screening programsat best detect 95% of cases with a false-positive rate of 5%, implyingthat the risk of pregnancy loss is greater than 1/10,000 among normalpregnancies that undergo screening with subsequent invasive testing.

Significance

Noninvasive prenatal testing (NIPT) using sequencing of fetalcell-free DNA from maternal plasma has allowed accurateprenatal diagnosis of aneuploidy without requirement of aninvasive procedure; this approach has gained increasing clinicalacceptance. In this manuscript, we evaluate whether NIPT us-ing semiconductor sequencing platform (SSP) could reliablydetect subchromosomal deletions/duplications in women car-rying a high-risk fetus. We show that detection of fetal sub-chromosomal abnormalities is a viable extension of NIPT usingSSP. Given the recent strong interest in NIPT using cell-freeDNA, these results should be of broad interest and immediateclinical importance, even beyond the NIPT field, such as fordetection of cell-free tumor DNA.

Author contributions: A.-h.Y., Q.-f.H., M. Karin, and K.Z. designed research; C.-f.P.,X.Z., J.-x.Y., J.L., W.-w.H., C.L., D.-h.L., J.W., R.Q.X., M.-q.M., F.-f.G., Y.-m.Q., D.-m.W.,and Y.-n.W. performed research; C.-f.P., H.-l.L., and Y.-y.C. contributed new reagents/analytic tools; A.-h.Y., B.A.C., H.-l.L., R.H., M.Z., M.A., L.Z., M. Krawczyk, D.Z., and K.Z.analyzed data; and A.-h.Y., B.A.C., H.-l.L., M. Karin, and K.Z. wrote the paper.

The authors declare no conflict of interest.1A.-h.Y., C.-f.P., and X.Z. contributed equally to this work.2To whom correspondence may be addressed. Email: yinaiwa@vip.126.com, qfhuang@capitalgenomics.com, mkarin@ucsd.edu, or kang.zhang@gmail.com.

This article contains supporting information online at www.pnas.org/lookup/suppl/doi:10.1073/pnas.1518151112/-/DCSupplemental.

14670–14675 | PNAS | November 24, 2015 | vol. 112 | no. 47 www.pnas.org/cgi/doi/10.1073/pnas.1518151112

There are no serum-based tests in clinical use that specificallyscreen for subchromosomal abnormalities.Recently published international guidelines (10) recommend

the clinical use of noninvasive prenatal testing (NIPT) for an-euploidy screening only among pregnant women whose fetusesare deemed at high risk; NIPT has become increasing adoptedfor this purpose. Studies, such as the study by Bianchi et al. (11),have shown the potential of using cell-free DNA (cfDNA) forfetal aneuploidy testing in average-risk women, in particularbecause of reductions in the false-positive rate compared withstandard prenatal testing. However, NIPT has not yet beenadopted as standard practice for average-risk women, and con-cerns persist regarding the relatively high cost per detected casein this population (10). NIPT methods currently in clinical usedepend primarily on Illumina sequencing technologies. We re-cently published a study (12) that showed the viability of using asemiconductor sequencing platform (SSP) for NIPT. SSP offersthe potential advantages of lower cost and faster sequencing, andthe small footprints of these machines may make them easier todeploy in clinical laboratories.NIPT through either platform offers several advantages over

current screening practices that rely on serum and sonographicmarkers. NIPT has the potential to avoid most invasive antenataldiagnostic procedures performed secondary to positive trisomyscreen, abnormal ultrasound, or advanced maternal age becauseof greatly reduced false-positive rates. NIPT can generally beperformed in the late first trimester or around 12–13 wk (13),and the sequencing turnaround may be faster than methods re-lying on cell culture. Thus, a pregnant woman who chooses toterminate a pregnancy based on the result will have the option todo so earlier in her pregnancy compared with the most accuratestandard screening (10). Elective termination at an earlier ges-tational age is safer for the woman (14) as well as enables greaterprivacy and less emotional distress.Previous studies have shown the potential of extending NIPT to

detect fetal deletion/duplication syndromes from maternal plasma.Peters et al. (15) reported that maternal plasma sequencing with243 million reads could identify a 4-Mb deletion at 35 wk ofgestation, and more recently, Srinivasan et al. (16) identifiedsubchromosomal duplications and deletions, translocations, mo-saicism, and trisomy 20 by maternal plasma sequencing in sevencases, including a 300-kb microdeletion. Unfortunately, both ofthese studies required deep sequencing, increasing both cost andtime required, and they were limited to only a handful of patients(15, 16). No study has yet been published that evaluates the po-tential of NIPT to detect subchromosomal abnormalities in a largepopulation of pregnant women.The fetal DNA fraction is a critical parameter affecting the

accuracy of NIPT using cfDNA in maternal plasma. SequencingSNPs was considered as the primary method of determining fetalDNA fraction (17). Recently, Lo et al. (18) developed a novelmethod that relies on the fact that circulating fetal DNA mole-cules are generally shorter than the corresponding maternalDNA molecules to estimate the fraction of fetal DNA in ma-ternal plasma. This approach can be adapted for an SSP, which iscapable of sequencing entire molecules of fetal DNA to obtainthe size of the DNA fragments.Based on the potential for NIPT to provide a fetal karyotype,

we investigated whether we could further develop our techniqueof NIPT with SSP to detect subchromosomal deletions/duplica-tions in a population of women with high-risk pregnancies. Thesliding window strategy has been reported to accurately detectcopy number variations in the human genome with relativity low-coverage whole-genome sequencing (19). Thus, this strategycould enable the detection of deletions/duplications in fetalcfDNA from maternal plasma using less sequencing depth and atlower cost. Our study highlights the prospect of universal andpractical noninvasive screening for deletions/duplications in the

fetal genome that can be incorporated into the current programof noninvasive prenatal detection of fetal aneuploidy.

ResultsOptimization of Fetal Deletion/Duplication Detection by SSP. To testthe accuracy of fetal deletion/duplication detection by SSP, DNAderived from the blood of 19 newborns with known deletions/duplications diagnosed by aCGH (Table S1) was mixed withDNA from normal samples at a range of concentrations [5%, 10%,15%, 20%, and 30% (mass/mass)]. Each mixed sample was se-quenced at various depths (3.5, 5, 10, or 15 million reads). Samples14,452, 14,095, and 14,181 harbored two deletions or duplicationseach, whereas the remaining samples had one deletion/duplication.In total, 22 deletions/duplications were analyzed. Stouffer’s Z scoreswere generated as described in Materials and Methods for eachsample at every concentration and sequencing depth; Z scores in-creased with both concentration and sequencing depth, with theaverage Z scores for all samples shown in Fig. 1. The Z scores ofsample 14,945, with DNA that contains a 21-Mb deletion, areshown in Fig. S1 A and B as an example. Fig. S1B shows that, at aconstant sequencing depth of 5 million reads, Z-score magnitudeincreases with greater abnormal DNA concentration. Fig. S1Bshows increasing Z-score magnitude with increasing sequencingdepth at a constant 15% abnormal DNA concentration. The aCGHplot for sample 14,945 is shown in Fig. S1C as a comparison.Increasing the fraction of abnormal DNA as well as increasing

the sequencing depth resulted in improved detection and de-lineation of subchromosomal abnormalities (Fig. 2). At the lowestconcentration of 5% and 3.5 million reads, the smallest abnormalitydetected was 20 Mb. Increasing sequencing depth at this concen-tration dramatically improved detection of deletions/duplications toas small as 3 Mb with 15 million reads. The rates of false-positiveand false-negative events decreased with increasing concentrationas well as increasing sequencing depth. These results show thefeasibility of using Stouffer’s Z scores generated with a slidingwindow strategy on SSP to detect subchromosomal abnormalities.

Fetal Fraction Estimation in Maternal Plasma DNA by Size Analysis.Given that the concentration of fetal DNA seems to affect theaccuracy of NIPT, we investigated a technique using sequencing

Fig. 1. Average absolute Z scores of subchromosomal abnormalities for allmixed abnormal and normal DNA samples at various concentrations andsequencing depths. Error bars represent ±1.

Yin et al. PNAS | November 24, 2015 | vol. 112 | no. 47 | 14671

MED

ICALSC

IENCE

S

to estimate the fraction of fetal cfDNA in maternal plasma.Because fetal cfDNA is generally shorter than maternal cfDNA,maternal plasma samples with a higher fetal DNA concentrationwould have a higher proportion of short DNA fragments. Wepostulated that, if the proportions of short and long DNA frag-ments in maternal plasma were correlated with the fetal DNAfraction, we would be able to determine the fetal DNA fractionusing maternal plasma DNA size analysis alone. We analyzedplasma samples from 1,456 pregnant women, 712 of whom car-ried a male fetus, performing SSP at 400 flows to sequence theentire length of fetal DNA fragments. The overall DNA fragmentsize distribution for each maternal plasma sample was generated.Characteristic size distributions from six individual maternalplasma samples with various fetal DNA fractions are shown in Fig.3A. The aggregated distributions of all samples are shown in Fig.3B as the average reads ratio at each fragment size, with thevariance around each point shown by the gray region.Samples from 712 women carrying male fetuses were randomly

divided into a training group and a validation group, each con-taining 356 samples. The fetal DNA concentration was deter-mined from the proportion of Y-chromosome sequences inmaternal plasma. We then examined the relationship between thereads ratio of different fragment sizes and fetal DNA fraction inthe training group. We observed a correlation between the fetalDNA fraction estimated by the Z score of Y chromosome and thereads ratio of two DNA fragment size regions: 130–140 bp (regionA in Fig. 4A) and 155–175 bp (region B in Fig. 4B). Fig. 3 showsregions A and B in relation to the size distributions of the ma-ternal cfDNA fragments denoted by vertical blue or red bands,respectively. As described in Materials and Methods, we used thiscorrelation to derive an equation for estimating the fetal DNAconcentration. Next, we estimated the fetal DNA fractions for 356samples in the validation group using the regression equationobtained from the training group. For each size region, the size-deduced fetal DNA fractions were highly concordant with thosedetermined using the proportion of Y chromosome (region A: r =0.839, P < 2.2e-16 and region B: r = −0.83, P < 2.2e-16 by linearregression). The estimated fetal fraction was highly concordantbetween each region as well (Fig. 4C) (r = 0.818, P < 2.2e-16 bylinear regression). As described in Materials and Methods, the es-timated fetal DNA fraction was considered unreliable for a smallnumber of samples, in which the estimated fetal fraction was in-consistent between the different size regions.

Performance of SSP for Detecting Fetal Deletions/Duplications. Weprospectively recruited a cohort of 1,476 pregnant women pre-senting with a fetal indication of structural abnormalities de-tected on ultrasonography for invasive prenatal testing. Theaverage maternal age of the cohort was 29.7 ± 5.4 y old (range =17–46 y old), and the average gestational age was 24 ± 6 wk(range = 12–37 wk). Maternal plasma was collected concurrentlywith fetal DNA obtained using an invasive procedure, such asamniocentesis or chorionic villous sampling. Invasively collectedfetal DNA was tested by aCGH, whereas DNA derived frommaternal plasma was sequenced by SSP. We applied our methoddescribed above to estimate the fetal DNA fraction in maternal

Fig. 2. Smallest size of deletions/duplications detected in mixed abnormaland normal DNA samples at various sequencing depths and concentrationsof abnormal DNA.

A

B

Fig. 3. Size distribution of maternal plasma cfDNA fragments sequenced bySSP in a population of pregnant women. The blue region encompasses 130–140 bp (region A), and the red region encompasses 155–175 bp (region B).(A) Representative examples from maternal plasma with different fetal DNAfractions. (B) Aggregate of all samples. The blue line represents the meanread ratio of all samples, and the gray region represents ±1 SD.

14672 | www.pnas.org/cgi/doi/10.1073/pnas.1518151112 Yin et al.

plasma for the entire cohort (Fig. S2); 88.5% samples were es-timated to have a fetal DNA fraction of >10%. Although theaverage fetal DNA fraction was greater than 10% for all gesta-tional ages, fetal DNA fraction increased significantly with ges-tational age (r = 0.5863, P value = 3.0190e-27); 66.7% of sampleswith a gestational age of 12–16 wk had a fetal DNA fractiongreater than 10% compared with 95% of samples with gesta-tional age greater than 20 wk.The same sequencing depth of 3.5 million reads was used like

in our previous study on trisomy 21, 18, and 13 and sex chro-mosome aneuploidies (12). To confirm that our technique pro-vided results consistent with our previous study, we first assessedthe ability of SSP to detect aneuploidy in our cohort. In 1,476samples, 73 cases of chromosomal aneuploidy were identified,with the extra chromosome clearly detected in the 1-Mb bincurve of the SSP analysis. Compared with aCGH performed oninvasively derived fetal DNA, the false-positive rate for aneu-ploidy was 0.63% (nine false-positive samples), and the false-negative rate was 0%, consistent with our previous study andother previously published results (11–13, 20). The most com-mon etiologies of false-positive results in NIPT for aneuploidyare placental mosaicism and a demised twin fetus (21, 22). Smallduplications present in maternal DNA may also lead to relativeoverrepresentation of a chromosome in cfDNA, leading to afalse-positive result (23), although we did not detect any in-stances of this in our study.We then applied our method to detect subchromosomal

deletions and duplications (Table S2). Seventy-eight sub-chromosomal deletions and duplications ranging from 0.52 to 84 Mbin size were identified by aCGH in 57 samples, many corre-sponding to known deletion or duplication syndromes; 56 of 78(71.8%) abnormalities were detected by our method at 3.5 mil-lion reads. The size of the deletion/duplication seemed to be themajor determinant of the performance of our method; 42 of 45(93.3%) deletions/duplications larger than 5 Mb and 35 of 35(100%) larger than 10 Mb in size were detected at 3.5 millionreads. In contrast, only 14 of 34 (41.2%) of samples less than5 Mb in size and none less than 1 Mb in size were detected at3.5 million reads. However, for abnormalities greater than 1 Mbin size, sensitivity improved to 69 of 73 (94.5%) with increasingsequencing depth up to 10 million reads.Several false-negative samples (A0133, A0947, 0899, A0310,

A0686, and A0769) had a low estimated fetal DNA fraction, orthe fraction could not be reliably calculated, likely making de-tection of abnormalities more difficult. Furthermore, certainregions of the genome, primarily those with highly repetitiveDNA sequences, such as near centrosomes, showed significantlygreater covariance and thus, made detection of true abnormali-ties in these regions more challenging. In our analysis, wemasked some repetitive sequences before read mapping to

increase accuracy of read counting in each bin, which may haveresulted in false negatives in these repetitive sequence-richregions. For example, the deletion in sample A0769 (chr22:17299942–19770514) is close to centrosome, and the deletion inA0133 (chr14: 1.01E+08–1.07E+08) is located at the end ofchromosome 14, resulting in false negatives, despite increasedsequencing depth.Our assay produced 58 false positives from 55 samples at a

sequencing depth of 3.5 million reads for a false-positive rate of3.8%; 39 of 58 (67.2%) false positives predicted a deletion/duplication less than 5 Mb in size. We hypothesized that many ofthese false positives might be caused by deletions/duplications inmaternal DNA and therefore, would be detected in the maternalplasma. To investigate this hypothesis, we performed aCGH onall 55 samples of maternal DNA to assess for the presence ofdeletions and duplications in the maternal DNA; 35 of 55(63.6%) samples were validated as maternal deletions/duplica-tions, suggesting that the majority of false positives using thismethod can be accounted for by abnormalities present in themother’s genome. This finding has significant implications forthe clinical utility of NIPT to detect subchromosomal abnor-malities, because a positive test using this assay may require afollow-up test to rule out deletions and duplications in thematernal DNA before the fetus can be diagnosed. For theremaining 20 false-positive samples, repeat sequencing andanalysis revealed no repeat false-positive results. Thus, thesefalse positives seem to be the random result of our sequencingtechnique and statistical model.Detection of subchromosomal abnormalities would most likely

be applied as an extension of current methods of NIPT to detectaneuploidy. When results for aneuploidy and subchromosomalabnormalities are considered together, our assay detected 129of 151 karyotype abnormalities of all sizes in 1,476 samples at3.5 million reads for an overall sensitivity of 85.4% comparedwith aCGH. There were 64 false-positive samples total for anoverall specificity of 95.7%. If size is restricted to abnormalitiesgreater than 5 Mb, 115 of 117 abnormalities in 1,476 sampleswere detected for an overall sensitivity and specificity of 98.3%and 98.1%, respectively. These results show that detection ofsubchromosomal abnormalities is a viable extension of NIPTusing SSP.

DiscussionThis study shows that sequencing maternal plasma DNA canreliably produce a fetal karyotype of chromosomal deletions andduplications with relatively high resolution while avoiding thenecessity of an invasive procedure. Although our NIPT methodis less accurate than CMA, especially for deletions/duplicationssmaller than 5 Mb, it offers the advantage of unbiased testingand is substantially more accurate than current blood-based

A B C

Fig. 4. Estimation of the fetal DNA concentration in maternal plasma using the size distribution of cfDNA fragments. (A and B) Correlation between thereads ratio of DNA fragments in (A) the 130- to 140-bp region or (B) the 155- to 175-bp region and the fetal DNA concentration predicted by the Z score of Ychromosome. (C) Correlation between fetal DNA concentration estimated by regions A and B.

Yin et al. PNAS | November 24, 2015 | vol. 112 | no. 47 | 14673

MED

ICALSC

IENCE

S

screening tests for aneuploidy. In our study population, we de-tected 75% more abnormal karyotypes than detection of aneu-ploidy alone at a sequencing depth of 3.5 million reads, withgreater sequencing depth significantly increasing sensitivity. Thus,because no specific prenatal screening exists for subchromosomalabnormalities, the ability of a single test to detect these abnor-malities in addition to aneuploidy represents a powerful extensionof NIPT.Our false-positive rate of 3.8% for subchromosomal abnor-

malities was relatively high and would potentially limit the clinicalutility of our method as a stand-alone screening test. However,64% of false positives were found to be because of abnormalitiespresent in maternal DNA, suggesting that the false-positive ratecould be greatly reduced with a reflex follow-up test of maternalDNA collected simultaneously with maternal plasma. A reflex testof maternal DNA would also identify inherited variants, whichmay represent a significant fraction of true positives as well. Mostof these variants are polymorphisms and unlikely to be detri-mental for the mother. Further lowering of the false-positive ratemay be possible with greater sequencing depth; additional study isrequired to investigate these possibilities adequately.Our study population was drawn from pregnancies with ana-

tomic abnormalities found on ultrasound to generate a high in-cidence of karyotype abnormalities, resulting in tested pregnanciesranging in gestational age from 12 to 37 wk, with an average age of24 wk, much later than NIPT would ideally be applied. In a futurestudy, detection of subchromosomal abnormalities could be ap-plied to a population with lower but still elevated risk (i.e., ad-vanced maternal age) at a gestational age of 12–16 wk. This timingwould offer the advantage of safer, less expensive, and more privateelective termination and would be consistent with current prenatalscreening practices. In addition, many governments limit the legalgestational age at which an elective termination may be performed.Therefore, accurate diagnosis of chromosomal abnormalities atearlier gestational ages is an important goal of prenatal testing.Because the fetal concentration of DNA is lower at a gestationalage of 12–16 wk, with only 66.7% having a fetal DNA fractiongreater than 10% in our study, deeper sequencing than 3.5 millionreads is likely required for accurate detection of subchromosomalabnormalities. Increasing sequencing depth is likely the most directway of improving accuracy of diagnosis in general and may beachieved in the near future with additional reductions in sequencingcosts over time. However, certain parts of the genome may not beamenable to adequate coverage because of the inherent nature ofchromosome structure or features of DNA (for example, hetero-chromatin regions or repetitive DNA sequences).One crucial challenge is that our technique may identify deletions

and duplications of unknown clinical significance, with incompletepenetrance, or with mild or unpredictable clinical consequences.Many of these abnormalities may be normal inherited variants.Thus, if our extension of NIPT is put into clinical practice, greatcare must be taken in presenting results and providing appropriatecounseling to patients. We expect that the utility of NIPT for sub-chromosomal abnormalities will increase with greater under-standing of genomic disorders. Databases compiling informationon these deletions/duplications are being developed, such asDECIPHER (24). With increased sample sizes and additionalclinical correlations, such databases will undergo validation andbecome more useful in the future. At minimum, detecting suchabnormalities will alert the child’s providers to the need for closesurveillance for sequelae, such as neurodevelopmental delay orthe consequences of a known genomic syndrome.NIPT using SSP is able to reliably estimate the fetal DNA frac-

tion in maternal plasma, which we showed with a large cohort ofpregnant women with normal fetuses. We applied this technique toidentify samples with a lower fetal DNA fraction that, thus, mightprovide a less reliable result, although the presence of a sub-chromosomal abnormality was correctly identified in many samples

with fetal DNA fractions less than 10%. As noted above, this effectmay be more significant and therefore, estimation of the fetalfraction may be more useful in trials studying women receivingNIPT at earlier gestational ages. In a clinical setting, this anal-ysis may identify samples that require either deeper sequencing,which may ameliorate the effect of a low fetal DNA fraction, orrepeat testing at a later gestational age when the fetal fraction ofDNA has increased.The current cost of sequencing a human genome with 0.5×

coverage by single reads, approximately the coverage in our studyat 3.5 million reads, is comparable with the cost of CMA whileavoiding an invasive procedure. This depth allowed excellent de-tection of deletions/duplications greater than 5 Mb in size in ourstudy. Deeper sequencing allows for finer resolution at an addi-tional cost. Already, many obstetric practices in the United Statesare offering NIPT as a replacement for invasive testing for de-tection of common chromosomal aneuploidies and select commondeletions/duplications. Our method offers unbiased detection andgreater resolution relative to current NIPT methods, with thepotential for lower costs because of lower sequencing require-ments. SSP has the potential to be a cheaper and faster technologyto use for this purpose compared with Illumina technology; how-ever, both technologies continue to improve, with increasing speedand accuracy and falling costs. Crucially, this type of analysis hasthe potential to be implemented for all high-risk pregnancies andeventually, as standard of care prenatal screening for all women,replacing current screening methods.The methods developed here can be applied to any mixed bi-

ological sample to determine the subchromosomal abnormalitiesin the minor component, even when this fraction represents only asmall percentage of the total DNA. In prenatal diagnostics, sam-ples obtained from chorionic villi could be analyzed for mosaickaryotypes or maternal contamination. Furthermore, many dif-ferent cancers have been associated with copy number changesthat could potentially be detected from serum cfDNA or a solidtumor sample that contains both normal and cancer cells (25, 26).Because the cost of next generation sequencing continues to drop,we expect that this technique will become applicable for a broadrange of diseases.

Materials and MethodsSubject Recruitment. This study was approved by the Institutional ReviewBoards of Guangdong Women and Children Hospital. Informed consent wasobtained from all participants. Women were offered participation in ourstudy if they presented with a singleton gestation, in which either CVS oramniocentesis was indicated for structural anomalies detected on ultraso-nography. Those choosing to participate provided written informed consentafter discussion of the potential advantages and risks of chromosomalmicroarray testing, including the possibility of findings of uncertain clinicalsignificance and the identification of genetic variants in the fetus or a parentthat cause adult-onset disorders. CVS was performed in the usual fashion. Forwomen undergoing amniocentesis, an additional 10 mL amniotic fluidwas retrieved.

Microarray CGH Laboratory Procedures. After DNA extraction, an aCGH analysiswas performed using Agilent’s 8 × 60K commercial arrays (Agilent Technolo-gies) according to the procedures described in our previous paper (12). Thedata were analyzed with Agilent Genomic Workbench Lite Edition 6.5.0.18software (Agilent Technologies). Aberrations were identified using the AgilentGenomic Workbench Lite Edition 6.5.0.18 software through the AberrationDetection Method-2 Algorithm, with a sensitivity threshold of 6.0 and a datafilter that rejected aberrations that did not include at least five probes with alog2 set of 0.25. All quality control metrics passed.

cfDNA Preparation and Sequencing. Preparation and sequencing of cfDNAobtained frommaternal plasmawere performed as previously described (12).To obtain complete sequences of cfDNA fragments, sequencing was per-formed using an Ion Proton Sequencer (Life Technologies) at 400 flows, andBaseCalling options were modified as follows: BaseCaller–disable-all-filters

14674 | www.pnas.org/cgi/doi/10.1073/pnas.1518151112 Yin et al.

off–keypass-filter off–phasing-residual-filter=2.0–trim-qual-cutoff 100–trim-qual-window-size 30–trim-adapter-cutoff 16–num-unfiltered 1000.

Data Preprocessing. Reads mapping and filtering procedures were performedaccording to our previous paper (12). The retained reads were aligned to thehuman genomic reference sequences (hg19) using the BWA (27). Reads thatwere unmapped or had multiple primary alignment records were filtered by“FLAG” field in the alignment file using an in-house Perl script. Duplicatereads were identified by Picard (broadinstitute.github.io/picard/) with theparameters java -jar MarkDuplicates.jar M=picard_duplication_metricsREMOVE_DUPLICATES=true ASSUME_SORTED=true 1 and removed by anin-house Perl script. The remaining reads were considered unique readsfor additional analysis.

Calculation of Fetal DNA Fraction. To identify the full sequences of cfDNA, weused in-house Perl scripts to exclude incomplete reads and determine the sizedistributions of plasma cfDNA molecules. Fetal DNA is generally shorter thanmaternal DNA (28), resulting in the feasibility of predicting the fetal DNAconcentration in maternal plasma (18). Plasma samples with a higher fetalDNA fraction would have a higher proportion of short plasma DNA frag-ments (∼130–140 bp; region A) and a lower proportion of long plasma DNAfragments (∼155–175 bp; region B). Locally weighted scatterplot smoothing(LOESS) regression was applied to fit the fetal fraction against reads ratio infeatures A and B. We obtain the LOESS fit-predicted fetal fraction PA forfeature A and PB for feature B. Because both PA and PB predict the fetal DNAfraction, PA and PB should also closely correlate. Therefore, we used refer-ence samples to compare PA and PB and thus, identify instances of poorcorrelation. If Pdiff = ðPA −PBÞ× 2=ðPA +PBÞ is larger than 0.40 (larger than99% normal samples), PA and PB are inconsistent, and the fetal fraction isconsidered unpredictable. Otherwise, we calculated the final predicted fetalfraction using P = ðPA + PBÞ=2.

Identification of Subchromosomal Abnormalities.After reads mapping, guanosine-cytosine (GC) correction was performed according to our previous paper (12). Inbrief, the number of unique reads from each 20-kb bin on each chromosome

was counted. To eliminate the effect of GC bias in different samples, an in-tegrated GC correction methodwas applied, in which LOESS regression was usedto compute the corrected number of unique reads in each 20-kb bin dependingon the GC content of the genomic sequence in the corresponding bin. Thecorrected reads number of each chromosome was determined by summing theweighted values of all 20-kb bins of each specific chromosome in a sample.

Other than GC bias, some genome blocks behaved similarly. Straver et al.(29) applied a “within-sample” comparison method to similar chromosomeregions to detect subchromosomal aberrations. Thus, reads ratio valueswere obtained from 500 normal samples as within-sample reference bins.Corrected reads ratio Ri for each 1-Mb region in each sample and then, theSD, mean, and coefficient of variation (CV) values for each 1-Mb regionwere calculated. Consequently, for block i and sample j, we obtain the Zscore Zij (Zij =Rij −Riref=SDiref). For block i of sample j, we obtain a Z score Zijin each independent test. Combining the Z scores of adjacent blocks wouldincrease the precision to detect subchromosome aberrations usingStouffer’s Z-score method (29): Stouffer’s  Z =

Pki=1Zi=

ffiffiffik

p. When Stouffer’s Z

score is larger than 3, we classify it as microduplication, whereas when it isless than −3, we classify it as microdeletion. We tested different k valuesand found that k = 7 performed best. Lastly, if the Stouffers’ Z scores ofadjacent blocks are similar, we would merge these blocks together. If thesubchromosomal aberration region is larger than 1 Mb, we can use theStouffer’s Z score to enlarge the aberration signal and detect abnormali-ties. To avoid false positives, the Z scores of each bin were filtered by a thresholdbased on fetal DNA fraction and sequencing depth. We used the formulaZij =Co exp ratioij −Co exp ratioi=SD=±fc%=2×Co exp ratioi=SD=± fc%

2*CV. In

the formula, CV (CV in each bin) was related to sequencing depth. We foundthat the precision of subchromosomal deletion/duplication detection is closelyrelated to the fetal DNA fraction in maternal plasma. When the fetal DNAfraction is less than 10%, Zij may be less than 3, resulting in a false negative.

ACKNOWLEDGMENTS. This work was supported, in part, by the NationalScience Foundation for Young Scholars of China Grant 81000255, NationalNatural Science Foundation of China.

1. Beckmann JS, Estivill X, Antonarakis SE (2007) Copy number variants and genetictraits: Closer to the resolution of phenotypic to genotypic variability. Nat Rev Genet8(8):639–646.

2. Lejeune J, et al. (1963) 3 Cases of partial deletion of the short arm of a 5 chromosome.C R Hebd Seances Acad Sci 257:3098–3102.

3. Lammer EJ, Opitz JM (1986) The DiGeorge anomaly as a developmental field defect.Am J Med Genet Suppl 2:113–127.

4. Bianchi DW (2012) From prenatal genomic diagnosis to fetal personalized medicine:Progress and challenges. Nat Med 18(7):1041–1051.

5. Lee CN, et al. (2012) Clinical utility of array comparative genomic hybridisation forprenatal diagnosis: A cohort study of 3171 pregnancies. BJOG 119(5):614–625.

6. Liu J, Bernier F, Lauzon J, Lowry RB, Chernos J (2011) Application of microarray-basedcomparative genomic hybridization in prenatal and postnatal settings: Three casereports. Genet Res Int 2011:976398.

7. Miller DT, et al. (2010) Consensus statement: Chromosomal microarray is a first-tierclinical diagnostic test for individuals with developmental disabilities or congenitalanomalies. Am J Hum Genet 86(5):749–764.

8. American College of Obstetricians and Gynecologists Committee on Genetics (2013)Committee Opinion No. 581: The use of chromosomal microarray analysis in prenataldiagnosis. Obstet Gynecol 122(6):1374–1377.

9. Mujezinovic F, Alfirevic Z (2007) Procedure-related complications of amniocentesisand chorionic villous sampling: A systematic review. Obstet Gynecol 110(3):687–694.

10. Benn P, et al. (2013) Position statement from the Aneuploidy Screening Committee onbehalf of the Board of the International Society for Prenatal Diagnosis. Prenat Diagn33(7):622–629.

11. Bianchi DW, et al.; CARE Study Group (2014) DNA sequencing versus standard pre-natal aneuploidy screening. N Engl J Med 370(9):799–808.

12. Liao C, et al. (2014) Noninvasive prenatal diagnosis of common aneuploidies bysemiconductor sequencing. Proc Natl Acad Sci USA 111(20):7415–7420.

13. Chiu RW, et al. (2011) Non-invasive prenatal assessment of trisomy 21 by multiplexedmaternal plasma DNA sequencing: Large scale validity study. BMJ 342:c7401.

14. Bartlett LA, et al. (2004) Risk factors for legal induced abortion-related mortality inthe United States. Obstet Gynecol 103(4):729–737.

15. Peters D, et al. (2011) Noninvasive prenatal diagnosis of a fetal microdeletion syn-drome. N Engl J Med 365(19):1847–1848.

16. Srinivasan A, Bianchi DW, Huang H, Sehnert AJ, Rava RP (2013) Noninvasive detectionof fetal subchromosome abnormalities via deep sequencing of maternal plasma. Am JHum Genet 92(2):167–176.

17. Lo YM, et al. (2010) Maternal plasma DNA sequencing reveals the genome-widegenetic and mutational profile of the fetus. Sci Transl Med 2(61):61ra91.

18. Yu SC, et al. (2014) Size-based molecular diagnostics using plasma DNA for non-invasive prenatal testing. Proc Natl Acad Sci USA 111(23):8583–8588.

19. Chiang DY, et al. (2009) High-resolution mapping of copy-number alterations withmassively parallel sequencing. Nat Methods 6(1):99–103.

20. Palomaki GE, et al. (2011) DNA sequencing of maternal plasma to detect Downsyndrome: An international clinical validation study. Genet Med 13(11):913–920.

21. Liao C, Zhengfeng X, Zhang K (2014) DNA sequencing versus standard prenatal an-euploidy screening. N Engl J Med 371(6):577–578.

22. Lutgendorf MA, Stoll KA, Knutzen DM, Foglia LM (2014) Noninvasive prenatal test-ing: Limitations and unanswered questions. Genet Med 16(4):281–285.

23. Snyder MW, et al. (2015) Copy-number variation and false positive prenatal aneu-ploidy screening results. N Engl J Med 372(17):1639–1645.

24. Firth HV, et al. (2009) DECIPHER: Database of Chromosomal Imbalance and Phenotypein Humans Using Ensembl Resources. Am J Hum Genet 84(4):524–533.

25. Karachaliou N, et al.; Spanish Lung Cancer Group (2015) Association of EGFR L858Rmutation in circulating free DNA with survival in the EURTAC Trial. JAMA Oncol 1(2):149–157.

26. Bianchi DW, et al. (2015) Noninvasive prenatal testing and incidental detection ofoccult maternal malignancies. JAMA 314(2):162–169.

27. Li H, Durbin R (2009) Fast and accurate short read alignment with Burrows-Wheelertransform. Bioinformatics 25(14):1754–1760.

28. Chan KC, et al. (2004) Size distributions of maternal and fetal DNA in maternalplasma. Clin Chem 50(1):88–92.

29. Straver R, et al. (2014) WISECONDOR: Detection of fetal aberrations from shallowsequencing maternal plasma based on a within-sample comparison scheme. NucleicAcids Res 42(5):e31.

Yin et al. PNAS | November 24, 2015 | vol. 112 | no. 47 | 14675

MED

ICALSC

IENCE

S