nature biotechnology  VOLUME 28 NUMBER 10 OCTOBER 2010 1115

A rt i c l e s

Little is known about the basic pathways and events of early human embryo development, including factors that would aid in predicting success or failure to develop. Consequently, to increase the chances of pregnancy through IVF, multiple embryos are often transferred to the uterus, despite the potential for well-documented adverse outcomes. Development of the human embryo begins with the fusion of sperm and egg, the epigenetic reprogramming of the gametic pronuclei and a series of cleavage divisions that culminate with activation of the embryo-nic genome by day 3 of development1. The embryo compacts to form a morula and subsequently a blastocyst, containing the outer troph-ectoderm and inner cell mass1. Although development of the human embryo shares many features with other species, there are also some notable differences, including unique gene-expression and epigenetic patterns and a protracted period of transcriptional silence through the first 3 d after fertilization1–9. In the mouse, by contrast, activation of the zygotic genome is initiated concurrent with the first cleavage division on day 1 (refs. 7,8). Human embryo development is also more fragile than that of many other species. Human fecundity rates are relatively low, largely due to pre- and post-implantation embryo loss10,11. In vitro, 50–70% of IVF embryos fail to reach the blastocyst stage12,13.

Most human embryo research has been based on a small number of samples generated under diverse experimental conditions1,14–17. Studies that involve imaging have been limited to measurements of early development, such as pronuclear formation and fusion and time to first cleavage18–21, and molecular profiling studies have generally

required pooling of oocytes, embryos or blastomeres, which masks differences in gene expression between embryos or between single blastomeres within an embryo15–17,22,23. Here we sought to overcome these limitations and to define critical pathways and events of human embryo development by correlating imaging profiles and molecular data throughout preimplantation development from the zygote to the blastocyst stage. We studied a large set of supernumerary IVF embryos that had been cryopreserved at the zygote stage 12–18 h after fertilization (Fig. 1). The embryos appeared representative of the typical IVF population, as they were frozen at the two-pronucleate (2PN) stage and thus indiscriminately selected for cryopreservation relative to those selected for culture. This is in contrast to embryos cryopreserved at the 8-cell stage or later, which are not selected for transfer during fresh IVF cycles and may therefore be of lower quality. With this unique set of embryos, we carried out a large-scale study that correlated time-lapse image analysis and gene expression profil-ing to show that successful development to the blastocyst stage can be predicted by the 4-cell stage, before EGA.

RESULTSCytokinesis as an embryo quality markerA normal human zygote undergoes the first cleavage division early on day 2, at ~24–27 h after fertilization18–20,24 (Fig. 2a, embryo H in Supplementary Video 1). Subsequently, the embryo cleaves to a 4- and 8-cell embryo on days 2 and 3, respectively, before compacting

Non-invasive imaging of human embryos before embryonic genome activation predicts development to the blastocyst stageConnie C Wong1,2,7, Kevin E Loewke1–3,6,7, Nancy L Bossert4, Barry Behr2, Christopher J De Jonge4, Thomas M Baer5 & Renee A Reijo Pera1,2

We report studies of preimplantation human embryo development that correlate time-lapse image analysis and gene expression profiling. By examining a large set of zygotes from in vitro fertilization (IVF), we find that success in progression to the blastocyst stage can be predicted with >93% sensitivity and specificity by measuring three dynamic, noninvasive imaging parameters by day 2 after fertilization, before embryonic genome activation (EGA). These parameters can be reliably monitored by automated image analysis, confirming that successful development follows a set of carefully orchestrated and predictable events. Moreover, we show that imaging phenotypes reflect molecular programs of the embryo and of individual blastomeres. Single-cell gene expression analysis reveals that blastomeres develop cell autonomously, with some cells advancing to EGA and others arresting. These studies indicate that success and failure in human embryo development is largely determined before EGA. Our methods and algorithms may provide an approach for early diagnosis of embryo potential in assisted reproduction.

1Institute for Stem Cell Biology and Regenerative Medicine, School of Medicine, Stanford University, Stanford, California, USA. 2Department of Obstetrics and Gynecology, School of Medicine, Stanford University, Stanford, California, USA. 3Department of Mechanical Engineering, Stanford University, Stanford, California, USA. 4Reproductive Medicine Center, University of Minnesota, Minneapolis, Minnesota, USA. 5Stanford Photonics Research Center, Department of Applied Physics, Stanford University, Stanford, California, USA. 6Present address: Auxogyn, Inc., Menlo Park, California, USA. 7These authors contributed equally to this work. Correspondence should be addressed to R.A.R.P. (reneer@stanford.edu).

Received 5 April; accepted 3 September; published online 3 October 2010; doi:10.1038/nbt.1686

© 2

010

Nat

ure

Am

eric

a, In

c. A

ll ri

gh

ts r

eser

ved

.

1116  VOLUME 28 NUMBER 10 OCTOBER 2010 nature biotechnology

A rt i c l e s

into a morula on day 4 and forming a blastocyst on days 5 to 6. For the purposes of this study, embryos that reached the blastocyst stage were considered developmentally competent and designated ‘normal’, whereas embryos that arrested at a stage before the blastocyst stage were considered developmentally incompetent and designated ‘abnormal’.

We tracked the development of 242 IVF embryos in four independ-ent experimental sets using multiple time-lapse microscopes equipped with low-power, dark-field illumination. Of the 242 embryos, 100 were cultured to day 5 or 6, whereas the remaining 142 were removed at various stages for quantitative real-time (qRT) PCR gene expres-sion analysis. Among the 100 embryos cultured to day 5 or 6, 33–53% formed blastocysts (Fig. 2b), and the remaining embryos arrested at different developmental stages, usually between the 2- and 8-cell stages. To identify quantitative imaging parameters that would pre-dict success in development to the blastocyst stage, we extracted and analyzed several parameters from the time-lapse videos, including blastomere size, thickness of the zona pellucida, degree of fragmen-tation, length of the first cell cycles, time intervals between the first few mitoses and duration of the first cytokinesis. As the embryos in this study were cryopreserved 12–18 h after fertilization, we did not measure parameters before the onset of the first cytokinesis, such as time to first cleavage or length of the first cell cycle, properties that have been evaluated previously18–20.

Out of the set of parameters measured, three collectively predicted blastocyst formation: (i) duration of the first cytokinesis (the very brief last step in mitosis that physically separates the two daughter cells), (ii) time interval between the end of the first mitosis and the initiation of the second and (iii) the time interval between the second and third mitoses (the time between the appearance of the cleav-age furrows of the second and third mitoses) (Fig. 2c). The third

parameter represents the synchronicity in the formation of the two sets of granddaughter cells. The mean values and s.d. for these three parameters for the embryos that developed to the blastocyst stage were (i) 14.3 ± 6.0 min, (ii) 11.1 ± 2.2 h and (iii) 1.0 ± 1.6 h, respec-tively. It is important to note that the first three mitotic events yield a 4-cell embryo from a 1-cell embryo, as opposed to the first three cleavage divisions, which yield an 8-cell embryo (Supplementary Fig. 1). Embryos that reached the blastocyst stage could be predicted, with a sensitivity and specificity of 94% and 93%, respectively, by having a first cytokinesis of 0–33 min, a time between first and second mitoses of 7.8–14.3 h and a time between second and third mitoses of 0–5.8 h (Fig. 2d, Supplementary Fig. 2 and Supplementary Data Set 1). Conversely, embryos that exhibited values outside of one or more of these windows were predicted to arrest.

We further examined the behavior of cytokinesis in both normal and abnormal embryos. Embryos that reached the blastocyst stage initiated and completed cytokinesis in a smooth, controlled manner over a narrow time window of 14.3 ± 6.0 min (n = 36), from appear-ance of the cleavage furrows to complete separation of the daughter cells (Fig. 2e, first panel, and Supplementary Video 2). In contrast, abnormal embryos showed a diverse range of behaviors that can be classified into three aberrant cytokinesis phenotypes (Supplementary Fig. 3). In the least frequent and mildest phenotype, the morphology and mechanism of cytokinesis appears normal, but the time required to complete the process is increased by a few minutes to an hour (Fig. 2e, second panel, Supplementary Video 3 and Supplementary Fig. 3, top panel). A small fraction of the embryos that underwent a slightly prolonged cytokinesis still developed into a blastocyst. In the second phenotype, embryos formed a unipolar cleavage furrow and displayed unusual morphological behavior for several hours before finally cleaving and fragmenting into smaller pieces (Fig. 2e, third panel, Supplementary Video 4 and Supplementary Fig. 3, middle panel). In the third phenotype, embryos displayed membrane ruffling and/or multiple cleavage furrows before cleaving and fragmenting into smaller pieces (Supplementary Fig. 3, bottom panel). Together, the second and third abnormal cytokinesis phenotypes confirm that abnormal cytokinesis is one of the mechanisms for embryo fragmentation, a common observation in abnormal human embryo development. Moreover, we observed that fragmentation in abnormal embryos rarely reversed, whereas moderate fragmentation in normal embryos sometimes reversed at the 2-cell stage before the second mitosis (Supplementary Fig. 4).

To determine whether cryopreservation and thawing altered the kinet-ics of development, we also imaged a small set (n = 10) of embryos that had not been cryopreserved (Fig. 2e, fourth panel, Supplementary Video 5 and Online Methods). Analysis of our three dynamic imaging parameters suggested that cryopreserved embryos are not developmentally delayed by the cryopreservation process.

Validation of imaging parameters by automated analysisOur time-lapse imaging data showed that human embryo devel-opment varies substantially between embryos within a cohort and that embryos exhibit a wide range of behaviors during cell division. However, characterization of developmental events, such as the dura-tion of cytokinesis, by human observers may be distorted by subjec-tive interpretation. To validate our method for predicting blastocyst formation, we developed an algorithm for automated tracking of cell divisions up to the 4-cell stage. Our tracking algorithm employs a probabilistic model estimation technique based on sequential Monte Carlo methods. This technique works by generating distributions of hypothesized embryo models, simulating images based on a simple

Day 1: Thaw 1-cellhuman embryos

Expt. 1: n = 61Expt. 2: n = 80Expt. 3: n = 64Expt. 4: n = 37

Single embryos

Single blastomeres

High-throughput, single-cell qPCR analysis

Harvest a mixture of normaland arrested embryos on

consecutive days

Time-lapseimaging

on multiplemicroscopes

2

34

Day1–2

Day5–6

Day 3 Day 4

1

Figure 1 Experimental plan. We tracked the development of 242 two-pronucleate stage embryos in four experimental sets (containing 61, 80, 64 and 37 embryos, respectively). In each set of experiments, human zygotes were thawed on day 1 and cultured in small groups on multiple plates. Each plate was observed independently with time-lapse microscopy under dark-field illumination on separate imaging stations. At ~24 h intervals, one plate of embryos was removed from the imaging system and collected as either single embryos or single cells (blastomeres) for high-throughput qRT-PCR gene expression analysis. Each plate typically contained a mixture of embryos that reached the expected developmental stage at the time of harvest (termed ‘normal’) and those that were arrested or delayed at earlier development stages, or fragmented extensively (termed ‘abnormal’). Gene expression analysis was carried out on single intact embryos or on single blastomeres of dissociated embryos. One hundred of the 242 embryos were imaged until day 5 or 6 to monitor blastocyst formation.

© 2

010

Nat

ure

Am

eric

a, In

c. A

ll ri

gh

ts r

eser

ved

.

nature biotechnology  VOLUME 28 NUMBER 10 OCTOBER 2010 1117

A rt i c l e s

optical model and comparing these simulations to the observed image data (Fig. 3a and Supplementary Video 6).

Embryos were modeled as a collection of ellipses with position, orientation and overlap indices (to represent the relative heights of the cells). With these models, the duration of cytokinesis and time between mitoses can be extracted. Cytokinesis is typically defined by the first appearance of the cytokinesis furrow (where bipolar inden-tations form along the cleavage axis) to the complete separation of daughter cells. We simplified the problem by approximating cytokine-sis as the duration of cell elongation before a 1-cell to 2-cell division. A cell is considered elongated if its major axis has increased by >15% (chosen empirically). The time between mitoses is straightforward to extract by counting the number of cells in each model.

We tested our algorithm on 14 human embryos from the set of 100 that were imaged up to the blastocyst stage (Fig. 3b and Supplementary Video 7) and compared the automated measurements to manual image

analysis (Fig. 3c). In this data set, eight embryos reached the blastocyst stage with good morphology (Fig. 3d, top). The automated measure-ments were closely matched to the manual measurements, and all eight embryos were correctly predicted to reach the blastocyst stage by both methods. Two embryos reached the blastocyst stage with poor morphology (poor quality of inner cell mass; Fig. 3d, bottom). For these embryos, manual assessment indicated that one would reach the blastocyst stage and one would arrest, whereas the automated assessment predicted that both would arrest. Finally, four embryos arrested before the blastocyst stage; all four were correctly predicted to arrest by both methods. These results suggest that a systematic, automated prediction of blastocyst formation can be achieved as early as the 4-cell stage.

Gene expression and cytokinesisTo assess whether imaging parameters that predict success or fail-ure of development are associated with transcriptional patterns, we

Blastocyst8-cell4- to 7-cell2- to 3-cell1-cell

n = 25

100%

80%

60%

40%

20%

0%

1st mitosis

Time between 1st and 2nd mitoses

Duration of 1st cytokinesis Synchronicity of 2nd and 3rd mitoses

2nd mitosis

–0:05

–0:05

–0:05

–0:05

0:00

0:00

0:00

0:00

0:05

0:05

1:00

0:05

0:10

0:10

2:00

0:15

3:00

0:20

4:00

0:25

5:00

0:10

3rd mitosis

Normal

Abnormal

Expt 1 blastocyst (n = 9)Expt 2 blastocyst (n = 17)Expt 3 blastocyst (n = 3)Expt 4 blastocyst (n = 7)Expt 1 arrested (n = 14)Expt 2 arrested (n = 18)Expt 3 arrested (n = 6)Expt 4 arrested (n = 20)

0

2

4

1st cytokinesis duration (h)Time between

1st and 2nd mitoses (h)640

30

20

10

0

Tim

e be

twee

n2n

d an

d 3r

d m

itose

s (h

)0

10

20

30

40

50

n = 39 n = 9 n = 27

Expt. 1 Expt. 2 Expt. 3 Expt. 4

Day 1 Day 2 a.m. Day 2 p.m.

Day 3 Day 4 Day 5

Day 6

a b f

d

c

e

Figure 2 Abnormal embryos exhibit abnormal cytokinesis and mitosis timing during the first divisions. (a) The developmental time line of a healthy human preimplantation embryo. Scale bar, 50 μm. (b) The distribution of normal and arrested embryos among samples that were cultured to day 5 or 6. (c) Cytokinesis duration was measured from the appearance of a cleavage furrow to complete daughter-cell separation during the first division. Time between the first and second mitoses was measured from the completion of the first mitosis to the appearance of cleavage furrow of the second mitosis. Synchronicity of the second and third mitoses was defined as the time between the appearance of the cleavage furrows of the second and third mitoses. (d) Normal embryos followed strict timing in cytokinesis and mitosis during early divisions, before EGA begins. Out of the 100 embryos imaged to day 5 or 6, six were excluded from subsequent image analysis due to technical issues (e.g., inability to track identity after media change, or loss of image focus). Raw data for this plot are included as Supplementary Data Set 1, and additional views can be seen in Supplementary Figure 2. (e) Normal cytokinesis (first row) was typically completed in 14.3 ± 6.0 min in a smooth, controlled manner. In the mild phenotype (second row), the cytokinesis mechanism appears normal although it is slightly prolonged. In the severe phenotype (third row), a one-sided cytokinesis furrow is formed, accompanied by unusual ruffling of cell membranes for a prolonged period of time. Cytokinesis was defined by the first appearance of the cytokinesis furrow (arrows) to the complete separation of daughter cells. Imaging was also performed on a subset of triploid embryos (fourth row), which exhibited a distinct phenotype of dividing into three cells in a single event. Scale bar, 50 μm. (f) Embryos that underwent abnormal development and behavior (right) would occasionally appear morphologically similar to normal embryos (left) at the time of sample collection. In this particular case, time-lapse video data showed that what appeared to be a six to eight-cell embryo (right) was in fact the product of a highly aberrant cell division (Supplementary Video 10). Thus, the correlated imaging data served to ensure the accuracy of sample selection and identification for the gene expression analysis.

© 2

010

Nat

ure

Am

eric

a, In

c. A

ll ri

gh

ts r

eser

ved

.

1118  VOLUME 28 NUMBER 10 OCTOBER 2010 nature biotechnology

A rt i c l e s

analyzed the expression of nine putative cytokinesis-related genes in both normal and arrested embryos (Supplementary Table 1 and Supplementary Data Set 2). Aberrant cytokinesis seen in the time-lapse image data correlated strongly with reduced expression of key cytokinesis genes. Like their morphological phenotypes, the gene expression profiles of embryos that arrested were diverse and vari-able. For example, an arrested 2-cell embryo that displayed a slightly prolonged cytokinesis and an unusual plasma membrane ruffling (Supplementary Video 8) expressed all nine cytokinesis genes exam-ined at significantly lower levels (P < 0.05) compared with develop-mentally normal embryos at the same stage (Fig. 4a). On the other hand, an arrested 4-cell embryo (Fig. 4b) that underwent prolonged, unipolar cytokinesis during its first division (Supplementary Video 9) showed significantly reduced expression (P < 0.05) of only two cyto-kinesis genes, ANLN and ECT2.

We also examined genes in categories other than cytokinesis in arrested and normal embryos at the 1- and 2-cell stage. For this purpose we calculated average expression levels for each of 52 additional

genes that included housekeeping genes, germ cell markers, maternal factors, EGA markers, trophoblast markers, inner cell mass markers, pluripotency markers, epigenetics regulators, transcription factors, hormone receptors and others based primarily on published data in model organisms1,7,8. Normal 1-cell embryos were identified as having undergone successful fusion of the two pronuclei (syngamy) on day 1 and displaying a round, firm appearance. Eighteen of the 52 genes showed statistically significant differences in expression between normal and arrested embryos (P < 0.05), with certain gene categories affected more severely than others (Fig. 4c). In abnormal embryos, expression of most of the housekeeping genes, hormone receptors and maternal factors was not appreciably altered, but many genes involved in cytokinesis and in microRNA (miRNA) biogenesis, such as DGCR8, DICER1 and TARBP2, were expressed at highly reduced levels. Two of the most severely affected genes, CPEB1 and SYMPK, belong to the same molecular pathway, which regulates maternal mRNA storage and reactivation by modulating the length of poly(A) tails on oocyte/embryo transcripts25.

Window for blastManualAutomatic

2 4 6 8 10 12 14

Poor-morphologyblastocyst (embryo no.9)

Good-morphologyblastocyst (embryo no.6)

Frame 15 Frame 125 Frame 127 Frame 269 Frame 270 Frame 276 Frame 314

20

5

10

Tim

e be

twee

n1s

t and

2nd

mito

ses

(h)

15

20

0

50

100

Dur

atio

n of

firs

t cyt

okin

esis

(m

in)

150

200

4

Good-morphologyblastocyst

Poor-morphologyblastocyst

Arrested

6 8 10 12 14

a

b c d

Figure 3 Automated image analysis confirms the utility of the imaging parameters to predict blastocyst formation. (a) Results of tracking algorithm for a single embryo. Images were captured every 5 min, and only a select group is displayed. The top row shows frames from the original time-lapse image sequence, and the bottom row shows the overlaid tracking results. (b) Set of 14 embryos that were analyzed (Supplementary Video 6). One embryo was excluded as it was floating and out of focus. (c) Comparison of image analysis by a human observer and automated analysis of the duration of cytokinesis (top) and of the time between first and second mitoses (bottom). There is excellent agreement between the two methods for embryos that reached the blastocyst stage with good morphology. The few cases of disagreement occurred mostly for abnormal embryos and were caused by unusual behavior that is difficult to characterize by both methods. The gray shade region shows the window for blastocyst prediction. The two methods agreed on blastocyst prediction except in the case of embryo 10, which was predicted as abnormal by the automated method and normal by the manual method. (d) Comparison of blastocysts with good (top) and bad (bottom) morphology.

© 2

010

Nat

ure

Am

eric

a, In

c. A

ll ri

gh

ts r

eser

ved

.

nature biotechnology  VOLUME 28 NUMBER 10 OCTOBER 2010 1119

A rt i c l e s

Embryonic stage–specific patternsTo further correlate our three imaging parameters with gene expres-sion, we measured expression of two slightly different but overlapping sets of 96 genes (Supplementary Table 1) at multiple time points of embryo development. Time-lapse imaging was used to aid the identi-fication and classification of normal and abnormal embryos because occasionally embryos that developed and behaved abnormally would appear morphologically normal at the time of sample collection (Fig. 2f and Supplementary Video 10). By analyzing the gene expression pat-terns of 141 of the 242 embryos that had apparently normal develop-ment as assessed by imaging (and without any prior assumptions), we derived four unique embryonic stage–specific patterns (ESSPs) of gene expression (Fig. 5a, Supplementary Fig. 5 and Supplementary Table 2). ESSP1 describes maternally inherited oocyte mRNAs destined for degradation. These transcripts were expressed at high levels at the zygote stage and declined during development to the blastocyst stage. Their half-life was just 21 h (Supplementary Fig. 6). ESSP2 includes embryonic-activated genes, first transcribed on day 3, at approximately the 8-cell stage. ESSP3 comprises genes not expressed until the blas-tocyst stage. Finally, ESSP4 includes persistent transcripts that main-tained stable expression relative to the reference genes from the zygote to blastocyst stages. The half-life of ESSP4 genes was 193 h, more than nine times longer than that of ESSP1 genes (21 h) (Supplementary Fig. 6). Fourteen of the 96 genes analyzed did not fit into any of the four ESSP patterns and were labeled ‘undefined’ (Supplementary Table 2). We confirmed the four patterns of gene expression in two additional inde-pendent experimental sets using both single, intact normal embryos and isolated single blastomeres (Supplementary Fig. 7).

We compared our qRT-PCR data in 1-cell and 2-cell embryos to published microarray data on human oocytes23 (Supplementary Data Set 3). In ref. 19, the expression values for individual genes in the microarray data were normalized against the geometric mean of GAPDH and RPLP0, which were the same reference genes used in our studies. Among the 86 genes that we analyzed, almost every gene that was expressed in 1- and 2-cell embryos was also expressed or upregulated in oocytes, with the exception of TACC3 and H2AFZ. In addition, by dividing the genes into low-, medium- and high-expression genes, we observed good correla-tion between the two data sets among all gene sets, especially between the highly expressed genes. A comparison of our data to a study of cell cycle genes expressed in the 8-cell human embryo26 showed agreement for the two genes that were assayed in both studies, AURKA and CCNA1.

Individual blastomeres show cell autonomyIndividual blastomeres in an intact early human embryo are usually assumed to be synchronized in constitution and devel-opmental programming, and developmental success or failure is considered a property of the whole embryo. We measured expres-sion of ten maternal transcript genes and ten embryonic genes in single blastomeres of 36 normal and abnormal embryos between the 2- and 10-cell stage. Notably, this experiment revealed a subset of normal embryos that contained blastomeres whose gene expres-sion signatures corresponded to different developmental ages. Among 24 morphologically and developmentally normal embryos, 6 (25%) contained blastomeres of different transcriptional ages

–0:05 0:00 0:05 0:10 0:15

Arrested 2-cell embryoNormal 2-cell embryos

(n = 9)

(n = 12)

10

8

Rel

ativ

e ex

pres

sion

6

4

2

0

10

8

Rel

ativ

e ex

pres

sion

6

4

2

0

ANLN

CFL1

DIAPH1

DIAPH2

DNM2

ECT2

MKLP2

MYLC2

RHOA

ANLN

CFL1

DIAPH1

DIAPH2

DNM2

ECT2

MKLP2

MYLC2

RHOA

0:20 0:25

0:00 1:00 3:00 5:00 7:00 9:00 11:00

Arrested 4-cell embryoNormal 4-cell embryos

Transcriptionfactor

Receptor

Pluripotency

Housekeeping

Maternaleffect

RNAprocessing

miRNAbiogenesis

Cytokinesis

ANLN

100

10

1

DIAPH2DNM2

MKLP2

RHOA

XPO5

RNASEN

TACC3

AURKA

PARNCCR4

DAZLVASA

BNC2GDF9

HSF1NLRP5

PDCD5

ZP1ACTB

GAPDHHPRT1

RPLPOTBP

NANOG

POU5F1

FGFR1

FGFR2

IGFR1

ATF7IP2

GABPB2GTF2A1

ATF4

TAF4 CFL1*NELF* DIAPH1*

ECT2*

MYLC2*

DGCR8*

DICER1*

TARBP2*

CPEB1*SYMPK*

ZAR1*

CTNNB1*

DNMT3B*

TERT*

YY1*

IGFR2*

BTF3*

YBX2*

Normal 1- and 2-cell embryos (n = 5)Arrested 1- and 2-cell embryos (n = 6)

a c

b

Figure 4 Distinct gene expression profiles of developmentally delayed or arrested embryos. (a) An arrested 2-cell embryo that showed abnormal membrane ruffling during the first cytokinesis had significantly (P < 0.05) reduced expression level of all cytokinesis genes tested. Scale bar, 50 μm. (b) An arrested 4-cell embryo that underwent aberrant cytokinesis with a one-sided cytokinesis furrow and extremely prolonged cytokinesis during the first division showed lower expression of ANLN and ECT2. Scale bar, 50 μm. (c) The average expression level of 52 genes from six abnormal 1- to 2-cell embryos and five normal 1- to 2-cell embryos were plotted in a radar graph on a logarithmic scale. Arrested embryos in general expressed less mRNA than normal embryos, with genes related to cytokinesis, RNA processing and miRNA biogenesis most severely affected. Genes highlighted in orange with an asterisk indicate a statistically significant difference (P < 0.05) between normal and abnormal embryos as determined by the Mann-Whitney test.

© 2

010

Nat

ure

Am

eric

a, In

c. A

ll ri

gh

ts r

eser

ved

.

1120  VOLUME 28 NUMBER 10 OCTOBER 2010 nature biotechnology

A rt i c l e s

(Fig. 5b). Among 12 abnormal embryos arrested between the 2- and 10-cell stage, this phenomenon was detected in 8 (66%). In some cases (e.g., Fig. 5b, right), the transcriptional profile of indi-vidual blastomeres varied to such an extent that some blastomeres in both normal and abnormal embryos may have been arrested for a considerable amount of time while the others progressed in development.

DISCUSSIONWe have carried out a large-scale study of preimplantation human embryos that correlated time-lapse image analysis and gene expres-sion profiling with development from the zygote to the blastocyst. Our results shed light on human embryo develop-ment and provide an approach for predict-ing which embryos will reach the blastocyst stage using three dynamic imaging para-meters (Fig. 6). First, we showed that human embryos that develop to the blastocyst stage follow a strict and predictable developmen-tal timeline that is correlated with predict-able gene expression patterns. This timeline

enabled us to derive an algorithm to automatically measure our imag-ing parameters and to predict blastocyst formation systematically and reliably by the 4-cell stage, before EGA. The finding that embryo development can be predicted at this early stage suggests that success or failure is likely to be determined at least in part by inheritance of maternal transcripts, which we observed to be expressed at altered levels in abnormal embryos. Other factors that may contribute to abnormal development before EGA include inherited genetic muta-tions, aneuploidy, environmental insult to germ cells, events during fertilization and sperm-related factors27–29.

Second, we found that gene expression in preimplantation human embryos is cell autonomous and follows four distinct patterns. Maternally inherited transcripts in ESSP1 have a half-life after ferti-lization of ~21 h; ESSP2 and ESSP3 are expressed at EGA and thereaf-ter, respectively; and ESSP4 genes are stably expressed, with a half-life of ~193 h. Previous studies of gene expression in the human embryo from the oocyte to day 3 have not analyzed single blastomeres, primar-ily because of the technical difficulty of obtaining single-cell data1,26. At the whole-embryo level, maintenance of maternal mRNA expres-sion profiles and failure to progress to EGA has never been observed past the first cell division1. Our single-cell gene expression analy-sis shows that individual blastomeres in an embryo can differ, with some maintaining maternal mRNAs whereas others progress to EGA. The frequency of this observation (in >25% of embryos) indicates that individual blastomeres in human embryos are cell autonomous. Moreover, the observation that maternal mRNAs can be maintained even after 3 days of development suggests that the degradation of the maternal programs is not simply a passive process.

2-cell

Mat

erna

l/em

bryo

nic

ratio

Examples of abnormal embryos4-cell 6-cell 8-cell

EmbryonicMaternal

EmbryonicMaternal

1c 3c2c 4c 6c 8c 9c M B0

0

2

4

6

8

012345678

10

12

Rel

ativ

e ex

pres

sion

(n

= 1

41)

1

2

3

4

0

0

1

2

3

4ESSP3 ESSP4

ESSP1 ESSP2

1c 3c2c 4c 6c 8c 9c M B

a

b

Figure 5 Gene expression analysis of single human embryos and blastomeres. (a) Genes analyzed in human embryos are defined by four distinct ESSPs. Relative expression level of an ESSP was calculated by averaging the expression levels of genes with similar expression patterns. (b) The ratio of maternal to embryonic genes in embryos changes during preimplantation development (left). Some embryos contained blastomeres of different developmental ages (right). The expression levels of embryonic and maternal programs were calculated by averaging the relative expression of ten ESSP1 and ten ESSP2 markers, respectively.

Transfer prior to EGA

Embryonicgene activation

(ESSP2)

Blastomeres are cell autonomous

24 h24 h18 h1 h11 h15 min24 hTime line

Stage

Molecular

Imaging

Automatedtracking

Featureextraction

Embryotransfer

Each blastomereinherits half of

stable mRNA (ESSP4)

Onset ofdegradationof ESSP1

mRNA

OocyteprovidesmRNAs

ESSP1

Time betweenmitoses

Duration of1st cytokinesis

ESSP2

Figure 6 Proposed model for human embryo development. Human embryos begin life with a set of oocyte RNAs inherited from the mother. After fertilization, a subset of maternal RNAs specific to the egg (ESSP1) must be degraded as the transition from oocyte to embryo begins. As development continues, other RNAs are partitioned equally to each blastomere (ESSP4). At EGA, ESSP2 genes are transcribed in a cell-autonomous manner. During the cleavage divisions, embryonic blastomeres may arrest or progress independently. ‘Feature extraction’ indicates the three imaging parameters for predicting successful development to the blastocyst stage: cytokinesis, the time between 1st and 2nd mitoses, and the time between 2nd and 3rd mitoses.

© 2

010

Nat

ure

Am

eric

a, In

c. A

ll ri

gh

ts r

eser

ved

.

nature biotechnology  VOLUME 28 NUMBER 10 OCTOBER 2010 1121

A rt i c l e s

Previous studies in the mouse have sought to understand how and when the first cell fate decisions in the embryo are established. It has been suggested that the first cleavage division itself determines the blas-tocyst axis and that subsequently all blastomeres are not equivalent30–31. These studies may imply that cell lineage fate is determined in the first division32; more likely is that the first cleavage division affects the probabilities of fates in subsequent divisions33. Our results support the conclusion that some aspects of embryo fate, especially success or failure to reach the blastocyst stage, are determined very early in development and likely inherited from the oocyte, as described above. Moreover, they imply that each cell-autonomous blastomere is capable of contributing, or not, to subsequent lineages.

Third and finally, given that embryo developmental potential can be assessed with a combination of cytokinetic and mitotic parameters in the first two cleavage divisions, it may be feasible to translate these basic studies to clinical applications. Current morphological and growth criteria that are commonly used to assess embryo viability on day 3 in assisted reproduction clinics may both underestimate and overesti-mate embryo potential, with well-documented consequences, such as multiple births, the need for fetal reduction and miscarriage34. Given the uncertainties associated with evaluation at day 3, some clinics have turned to longer culture to assess embryo potential, as embryos trans-ferred at the blastocyst stage have a higher implantation rate compared with embryos transferred at day 3 (refs. 13,35–38). However, this prac-tice involves prolonged in vitro culture and may increase the chance of altered gene expression and epigenetic inheritance39–41. Thus, a method to predict blastocyst formation at day 2 could improve IVF outcomes by increasing pregnancy rates while reducing the risk of multiple gesta-tions. This question will be evaluated in future clinical studies.

METhODSMethods and any associated references are available in the online version of the paper at http://www.nature.com/naturebiotechnology/.

Note: Supplementary information is available on the Nature Biotechnology website.

ACKNoWLEDgMENTsWe thank R. Raja for help with the microarray analysis and early imaging experiments, the members of the Reijo Pera laboratory for technical assistance and discussions, S. Walker for advice regarding the cell tracking algorithm and K. Salisbury for providing K.E.L. with hardware and software resources. We acknowledge funding contributions from the Stanford Institute for Stem Cell Biology and Regenerative Medicine, a generous, anonymous donor and the March of Dimes (6-FY06-326).

AUTHoR CoNTRIBUTIoNsC.C.W. and K.E.L. performed and designed experiments, analyzed data and assisted in writing and editing of the manuscript. K.E.L. designed cell tracking algorithms. N.L.B. assisted in performing the experiments. B.B., N.L.B. and C.J.D.J. assisted in analyzing data and editing the manuscript. T.M.B. and K.E.L. designed and built the imaging instrumentation. T.M.B. and R.A.R.P. designed experiments, interpreted results and assisted in writing and editing the manuscript.

CoMPETINg FINANCIAL INTEREsTsThe authors declare competing financial interests: details accompany the full-text HTML version of the paper at http://www.nature.com/naturebiotechnology/.

Published online at http://www.nature.com/naturebiotechnology/. reprints and permissions information is available online at http://npg.nature.com/reprintsandpermissions/.

1. Dobson, A.T. et al. The unique transcriptome through day 3 of human preimplantation development. Hum. Mol. Genet. 13, 1461–1470 (2004).

2. Braude, P., Bolton, V. & Moore, S. Human gene expression first occurs between the four- and eight-cell stages of preimplantation development. Nature 332, 459–461 (1988).

3. Memili, E. & First, N.L. Zygotic and embryonic expression in cow: a review of timing and mechanisms of early gene expression as compared with other species. Zygote 8, 87–96 (2000).

4. Beaujean, N. et al. Effect of limited DNA methylation reprogramming in the normal sheep embryo on somatic cell nuclear transfer. Biol. Reprod. 71, 185–193 (2004).

5. Fulka, H., Mrazek, M., Tepla, O. & Fulka, J. Jr. DNA methylation pattern in human zygotes and developing embryos. Reproduction 128, 703–708 (2004).

6. Duranthon, V., Watson, A.J. & Lonergan, P. Preimplantation embryo programming: transcription, epigenetics, and culture environment. Reproduction 135, 141–150 (2008).

7. Wang, Q.T. et al. A genome-wide study of gene activity reveals developmental signaling pathways in the preimplantation mouse embryo. Dev. Cell 6, 133–144 (2004).

8. Zeng, F. & Schultz, R. RNA transcript profiling during zygotic gene activation in the preimplantation mouse embryo. Dev. Biol. 283, 40–57 (2005).

9. Vanneste, E. et al. Chromosome instability is common in human cleavage-stage embryos. Nat. Med. 15, 577–583 (2009).

10. Macklon, N.S., Geraedts, J.P.M. & Fauser, B.C.J.M. Conception to ongoing pregnancy: the “black box” of early pregnancy loss. Hum. Reprod. Update 8, 333–343 (2002).

11. Evers, J.L. Female subfertility. Lancet 360, 151–159 (2002).12. French, D.B., Sabanegh, E.S. Jr., Goldfarb, J. & Desai, N. Does severe

teratozoospermia affect blastocyst formation, live birth rate, and other clinical outcome parameters in ICSI cycles? Fertil. Steril. 93, 1097–1103 (2010).

13. Gardner, D.K., Lane, M. & Schoolcraft, W. Culture and transfer of viable blastocysts: a feasible proposition for human IVF. Hum. Reprod. 15 (Suppl 6), 9–23 (2000).

14. Payne, D., Flaherty, S.P., Barry, M.F. & Matthews, C.D. Preliminary observations on polar body extrusion and pronuclear formation in human oocytes using time-lapse video cinematography. Hum. Reprod. 12, 532–541 (1997).

15. Adjaye, J., Bolton, V. & Monk, M. Developmental expression of specific genes detected in high-quality cDNA libraries from single human preimplantation embryos. Gene 237, 373–383 (1999).

16. Assou, S. et al. The human cumulus—oocyte complex gene-expression profile. Hum. Reprod. 21, 1705–1719 (2006).

17. Kimber, S.J. et al. Expression of genes involved in early cell fate decisions in human embryos and their regulation by growth factors. Reprod. 135, 635–647 (2008).

18. Nagy, Z.P., Liu, J., Joris, H., Devroey, P. & Steirteghem, A.V. Time-course of oocyte activation, pronucleus formation and cleavage in human oocytes fertilized by intracytoplasmic sperm injection. Hum. Reprod. 9, 1743–1748 (1994).

19. Fenwick, J., Platteau, P., Murdoch, A.P. & Herbert, M. Time from insemination to first cleavage predicts developmental competence of human preimplantation embryos in vitro. Hum. Reprod. 17, 407–412 (2002).

20. Lundin, K., Bergh, C. & Hardarson, T. Early embryo cleavage is a strong indicator of embryo quality in human IVF. Hum. Reprod. 16, 2652–2657 (2001).

21. Lemmen, J.G., Agerholm, I. & Ziebe, S. Kinetic markers of human embryo quality using time-lapse recordings of IVF/ICSI-fertilized oocytes. Reprod. Biomed. Online 17, 385–391 (2008).

22. Bermudez, M.G. et al. Expression profiles of individual human oocytes using microarray technology. Reprod. Biomed. Online 8, 325–337 (2004).

23. Kocabas, A.M. et al. The transcriptome of human oocytes. Proc. Natl. Acad. Sci. USA 103, 14027–14032 (2006).

24. Rienzi, L. et al. Significance of morphological attributes of the early embryo. Reprod. Biomed. Online 10, 669–681 (2005).

25. Bettegowda, A. & Smith, G.W. Mechanisms of maternal mRNA regulation: implications for mammalian early embryonic development. Front. Biosci. 12, 3713–3726 (2007).

26. Kiessling, A.A. et al. Evidence that human blastomere cleavage is under unique cell cycle control. J. Assist. Reprod. Genet. 26, 187–195 (2009).

27. Schatten, H. & Sun, Q. The role of centrosomes in fertilization, cell division and establishment of asymmetry during embryo development. Semin. Cell Dev. Biol. 21, 174–184 (2010).

28. Ostermeier, G.C., Miller, D., Huntriss, J.D., Diamond, M.P. & Krawetz, S.A. Reproductive biology: delivering spermatozoan RNA to the oocyte. Nature 429, 154 (2004).

29. Hammoud, S.S. et al. Distinctive chromatin in human sperm packages genes for embryo development. Nature 460, 473–478 (2009).

30. Zernicka-Goetz, M. Patterning of the embryo: the first spatial decisions in the life of a mouse Development 129, 815–829 (2002).

31. Plusa, B. et al. The first cleavage of the mouse zygote predicts the blastocyst axis. Nature 434, 391–395 (2005).

32. Hiiragi, T., Louvet-Vallee, S., Solter, D. & Maro, B. Embryology: does prepatterning occur in the mouse egg? Nature 442, E3–4 (2006).

33. Zernicka-Goetz, M. The first cell-fate decisions in the mouse embryo: destiny is a matter of both chance and choice Curr. Opin. Genet. Dev. 16, 406–412 (2006).

34. Racowsky, C. High rates of embryonic loss, yet high incidence of multiple births in human ART: Is this paradoxical? Theriogenology 57, 87–96 (2002).

35. Milki, A.A., Hinckley, M., Fisch, J., Dasig, D. & Behr, B. Comparison of blastocyst transfer with day 3 embryo transfer in similar patient populations. Fertil. Steril. 73, 126–129 (2000).

36. Gardner, D.K., Lane, M., Stevens, J., Schlenker, T. & Schoolcraft, W.B. Blastocyst score affects implantation and pregnancy outcome: towards a single blastocyst transfer. Fertil. Steril. 73, 1155–1158 (2000).

37. Gardner, D.K. & Lane, M. Towards a single embryo transfer. Reprod. Biomed. Online 6, 470–481 (2003).

38. Gardner, D.K. et al. Single blastocyst transfer: a prospective randomized trial. Fertil. Steril. 81, 551–555 (2004).

39. Manipalviratn, S., DeCherney, A. & Segars, J. Imprinting disorders and assisted reproductive technology. Fertil. Steril. 91, 305–315 (2009).

40. Niemitz, E.L. & Feinberg, A. Epigenetics and assisted reproductive technology: a call for investigation. Am. J. Hum. Genet. 74, 599–609 (2004).

41. Horsthemke, B. & Ludwig, M. Assisted reproduction: the epigenetic perspective. Hum. Reprod. Update 11, 473–482 (2005).

© 2

010

Nat

ure

Am

eric

a, In

c. A

ll ri

gh

ts r

eser

ved

.

nature biotechnology doi:10.1038/nbt.1686

ONLINE METhODSSample source. All embryos used in this study were supernumerary embryos from the Lutheran General Hospital IVF Program that were donated to research by informed consent. Embryos were moved to the Reproductive Medicine Center at the University of Minnesota after the Lutheran General Hospital IVF Program closed in 2002. Before the program closed, the embryos were collected over several years. Oocytes were fertilized and cryopreserved by multiple embryologists. The average number of embryos per patient in our study was 3, and all age groups encountered in a routine IVF center were included. All embryos were generated by IVF, not intracytoplasmic sperm injection, so they were derived from sperm able to penetrate the cumulus, zona and oolemma and form a pronuclei. Stimulation protocols were standard long lupron protocols. The embryos were cryopreserved by placing them in freezing medium (1.5 M 1,2 propanediol + 0.2 M sucrose) for 25 min at 22 + 2 °C and then freezing them using a slow-freeze protocol (−1 °C/min to −6.5 °C; hold for 5 min; seed; hold for 5 min; −0.5 °C/min to −80 °C; plunge in liquid nitrogen). The embryos were approved for research by the University of Minnesota Internal Review Board and the Stanford University Internal Review Board and Stem Cell Research Oversight Committee. No protected health information could be associated with the embryos.

We chose to use this embryo set after consideration of alternatives. Three sources of IVF embryos are theoretically available: (i) day 1 embryos obtained from a clinic for immediate analysis without cryopreservation, (ii) clinical embryos destined for transfer for reproductive purposes if our imaging system was set up in a clinic and (iii) cryopreserved embryos that are validated with regard to key developmental landmarks. Clinical practices and general guide-lines pose considerable practical obstacles to alternatives (i) and (ii). Moreover, ‘fresh’ embryos donated for research are generally abnormal in development. We therefore chose to study a large set of cryopreserved zygotes available for research. The following considerations suggest that cryopreservation did not adversely affect our results. First, the timing of developmental landmarks was similar to that of normal embryos, including cleavage to 2 cells (occurred early day 2), onset of RNA degradation (occurred on days 1 to 3), cleavage to 4 and 8 cells (occurred on late day 2 and day 3, respectively), EGA (on day 3 at the 8-cell stage) and formation of the morula and blastocyst (occurred on days 4 and 5, respectively)1,2. Second, the fraction of embryos that reached the blastocyst stage is typical of IVF embryos in a clinical setting12,13,42. This is most likely because the embryos were cryopreserved at the 2PN stage and represented the spectrum of embryos encountered in an IVF clinic. No triage was done before cryopreservation. Third, embryos frozen at the 2PN stage have been shown to possess similar potential for development, implantation, clinical pregnancy and delivery compared with fresh embryos43–45. Other studies have also shown similar results for frozen oocytes24,46. Fourth, we focused on parameters that were not dependent on time of fertilization or thaw time. As described in the manuscript, the first parameter, duration of the first cytokinesis, is short (~10–15 min) and is not dependent on the time of fertilization. The other parameters we measured are relative to this initial measurement point and compared between embryos that succeed in develop-ing to blastocyst and those that do not.

Control for cryopreservation. In addition to the observations described above, which support the use of cryopreserved embryos, we also examined a small set of embryos obtained from the Stanford IVF clinic that were not cryopreserved as a control. These embryos were 3PN (triploid) starting at the single-cell stage. 3PN embryos have been shown to follow the same time line of landmark events as normal fresh embryos through at least the first three cell cycles47–49. These embryos were imaged before our main experiments to validate the imaging systems (but for technical reasons were not followed out to blastocyst). Out of this set of fresh embryos, three of the embryos fol-lowed a similar time line of events as our cryopreserved 2PN embryos, with duration of cytokinesis ranging from 15 to 30 min, time between first and second mitoses ranging from 9.6 to 13.8 h and time between second and third mitoses ranging from 0.3 to 1.0 h. However, in seven of the embryos we observed a unique cytokinesis phenotype that was characterized by the simultaneous appearance of three cleavage furrows, a slightly prolonged cyto-kinesis and ultimate separation into three daughter cells (Fig. 2e, fourth panel, and Supplementary Video 5). These embryos had a duration of cytokinesis

ranging from 15 to 70 min (characterized as the time between the initiation of the cleavage furrows until complete separation into three daughter cells), time between first and second mitoses (3-cell to 4-cell) ranging from 8.7 to 12.7 h, and time between second and third mitoses (4-cell to 5-cell) ranging from 0.3 to 2.6 h. This observation, together with the diverse range of cytokinesis phenotypes displayed by abnormal embryos, suggests that our cryopreserved embryos are not developmentally delayed by the cryopreservation process and behave similarly to fresh zygotes that cleave to two blastomeres. The data also demonstrate that abnormal cytokinesis may be associated with underlying abnormalities in chromosomal composition (as demonstrated by zygotes that form three blastomeres initially). This hypothesis is consistent with previous observations correlating aneuploidy with morphology50–52.

Human embryo culture and microscopy. Human embryos were thawed by removing the cryovials from the liquid nitrogen storage tank and thawing them at 22 + 2 °C. Once a vial was thawed, it was opened and the embryos were visualized under a dissecting microscope. The contents of the vial were then poured into the bottom of a 3003 culture dish. The embryos were located in the drop and the survival of each embryo was assessed and recorded. At 22 + 2 °C, the embryos were transferred to a 3037 culture dish containing 1.0 M 1,2 propanediol + 0.2 M sucrose for 5 min, then 0.5 M 1,2 propanediol + 0.2 M sucrose for 5 min and 0.0 M 1,2 propanediol + 0.2M sucrose for 5 min. Subsequently, embryos were cultured in Quinn’s Advantage Cleavage Medium (Cooper Surgical) supplemented with 10% Quinn’s Advantage Serum Protein Substitute (SPS; Cooper Surgical) between day 1 to 3, and Quinn’s Advantage Blastocyst Medium (Cooper Surgical) with 10% SPS after day 3 using micro-drops under oil. All of the experiments used the same type of cleavage-stage medium, except for two stations during the first experiment, which used a Global medium (LifeGlobal). In this small subset (12 embryos), the embryos exhibited a slightly lower blastocyst formation rate (3 out of 12, or 25%) but the sensitivity and specificity of our predictive parameters were both 100% for this group.

Time-lapse imaging was performed on multiple systems to accommodate concurrent analysis of multiple samples as well as to validate the consistency of the data across different platforms. The systems consisted of seven individual microscopes: (i) two modified Olympus IX-70/71 microscopes equipped with Tokai Hit heated stages, white-light Luxeon LEDs, and an aperture for dark-field illumination; (ii) two modified Olympus CKX-40/41 micro-scopes equipped with heated stages, white-light Luxeon LEDs, and Hoffman Modulation Contrast illumination (note: these systems were used only during the first of four experiments after it was decided that dark-field illumina-tion was preferable for measuring the parameters); and (iii) a custom built three-channel miniature microscope array that fits inside a standard incubator, equipped with white-light Luxeon LEDs and apertures for darkfield illumina-tion. We observed no important difference in developmental behavior, blasto-cyst formation rate or gene expression profiles between embryos cultured on these different systems; indeed, our parameters for blastocyst prediction were consistent across multiple systems and experiments.

The light intensity for all systems was considerably lower than the light typically used on an assisted reproduction microscope due to the low-power of the LEDs (relative to a typical 100 W Halogen bulb) and high sensitivity of the camera sensors. Using an optical power meter, we determined that the power of a typical assisted-reproduction microscope (Olympus IX-71 Hoffman Modulation Contrast) at a wavelength of 473 nm ranges from roughly 7 to 10 mW depending on the magnification, whereas the power of our imaging systems were measured to be between 0.2 and 0.3 mW at the same wavelength. Images were captured at a 1 s exposure time every 5 min for up to 5 or 6 d, resulting in ~24 min of continuous light exposure. At a power of 0.3 mW, this is equivalent to roughly 1 min of exposure under a typical assisted-reproduc-tion microscope.

To track the identity of each embryo during correlated imaging and gene expression experiment, we installed a video camera on the stereomicroscope and recorded the process of sample transfer during media change and sample collection. We performed control experiments with mouse preimplantation embryos (n = 56) and a small subset of human embryos (n = 22), and observed no significant difference (P = 0.96) in the blastocyst formation rate between imaged and control embryos.

© 2

010

Nat

ure

Am

eric

a, In

c. A

ll ri

gh

ts r

eser

ved

.

nature biotechnologydoi:10.1038/nbt.1686

High-throughput qRT-PCR analysis. For single embryo or single blastomere qRT-PCR analysis, embryos were first treated with Acid Tyrode’s solution to remove the zona pellucida. To collect single blastomeres, the embryos were incubated in Quinn’s Advantage Ca2+ Mg2+–free medium with HEPES (Cooper Surgical) for 5–20 min at 37 °C with rigorous pipetting. Samples were collected directly into 10 μl of reaction buffer; subsequent one-step reverse transcription/pre-amplification reaction was performed as previ-ously described53. Pooled 20× ABI assay-on-demand qRT-PCR primer and probe mix (Applied Biosystems) were used as gene-specific primers during the reverse transcription and pre-amplification reactions. High throughput qRT-PCR reactions were performed with Fluidigm Biomark 96.96 Dynamic Arrays as previously described53 using the ABI assay-on-demand qRT-PCR probes listed in Supplementary Data Set 4. All samples were loaded in three or four technical replicates. qRT-PCR data analysis was performed with qBase-Plus (Biogazelle), Microsoft Excel and a custom-built software. Certain genes were omitted from data analysis owing to either poor data quality (e.g., poor PCR amplification curves) or consistent low to no expression in the embryos assessed. For the analysis of blastomere age, the maternal transcript panel used includes DAZL, GDF3, IFITM1, STELLAR, SYCP3, VASA, GDF9, PDCD5, ZAR1 and ZP1, whereas the embryonic gene panel includes ATF7IP, CCNA1, EIF1AX, EIF4A3, H2AFZ, HSPA1B, JARID1B, LSM3, PABPC1 and SERTAD1. The expression value of each gene relative to the reference genes GAPDH and RPLP0, as well as relative to the gene average, was calculated using the geNorm and ΔΔCt methods (Supplementary Data Sets 5–7)54,55. GAPDH and RPLP0 were selected as the reference genes for this study empirically based on the gene stability value and coefficient of variation: 1.18 and 46% for GAPDH and 1.18 and 34% for RPLP0. These were the most stable among the ten house-keeping genes that we tested and well within range of a typical heterogeneous sample set56. Second, we observed that in single blastomeres, as expected, the amount of RPLP0 and GAPDH transcripts decreased by ~1 Ct value per divi-sion between 1-cell and 8-cell stage, congruent with expectations that each cell inherits approximately one-half of the pool of mRNA with each cleavage division, in the absence of new transcripts before EGA during the first 3 d of human development (Supplementary Fig. 8). Third, we noted that the expres-sion level of these reference genes in single blastomeres remained stable from the 8-cell to morula stage, after EGA began. At the intact embryo level, the Ct values of both RPLP0 and GAPDH remained largely constant throughout development until the morula stage with a slight increase following in the blas-tocyst stage perhaps due to increased transcript levels in the greater numbers of blastomeres present. Most of the gene expression analysis performed in this study focused on developmental stages before the morula stage, however, when the expression level of the reference genes was extremely stable.

Normal and arrested embryos. In the experiments described in this work, arrested embryos were defined as embryos that did not reach the expected developmental stage at the time of sample collection. For example, when a plate of embryos was removed from the imaging station on late day 2 for sample col-lection, any embryo that had reached 4-cell stage and beyond would be identified as normal, whereas those that failed to reach 4-cell stage would be labeled as arrested. These arrested embryos were categorized by the developmental stage at which they became arrested, such that an embryo with only 2 blastomeres on late day 2 would be analyzed as an arrested 2-cell embryo. Care was taken to exclude embryos that morphologically appeared to be dead and porous at the time of sample collection (e.g., degenerate blastomeres). Only embryos that appeared alive (for both normal and arrested) were used for gene expression analysis. However, it is possible that embryos that appeared normal during the time of collection might ultimately arrest if they were allowed to grow to a later stage. Multiplex qRT-PCR reactions of up to 96 genes belonging to different categories were assayed per sample, including housekeeping genes, cytokinesis compo-nents, germ cell markers, maternal factors, embryonic genome activation (EGA) markers, trophoblast markers, inner cell mass markers, pluripotency markers, epigenetics regulators, transcription factors, hormone receptors and others. Two slightly different but overlapping sets of genes were assayed in three different experimental sets (Supplementary Table 1 and Supplementary Data Set 2).

Individual blastomeres and maternal and embryonic programs. Based on our data, we defined a maternal program as the average expression values of the

ten markers most quantitatively representative of the ESSP1 genes (maternal transcripts), and the embryonic program by the average expression values of ten markers most representative of the ESSP2 genes (embryonic transcripts). Thus, the gene expression profile of a human embryo at any given time is the sum of maternal mRNA degradation and embryonic or EGA transcripts. Maternal transcripts typically comprise the bulk of transcripts in a young blastomere of early developmental age relative to EGA transcripts, and the opposite holds true for an older blastomere at a more advanced developmental age (Fig. 5b).

Automated cell tracking. Our cell tracking algorithm uses a probabilistic framework based on sequential Monte Carlo methods, which in the field of computer vision is often referred to as the particle filter. The particle filter tracks the propagation of three main variables over time: the state, the con-trol and the measurement. The state variable is a model of an embryo and is represented as a collection of ellipses. The control variable is an input that transforms the state variable and consists of our cell propagation and division model. The measurement variable is an observation of the state and consists of our images acquired by the time-lapse microscope. Our estimate of the current state at each time step is represented with a posterior probability distribution, which is approximated by a set of weighted samples called particles. We use the terms particles and embryo models interchangeably, where a particle is one hypothesis of an embryo model at a given time. After initialization, the particle filter repeatedly applies three steps: prediction, measurement and update.

Prediction. Cells are represented as ellipses in two-dimensional space, and each cell has an orientation and overlap index. The overlap index specifies the relative height of the cells. In general, there are two types of behavior that we want to predict: cell motion and cell division. For cell motion, our control input takes a particle and randomly perturbs each parameter for each cell, including position, orientation, and length of major and minor axes. The perturbation is randomly sampled from a normal distribution with relatively small variance (5% of the initialized values). For cell division, we use the following approach. At a given point in time, for each particle, we assign a 50% probability that one of the cells will divide. This value was chosen empirically, and spans a wide range of possible cell divisions while maintaining good coverage of the current configuration. If a division is predicted, then the dividing cell is chosen ran-domly. When a cell is chosen to divide (Supplementary Fig. 9, left), we apply a symmetric division along the major axis of the ellipse, producing two daughter cells of equal size and shape (Supplementary Fig. 9, middle). We then ran-domly perturb each value for the daughter cells (Supplementary Fig. 9, right). Finally, we randomly select the overlap indices of the two daughter cells while maintaining their collective overlap relative to the rest of the cells.

After applying the control input, we convert each particle into a simulated image. This is achieved by projecting the elliptical shape of each cell onto the simulated image using the overlap index. The corresponding pixel values are set to a binary value of 1 and dilated to create a membrane thickness compa-rable to the observed image data. Because the embryos are partially transpar-ent and out-of-focus light is collected, cell membranes at the bottom of the embryo are only visible sometimes. Accordingly, occluded cell membranes are added with 10% probability. In practice, we have found that these occluded membrane points are crucial for accurate shape modeling, but it is important to make them sparse enough so that they do not resemble a visible edge.

Measurement. Once we have generated a distribution of hypothesized models, the corresponding simulated images are compared to the actual microscope image. The microscope image (Supplementary Fig. 10, left) is pre-processed to create a binary image of cell membranes using a principle curvature-based method (Supplementary Fig. 10, middle) followed by thresholding (Supplementary Fig. 10, right). The accuracy of the comparison is evaluated using a symmetric truncated chamfer distance, which is then used to assign a weight, or likelihood, to each particle.

Update. After weights are assigned, particles are selected in proportion to these weights to create a new set of particles for the next iteration. This focuses the particle distribution in the region of highest probability. Particles with low probability are discarded, whereas particles with high probability are multi-plied. Particle re-sampling is performed using the low variance method.

© 2

010

Nat

ure

Am

eric

a, In

c. A

ll ri

gh

ts r

eser

ved

.

nature biotechnology doi:10.1038/nbt.1686

Once the embryos have been modeled, we can extract the dynamic imaging parameters such as duration of cytokinesis and time between mitoses, as dis-cussed in the main text. Our cell tracking software was previously implemented in Matlab, and computation times ranged from a couple seconds to half a minute for each image depending on the number of particles. Our current version of the software is implemented in the programming languages C/C++, and computation times range from 1 to 5 s depending on the number of particles.

Discrepancies. In cases where there were discrepancies between manual and automated measurements, abnormal embryos were correctly assessed by the automated software. Moreover, our results indicated that the use of automated image analysis software may potentially provide improved ability to differenti-ate not only the success or failure to reach blastocyst but also blastocyst quality; however, further data and analysis are needed to support this observation.

Applications. Our imaging analysis was validated through the development of a novel cell tracking algorithm that can automatically extract the dynamic imaging parameters. We believe this to be the first documented report with human embryo development subjected to diagnosis by automated image analysis algorithms. Recently, cell tracking algorithms have generated much interest in areas such as predicting stem cell fate57, characterizing subcellular and membrane dynamics58,59, and lineage tracing during embryogenesis in Caenorhabditis elegans60. We anticipate that our algorithms, and in particular the underlying probabilistic model estimation technique, will be useful in other applications of time-lapse image analysis that deal with arbitrary motion dynamics and measurement uncertainties.

44. Miller, K.F. & Goldberg, J.M. In vitro development and implantation rates of fresh and cryopreserved sibling zygotes. Obstet. Gynecol. 85, 999–1002 (1995).

45. Damario, M.A., Hammitt, D.G., Galanits, T.M., Session, D.R. & Dumesic, D.A. Pronuclear stage cryopreservation after intracytoplasmic sperm injection and conventional IVF: implications for timing of the freeze. Fertil. Steril. 72, 1049–1054 (1999).

46. Vajta, G., Nagy, Z., Cobo, A., Conceicao, J. & Yovich, J. Vitrification in assisted reproduction: myths, mistakes, disbeliefs and confusion. Reprod. Biomed. Online 19, 1–7 (2009).

47. Liebermann, J. et al. Blastocyst development after vitrification of multipronuclear zygotes using the Flexipet denuding pipette. Reprod. Biomed. Online 4, 146–150 (2002).

48. Tarin, J.J., Trounson, A. & Sathananthan, H. Origin and ploidy of multipronuclear zygotes. Reprod. Fertil. Dev. 11, 273–279 (1999).

49. Sathananthan, A.H. et al. Development of the human dispermic embryo (CD-ROM). Hum. Reprod. Update 5, 553–560 (1999).

50. Baltaci, V. et al. Relationship between embryo quality and aneuploidies. Reprod. Biomed. Online 12, 77–82 (2006).

51. Fino, E. et al. How good is embryo morphology at predicting chromosomal integrity? When is aneuploidy PGD useful? Fertil. Steril. 84, S98–S99 (2005).

52. Kearns, W. et al. Aneuploidy rates of human preimplantation embryos in relation to morphology and development. Fertil. Steril. 86, S474 (2006).

53. Foygel, K. et al. A novel and critical role for Oct4 as a regulator of the maternal-embryonic transition. PLoS ONE 3, e4109 (2008).

54. Livak, K.J. & Schmittgen, K.D. Analysis of relative gene expression data using real time quantitative PCR and the 2-deltaCT Method. Methods 25, 402–408 (2001).

55. Vandesompele, J. et al. Accurate normalization of real-time quantitative RT-PCR data by geometric averaging of multiple internal control genes. Genome Biol. 3, 0034.1–0034.12 (2002).

56. Hellemans, J., Mortier, G., De Paepe, A., Speleman, F. & Vandesompele, J. qBase relative quantification framework and software for management and automated analysis of real-time quantitative PCR data. Genome Biol. 8, R19.1–R19.14 (2007).

57. Cohen, A.R., Gomes, F.L., Roysam, B. & Cayouette, M. Computational prediction of neural progenitor cell fates. Nat. Methods 7, 213–218 (2010).

58. Jaqaman, K. et al. Robust single-particle tracking in live-cell time-lapse sequences. Nat. Methods 5, 695–702 (2008).

59. Sergé, A., Bertaux, N., Rigneault, H. & Marguet, D. Dynamic multiple-target tracing to probe spatiotemporal cartography of cell membranes. Nat. Methods 5, 687–694 (2008).

60. Bao, Z. et al. Automated cell lineage tracing in Caenorhabditis elegans. Proc. Natl. Acad. Sci. USA 103, 2707–2712 (2006).

42. Zhang, J.Q. et al. Reduction in exposure of human embryos outside the incubator enhances embryo quality and blastulation rate. Reprod. Biomed. Online 20, 510–515 (2010).

43. Veeck, L.L. et al. Significantly enhanced pregnancy rates per cycle through cryopreservation and thaw of pronuclear stage oocytes. Fertil. Steril. 59, 1202–1207 (1993).

© 2

010

Nat

ure

Am

eric

a, In

c. A

ll ri

gh

ts r

eser

ved

.