Common variation at 6q16 within HACE1 and LIN28B influences susceptibility to neuroblastoma
-
Upload
independent -
Category
Documents
-
view
3 -
download
0
Transcript of Common variation at 6q16 within HACE1 and LIN28B influences susceptibility to neuroblastoma
©20
12 N
atu
re A
mer
ica,
Inc.
All
rig
hts
res
erve
d.
Nature GeNetics ADVANCE ONLINE PUBLICATION �
l e t t e r s
Neuroblastoma is a cancer of the sympathetic nervous system that accounts for approximately �0% of all pediatric oncology deaths�. Here, we report a genome-wide association study of 2,8�7 neuroblastoma cases and 7,473 controls. We identified two new associations at 6q�6, the first within HACE1 (rs4336470; combined P = 2.7 × �0−��; odds ratio �.26, 95% confidence interval (CI) �.�8–�.35) and the second within LIN28B (rs�70654�7; combined P = �.2 × �0−8; odds ratio �.38, 95% CI �.23–�.54). Expression of LIN28B and let-7 miRNA correlated with rs�70654�7 genotype in neuroblastoma cell lines, and we observed significant growth inhibition upon depletion of LIN28B, specifically in neuroblastoma cells that were homozygous for the risk allele. Low HACE1 and high LIN28B expression in diagnostic primary neuroblastomas were associated with worse overall survival (P = 0.008 and 0.0�4, respectively). Taken together, these data show that common variants in HACE1 and LIN28B influence neuroblastoma susceptibility and indicate that both genes likely have a role in disease progression.
Neuroblastoma is a malignancy derived from the developing sympathetic nervous system. The median age at diagnosis is 17 months, and the survival rate for the most aggressive subset remains approximately 50%, despite intensive multimodal cytotoxic therapy1 and recent advances in immunotherapy2. The genetic etiology of familial neuroblastoma, which accounts for approximately 1% of cases, has recently come into focus3–6; however, the genetic and environmental factors that cause sporadic neuroblastoma remain largely unknown. We have recently reported common SNPs within or upstream of LINC00340 and LOC729177 (also known as FLJ44180), BARD1, LMO1, DUSP12, HSD17B12 and DDX4IL31RA and a common copynumber variation (CNV) within NBPF23 are each highly associated with neuroblastoma7–11. Collectively, however, these variants still account for
only a small portion of neuroblastoma heritability, and it is likely that additional predisposition loci remain to be discovered.
To identify additional variants associated with neuroblastoma, we expanded our previous genomewide association study (GWAS) discovery cohort and analyzed 2,101 neuroblastoma cases accrued through the North America–based Children’s Oncology Group (Supplementary Table 1) with 4,202 control subjects of European ancestry who were matched genetically and by genotyping array version to minimize genomic inflation (Online Methods and Supplementary Figs. 1 and 2). All subjects were genotyped using the Illumina HumanHap550 or Quad610 BeadChips. We restricted our analysis to the SNPs present on both platforms that passed our quality control metrics; the genomic control inflation factor was 1.14 (Supplementary Fig. 3). Evaluation of the first three principal components in cases and controls confirmed that the slightly high inflation factor was not due to gross population stratification (Supplementary Fig. 4). Clusters of SNPs from six genomic loci reached genomewide significance (P values ranged from 7.8 × 10−16 to 4.8 × 10−8; Supplementary Fig. 1 and Supplementary Table 2), including three SNPs within LINC00340 and LOC729177 (FLJ44180) at 6p22 (P values ranged from 7.8 × 10−16 to 1.7 × 0−14), ten SNPs within or near BARD1 at 2q35 (P values ranged from 4.1 × 10−14 to 3.7 × 10−8), two SNPs within LMO1 at 11p15 (P values ranged from 1.2 × 10−13 to 3.8 × 10−10) and one SNP within HSD17B12 at 11p11 (P = 4.8 × 10−8), further confirming our previous reports7–10. In addition, we identified one SNP (rs4696715) at chromosome 4p16 and two SNPs (rs4336470 and rs9404576) at 6q16 that have not been reported previously (P values ranged from 1.8 × 10−8 to 3.4 × 10−8; Table 1). Several genotyped SNPs in strong linkage disequilibrium (LD) with rs4696715 at 4p16 did not show evidence for association with neuroblastoma; therefore, this SNP was not considered further.
Closer examination of the 6q16 locus identified four additional SNPs that showed association (P < 1 × 10−4) with neuroblastoma
Common variation at 6q16 within HACE1 and LIN28B influences susceptibility to neuroblastomaSharon J Diskin1–3,11, Mario Capasso4,5, Robert W Schnepp1,2, Kristina A Cole1–3, Edward F Attiyeh1–3, Cuiping Hou6, Maura Diamond1,2, Erica L Carpenter1,2, Cynthia Winter1,2, Hanna Lee1,2, Jayanti Jagannathan1,2, Valeria Latorre7,8, Achille Iolascon4,5, Hakon Hakonarson3,6,7, Marcella Devoto3,7,9,10 & John M Maris1–3,11
1Division of Oncology, Children’s Hospital of Philadelphia, Philadelphia, Pennsylvania, USA. 2Center for Childhood Cancer Research, Children’s Hospital of Philadelphia, Philadelphia, Pennsylvania, USA. 3Department of Pediatrics, Perelman School of Medicine, University of Pennsylvania, Philadelphia, Pennsylvania, USA. 4Dipartimento di Biochimicae Biotecnologie Mediche, Università degli Studi di Napoli Federico II, Naples, Italy. 5Ceinge–Biotecnologie Avanzate, Naples, Italy. 6The Center for Applied Genomics, Children’s Hospital of Philadelphia, Philadelphia, Pennsylvania, USA. 7Division of Genetics, Children’s Hospital of Philadelphia, Philadelphia, Pennsylvania, USA. 8Department of Cell Biology, Università della Calabria, Rende, Italy. 9Department of Molecular Medicine, University of Rome La Sapienza, Rome, Italy. 10Department of Biostatistics and Epidemiology, Perelman School of Medicine, University of Pennsylvania, Philadelphia, Pennsylvania, USA. 11Abramson Family Cancer Research Institute, Perelman School of Medicine, University of Pennsylvania, Philadelphia, Pennsylvania, USA. Correspondence should be addressed to J.M.M. ([email protected]).
Received 16 March; accepted 24 July; published online 2 September 2012; doi:10.1038/ng.2387
©20
12 N
atu
re A
mer
ica,
Inc.
All
rig
hts
res
erve
d.
2 ADVANCE ONLINE PUBLICATION Nature GeNetics
l e t t e r s
(Table 1 and Supplementary Table 3). Three of these SNPs mapped to introns of the HACE1 gene (encoding HECT domain– and ankyrin repeat–containing E3 ubiquitin protein ligase 1) and exhibited a moderate degree of LD with rs4336470. The fourth SNP (rs17065417; P = 1.8 × 10−7) mapped to an intron of the LIN28B gene (encoding lin28 homolog B) and showed very little evidence for LD with rs4336470 in multiple HapMap populations (Supplementary Fig. 5). To ensure that these results were not influenced by subtle substructure in the discovery phase, we included the first 20 principal components as covariates in a logistic regression analysis. This reduced the inflation factor to 1.04 but did not alter our conclusions regarding regions that were associated at genomewide significance, and the P values for SNPs at the newly identified 6q16 locus were essentially unchanged (Supplementary Table 4). Finally, we conditioned the analysis of the 6q16 SNPs on rs4336470 to investigate whether there might be more than one independent association signal at 6q16. As expected, the associations for SNPs in modest LD with rs4336470 were no longer statistically significant after conditioning, and these SNPs clearly represent one signal (Supplementary Table 5). In contrast, whereas the signal at rs17065417 was attenuated, it remained significant (P = 2.5 × 10−4; Supplementary Table 5), suggesting that rs4336470 and rs17065417 may contribute independently to neuroblastoma risk. No association was observed between rs4336470 or rs17065417 genotype and clinical and biological covariates (Supplementary Tables 6 and 7), and we observed only weak evidence for epistasis at the other significant loci (Supplementary Tables 8 and 9), suggesting that each may act independently to confer risk.
We next sought to replicate the rs4336470 and rs17065417 associations in an Italian cohort of 351 cases and 780 controls using PCRbased genotyping. Both SNPs showed evidence for association in the same direction as seen in the discovery effort (Table 1 and Supplementary Table 3). To assess whether these variants influence susceptibility in other ancestry groups and to seek additional replication, we analyzed a third independent case series comprised of 365 AfricanAmerican neuroblastoma cases and 2,491 genetically matched controls, all genotyped on Illumina HumanHap550 or Quad610 BeadChips. After accounting for African admixture in logistic regression analysis, the genomic inflation factor was 1.02, as
described previously12 (Online Methods and Supplementary Fig. 6). Consistent with the lower incidence rate of neuroblastoma in AfricanAmericans13, the rs4336470[T] protective allele is actually the major allele in AfricanAmericans, whereas it is the minor allele in individuals of European ancestry. Despite these observed differences in allele frequencies, all SNPs mapped to HACE1 replicated robustly in the AfricanAmerican cohort (P values ranged from 1.3 × 10−4 to 1.4 × 10−3; Table 1 and Supplementary Table 3), and all associations were in the same direction. Allele frequencies for the LIN28B SNP rs17065417 were comparable across populations of different ancestry and showed a trend toward association in AfricanAmericans in the same direction as in the EuropeanAmerican discovery and Italian replication cohorts (Table 1 and Supplementary Table 3). Combined analysis using the weighted inversevariance method in METAL14 showed that all six genotyped SNPs had P values beyond or approaching the conservative Bonferroniadjusted genomewide
10a
b
8
6
4
2
0
105.0 105.1 105.2 105.3Position on chr. 6 (Mb)
105.4 105.5 105.6
100
80
60
40
20
0
Recom
bination rate (cM/M
b)
–log
10 (
P v
alue
)
rs4336470P = 1.8 × 10–8
r2
0.80.60.40.2
HACE1 LIN28B BVESPOPDC3
105.0 105.1 105.2 105.3
Position on chr. 6 (Mb)
105.4 105.5 105.6
HACE1 LIN28B BVESPOPDC3
10
8
6
4
2
0
–log
10 (
P v
alue
)
rs17065417P = 1.8 × 10–7
100
80
Recom
bination rate (cM/M
b)
60
40
20
0
r2
0.80.60.40.2
Figure 1 Regional association plots at the HACE1 and LIN28B loci. (a,b) Regional association plots including both genotyped and imputed SNPs for the HACE1 (a) and LIN28B (b) loci generated by LocusZoom36. Plotted are the significance of association (−log10-transformed P values) and the recombination rate. SNPs are colored to reflect pairwise LD (r2) with the most significantly associated genotyped SNP in the 1000 Genomes Project Phase 1 interim release EUR population genotypes. The most significant genotyped SNPs are labeled and shown in purple.
table 1 significantly associated genotyped sNPs at the HACE1 and LIN28B loci at 6q16Discovery cohorta Italian replicationa African-American replicationa Combined
SNP
A1/A2 (protective/ risk allele)
Freq. A1 cases
(n = 2,101)
Freq. A1 controls
(n = 4,202) P b
Freq. A1 cases
(n = 351)
Freq. A1 controls
(n = 780) Pb
Freq. A1 cases
(n = 365)
Freq. A1 controls
(n = 2,491) Pc Pmetad ORe (95% CI)
HACE1
rs4336470 T/C 0.30 0.35 1.8 × 10−8 0.30 0.34 0.060 0.62 0.70 1.4 × 10−3 2.7 × 10−11 1.26 (1.18–1.35)
rs9404576 G/T 0.30 0.35 3.4 × 10−8 – – – 0.62 0.70 1.3 × 10−3 1.8 × 10−10 1.27 (1.18–1.36)
rs4079063 G/A 0.43 0.47 4.0 × 10−5 – – – 0.70 0.78 1.3 × 10−4 1.3 × 10−7 1.20 (1.12–1.29)
rs2499663 C/T 0.43 0.47 4.5 × 10−5 – – – 0.70 0.78 1.5 × 10−4 1.6 × 10−7 1.21 (1.13–1.29)
rs2499667 G/A 0.43 0.47 2.6 × 10−5 – – – 0.71 0.78 2.6 × 10−4 1.2 × 10−7 1.21 (1.13–1.29)
LIN28B
rs17065417 C/A 0.08 0.11 1.8 × 10−7 0.08 0.11 0.033 0.09 0.11 0.129 1.2 × 10−8 1.38 (1.23–1.54)
CHOP, Children’s Hospital of Philadelphia; freq., frequency.aNo deviations from Hardy-Weinberg equilibrium were observed (P > 0.001) in all cohorts. bP values were calculated by allelic test. cP values were calculated by logistic regression, with percent African admixture as covariate12. dMeta-analysis P value calculated using METAL14. eOR, odds ratio of risk allele based on meta-analysis.
©20
12 N
atu
re A
mer
ica,
Inc.
All
rig
hts
res
erve
d.
Nature GeNetics ADVANCE ONLINE PUBLICATION 3
l e t t e r s
significance threshold (P values ranged from 2.7 × 10−11 to 1.6 × 10−7; Table 1).
To identify variants at the 6q16 locus that were not assayed directly on the Illumina SNP arrays, we performed genotype imputation in our discovery cohort using data from the 1000 Genomes Project. This analysis identified ten additional SNPs associated with genomewide significance (P values ranged from 4.4 × 10−9 to 5.7 × 10−8; Fig. 1). Five of these imputed SNPs were in strong LD with rs4336470 (r2 > 0.8 in the 1000 Genomes Project European (EUR) population); three mapped to introns of HACE1 and the other two were located just downstream of HACE1 (Fig. 1a). The remaining five imputed SNPs reaching genomewide significance were in strong LD with rs17065417 (r2 > 0.8 in the 1000 Genomes Project EUR population) and were located within introns of LIN28B (Fig. 1b). In total, we identified 46 imputed SNPs that showed strong evidence for association (P < 1.0 × 10−6; Supplementary Table 10). To further evaluate whether two independent association signals exist at 6q16, we conditioned the regional association analysis on either rs4336470 or rs17065417,
finding that neither could fully account for the observed associations (Supplementary Fig. 7). The signal at 6q16 was only abolished across the entire region after conditioning on both rs4336470 and rs17065417 (Supplementary Fig. 8). Although we cannot rule out the possibility that both SNPs may be tagging the same underlying risk variant, these data are consistent with the presence of two independent association signals at 6q16, one implicating HACE1 and the other LIN28B.
The HACE1 gene encodes an E3 ubiquitin protein ligase that was first identified in a sporadic Wilms’ tumor that harbored a t(6:15)(q21;q21) translocation15. The rearrangement was associated with decreased HACE1 expression, and further study showed that
a
SK
NA
S
SK
NF
I
NB
1691
BE
2C
CH
P13
4
IMR
5
NB
1771
NB
LS
SK
ND
Z
NG
P
KE
LLY
0
1,000
2,000
3,000
4,000
5,000
LIN
28B
mR
NA
exp
ress
ion
SH
-SY
5Y
CA AA
P = 0.02
0
1,000
2,000
3,000
4,000
5,000
6,000
MYC
N m
RN
A e
xpre
ssio
n
b
SK
NA
S
SK
NF
I
NB
1691
BE
2C
CH
P13
4
IMR
5
NB
1771
NB
LS
SK
ND
Z
NG
P
KE
LLY
SH
-SY
5Y
CA AA
P = 0.0003
–2.0
hsa-
let-7
b-3p
hsa-
let-7
a-5p
hsa-
let-7
a-3p
hsa-
let-7
b-5p
hsa-
let-7
c-5p
hsa-
let-7
c-3p
hsa-
let-7
d-5p
hsa-
let-7
d-3p
hsa-
let-7
e-5p
hsa-
let-7
e-3p
hsa-
let-7
f-1-3
p
hsa-
let-7
f-5p
hsa-
let-7
f-2-3
p
hsa-
let-7
g-5p
hsa-
let-7
g-3p
hsa-
let-7
i-5p
hsa-
let-7
i-3p
–1.5
–1.0
–0.5
0
0.5
1.0
1.5
2.0SKNAS (CA) SH-SY5Y (CA) NGP (AA) KELLY (AA)d
miR
NA
exp
ress
ion
(s.d
. fro
m m
ean)
c
Actin
LIN28B
SK
NA
S
SH
-SY
5Y
NG
P
KE
LLY
CA AA
32 kDa
LIN
28B
pro
tein
43 kDa
0
1.0
1.5
0.5
P = 0.04
Figure 2 LIN28B risk alleles correlate with increased LIN28B expression and decreased let-7 miRNA expression. (a) LIN28B mRNA expression is significantly higher in neuroblastoma cell lines homozygous for the rs17065417[A] risk allele (AA) compared to neuroblastoma cell lines heterozygous for the risk allele (CA). Given the MAF of rs17065417, we did not identify any cell lines homozygous for the rs17065417[C] protective allele. (b) Neuroblastoma cell lines homozygous for the rs17065417 risk allele show high expression of MYCN. (c) Protein blot and corresponding densitometry analysis confirms higher amounts of LIN28B protein in neuroblastoma cell lines homozygous for the rs17065417 risk allele. (d) Neuroblastoma cell lines homozygous for the rs17065417 risk allele that have high LIN28B expression show decreased or absent expression of the let-7 miRNAs.
0 24 48 72 96 120 144
20
40
60
80
100
SKNDZ (AA)
Time (h)
b
Gro
wth
(%
)
0
Time (h)
0 12 24 36 48 60 72 84
20
40
60
80
100
NGP (AA)d
Gro
wth
(%
)
00 24 48 72 96 120
20
40
60
80
100
Time (h)
c
Gro
wth
(%
)
0
BE-2C (AA)
0 12 24 36 48 60 72 84 96 108
Gro
wth
(%
)
e
20
40
60
80
100
0
Time (h)
SK-N-AS (AC)
50
0
100
Exp
ress
ion
(%)
LIN28BNC
LIN28B
Actin
NGP (AA)
50
0
100
Exp
ress
ion
(%)
LIN28B
Actin
LIN28BNC
BE-2C (AA)
Exp
ress
ion
(%)
LIN28B
Actin
50
100
0LIN28BNC
SK-N-AS (AC)
LIN28B
Actin
Exp
ress
ion
(%)
LIN28BNC
50
0
100
SKNDZ (AA)
Exp
ress
ion
(%)
100
0
50
LIN28BNC
LIN28B
Actin
Kelly (AA)f
20
40
60
80
100
0 12 24 36 48 60 72 84 96 108
Kelly (AA)
NC
Gro
wth
(%
)
a
0
LIN28B
Time (h)
Figure 3 Transient knockdown of LIN28B influences neuroblastoma cell growth in an expression-specific manner. (a–d) In cells homozygous for neuroblastoma risk alleles that have higher LIN28B expression levels, LIN28B knockdown leads to significant growth inhibition, including in Kelly (a), SKNDZ (b), BE-2C (c) and NGP (d) cells. NC, negative control small interfering RNA (siRNA). Experiments were performed in triplicate. Error bars, range of growth observed. (e) In cells heterozygous for the risk allele (carrying one protective allele) that have low LIN28B expression, LIN28B knockdown does not affect cell growth. (f) LIN28B knockdown relative to NC, as measured by quantitative RT-PCR (top) and protein blot (bottom) for experiments in a–e.
©20
12 N
atu
re A
mer
ica,
Inc.
All
rig
hts
res
erve
d.
4 ADVANCE ONLINE PUBLICATION Nature GeNetics
l e t t e r s
HACE1 is silenced in the majority of Wilms’ tumors via hypermethylation of two CpG islands upstream of the transcriptional start site15. Similar epigenetic silencing has been reported in advanced colorectal cancer16 and gastric carcinoma17. Indeed, HACE1 is downregulated in multiple human tumors and maps to a region of common deletion or loss of heterozygosity (LOH), consistent with a tumor suppressor function. Hace1null mice form spontaneous tumors in a wide array of tissues and are susceptible to additional cancer triggers, both genetic and environmental18. Through its E3 ubiquitin ligase function, HACE1 has been shown to suppress cell growth and anchorage independence of human tumor cells, including in the neuroblastoma cell line IMR32 (ref. 18). Studies suggest that HACE1 inhibits cell cycle progression during stress via regulation of cyclin D1 degradation18 and that HACE1 also regulates retinoic acid receptor (RAR) activity19. In addition to multiple somatic alterations involving the HACE1 locus, a constitutional t(5;6)(q21;q21) translocation that disrupts HACE1 was recently identified, making it a putative Wilms’ tumor susceptibility gene20.
The LIN28B gene encodes a developmentally regulated RNA binding protein and is a key repressor of the let7 family of microRNAs (miRNAs)21. Both LIN28A and LIN28B are oncogenes that promote cellular transformation when ectopically overexpressed22–24, and high levels of expression of LIN28A or LIN28B have been observed in several human cancers and correlate with low let7 levels22,24. Posttranscriptional regulation of let7 by LIN28A is required for normal development and contributes to the pluripotent state by preventing let7–mediated differentiation of embryonic stem cells21,25,26. Overexpressing LIN28A or LIN28B or inhibiting let7 with antisense RNA promotes reprogramming of human and mouse fibroblasts into pluripotent stem cells26,27. In a panel of 60 pediatric cancer cell lines28, LIN28B was consistently expressed at high levels in neuroblastoma (Supplementary Fig. 9). Recently, LIN28B and let7 were also identified as key regulators of glucose homeostatis29 and hematopoiesis30. To date, GWAS have identified variants in LIN28B associated with human height31 and the ages of onset of puberty32 and menarche33.
In addition, a recent candidate gene study identified a putative association with epithelial ovarian cancer34.
To investigate the functional relevance of neuroblastomaassociated SNPs within HACE1 and LIN28B, we first analyzed a set of 12 neuroblastoma cell lines with matched genomewide SNP genotyping and mRNA expression data. No correlation between rs4336470 and HACE1 mRNA expression was observed (P = 0.30), but LIN28B expression was significantly higher in cell lines homozygous for the rs17065417[A] risk allele compared to heterozygous cell lines (P = 0.02; Fig. 2a). No cell lines tested were homozygous for the rs17065417[C] protective allele. LIN28A was not expressed in neuroblastoma cell lines, consistent with recent reports that LIN28A and LIN28B may be mutually exclusive in terms of their expression in cancer cells21. In addition to a correlation between rs17065417 genotype and LIN28B expression, we observed a strong positive correlation with MYCN expression (P = 0.0003; Fig. 2b). We next confirmed LIN28B protein levels by protein blot in four cell lines at the extremes of LIN28B mRNA expression (Fig. 2c). Cell lines expressing high amounts of LIN28B showed lower let7 expression across the entire miRNA family (Fig. 2d). Transient knockdown of LIN28B resulted in significant growth inhibition in neuroblastoma cells homozygous for the rs17065417 risk allele that had high LIN28B expression (Fig. 3a–d) but not in the heterozygous cell line SKNAS that had low LIN28B expression (Fig. 3e,f). Taken together, these data are consistent with the hypothesis that LIN28B promotes neuroblastoma tumorigenesis, in part through repression of let7 family miRNAs, and that the risk alleles are associated with growth advantage through increased LIN28B expression; however, additional studies with larger sample sizes are needed to confirm this model.
To examine the relevance of HACE1 and LIN28B in tumor samples, we assayed expression of both genes by quantitative RTPCR in a representative set of 87 primary tumors obtained at time of diagnosis (Supplementary Table 1). HACE1 expression was significantly lower (P = 0.002; Fig. 4a), and LIN28B expression was significantly higher (P = 0.032; Fig. 4b) in the highrisk tumors compared to the
a**
*
Low ri
sk (2
6)
Inte
rmed
iate
risk (
18)
High ri
sk (4
3)
HA
CE
1/H
PR
T ra
tio
0
0.5
1.0
1.5
2.0
NBL primary tumorsby COG group
b
Low ri
sk (2
6)
Inte
rmed
iate
risk (
18)
High ri
sk (4
3)
**
LIN
28B
/HP
RT
ratio
*
1.0
2.0
NBL primary tumorsby COG group
0.5
0
1.5
e
7.0
7.5
8.0
8.5
9.0
HA
CE
1 m
RN
A e
xpre
ssio
n
1(8)
2(15)
3(13)
4(40)
NBL primarytumors by stage
****
*****
f
1(8)
2(15)
3(13)
4(40)
NBL primarytumors by stage
LIN
28B
mR
NA
exp
ress
ion
*
*****
7.0
8.0
9.0
8.5
7.5Ove
rall
surv
ival
P = 0.008
High HACE1 (28)
IntermediateHACE1 (28)
Low HACE1 (28)
c
0
0.2
0.4
0.6
0.8
1.0
01,
000
2,00
03,
000
4,00
05,
000
Time (d)
d
P = 0.015
Time (d)
0
0.2
0.4
0.6
0.8
1.0
High LIN28B (28)
IntermediateLIN28B (28)
Low LIN28B (28)
01,
000
2,00
03,
000
4,00
05,
000
g
Ove
rall
surv
ival
Low HACE1 (25)
IntermediateHACE1 (26)
High HACE1 (25)
P < 0.00010
0.2
0.4
0.6
0.8
1.0
Time (d)
0 2,000 4,000 6,000
h
0
0.2
0.4
0.6
0.8
1.0
High LIN28B (25)
LowLIN28B (25)
IntermediateLIN28B (26)
Time (d)
0 2,000 4,000 6,000
P = 0.001
Figure 4 HACE1 and LIN28B expression are associated with advanced disease and poor outcome in neuroblastoma. (a,b) mRNA expression for Children’s Oncology Group (COG) risk groups of HACE1 (a) and LIN28B (b). The number of tumors in each group is indicated in parentheses. Error bars, s.e.m. NBL, neuroblastoma. (c,d) Kaplan-Meier analysis is shown, with individuals grouped by tertiles of expression of HACE1 (c) and LIN28B (d). Log-rank P values are shown. (e,f) Replication of changes in expression for HACE1 (e) and LIN28B (f) in advanced-stage neuroblastoma using published Affymetrix U133 plus v2 array data37 (R2 bioinformatics tool). Data are shown for International Neuroblastoma Staging System (INSS) stages 1–4. The number of tumors in each group is indicated in parentheses. Error bars, s.e.m. (g,h) Kaplan-Meier analysis is shown, with individuals grouped by tertiles of expression of HACE1 (g) and LIN28B (h). Log-rank P values are shown. Data sets for g and h correspond to those in e and f, respectively. ****P < 0.0001; ***P < 0.001; **P < 0.01; *P < 0.05.
©20
12 N
atu
re A
mer
ica,
Inc.
All
rig
hts
res
erve
d.
Nature GeNetics ADVANCE ONLINE PUBLICATION 5
l e t t e r s
lowrisk tumors. Accordingly, low HACE1 expression was associated with worse overall survival (P = 0.008; Fig. 4c), as was high LIN28B expression (P = 0.015; Fig. 4d). Analysis of four independent mRNA expression array data sets comprising 517 neuroblastoma tumors provided robust replication of these observations (Fig. 4e–h and Supplementary Figs. 10–12). These data support the hypothesis that HACE1 may function as a tumor suppressor and LIN28B as an oncogene in advanced neuroblastomas.
In conclusion, we have identified common variants within HACE1 and LIN28B that are associated with neuroblastoma. As with other genes identified in our ongoing GWAS efforts8,9,35, it is likely that the germline susceptibility variants identified here have an important role, not only in tumor initiation, but also in disease progression via cis effects on major cancerassociated genes. Further study of HACE1 and LIN28B in neuroblastoma is warranted and may lead to new insights into the genetic and epigenetic mechanisms underlying an aggressive clinical phenotype.
URLs. Database of Genotypes and Phenotypes (dbGaP), http://www.ncbi.nlm.nih.gov/gap; LocusZoom, http://csg.sph.umich.edu/locuszoom/; 1000 Genomes Project, http://www.1000genomes.org/; R2 bioinformatics tool, http://hgserver1.amc.nl/cgibin/r2/main.cgi; National Cancer Institute (NCI) Oncogenomics, http://home.ccr.cancer.gov/oncology/oncogenomics/; LiftOver tool, http://genome.ucsc.edu/cgibin/hgLiftOver; 1000 Genome Project Phase 1 interim data, http://mathgen.stats.ox.ac.uk/impute/data_download_1000G_phase1_interim.html.
METHodsMethods and any associated references are available in the online version of the paper.
Accession codes. Genotyping data have been deposited at dbGaP under accession phs000124.
Note: Supplementary information is available in the online version of the paper.
ACKnoWLEDgMEntSWe thank P. Sleiman for useful discussions regarding metaanalysis using METAL. This work was supported in part by US National Institutes of Health (NIH) grants R01CA124709 (J.M.M.), K99CA151869 (S.J.D.), P30HD026979 (M. Devoto) and K08CA136979 (K.A.C.); the Giulio D’Angio Endowed Chair (J.M.M.); the Alex’s Lemonade Stand Foundation (J.M.M.); Andrew’s Army Foundation (J.M.M.); the PressOn Foundation (J.M.M.); the Abramson Family Cancer Research Institute (J.M.M.); Fondazione Italiana per la Lotta al Neuroblastoma and Associazione Italiana per la Ricerca sul Cancro (M.C.); and the Center for Applied Genomics at the Children’s Hospital of Philadelphia Research Institute (H.H.).
AUtHoR ContRIBUtIonSS.J.D. and J.M.M. designed the experiment and drafted the manuscript. S.J.D. analyzed SNP data, performed SNP association studies, and analyzed mRNA and miRNA expression data. M.C. and A.I. replicated SNP associations in the Italian cohort. V.L. and M.D. replicated SNP associations in the AfricanAmerican cohort. E.L.C. and H.L. confirmed LIN28B protein expression by protein blot. R.W.S. and C.W. performed experiments with siRNAmediated knockdown of LIN28B. E.F.A. generated miRNA expression array data, including lowlevel summary values. K.A.C. performed RTPCR in primary tumors. M. Diamond and C.H. organized samples and genotyped cases. J.J. participated in expression array analyses. H.H. generated and provided all control data for the GWAS. M. Devoto and H.H. contributed to overall study design. All authors commented on or contributed to the current manuscript.
CoMPEtIng FInAnCIAL IntEREStSThe authors declare no competing financial interests.
Published online at http://www.nature.com/doifinder/10.1038/ng.2387. Reprints and permissions information is available online at http://www.nature.com/reprints/index.html.
1. Maris, J.M. Recent advances in neuroblastoma. N. Engl. J. Med. 362, 2202–2211 (2010).
2. Yu, A.L. et al. Anti-GD2 antibody with GM-CSF, interleukin-2, and isotretinoin for neuroblastoma. N. Engl. J. Med. 363, 1324–1334 (2010).
3. Mossé, Y.P. et al. Identification of ALK as a major familial neuroblastoma predisposition gene. Nature 455, 930–935 (2008).
4. Janoueix-Lerosey, I. et al. Somatic and germline activating mutations of the ALK kinase receptor in neuroblastoma. Nature 455, 967–970 (2008).
5. Trochet, D. et al. Germline mutations of the paired-like homeobox 2B (PHOX2B) gene in neuroblastoma. Am. J. Hum. Genet. 74, 761–764 (2004).
6. Mosse, Y.P. et al. Germline PHOX2B mutation in hereditary neuroblastoma. Am. J. Hum. Genet. 75, 727–730 (2004).
7. Maris, J.M. et al. Chromosome 6p22 locus associated with clinically aggressive neuroblastoma. N. Engl. J. Med. 358, 2585–2593 (2008).
8. Capasso, M. et al. Common variations in BARD1 influence susceptibility to high-risk neuroblastoma. Nat. Genet. 41, 718–723 (2009).
9. Wang, K. et al. Integrative genomics identifies LMO1 as a neuroblastoma oncogene. Nature 469, 216–220 (2011).
10. Nguyen, B. et al. Phenotype restricted genome-wide association study using a gene-centric approach identifies three low-risk neuroblastoma susceptibility loci. PLoS Genet. 7, e1002026 (2011).
11. Diskin, S.J. et al. Copy number variation at 1q21.1 associated with neuroblastoma. Nature 459, 987–991 (2009).
12. Latorre, V. et al. Replication of neuroblastoma SNP association at the BARD1 locus in African-Americans. Cancer Epidemiol. Biomarkers Prev. 21, 658–663 (2012).
13. Henderson, T.O. et al. Racial and ethnic disparities in risk and survival in children with neuroblastoma: a Children’s Oncology Group study. J. Clin. Oncol. 29, 76–82 (2011).
14. Willer, C.J., Li, Y. & Abecasis, G.R. METAL: fast and efficient meta-analysis of genomewide association scans. Bioinformatics 26, 2190–2191 (2010).
15. Anglesio, M.S. et al. Differential expression of a novel ankyrin containing E3 ubiquitin-protein ligase, Hace1, in sporadic Wilms’ tumor versus normal kidney. Hum. Mol. Genet. 13, 2061–2074 (2004).
16. Hibi, K. et al. Aberrant methylation of the HACE1 gene is frequently detected in advanced colorectal cancer. Anticancer Res. 28, 1581–1584 (2008).
17. Sakata, M. et al. Methylation of HACE1 in gastric carcinoma. Anticancer Res. 29, 2231–2233 (2009).
18. Zhang, L. et al. The E3 ligase HACE1 is a critical chromosome 6q21 tumor suppressor involved in multiple cancers. Nat. Med. 13, 1060–1069 (2007).
19. Zhao, J., Zhang, Z., Vucetic, Z., Soprano, K.J. & Soprano, D.R. HACE1: a novel repressor of RAR transcriptional activity. J. Cell. Biochem. 107, 482–493 (2009).
20. Slade, I. et al. Constitutional translocation breakpoint mapping by genome-wide paired-end sequencing identifies HACE1 as a putative Wilms tumour susceptibility gene. J. Med. Genet. 47, 342–347 (2010).
21. Piskounova, E. et al. Lin28A and Lin28B inhibit let-7 microRNA biogenesis by distinct mechanisms. Cell 147, 1066–1079 (2011).
22. Iliopoulos, D., Hirsch, H.A. & Struhl, K. An epigenetic switch involving NF-κB, Lin28, Let-7 microRNA, and IL6 links inflammation to cell transformation. Cell 139, 693–706 (2009).
23. West, J.A. et al. A role for Lin28 in primordial germ-cell development and germ-cell malignancy. Nature 460, 909–913 (2009).
24. Viswanathan, S.R. et al. Lin28 promotes transformation and is associated with advanced human malignancies. Nat. Genet. 41, 843–848 (2009).
25. Martinez, N.J. & Gregory, R.I. MicroRNA gene regulatory pathways in the establishment and maintenance of ESC identity. Cell Stem Cell 7, 31–35 (2010).
26. Melton, C., Judson, R.L. & Blelloch, R. Opposing microRNA families regulate self-renewal in mouse embryonic stem cells. Nature 463, 621–626 (2010).
27. Yu, J. et al. Induced pluripotent stem cell lines derived from human somatic cells. Science 318, 1917–1920 (2007).
28. Neale, G. et al. Molecular characterization of the pediatric preclinical testing panel. Clin. Cancer Res. 14, 4572–4583 (2008).
29. Zhu, H. et al. The Lin28/let-7 axis regulates glucose metabolism. Cell 147, 81–94 (2011).
30. Yuan, J., Nguyen, C.K., Liu, X., Kanellopoulou, C. & Muljo, S.A. Lin28b reprograms adult bone marrow hematopoietic progenitors to mediate fetal-like lymphopoiesis. Science 335, 1195–1200 (2012).
31. Widén, E. et al. Distinct variants at LIN28B influence growth in height from birth to adulthood. Am. J. Hum. Genet. 86, 773–782 (2010).
32. Ong, K.K. et al. Genetic variation in LIN28B is associated with the timing of puberty. Nat. Genet. 41, 729–733 (2009).
33. Sulem, P. et al. Genome-wide association study identifies sequence variants on 6q21 associated with age at menarche. Nat. Genet. 41, 734–738 (2009).
34. Permuth-Wey, J. et al. LIN28B polymorphisms influence susceptibility to epithelial ovarian cancer. Cancer Res. 71, 3896–3903 (2011).
35. Bosse, K.R. et al. Common variation at BARD1 results in the expression of an oncogenic isoform that influences neuroblastoma susceptibility and oncogenicity. Cancer Res. 72, 2068–2078 (2012).
36. Pruim, R.J. et al. LocusZoom: regional visualization of genome-wide association scan results. Bioinformatics 26, 2336–2337 (2010).
37. Molenaar, J.J. et al. Copy number defects of G1-cell cycle genes in neuroblastoma are frequent and correlate with high expression of E2F target genes and a poor prognosis. Genes Chromosomes Cancer 51, 10–19 (2012).
©20
12 N
atu
re A
mer
ica,
Inc.
All
rig
hts
res
erve
d.
Nature GeNetics doi:10.1038/ng.2387
oNLINE METHodsSample selection. Cases were defined as subjects who were diagnosed with neuroblastoma or ganglioneuroblastoma and registered through COG. All specimens were obtained at time of diagnosis, and the majority were annotated with clinical and genomic information that included age at diagnosis, site of origin, INSS disease stage38, International Neuroblastoma Pathology Classification (INPC)39, MYCN oncogene copy number40, DNA index (ploidy)41, registration in clinical trial(s), eventfree and overall survival and second malignancies.
Control subjects were recruited from the Philadelphia region through the CHOP Health Care Network, including four primarycare clinics and several group practices and outpatient practices that included wellchild visits. Eligibility criteria for control subjects were (i) selfreporting as Caucasian; (ii) availability of 1.5 µg of highquality DNA from peripheral blood mononuclear cells; and (iii) no serious underlying medical disorder, including cancer. The Research Ethics Board of the Children’s Hospital of Philadelphia (CHOP) approved the study, and written informed consent was obtained from all subjects by nursing and medical assistant staff under the direction of CHOP clinicians.
Genome-wide SNP genotyping. Genotyping for both discovery and replication phases was performed using the Illumina Infinium II HumanHap550 and Human Quad610 BeadChips according to methods detailed elsewhere11,42,43 and summarized here. DNA samples were surveyed for quality, both by optical density spectrophotometry and PicoGreen assay (Invitrogen). A total of 750 ng of DNA isolated from blood was used to genotype each sample according to the manufacturer’s guidelines. On day 1, genomic DNA was amplified 1,000–1,500fold. On day 2, amplified DNA was fragmented to ~300–600 bp in size and was precipitated and resuspended. Fragments were then hybridized to the BeadChip arrays. Singlebase extension (SBE) used a single probe sequence of approximately 50 bp designed to hybridize immediately adjacent to the SNP query site. Following targeted hybridization to the array, locusspecific primers for the arrayed SNP (attached to the beads) were extended with a single hapten labeled dideoxynucleotide in the SBE reaction. The haptens were subsequently detected by a multilayer immunohistochemical sandwich assay. The Illumina BeadArray Reader scanned each BeadChip at two wavelengths and created an image file. As BeadChip images were collected, intensity values were determined for all instances of each bead type, and data files were created that summarized intensity values for each bead type. These files consisted of intensity data that were loaded directly into the Illumina GenomeStudio genotype analysis software. Once normalization was complete, the clustering algorithm was run to evaluate cluster positions for each locus and assign individual genotypes. Only samples yielding a genotype call rate of ≥95% were considered for inclusion in this study.
Quality control and association testing for the discovery cohort. Overlap of the HumanHap550 v1, HumanHap550 v3 and Quad610 arrays. The discovery cohort consisted of individuals genotyped on the HumanHap550 v1, HumanHap550 v3 and the Human Quad610 arrays. Our analysis only considered markers shared by all three arrays. The HumanHap550 v1 array contains 555,175 markers, the HumanHap550 v3 array contains 561,288 markers and the Quad610 array contains 620,901 markers. Overall, 535,752 markers are shared by all three arrays and were analyzed in this study.
Low genotype call rate (<95%). Call rate was calculated on the basis of the number of ‘no call’ genotypes, with default genotyping calling algorithm implemented in Illumina GenomeStudio software. We did not consider any sample with a call rate of <95% for inclusion in this study. The call rate per individual was assessed by PLINK software and was confirmed to exceed 95% for all individuals, with an average genotyping rate of 99.85% across included individuals.
Inferring population structure for individuals of European ancestry. Multidimensional scaling (MDS), as implemented in PLINK software, was used to infer population structure in the neuroblastoma data set. Comparing selfidentified ancestry with MDSinferred ancestry confirmed the reliability of MDS in identifying genetically inferred individuals of European ancestry (Supplementary Fig. 2).
Detection and elimination of cryptic relatedness and duplicate genotypes. To detect cryptic relatedness and potential duplicate genotypes within our data, we applied a twostep procedure to calculate pairwise identitybydescent (IBD) estimates between all individual case and control subjects. First, we examined the initial MDS and retained only those individuals of inferred European ancestry with call rates greater than 95%. Second, we recalculated genomewide identitybystate (IBS) estimates and IBD estimates in the remaining individuals of European ancestry using PLINK software. This twostep procedure ensures that allele frequency differences between populations do not lead to biases in IBD estimation. We applied a stringent threshold for detecting cryptic relatedness: any pairs of subjects with IBD of >0.15 were flagged, and one individual was removed, so that only unrelated subject remained in the final association test. A total of 11 neuroblastoma cases were excluded.
Low call rate per marker (<99%). Markers with call rates of less than 99% were excluded from analysis. Call rates were calculated by PLINK software. A total of 49,130 markers were excluded from association analysis in this step.
Minor allele frequency (individuals of European ancestry). Markers with minor allele frequency (MAF) of less than 5% were excluded from our analysis. MAFs were calculated by PLINK software. A total of 50,159 markers were excluded from association analysis in this step.
Hardy-Weinberg equilibrium (individuals of European ancestry). Markers with HardyWeinberg equilibrium P values of less than 0.001 were excluded from analysis. A total of 11,953 markers were excluded from association analysis in this step.
Matching controls. On the basis of genomewide IBS estimates for all pairwise comparisons between case and control subjects, we identified two matched controls for each case to correct for the potential effects of population structure, as described previously9.
Final counts of subjects and markers passing quality control. Applying the quality control measures mentioned, we were left with 2,101 cases, 4,202 matching controls and 426,697 SNPs for association analysis.
Quality control and replication in the African-American cohort. The quality control procedure for the AfricanAmerican replication cohort is largely similar to those performed on the discovery cohort and is described in detail elsewhere12. The replication cohort after quality control consisted of a total of 365 cases and 2,491 control subjects who were genotyped on the Illumina HumanHap550 or Human610Quad BeadChips. Significant SNPs from the discovery effort (P < 1 × 10−4) at the 6q16 loci were tested for association with neuroblastoma by logistic regression, including admixture estimate as a covariate.
Replication in Italian cohort. Genotyping of the top SNPs at HACE1 (rs4336470) and LIN28B (rs17065417) was performed using TaqMan SNP genotyping assays (Life Technology) for 350 neuroblastoma cases and 780 controls. Twenty DNA samples were also genotyped by Sanger sequencing for a further validation.
Meta-analysis. Metaanalysis was performed using the inversevariance method within the METAL14 software package, and a fixedeffects model was assumed. The Cochran Q test was used to assess evidence of betweenstudy heterogeneity of effect sizes.
Genotype imputation. Imputation was performed with IMPUTE2 (ref. 44) using the worldwide 1000 Genomes Project Phase 1 interim data as reference (June 2011 release). Rather than preselecting a reference population, we elected to follow the approach of Howie et al.45 and use a multipopulation reference panel with IMPUTE2. Genotypes for markers located on chromosome 6 were extracted and mapped to the hg19 human reference genome using the LiftOver tool. Multipopulation haplotype data from the 1000 Genomes Project Phase 1 interim release (June 2011) were downloaded from the 1000 Genomes Project website (see URLs). IMPUTE2 was applied with default parameters and
©20
12 N
atu
re A
mer
ica,
Inc.
All
rig
hts
res
erve
d.
Nature GeNeticsdoi:10.1038/ng.2387
Ne = 20,000. Following imputation, SNPs with MAF of <1% and/or IMPUTE2–info quality score of <0.8 were removed. To correctly account for uncertainty in the data resulting from the imputation process, the remaining SNPs were tested for association with neuroblastoma using the frequentist association test under the additive model implemented in SNPTEST46.
Genome-wide mRNA expression profiling of neuroblastoma cell lines. Genomewide mRNA expression profiling in neuroblastoma cell lines was performed using the Illumina WG6 expression array according to the manufacturer’s specifications. Data were normalized using the average normalization method provided in Illumina GenomeStudio software. Twosided t tests were performed at the gene level to assess differential expression in cell lines. P < 0.05 was considered significant.
miRNA expression profiling of neuroblastoma cell lines. RNA from samples was labeled with oligonucleotides from the Illumina MicroRNA Expression Profiling Assay (version 1) and then hybridized to a universal Sentrix Array Matrix according to the manufacturer’s specifications. Average signal values were normalized (rank invariant) using Illumina GenomeStudio software.
LIN28B protein detection. Neuroblastoma cell lines were grown in T75 flasks under standard cell culture conditions. Wholecell lysates were extracted with 100 µl of cell extraction buffer (Invitrogen) containing protease inhibitors (Sigma) and phenylmethyl sulfonyl fluoride, were briefly sonicated and were rotated for 1 h at 4 °C. After 30 min of centrifugation at 4 °C, the supernatant was removed, and protein quantification was performed using the Bradford method. Lysates (100 µg) were separated on 4–12% BisTris gradient gels and were transferred to PVDF membranes. Membranes were washed and incubated with antibodies directed against LIN28B (Cell Signaling Technology, 4196) and actin (Santa Cruz Biotechnology, sc1616).
LIN28B knockdown and monitoring of cell growth. For routine maintenance, cells were grown in RPMI 1640 complete medium (Gibco, 22400) containing 10% FBS (Hyclone, SH 3007303), 1× antibiotic antimycotic (Gibco, 15240062) and 2 mM lglutamine (Gibco, 25030). On day 0, cells were seeded in triplicate into antibioticfree medium in 96well RTCES plates (ACEA). On day 1, using Dharmafect (Dharmacon, T200102), cells were transiently transfected with 50 nM of either a nontargeting negative control siRNA (Dharmacon, D008101005) or an siRNA directed against LIN28B (L028584010005). Realtime cell growth was monitored every 30 min for at least 96 h using the RTCES system, as previously described3,47. Data presented are representative of at least three independent experiments. To monitor efficiency of LIN28B knockdown, transfection was performed as described, and, 48 h later, RNA was isolated using the Qiagen mini extraction kit. Total RNA (200 ng) was primed with oligo(dT) and reverse transcribed using Superscript
II reverse transcriptase (Invitrogen). Quantitative RTPCR using TaqMan gene expression assays (ABI) was performed. To monitor LIN28B protein knockdown, protein was isolated 72 h after transfection. Protein blotting analysis was performed using antibodies directed against LIN28B and actin.
RT-PCR in primary neuroblastomas. TaqMan Gene expression assays for HACE1 (Hs00410879_m1) and LIN28B (Hs01013729_m1) were purchased through Applied Biosystems. Reactions were set up in duplicate, using 10 ng of cDNA in a 10µl reaction that contained 200 nM of probe, 900 nM of each amplification primer and 1× Realtime PCR Master Mix (Applied Biosystems). Standard curves were generated using serial dilutions of the neuroblastoma cell line Kelly. Samples were amplified on an Applied Biosystems 7900HT Sequence Detection System using standard cycling conditions, and data were collected and analyzed with SDS 2.3 software.
Survival analysis. Survival analyses were performed using the methods of Kaplan and Meier48. For overall survival, time was defined as the time from diagnosis until the time of death from disease or until the time of last contact if death did not occur. Individuals who were alive were censored at the time they were last known to be alive. Logrank P values < 0.05 were considered significant.
38. Brodeur, G.M. et al. Revisions of the international criteria for neuroblastoma diagnosis, staging, and response to treatment. J. Clin. Oncol. 11, 1466–1477 (1993).
39. Shimada, H. et al. The International Neuroblastoma Pathology Classification (Shimada) System. Cancer 86, 364–372 (1999).
40. Mathew, P. et al. Detection of MYCN gene amplification in neuroblastoma by fluorescence in situ hybridization: a Pediatric Oncology Group study. Neoplasia 3, 105–109 (2001).
41. Look, A.T. et al. Clinical relevance of tumor cell ploidy and N-myc gene amplification in childhood neuroblastoma: a Pediatric Oncology Group study. J. Clin. Oncol. 9, 581–591 (1991).
42. Gunderson, K.L., Steemers, F.J., Lee, G., Mendoza, L.G. & Chee, M.S. A genome-wide scalable SNP genotyping assay using microarray technology. Nat. Genet. 37, 549–554 (2005).
43. Steemers, F.J. et al. Whole-genome genotyping with the single-base extension assay. Nat. Methods 3, 31–33 (2006).
44. Howie, B.N., Donnelly, P. & Marchini, J. A flexible and accurate genotype imputation method for the next generation of genome-wide association studies. PLoS Genet. 5, e1000529 (2009).
45. Howie, B., Marchini, J. & Stephens, M. Genotype imputation with thousands of genomes. G3 (Bethesda) 1, 457–470 (2011).
46. Marchini, J., Howie, B., Myers, S., McVean, G. & Donnelly, P. A new multipoint method for genome-wide association studies by imputation of genotypes. Nat. Genet. 39, 906–913 (2007).
47. Cole, K.A. et al. A functional screen identifies miR-34a as a candidate neuroblastoma tumor suppressor gene. Mol. Cancer Res. 6, 735–742 (2008).
48. Kaplan, E.L. & Meier, P. Nonparametric estimation from incomplete observations. J. Am. Stat. Assn. 53, 457–481 (1958).