Highly sensitive method for genomewide detection of allelic composition in nonpaired, primary tumor specimens by use of affymetrix single-nucleotide-polymorphism genotyping microarrays

Abstract

Loss of heterozygosity (LOH), either with or without accompanying copy-number loss, is a cardinal feature of cancer genomes that is tightly linked to cancer development. However, detection of LOH is frequently hampered by the presence of normal cell components within tumor specimens and the limitation in availability of constitutive DNA. Here, we describe a simple but highly sensitive method for genomewide detection of allelic composition, based on the Affymetrix single-nucleotide-polymorphism genotyping microarray platform, without dependence on the availability of constitutive DNA. By sensing subtle distortions in allele-specific signals caused by allelic imbalance with the use of anonymous controls, sensitive detection of LOH is enabled with accurate determination of allele-specific copy numbers, even in the presence of up to 70%-80% normal cell contamination. The performance of the new algorithm, called "AsCNAR" (allele-specific copy-number analysis using anonymous references), was demonstrated by detecting the copy-number neutral LOH, or uniparental disomy (UPD), in a large number of acute leukemia samples. We next applied this technique to detection of UPD involving the 9p arm in myeloproliferative disorders (MPDs), which is tightly associated with a homozygous JAK2 mutation. It revealed an unexpectedly high frequency of 9p UPD that otherwise would have been undetected and also disclosed the existence of multiple subpopulations having distinct 9p UPD within the same MPD specimen. In conclusion, AsCNAR should substantially improve our ability to dissect the complexity of cancer genomes and should contribute to our understanding of the genetic basis of human cancers.

Figure D1.

Expected concordance rate of SNP calls between normal samples. In the AsCNAR algorithm, SNP-specific signals of each SNP in a tumor sample were compared with those in reference samples that had a SNP call identical to that of the tumor sample. The probability of finding such concordant SNPs between a given tumor sample and a set of references was estimated as the function of the number of reference samples, by use of genotyping data from the 96 normal individuals. To do this, the latter were first divided into a test set and a reference set, each consisting of 48 individuals. Then, for each individual from the test set, the number of those SNP loci was enumerated that were identical to one or more SNPs within i references randomly selected from the reference set (i=1,2,3,…,10). No-call SNPs in test samples were excluded from the enumerations. The concordance rates were expressed as the mean ± SD for the 48 test samples. The concordance rate was separately estimated for heterozygous (hetero call) SNPs and for all SNPs in 50K Xba and 50K Hind arrays.

Figure 1.

AsCN analysis with or without paired DNA. DNA from a lung cancer cell line (NCI-H2171) was mixed with DNA from an LCL (NCI-BL2171) established from the same patient at the indicated percentages and was analyzed with GeneChip 50K Xba SNP arrays. AsCNs, as well as total CNs, were analyzed using either the paired reference sample (NCI-BL2171) (upper panels, A–C) or samples from unrelated individuals simultaneously processed with the tumor samples (middle and lower panels, D–I). On each panel, the upper two graphs represent total CNs and their moving averages for the adjacent 10 SNPs, whereas moving averages of AsCNs for the adjacent 10 SNPs are shown below (red and green lines). Green and pink bars in the middle are heterozygous (hetero) calls and discordant SNP calls between the tumor and its paired reference, respectively. At the bottom of each panel, LOH regions inferred from AsCNAR (orange), SNP call–based LOH inference of CNAG (blue), dChip (purple), and PLASQ (light green) are depicted. Asterisks (*) indicate the loci at which total CNs were confirmed by FISH analysis (data not shown). The calibrations of CN graphs are linearly adjusted so that the mean CNs of null and single alleles should be 0 and 1, respectively.

Figure 2.

Sensitivity and specificity of LOH detection for intentionally mixed tumor samples. Sensitivity of detection of LOH with or without CN loss (A and B) in different algorithms were compared using a mixture of the tumor sample (NCI-H2171) and the paired LCL sample (NCI-BL2171). The results for all LOH regions are shown in panel C, and the specificities of LOH detection are depicted in panel D. For precise estimation of sensitivity and specificity, we defined the SNPs truly positive and negative for LOH as follows. The tumor sample and the paired LCL sample were genotyped on the array three times independently, and we considered only SNPs that showed the identical genotype in the three experiments. SNPs that were heterozygous in the paired LCL sample and were homozygous in the tumor sample were considered to be truly positive for LOH, and SNPs that were heterozygous both in the paired LCL sample and in the tumor sample were considered to be truly negative. Proportions of heterozygous SNP calls (%hetero-call) that remained in LOH regions of each sample are also shown in panels A–C.

Figure 3.

The number of UPD regions for acute leukemia and MPD samples detected by either the SNP call–based method or AsCNAR. The number of UPD regions for ALL and AML samples detected by the SNP call–based method or by AsCNAR is shown in panel A, and the number of 9p UPDs for MPD samples detected by the two methods is shown in panel B. Some samples have more than one UPD region. Details of UPD regions are given in table 1. Significance between the numbers of UPDs detected by the SNP call–based method and by the AsCNAR method was tested by one-tailed binomial tests.

Figure 4.

Detection of AI in samples of primary AML and MPD. AsCN analyses disclosed the presence of a small population with 17p UPD in a primary AML specimen (W150673) (93% blasts in microscopic examination) with either a paired sample (A) or anonymous reference samples (B). The difference of the mean CNs of the two parental alleles is statistically different between panels A (0.38) and B (0.55) (P<.0001, by t test), which is explained by the residual tumor component within the bone marrow sample in complete remission (1% blast) used as a paired reference (W150673CR) (C). AI in the 9p arm was also sensitively detected in JAK2 mutation–positive MPD cases. UPD may be carried only by a very small population (∼20% estimated from the mean deviation of AsCNs in 9p) (IMF_10) (D), or by two discrete populations within the same case (PV_06), as indicated by two-phased dissociation of AsCN graphs (pink and green arrows) (F). AI in 9p is mainly caused by UPD but may be caused by gains of one parental allele without loss of the other allele (E), both of which are not discriminated by conventional allele measurements. Blue and pink bars are UPD and AI calls, respectively, from the HMM-based LOH detection algorithm. Other features are identical to those indicated in figure 1.

Figure 5.

Estimation of tumor populations carrying 9p UPD and the JAK2 mutation in MPD samples. The populations of 9p UPD–positive components in the 53 MPD cases were estimated by calculation of the mean difference of AsCNs within the UPD regions. Heterozygous (blue bars) or homozygous (red bars) JAK2 mutations in MPD samples were also estimated by measurement of JAK2 mutated alleles and UPD alleles, under the assumption that all the UPD alleles have a JAK2 mutation. Measurement of JAK2 mutated alleles was performed by allele-specific PCR. For three cases having trisomy components (orange bars), the duplicated allele was assumed to have a JAK2 mutation, which is the consistent interpretation of the observed fraction of trisomy and mutated JAK2 alleles for case PV_02 (table 2). mt = JAK2 mutated allele; wt = wild-type allele.

Figure 6.

Effects of the use of the different reference sets on signal-to-noise ratios (S/N) in CN analysis. The same DNA sample, containing 30% tumor (NCI-H2171) content, was analyzed on the 50K Xba SNP array in two different experiments by use of the identical reference set, including the paired LCL (NCI-BL2171). AsCN profiles obtained with the simultaneously processed reference set with the paired LCL (A and B) and with the anonymous references (E and F) show higher S/N than do those obtained with the same reference set but processed in different experiments by use of paired LCL (C and D) or anonymous references (G and H). S/N values are provided by the mean CN shift from baseline in the CN loss region divided by the SD of the diploid region.

Figure 7.

CN profile obtained with the use of a varying number of anonymous references. NCI-H2171 was analyzed with either one (A), three (B), or five (C) anonymous references, as well as its paired LCL (NCI-BL2171) (D) by use of the AsCNAR algorithm. Even though the expected concordance rate of SNP calls between the tumor and a single reference is only 37%, almost-equivalent CN profiles were obtained, regardless of the number of anonymous references used. Ch = chromosome.

Figure C1.

Inference of LOH on the basis of heterozygous SNP calls. A–C, The schema of determination of LOH blocks in inference of LOH. D, ROC curve for LOH determination. The sensitivity and specificity of LOH detection for pure tumor specimens were plotted for varying thresholds of Z(Ω) for LOH determination, where SNPs that were heterozygous in the paired LCL sample and were homozygous in the tumor sample were considered to be truly positive for LOH and SNPs that were heterozygous in both the paired LCL and tumor sample were considered to be truly negative.

Figure E1.

Sensitivity and specificity for determination of AI, LOH, and UPD. The sensitivity and specificity of detection of AI (A and B), LOH (i.e., decrease of the smaller allele in AI region) (C and D), and UPD (i.e., increase of the larger allele in LOH region) (E and F) in 10%, 20%, and 30% tumor (NCI-H2171) samples are plotted for varying threshold parameters of the HMM analysis. The sensitivity and specificity were calculated assuming that the analysis with the pure tumor and its paired LCL provides truly positive and truly negative results. In panel B, the specificity of the 10% tumor sample is <0.8. Asterisks (*) represent the thresholds used for determination of AI, LOH, and UPD.