![[Feedback]](/icons/banner/feedbackACT.gif){kind=icon}
![[For Subscribers]](/icons/banner/subscriberACT.gif){kind=icon}
![[Archive]](/icons/banner/archiveACT.gif){kind=icon}
![[Search]](/icons/banner/searchACT.gif){kind=icon}
![[Contents]](/icons/banner/tocACT.gif){kind=icon}
|
|
|
|
|
||
|
Published online before print
September 15, 2003, 10.1101/gr.1349003 Genome Res. 13:2291-2305, 2003 ©2003 by Cold Spring Harbor Laboratory Press; ISSN 1088-9051/03 $5.00
Letter Representational Oligonucleotide Microarray Analysis: A High-Resolution Method to Detect Genome Copy Number Variation1 Cold Spring Harbor Laboratory, Cold Spring Harbor, New York 11724, USA 2 Tularik Inc., Genomics Division, Greenlawn, New York 11740, USA 3 Department of Applied Math and Statistics, SUNY at Stony Brook, Stony Brook, New York 11794, USA 4 Memorial Sloan-Kettering Cancer Center, New York, New York 10021, USA
We have developed a methodology we call ROMA (representational oligonucleotide microarray analysis), for the detection of the genomic aberrations in cancer and normal humans. By arraying oligonucleotide probes designed from the human genome sequence, and hybridizing with "representations" from cancer and normal cells, we detect regions of the genome with altered "copy number." We achieve an average resolution of 30 kb throughout the genome, and resolutions as high as a probe every 15 kb are practical. We illustrate the characteristics of probes on the array and accuracy of measurements obtained using ROMA. Using this methodology, we identify variation between cancer and normal genomes, as well as between normal human genomes. In cancer genomes, we readily detect amplifications and large and small homozygous and hemizygous deletions. Between normal human genomes, we frequently detect large (100 kb to 1 Mb) deletions or duplications. Many of these changes encompass known genes. ROMA will assist in the discovery of genes and markers important in cancer, and the discovery of loci that may be important in inherited predispositions to disease.
Cancer is a disease caused, at least in part, by somatic and inherited mutations in genes called oncogenes and tumor suppressor genes. It is likely that we know only a minority of the critical genes that are commonly mutated in the major cancer types. The identification of these genes can lead to rational targets for chemotherapy. Moreover, in many cases, the knowledge of which genes have been mutated can predict the course of neoplasias, including their therapeutic vulnerabilities, if any. This knowledge is likely to become increasingly important as cancers, or suspected cancers, are detected at earlier and earlier stages.
Methods for finding cancer genes date back to the early 1980s, but general methods have only recently been developed. This problem is being addressed by a variety of evolving techniques, some capable of detecting the genetic losses and amplifications that often accompany the mutation of tumor suppressor genes or oncogenes, respectively. We describe here our success with ROMA (representational oligonucleotide microarray analysis), a technique that evolved from an earlier method, RDA (representational difference analysis; Lisitsyn et al. 1993
We developed RDA as one general approach to the cancer problem. RDA compares two genomes by subtractive hybridization. To apply RDA, the complexity of the two genomes must first be reduced so that hybridization can go nearly to completion. To achieve this, we use low-complexity representations, a PCR-based method (Lisitsyn et al. 1993
RDA has been successfully used to detect deletions and amplifications in tumors, and its use has led to the discovery of several candidate tumor suppressor genes and oncogenes (Li et al. 1997
Microarray analysis is a high-throughput method that has been widely used to profile gene expression in cancers (DeRisi et al. 1996
We previously demonstrated that complexity reduction of samples by representation improves signal-to-noise performance, and diminishes the amount of sample required for analysis, relative to other microarray hybridization methods (Lucito et al. 2000
Adopting microarrays of oligonucleotide probes solves these problems. Representations are based on amplification of short restriction endonuclease fragments, and hence are predictable from the nucleotide sequence of the genome. Therefore, with the publication of the rough draft of the human genome (Lander et al. 2001
There are many other advantages to oligonucleotide-based microarrays. Based on our experience with the earlier implementation of this method using fragment arrays, the quality and reproducibility of printed oligonucleotide arrays ("print format") are superior. Although there is a large initial capital outlay to purchase large sets of oligonucleotides, the printed arrays are very inexpensive per unit when costs are amortized, and laborious and expensive replication of an underlying collection is not required. Furthermore, "long" oligonucleotide probes can be synthesized directly on an array surface (photoprinted arrays), and we demonstrate herein the equivalence of the two formats. In the photoprint format, there is no underlying physical collection at all (Singh-Gasson et al. 1999 We show results from two array formats. The printed arrays are a format that is readily achievable. The regions that are represented on the array can be changed to suit the user. A whole-genome array can be printed with the desired resolution. Smaller ROMA arrays can be designed and printed to focus on specific regions of the genome if wished, the advantage being that less capital outlay would be required for a smaller set of oligonucleotides. Results from the second format used, photoprint array, were presented to demonstrate the power of high-resolution copy number analysis. In this paper, we demonstrate our system, illustrating results and analytical techniques, present high-resolution analysis of cancer genomes, and provide initial evidence for widespread copy number polymorphism in humans. We discuss applications of our method, compare our method to other methods for global genomic analysis, and outline likely future developments.
Overview This paper describes a complex procedure, observations, and methods of analysis that are highly interactive. We therefore give here an overview of our results to guide the reader through a sensible reading of this portion of the manuscript. The first section reviews the technique of representations, and in particular "depleted" representations. Next we describe the design and selection of probes selected to hybridize well to representations. We introduce the two array formats that we use. The third section illustrates how to use hybridization to depleted representations to validate the composition of an array design, and the fourth section illustrates the use of such hybridization data to characterize probes and model overall array performance. Next we view essentially raw data of tumor and normal genomes, using two very different array formats, and show that the data from both formats are highly comparable. In the next section, we demonstrate a new statistical approach to gene copy number analysis based on segmentation analysis, and apply the method to two cancer genomes. The clonal nature of the cancers appears evident, as does the highly turbulent nature of their genomic rearrangements. The concordance between copy number analysis and our mathematical model is re-examined. Then we look more closely at several genetic lesions detected by our arrays following our statistical processing. Several distinct types of lesions are illustrated, including large regions of amplification and very narrow regions of homozygous and hemizygous deletion. Different types of inferences that can be made by the method are demonstrated. In the final section, we find a surprising abundance of "normal" variation in copy number between two individuals, and illustrate the need to coordinate data about such variation with interpretation of cancer data.
Representations
In our present studies, we have limited ourselves to the use of representations made with BglII, an enzyme with a typical 6-bp recognition site. BglII is one of many restriction enzymes that satisfy these useful criteria: It is a robust enzyme; its cleavage site is not affected by CpG methylation; it leaves a four-base overhang; and its cleavage sites have a reasonably uniform distribution in the human genome. After cleavage with BglII, we ligate adapters, and use the resulting product as a template for a PCR reaction. Because PCR selects small fragments, BglII representations are made up of the short BglII fragments, generally smaller than 1.2 kb, and we estimate that there are For array characterization, we use "depleted" BglII representations. These are representations made according to the usual protocol, but prior to PCR (to selectively amplify small BglII fragments), the adaptor-ligated BglII fragments are cleaved with a second restriction endonuclease. Cleavage destroys the capacity of some fragments to be exponentially amplified. For example, a BglII representation-depleted by EcoRI would consist of all small BglII fragments of the genome that do not contain within them EcoRI sites. Depleted representations are used for probe validation and modeling performance because we can remove a known subset of fragments from the representation, and observe the consequence upon hybridization to those probes complementary to the depleted fragments. In all of the experiments described herein, we have used comparative hybridization of representations prepared in parallel. Our approach works best if the DNA from two samples being compared is prepared at the same time, from the same concentration of template, using the same protocols, reagents and thermal-cycler. This diminishes the "noise" created by variable yield upon PCR amplification.
Design and Selection of Probes, and Composition of Probes for Microarrays Formats Our probes are derived from the short BglII restriction endonuclease fragments that we predict to exist from analysis of the human genome sequence. We initially evaluated probes of length 30 through 70, using methods described in the next section. The signal-to-noise ratio was maximal for probes of 70 nt in length, and we chose that length as our standard.
We selected our probes to be as unique as possible within the human genome, and tried to minimize short homologies to all unrelated sequences. We devised algorithms by which we could annotate any sequence of the genome with its frequency of exact matches in the genome (Healy et al. 2003 We used two formats for constructing microarrays. In the first of these, the "print" format, we purchased nearly 10,000 oligonucleotides made with solid-phase chemistry, and printed them with quills on a glass surface. In the second format, "photoprint arrays," oligonucleotides were synthesized directly on a silica surface using laser-directed photochemistry by NimbleGen Systems Inc. The photoprint arrays were a gift of NimbleGen Systems Inc., and fabricated to our design. Many more probes can be synthesized per array with laser-directed photochemistry, and in these experiments our arrays contained 85,000 oligonucleotide probes.
The probe composition for the 85K set was determined by a combination of design and selection, as described below. Unlike oligonucleotide probes synthesized by standard phosphoramidite solid-phase chemistry, certain oligonucleotides synthesized by laser-directed photochemistry are made in poor yield. However, unlike probes synthesized by the solid-phase chemistry and then printed, the cost of testing a set of probes synthesized directly on a chip is no more than the cost of the chip itself. Therefore, we tested In both our 10K and 85K formats, probes are arrayed in a random order, to minimize the possibility that a geometric artifact during array hybridization will be incorrectly interpreted as a genomic lesion.
Validation of Printed Arrays With Depleted Representations
To illustrate this process with a 10K array, we show in Figure 1 results obtained with BglII representations depleted by HindIII. In Figure 1A, we graph the ratios of hybridization intensity of each probe along the Y-axis. (See Methods for a description of how we process raw scanned data. We perform no background subtraction, as that only increases noise.) Each experiment is performed in color reversal, and the geometric mean of ratios from the separate experiments is plotted. Probes that we predict to detect fragments in both the full and depleted representations, based on the published human sequence, are grouped on the left. There are
From the experiment shown in Figure 1A, we can infer that the promise of the method is largely fulfilled: The restriction profile of representational fragments is correctly predicted, the probes are correctly arrayed, and the probes detect the predicted fragment with acceptable signal intensity. To calculate the data shown in Figure 1A, each hybridization was performed in color reversal, and the geometric mean of ratios from the separate experiments was plotted. In Figure 1B, the agreement between the ratios of the color reversal experiments is graphed, as a log-log scatter plot, showing excellent correlation of the data regardless of the labeling choice.
Modeling Array Hybridization When, as here, there is significant variation in measurements, statistical methods need to be used for the most accurate interpretation of data. It is also often useful to construct a mathematical model that can simulate measurement. Moreover, a good model can help predict the limits of detection, and be of assistance in the design of experiments. In this section, we describe a mathematical model that fits the data, and in a later section we describe statistical methods for data analysis. The mathematical model is useful for individual probe characterization, a clearer interpretation of the data, and the sharpening of statistical tools.
There is always more than one way to model data, and various enhancements can be added, but for our arrays we have found that a simple equation and sampling technique creates a model with great predictive power. This model will be described in detail in a subsequent manuscript, but it is based on an equation for the intensity of the i-th probe in a given channel, I[i]:
 is a multiplicative system noise;  is an additive system noise that encompasses background hybridization; and  is the multiplicative noise created during parallel representation and labeling. By definition, both and have a mean of 1, and for a diploid genome, c[i] = 1. A[i] can be viewed as the "brightness" of the i-th probe, and is a major determinant of the signal-to-noise ratio. In principle, A[i] should depend on at least two factors: the proportionate amplification of the fragment complementary to the probe during representation; and the purity of the probe. For example, a probe that is complementary to a poorly amplified fragment will have a low A value. Conversely, a probe complementary to a well-amplified fragment should be "bright" and have a high signal-to-noise ratio. Similarly, a probe that is synthesized with poor yield will have a low intensity and a poor signal-to-noise ratio. Other factors may influence A, such as the secondary structure of the probe and its base composition.
In the actual data, the highest ratios are observed from the most intense probes (see Fig. 1C). According to the model, this is explained by a fairly constant nonspecific signal for most probes. That is,
The model makes additional predictions: First, actual ratios are linearly related to measured ratios, and second, the standard deviation of probe measurement is a strong function of ratio, being a minimum for ratios of unity. Using parameters derived from the experiments displayed in Figure 1, we illustrate these relationships in Figure 1D. We assume 15 sets of 600 probes with various copy numbers n/4, with n = 0-14, bracketed by 600 probes of diploid copy number (4/4) on either end, measured against a diploid genome (c[i] = 1), and measured in duplicate. Note that the mean measured ratio of a set of probes is a linear function of the "true" copy number, the number of gene copies per cell, and the mean measured ratio, RM, of a subset of probes reflects their true ratio, RT, by the following equation:
Views of Tumor Genomes at 10K and 85K Resolution
The samples from Figure 2A were derived by flow sorting the nuclei of a surgical biopsy into aneuploid and diploid fractions, and making representations from as few as 15,000 nuclei (  100 ng of DNA). We estimate that the aneuploid fraction has perhaps 10% contamination from diploid nuclei, whereas the diploid fraction is not expected to be completely normal. Nevertheless, highly interpretable data result. These data are in two formats: the 10K print format (Fig. 2A1,B1,C1) and an 85K photoprint format (Fig. 2A2,B2,C2). Unlike the 10K format, probes of the 85K format were also selected for performance, as described and justified in earlier sections. This selection procedure produces a slight bias, in that no probe from the 85K set will detect a small BglII fragment that is homozygously missing in J. Doe. The consequences of this bias can be seen in comparisons of the 10K print format with the 85K photoprint format. In results from the 10K print format, there are roughly equal numbers of extreme "singlets" above and below a copy number of 1 (most apparent in Fig. 2C1). In contrast to this, using the 85K format, more extreme singlets are below rather than above a copy number of 1 (Fig. 2C2). In Figure 2, A1, A2, B1, B2, C1, C2, increased copy number is indicated by a ratio above 1, and decreased copy number by a ratio below 1. Even at this global view, with all probes displayed, several interesting observations can be made. There are clearly profiles to the cancer genomes, large regions of amplification, some quite high, and large regions of deletion (Fig. 2A,B). The profiles of the cancer genomes are varied. In contrast, the profile of the normal-normal appears to be flat, although some features can be seen. These will be examined more closely below. There are, in all three genomes, many stand-alone probes detecting minor losses and gains, which we attribute to heterozygous BglII polymorphism. These are manifest in the normal-normal comparison (Fig. 2C2) as a "shell" of probes that approach ratios of 0.5 and 2.0 throughout the genome. In contrast, in the tumor-normal comparison, wherein the normal is matched, there is only one stand-alone probe detecting major gains, and the stand-alone probes detecting major losses are more or less confined to extensive regions showing minor loss. This pattern is consistent with a hypothesis of allelic polymorphism and loss of heterozygosity (LOH). For a patient with heterozygosity at a BglII fragment, with a large and a small fragment, loss of the small allele will result in the virtual loss of specific signal because the large allele will not be abundant in the representation. This will present as an apparent major loss. On the other hand, a loss of the large allele, for example, by gene conversion, would at most result in a twofold increase in ratio, appearing as a minor gain. It is evident, looking at the results of the 10K print and the 85K photoprint formats in Figure 2, A1, A2, B1, B2, C1, C2, that the two systems capture a similar view of the larger genomic features. A correspondence between the two formats can be seen quantitatively. We call probes "brothers" if they share complementarity to the same BglII fragment. Brothers do not necessarily have overlapping sequence, or may be complementary across their entire length. In Figure 2, A3, B3, C3, we plot the ratios of brothers from one format to ratios of their brothers from the other format. There are in excess of 7000 brother probes. For all three experiments, in spite of the fact that the probe sequences differ between formats, the order of arraying is different, the hybridization conditions differ, and the surfaces of the array are different, there is remarkable concordance between the ratios of brother probes regardless of format.
Automated Segmentation and Whole-Genome Analysis
Figure 3 illustrates some of the output for the analysis of the cancer cell line SK-BR-3 at 85K resolution. We show four chromosomes, the highly turbulent Chromosome 8, a somewhat less active Chromosome 17, Chromosome 5, and the X-chromosome. The segmentation profiles and segment means for the 10K and 85K sets are very similar (data not shown), but clearly are not identical. More features are seen with the 85K set. In the next section, we inspect some of the data more closely. The full data, and that for the other two genomes, can be viewed at our Web site (http://roma.cshl.org/
Once segmented, we can assign to every probe the mean ratio of the segment to which it belongs, and then view the assigned mean ratios in sorted order. We do this for the two cancer genomes in Figure 4, A (CHTN159) and C (SK-BR-3). It is evident from the figure the segment mean ratios within each genome are quantized, with major and minor plateaus of similar value. In fact, it is likely that we can deduce the copy number by counting. As determined by flow analysis, the tumor is subtriploid, and the cell line is tetraploid. Assuming each sample is roughly monoclonal, then the two major plateaus in the tumor would be two and three copies per cell, and the major plateaus in the cell line are likely to be three and four copies per cell.
We can then use the copy number assumptions of the major plateaus to solve the ploidy and SN for each experiment. Our method is to use a version of equation 2 for each plateau. We select RM, the mean measured ratio, as the average of the probes of the segments in the plateau. We first set RT to CN/P, where CN is the "true" copy number. CN is the number of gene copies per cell, assumed to be known and equal for the plateau. P is the ploidy of the tumor genome. The result is two equations and two unknowns, with the unknowns being P and SN. For the tumor biopsy experiment (Fig. 4A), we calculate the ploidy P to be 2.60, and SN to be 1.13. For the cell line experiment (Fig. 4C), we calculate that P is 3.93, and SN is 1.21. We can then use equation 2 again to calculate what mean ratios would be expected for higher and lower copy numbers. These expectations are marked on the respective graphs, from zero to a copy number of 12, with horizontal lines forming a "copy number lattice." The assigned mean-segment values for probes are displayed in genome order, embedded with the expected copy number lattice (Fig. 4B,D). The copy number lattice fits remarkably well the minor plateaus of the data, especially for the higher copy numbers. However, there appears to be error in the expected ratios for probes detecting loss. The assigned mean-segment ratios of probes detecting loss cluster around values somewhat below the predicted values. In other words, the array appears to perform better for deletions than predicted based on the major plateaus and our present model. This deviation might be explained if we reexamine our assumption of clonality, and will be investigated further.
Specific Illustrative Examples The first example is a closer inspection of a region of a break in the X-chromosome, seen in Figure 3D. SK-BR-3, which derives from a female, has been compared to an unrelated male. The expectation is that probes in the X-chromosome will have elevated ratios. This is the case through much of the long arm of Chromosome X. In the midst of Xq13.3, over a region spanning 27 kb, there is a sharp break in copy number, and for the remainder of the chromosome, ratios near 1 are observed (Fig. 5A). This example demonstrates the boundaries that can be drawn from the array data by segmentation. In our data there are other examples of sharp copy number transitions that must break genes.
There are three to four narrow amplifications in SK-BR-3, each containing two or fewer genes, among which are transmembrane receptors. But broad amplifications can also be informative. The second example comes from the highly turbulent Chromosome 8 (see Fig. 3B). Despite the abundance of aberrations, we can clearly discern distinct regions of amplification. One such region is shown in Figure 5B. The rightmost peak is approximately a 1-Mb stretch, comprised of 37 probes (probe coordinates 45099-45138, June 2002 assembly, or NCBI build 30 genome coordinates 126815070-128207342). Yet it contains a single Ref-Seq gene, c-myc.
There is a second very broad peak in SK-BR-3, ascending to the left of the c-myc peak, and off the graph. This broad peak has a broad shoulder on its right (probe coordinates 44994-45051, June 2002 assembly, or NCBI build 30 genome coordinates 123976563-125564705), with a very narrow peak in its midst. We can overlay on this the segmentation data from the tumor genome, CHTN159, which has an even broader peak encompassing c-myc (probe coordinates 44996-45131, June 2002 assembly, or NCBI build 30 genomic coordinates 124073565-127828283). The peak in CHTN159 also encompasses the shoulder of the second SK-BR-3 peak (Fig. 5B). Thus, the shoulder may contain candidate oncogenes that merit attention. Within that region, at the narrow peak, we find TRC8, the target of a translocation implicated in hereditary renal carcinoma (Gemmill et al. 1998
We next show an example of a narrow deletion that highlights the need for high-resolution arrays, and also raises additional questions. The lesion occurs on Chromosome 5. In Figure 5C, we show a combined 10K (red) and 85K (blue) view. We do not show segmentation, but show the copy number lattice. A deletion is evident at both 10K and 85K resolutions (probe coordinates 26496-26540, June 2002 assembly, or NCBI build 30 genomic coordinates 14231414-15591226), one we judge to be hemizygous loss, but which may represent the presence of one copy in a tetraploid genome. The boundaries are much more clearly resolved at 85K. This region contains TRIO, a protein having a GEF, SH3, and serine threonine kinase domain (Lin et al. 2000); ANKH, a transmembrane protein (Nurnberg et al. 2001 It is also clear from the data that the lesion does not appear "neat." In the middle of the deletion are four or five probes that report ratios near 1. We can consider several explanations for this result. First, the hybridization to those probes may have failed for a variety of reasons. For example, the probes might not have been completely synthesized, or their complementary BglII fragments might not have amplified well. However, the intensities of these probes are in the middle range for all probe intensities, which diminishes the likelihood of this hypothesis. Second, the human assembly may be in error, and the outlier probes have been incorrectly posted at this location. Third, the deletion event may indeed be complex, the result of a localized genomic instability. Our last example is a region of homozygous loss (Fig. 5D). In this example, a cluster of zinc-finger proteins on Chromosome 19 is affected (probe coordinates 77142-77198, June 2002 assembly, or NCBI build 30 genomic coordinates 21893948-24955961). These genes, having zinc-finger domains, may encode transcription factors, whose deletion may have a role in tumorigenesis. There are an abundance of narrow hemizygous and homozygous lesions. These are seen both in the analysis of the cancer cell line and the cancer biopsy. However, as described below, we must take caution in their interpretation. Our next examples will all be in the context of normal-normal variation.
Examining Normal Genomic Variation When the tumor DNA cannot be matched against normal DNA, and an unrelated normal DNA is used as a reference, the differences observed may be the result of polymorphic variation. This variation can be of two sorts, the run-of-the-mill point sequence variation, of the sort that creates or destroys a BglII fragment, SNPs for example, or actual copy number fluctuation present in the human gene pool. The former is relatively harmless, as it will produce scattered noise that can largely be filtered by statistical means. We illustrate the application of a very mild filtration algorithm: If a ratio is the most deviant of the surrounding four, we replace it with the closer ratio of its two neighbors. In Figure 2C2, we showed a normal-normal comparison. The data look flat, with a cloud of scattered polymorphisms. In Figure 6A (combined 10K and 85K sets), we have applied filtration. The data no longer look so flat, and the cloud of scattered polymorphism is lifted, revealing nonrandom clusters of deviant probe ratios. These clusters reflect large-scale genomic differences between normal individuals, and we will say more of this presently.
Polymorphic variation of the scattered variety can also be filtered by serial comparison of experiments. We illustrate such a process in Figure 6B. In this figure, we display data from SK-BR-3 compared with normal donor J. Doe, the 85K ratios displayed in blue circles, and the 10K in red. On the same graph we display the ratios of J. Doe compared with another normal, DNA from an African pygmy, in green triangles. This is a fairly typical field of view. We see three probes of extreme ratio in the SK-BR-3-normal hybridization that can be identified as polymorphisms by comparison to hybridization between the two normal individuals. The simplest interpretation is that J. Doe is +/+, pygmy +/-, and SK-BR-3 -/-, where + designates the presence of a small BglII fragment and designates the absence of a fragment (most likely a SNP at a BglII site). In general, pairwise comparisons of three genomes allow interpretable calls of allele status. Hence, we suggest that when a malignant genome cannot be paired to a matched normal, or perhaps even when it can, such genomes should be compared with a single reference normal donor, whose allele status can be firmly established by extensive comparisons against other normals. Polymorphism in copy number, however, presents a different sort of problem. In this case, many probes within a region will show a deviation from a ratio of unity, and the pattern will appear coherent, not scattered. Statistical means will not suppress this signal. But do such variations commonly exist, and are they likely to be a source of misinterpretation if ignored? The perhaps surprising answer is emphatically, yes. Figure 6A indicates that there are gross regional differences in the normal-normal comparison. Indeed, many regions that display altered copy number between the two normal individuals are revealed upon segmentation analysis. Close inspections of two such regions are displayed in Figure 6, C and D, with ratios as connected blue dots and copy number lattice values in orange. In Figure 6C, the abnormal region is 135 kb on Chromosome 6p21 (probe coordinates 32518-32524, June 2002 assembly, or NCBI build 30 genomic coordinates 35669083-35804705), and encompasses three known genes. In Figure 6D, the region is a 620-kb region from Chromosome 2p11 (probe coordinates 9927-9952, June 2002 assembly, or NCBI build 30 genomic coordinates 88787694-89385815) that contains a number of heavy chain variable regions. We observe on the order of a dozen such regions in any normal-normal comparison. They range from 100 kb to >1 Mb in length and are more frequently observed near telomeres and centromeres, but can apparently occur anywhere. They often encompass known genes. We are presently investigating this phenomenon more fully, and will report on them subsequently. For now, we show how they impact the interpretation of cancer-normal data. In Figure 6, C and D, we have overlain the segmentation values from the analysis of SK-BR-3 in green. The copy number lattice for SK-BR-3 is plotted as orange lines. Figure 6C illustrates a region in SK-BR-3 that would be called a deletion in comparison to the normal. In SK-BR-3 compared to normal, the flanking region occurs at a copy number that we judge to be two copies per cell, and within that region, copy number becomes reduced to one. But the same region appears in the comparison of pygmy DNA to the same normal. In Figure 6D, we observe an analogous condition on Chromosome 2p11. In this panel, we have also plotted segmentation data from the tumor. This region is evidently abnormal there as well. Hence, we are inclined to view this "lesion" as pre-existing in the normal cells of the patient.
Comparison of Methodologies for Global Genomic Analysis We have described a method, representational oligonucleotide microarray analysis, or ROMA, that is useful for detecting amplifications and deletions and sites of breakage in cancer and normal genomes. Detection of these events can in principle be used to discover genes involved in cancer and other diseases of genetic origin, and serve as markers or guides for the diagnosis and treatment of such diseases. Because our method is sensitive to even single nucleotide polymorphisms at restriction endonuclease sites, it could in principle also be used as a high-density array for detecting SNPs.
There are other methods for global analysis of cancers. Most well known is the gene expression microarray (Chee et al. 1996
There are other DNA-based methods for measuring changes in copy number in cancers. The oldest of these is fluorescent in situ hybridization (FISH), which is used clinically to evaluate amplification at the ErbB-2 locus in breast cancer, for example (Tkachuk et al. 1990
Another DNA-based method is the BAC array, which is a method that is more commonly known, and more widely practiced, than our method (Pinkel et al. 1998
cDNA arrays have also been used for measuring copy number mutations (Pollack et al. 1999
Is Our Knowledge of Cancer Complete? There are simple tests for the completeness of our understanding and knowledge of how cancers survive in and kill their hosts. If our knowledge of the genes were complete, we would see a plateau in the number of common mutant genes found in all cancers. If our understanding of the principles were complete, even advanced cancers with a large number of accumulated genetic lesions would show only a small number of commonly affected pathways. It follows from this, that if mutation in a single gene were sufficient to affect a given pathway, even advanced cancers would show only a small number of commonly affected genes, the remainder of lesions being highly sporadic. The microarray-based method we have just described can partially address these issues. We can readily identify loci in the genome that undergo amplification, deletion, and imbalanced breaks. Although there are many other possible mechanisms that alter critical genes, such as point mutations, balanced translocations, and possibly stable epigenetic changes, many if not most oncogenes and tumor suppressor genes will eventually be found in the types of lesions that we can readily detect. Moreover, if a region is commonly found altered in cancers, that region harbors a good candidate cancer gene. Therefore, the application of our method to a large series of cancers, and the comprehensive comparative analysis of such data, should reveal the existence and number of candidate cancer genes in cancers.
Sources of Cancer Genomes
The direct analysis of tumor material offers many opportunities. There is a virtually unlimited source of different samples, and they can often be matched to the same normal, easing somewhat the analytical burden of interpretation. It is in principle possible to determine whether there are clinical parameters, such as survival and drug responsiveness, that correlate with specific gene amplification, deletions, and breakage, or overall patterns of genomic instability. These correlations may find utility in the treatment of patients. The disadvantages of tumor material are also clear. Tumors are always contaminated with stroma, can be oligoclonal, poorly preserved, and available in limiting amounts. Fortunately, our method seems to be highly sensitive, and does not require vast amounts of starting material. We routinely start from 50 ng of sample, which corresponds to
Technological Critique
Because of the success of the human genome sequencing project, and the reproducibility of representations, we are able to design oligonucleotide probes that are complementary to a given representation, such as the BglII representations that we have used here. Because the human genome sequence is very reliable, at least locally, we are able to experimentally validate our computationally derived designs by exploiting the known restriction endonuclease sites in our fragments (see Fig. 1). In principle, we can thus calibrate every probe's performance. The detection of these
Of the 8000 probes predicted to hybridize to fragments not cleaved by HindIII (see Fig. 1), There are many advantages to an oligonucleotide microarray format. The composition of the microarray is precisely formulated, and hence entirely reproducible by others. The work presented here demonstrates the equivalence of measurements achieved by the printed and light directed microarray formats. Using printed arrays we can achieve densities of 30,000 probes per slide, and using in situ light-directed synthesis, we have achieved densities of 190,000, although only 85K data are illustrated here. The latter technique has many advantages over the printed array. Besides achieving higher density, the layout of probes and the choice of probes are flexible. Although the unit cost of printed arrays is presently below the costs of light-directed microarrays, with the latter there is no need for a large initial capital expenditure for the purchase of oligonucleotides. Our method is dependent on representations. Without complexity reduction, which increases the concentration of DNA complementary to the probes, signal intensity from specific hybridization is too weak to measure above background. Dependence on representations is a mixed blessing. Representations use PCR both for the amplification of sample and complexity reduction. As a consequence, very little sample is required. However, PCR does introduce noise, and this requires that the test sample be compared with a control sample that is prepared exactly in parallel. We find that if the starting DNAs of test and control are of comparable quantity and quality, then subsequent parallel sample preparation, from PCR to labeling, is usually sufficient to give data of the type that is illustrated in this report.
There are a finite number of repeat-free 70-mer-long oligonucleotide probes in the genome that are useful for measuring BglII representations. We estimate that there are on the order of 120,000 of these scattered about the genome in a Poisson-like distribution, and the distribution of probes does not reflect the distribution of genes. At present we only array
Data Interpretation Our present segmentation algorithm requires a minimum of three probes to define a lesion, but clearly this is conservative. For example, when our tumor sample is compared with a matched normal, polymorphisms are controlled, and even a single probe with an elevated copy number in the tumor is likely to be meaningful. Other approaches to data analysis should be pursued, and we are attempting to integrate polymorphism data, probe calibration data, and probe intensity data into a more comprehensive model. Our present methods are not finished, but they are clearly already useful. We expect that the borders of regions can be drawn very sharply, most often to within a single probe, and this is confirmed in modeling experiments (data not shown). We will report on our progress in statistical methods in subsequent publications. In the end, however, no statistical interpretation of a single experiment is certain, and only the accumulation of larger data sets and molecular confirmation can increase confidence in a conclusion.
Normal Polymorphic Variation There is another type of "polymorphism" that we see, which for now we call "copy number" polymorphism. This type is much more interesting, and more pernicious, than scattered polymorphism, and it is documented in Figure 6. A series of regionally clustered probes may display a consistently altered ratio in the comparison of one normal sample against another. We see these regions in every normal-normal comparison that we have made, and many of these lesions appear in cancer-normal comparisons. In fact, some of these regions may be prone to genomic instability (see Fig. 6D). They vary in size from <100 kb to in excess of 1 Mb, and in most cases encompass genes. Creating a large database of normal-normal comparisons may mitigate the misinterpretation of these lesions as somatic events occurring in cancer, and this is something we intend to do. Our present hypothesis is that these normal-normal variations are in fact copy number polymorphisms, genetic in origin, but this is by no means proven here, nor is it the only plausible hypothesis. For example, these variant regions might be caused by locally high sequence divergence, or the consequence of highly altered chromatin structure, affecting the yield of DNA during purification from nuclei. Additional experimentation is needed to resolve these questions, and work in progress strongly indicates that the majority of these normal variations are, indeed, alterations in the gene pool. If there is, in fact, widespread copy number variation in humans, such variation might well contribute to human traits, including disease susceptibility and resistance.
Reagents Oligonucleotides were synthesized by Illumina Inc. Human Cot-1 DNA (15279-011) and yeast tRNA (15401-029) were supplied by Invitrogen Inc. Restriction enzymes, ligase, and Klenow fragments (M0212M) were supplied by New England Biolabs. The Megaprime labeling kit, Cy3-conjugated dCTP, and Cy5-conjugated dCTP were supplied by Amersham-Pharmacia. Taq polymerase was supplied by Eppendorf. Centricon YM-30 filters were supplied by Amicon (42410), and formamide was supplied by Amresco (0606-500). Phenol:chloroform was supplied by Sigma (P2069). NimbleGen photoprint arrays were a gift from NimbleGen Systems Inc.
Representation
Probe Selection
Printed Arrays
Labeling
Slide Preparation
Hybridization | ||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||
Top
ABSTRACT
RESULTS
