Genomic analysis of tumors by array comparative genomic hybridization: more is better.

We read with interest the article by Baldwin et al. (1), which describes the use of an array of 32,433 bacterial artificial chromosome clones tiled across the genome (2, 3) to map genomic aberrations in 20 oral squamous cell carcinoma samples by array comparative genomic hybridization (array CGH). The authors conclude that improved detection sensitivity and resolution of tiling arrays is necessary for analysis of tumor genomes. No one would argue against the benefits of higher-resolution copy number measurements. Indeed, regions of aberration have been narrowed using arrays, providing high-density regional coverage (4, 5), and direct comparison shows that tiling arrays provide more information on aberrations that oftentimes are missed or detected by only single clones on lower-resolution arrays. Nevertheless, comparison with the study of Snijders et al. (6), which used bacterial artificial chromosome arrays with megabase resolution to analyze genomic aberrations in 89 oral squamous cell carcinoma samples, shows that another important contributor to the biological utility of an experiment is the inclusion of sufficient numbers of samples distributed over the breadth of tumor subtypes. Several examples indicate the interplay of array resolution and sample cohort in affecting biological conclusions. Snijders et al. (6) showed that oral squamous cell carcinomas differ widely in frequency of copy number aberrations, with hierarchical clustering based on genomic aberrations indicating at least two subtypes. Thus, the distribution of patients among these groups will affect overall estimates of aberration frequencies and is more likely to explain the difference in frequency of EGFR amplification reported by the two groups. Differences in the classification of loci as being ‘‘amplified’’ may also have contributed, but the statement by Baldwin et al. (1) that the higher frequency of EGFR amplification in their data set is evidence of increased detection sensitivity is unlikely to be the cause, because the array used by Snijders et al. contained a specific clone for the EGFR locus. Thus, none of these amplifications could be missed by the lower-resolution array. [Baldwin et al. also incorrectly cite the number of such amplifications found by Snijders et al. (6) as 4 of 89 instead of 10 of 89.] In addition, in at least one case, the analysis of more tumors by Snijders et al. (6) provided greater biologically relevant genomic resolution than Baldwin et al., narrowing the minimal amplified region at 11q22.3 to 0.8 Mb (excluding the matrix metalloproteinases distal to MMP7) compared with the 1.24 Mb interval reported by Baldwin et al.(1). On the other hand, the tiling array provided greater refinement of the CDK6 amplicon. We note that regardless of resolution, it may be necessary to validate some results by other means, particularly those involving single clones. For example, Baldwin et al. (1) reported that the gain of a single clone RP11-338N5 at 7p12.3-p13 (TENS1) is present in 60% of cases, potentially indicating the significance of this gene in oral cancer. Nevertheless, whereas Baldwin et al. (1) map RP11338N5 to 7p, the National Center for Biotechnology Information (NCBI) database places it on 3p, and we have it located at 8q24.3 on our tiling array (3). In our experience, ratio changes reported by this clone agree with location at 8q24.3. This position would also be consistent with the high frequency of gain (60%) reported by Baldwin et al. (1) because +8q is common in oral cancer. Clearly, measuring the largest possible number of tumors on the highest-resolution arrays will provide the greatest information. Given the range of aberration sizes in cancer, properly annotated primary data of any resolution are valuable if readily accessible and will allow meta-analyses using consistent analytic criteria to be done. Similarly, a number of groups are carrying out array CGH using the set of tiling array clones assembled by Kryzwinski et al. and access to data will facilitate community annotation of copy number variants, as well as updating mapping information. Snijders et al. published their data; unfortunately, Baldwin et al. did not. We urge the editors of Cancer Research to make database submission (e.g., ArrayExpress, NCBI GEO) of primary data and associated patient characteristics a requirement for publication of genomic data.