Specific Genomic Regions Are Differentially Affected by Copy Number Alterations across Distinct Cancer Types, in Aggregated Cytogenetic Data

in Aggregated Cytogenetic Data. PLoS ONE 7(8): e43689. doi:10.1371/journal.pone.0043689 Specific Genomic Regions Are Differentially Affected by Copy Number Alterations across Distinct Cancer Types, in Aggregated Cytogenetic Data Nitin Kumar 0 Haoyang Cai 0 Christian von Mering 0 Michael Baudis 0 Patrick Tan, Duke-National University of Singapore Graduate Medical School, Singapore 0 1 Institute of Molecular Life Sciences, University of Zurich , Zurich , Switzerland , 2 Swiss Institute of Bioinformatics , Quartier Sorge, Lausanne , Switzerland Background: Regional genomic copy number alterations (CNA) are observed in the vast majority of cancers. Besides specifically targeting well-known, canonical oncogenes, CNAs may also play more subtle roles in terms of modulating genetic potential and broad gene expression patterns of developing tumors. Any significant differences in the overall CNA patterns between different cancer types may thus point towards specific biological mechanisms acting in those cancers. In addition, differences among CNA profiles may prove valuable for cancer classifications beyond existing annotation systems. Principal Findings: We have analyzed molecular-cytogenetic data from 25579 tumors samples, which were classified into 160 cancer types according to the International Classification of Disease (ICD) coding system. When correcting for differences in the overall CNA frequencies between cancer types, related cancers were often found to cluster together according to similarities in their CNA profiles. Based on a randomization approach, distance measures from the cluster dendrograms were used to identify those specific genomic regions that contributed significantly to this signal. This approach identified 43 non-neutral genomic regions whose propensity for the occurrence of copy number alterations varied with the type of cancer at hand. Only a subset of these identified loci overlapped with previously implied, highly recurrent (hot-spot) cytogenetic imbalance regions. Conclusions: Thus, for many genomic regions, a simple null-hypothesis of independence between cancer type and relative copy number alteration frequency can be rejected. Since a subset of these regions display relatively low overall CNA frequencies, they may point towards second-tier genomic targets that are adaptively relevant but not necessarily essential for cancer development. - . These authors contributed equally to this work. Genetic changes such as point mutations, regional copy number alterations/aberrations (CNA) and structural changes (e.g. gene fusion events) are all hallmarks of cancer. CNAs arise as somatic changes in the tumor cell genome through a variety of mechanisms and can be observed in virtually all types of cancer, to a varying extent. So far, the most widely used methods for the detection of CNAs have been chromosomal and array-based Comparative Genomic Hybridization (CGH) techniques [14]. Localized, recurring CNAs (hot-spots) have been shown to target canonical oncogenes (e.g. duplications/amplifications of the MYC, MYCN, REL loci) or tumor suppressor genes (e.g. deletions of the CDKN2A/B, TP53, ATM loci). Some regional CNAs such as gains on 8q and losses on 3p are present across multiple cancer types, whereas other imbalances may be largely restricted to a limited number of cancer entities [5]. Datasets integrated across multiple cancer types have previously been analyzed, to report regional hot-spots of frequent CNAs [5,6]. In a given set of individual tumor samples, the number and distribution of CNAs varies considerably [5] and this genetic heterogeneity has been used to detect and report co-occurring CNAs [7]. In principle, specific patterns and similarities in the individual and/or disease specific CNA profiles might point to distinct oncogenomic mechanisms acting in different cancer types and specimens, given a sufficiently large number of data points. Indeed, clustering of CNA patterns has been used to identify oncogenomic similarities [5,811]. The adaptation of clustering techniques to the analysis of CNA patterns has been subject of previous studies [1214]. With a few exceptions [5,14], however, sample-based clustering has been the main focus of such studies so far. In contrast, we here explore the clustering of cancer types, not of individual cancer samples. Both descriptive and clustering-based analyses of CNA across multiple cancer types suffer from a bias towards the more frequently occurring events. Due to the heterogeneity of the overall CNA signal, with greatly varying average frequencies of CNAs per cancer type (Figure 1a), clustering results may be distorted depending on the disease entities analyzed. This variation in overall CNA occurrence frequencies across cancer types may simply be owed to differences in the average time points of clinical detection or in different progression characteristics, and should be corrected for prior to clustering analyses. To the best of our knowledge, so far no implementation has been reported for a comprehensive, very large-scale clustering analysis of frequencynormalized cancer CNA profiles. Here, we focus on the identification of genomic regions that contribute meaningfully to the clustering of cancer types. From hereon we will refer to those as non-neutral regions. As the starting point of our analysis, we use hierarchical clustering to arrange cancer types on the basis of their CNA frequency profiles. We then employ a permutation approach to estimate the relative contribution of individual genomic regions to the quality of the clustering and to the derived relationship tree. The clustering quality is inferred from an intrinsic measure (summed branch lengths: tree height statistics), and genomic regions that reject the null hypothesis are termed non-neutral. Identified regions are compared to canonical CNA hot-spots (i.e. those that occur most frequently across the entire dataset). Our current analysis is based on data from a total of 25579 samples, which are classified into 160 different cancer entities (table S1) according to the International Classification of Disease in Oncology (ICD-O 3). Our approach is unique in that it a) focuses less on the clustering as such but more on the individual genomic regions that best support the clustering, b) uses an intrinsic quality measure coupled to a permutation strategy for validation, c) performs CNA frequency normalization prior to analysis, and d) is based on a very large data set, processed in a standardized setup. We aim for the identification of potential cancer-specific driver/modulator regions, which may not have been detected in earlier, largely hot-spot-focused approaches. All of the underlying cancer data is available through our Progenetix repository (www.progenetix.org; [15]). The average overall frequency of CNAs across the entire genome varies among different cancer types (Figure 1a). Since the relative weight of CNAs at individual genom (...truncated)