An algorithm for classifying tumors based on genomic aberrations and selecting representative tumor models

Xin Lu 0 Ke Zhang 0 Charles Van Sant 0 John Coon 0 Dimitri Semizarov 0 0 Global Pharmaceutical Research and Development , Abbott Laboratories, 100 Abbott Park Road, Building AP-10, Dep. R4CD, Abbott Park, IL 60064 , USA Background: Cancer is a heterogeneous disease caused by genomic aberrations and characterized by significant variability in clinical outcomes and response to therapies. Several subtypes of common cancers have been identified based on alterations of individual cancer genes, such as HER2, EGFR, and others. However, cancer is a complex disease driven by the interaction of multiple genes, so the copy number status of individual genes is not sufficient to define cancer subtypes and predict responses to treatments. A classification based on genome-wide copy number patterns would be better suited for this purpose. Method: To develop a more comprehensive cancer taxonomy based on genome-wide patterns of copy number abnormalities, we designed an unsupervised classification algorithm that identifies genomic subgroups of tumors. This algorithm is based on a modified genomic Non-negative Matrix Factorization (gNMF) algorithm and includes several additional components, namely a pilot hierarchical clustering procedure to determine the number of clusters, a multiple random initiation scheme, a new stop criterion for the core gNMF, as well as a 10-fold cross-validation stability test for quality assessment. Result: We applied our algorithm to identify genomic subgroups of three major cancer types: non-small cell lung carcinoma (NSCLC), colorectal cancer (CRC), and malignant melanoma. High-density SNP array datasets for patient tumors and established cell lines were used to define genomic subclasses of the diseases and identify cell lines representative of each genomic subtype. The algorithm was compared with several traditional clustering methods and showed improved performance. To validate our genomic taxonomy of NSCLC, we correlated the genomic classification with disease outcomes. Overall survival time and time to recurrence were shown to differ significantly between the genomic subtypes. Conclusions: We developed an algorithm for cancer classification based on genome-wide patterns of copy number aberrations and demonstrated its superiority to existing clustering methods. The algorithm was applied to define genomic subgroups of three cancer types and identify cell lines representative of these subgroups. Our data enabled the assembly of representative cell line panels for testing drug candidates. - Background Cancer is a disease of the genome that is characterized by substantial variability in the clinical course, outcome, and response to therapies. A key factor underlying this variability is the genomic heterogeneity of human tumors: individual tumors of the same histopathological subtype and anatomical origin typically carry different aberrations in their cellular DNA. Many of the most efficacious recent drugs target specific genetic aberrations rather than histological disease subtypes, for example trastuzumab and lapatinib for treating HER2-positive breast cancers [1], tamoxifen for treating ER-positive breast cancers[2,3], and gefitinib and erlotinib for non-small cell lung cancer with EGFR mutations [4-8]. Several subtypes of common cancers have been identified based on the aberrations of individual cancer genes, for example HER2-amplified breast cancer [1,9,10], EGFR-mutated and EGFR-amplified non-small-cell lung cancer [5,8], and others. However, cancer is a complex disease driven by the interaction of multiple genes and pathways [11,12]. Therefore, the copy number status of individual genes may not be sufficient to define cancer subtypes and predict the response to treatments. More comprehensive cancer taxonomy needs to be designed based on genome-wide patterns of DNA copy number abnormalities. Previous ground-breaking studies have reported molecular classifications for key cancer types based on their global patterns of gene expression [13-16]. As the high-density array technology became a reliable tool for copy number profiling, multiple gene copy number datasets were generated, revealing the genomic heterogeneity of key cancer types at the gene copy number level [17]. Various clustering methodologies have been applied to comparative genomic hybridization (CGH) data sets to classify cancers based on their copy number patterns and identify copy number aberration hotspots [17-23]. Taxonomies based on gene copy number have a number of advantages over gene expression-based classifications. In particular, copy number alterations are stable events, not affected by cell cycle or cytokine stimulation. Additionally, they show greater consistency between primary human tumors and cultured cell lines. Here we developed a copy number-based methodology for cancer classification in order to enable identification of genomic subgroups of major cancer types and facilitate rational selection of tumor models representative of individual subgroups. The methodology is based on the previously published genomic non-negative matrix factorization (gNMF) algorithm [23-26], with several major modifications to enhance the performance. We applied the algorithm to three major tumor types: nonsmall cell lung carcinoma (NSCLC), colorectal carcinoma (CRC), and malignant melanoma, identified distinct genomic subtypes for each cancer, and identified cell lines representative of each subtype. Our data enabled the assembly of representative cell line panels for testing drug candidates. The overall flow of our tumor classification methodology is illustrated in Fig. 1. After data pre-processing, a sample quality control procedure was applied to eliminate contaminated samples. For the remaining samples, a pilot hierarchical clustering was first applied to the segment smoothed tumor and cell line CGH data to determine the range of possible numbers of clusters, because the number of clusters needs to be fed into the gNMF algorithm, but is usually unknown for a given data set. The modified gNMF algorithm was then applied to the same set of segment smoothed CGH data to classify it into the initial numbers of clusters suggested by the hierarchical clustering. Using divergence as a stopping criterion and averaging results over multiple initiations, this modification significantly improved the accuracy of clustering at the cost of a higher computational complexity. To determine the best of the models built by gNMF algorithm with different numbers of clusters, we calculated the Cophenetic correlation coefficient and Bayesian Information Criterion (BIC) for these models, and then selected the one with the minimum BIC or the greatest decrease of Cophenetic correlation. In our study, the minimum BIC and greatest decrease of Cophenetic correlation often pointed to the same model. Finally, the 10fold stability test was performed on the selected model. Thus, the iteration procedure converges to the (...truncated)