Finding regions of aberrant DNA copy number associated with tumor phenotype

DNA copy number alterations are a hallmark of cancer. Understanding their role in tumor progression can help improve diagnosis, prognosis and therapy selection for cancer patients and can contribute to the development of personalised therapies. High-resolution, genome-wide measurements of DNA copy number changes for large cohorts of tumors are currently available, owing to the rapid development of technologies like microarray-based array comparative hibridization (arrayCGH). In this manuscript, we introduce a computational pipeline for statistical analysis of tumor cohorts, which can help extract relevant patterns of copy number aberrations and infer their association with various phenotypical indicators. The pipeline makes use of machine learning techniques for classification and feature selection, with emphasis on interpretable models (linear models with penalties, tree-based models). The main challenges that our methods meet are the high dimensionality of the arrays compared to the small number of tumor samples available, as well as the large correlations between copy number estimates measured at neighboring genomic locations. Consequently, feature selection is unstable, depending strongly on the set of training samples, leading to un-reproducible signatures across different clinical studies. We also show that the feature ranking given by several widely-used methods for feature selection is biased due to the large correlations between features. In order to correct for the bias and instability of the feature ranking, we introduce a dimension reduction step in our pipeline, consisting of multivariate segmentation of the set of arrays. We present three algorithms for multivariate segmentation, which are based on indentifying recurrent DNA breakpoints or DNA regions of constant copy number profile. The multivariate segmentation constitutes the basis for computing a smaller set of super-features, by summarizing the DNA copy number within the segmentation regions. Using the super-features for supervised classification, we improve the interpretability and stability of the models, where the baseline for comparison consists of classification models trained on probe data. We validated the methods by training models for prediction of the phenotype of breast cancers and neuroblastoma tumors. We show that the multivariate segmentation step affords higher model stability and it does not decrease the accuracy of the prediction. We obtain substantial dimension reduction (up to 200-fold less predictors), which recommends the multivariate segmentation procedures not only for the purpose of phenotype prediction, but also as preprocessing step for downstream integration with other data types. The interpretability of the models is also improved, revealing important associations between copy number aberrations and phenotype. For example, we show that a very informative predictor that distinguishes between inflammatory and non-inflammatory breast cancers with ERBB2 amplification is the co-amplification of the genomic region located in the immediate vicinity of the ERBB2 gene locus. Therefore, we conclude that the size of the amplicon is associated with the cancer subtype, a hypothesis present elsewhere in the literature. In the case of neuroblastoma tumors, we show that patients belonging to different age subgroups are characterized by distinct copy number patterns, especially when the subgroups are defined as older or younger than 16-18 months. Indeed, considering a large set of age cutoffs, our prediction models are most accurate if the cutoff is around 16-18 months. We thereby confirm the recommendation for a higher age cutoff than 12 months