Unsupervized identification of prognostic copy-numberalterations using segmentation and lasso regularization

Identifying copy-number alteration with prognostic impact is typically done in a supervised approach, were candidate regions are user-selected (chomosome arms, oncogenes, etc). Yet CNA events may range from whole chromosome alterations to small focal amplifications or deletions, with no available approach to combine the potential prognostic impact of different aberration ranges. We propose and compare different statistical models to integrate the effects of multi-scale CNA events by exploiting the longitudinal structure of the genome, and assume that the survival distribution follows a Cox-proportional hazard model. These methods are adaptable to any cohorts screened for CNA by genome-wide assays such as CGH-array or whole-genome sequencing technologies, and with sufficient follow-up time. We show that combining a segmentation in the survival odds strategy with a lasso-regularization selection approach provides the best results in terms of recovering the true significant CNA regions as well as predicting survival outcomes. In particular, as shown on a 551 Multiple Myeloma patient cohort, this method allows to refine previously identified regions to exhibit potential novel driver genes.