Section: New Results

A probabilistic method to characterize genomic alterations of tumors

Characterizing the genomic copy number alterations (CNA) in cancer is of major importance in order to develop personalized medicine. Single nucleotide polymorphism (SNP) arrays are still in use to measure CNA profiles. Among the methods for SNP-array analysis, the Genome Alteration Print (GAP) by Popova et al, based on a preliminary segmentation of SNP-array profiles, uses a deterministic approach to infer the absolute copy numbers profile. C. Keribin with Y. Liu, Y. Rozenholch and T. Popova developed a probabilistic model in [9] for GAP and define a Gaussian mixture model where centers are constrained to belong to a frame depending on unknown parameters such as the proportion of normal tissue. The estimation is performed using an expectation-maximization (EM) algorithm to recover the parameters characterizing the genomic alterations as well as the most probable copy number change of each segment and the unknown proportion of normal tissue. The tumor ploidy can be deduced from penalized model selection criterion. The model is tested on simulated and real data.

Surprisingly, the BIC selection criterion cannot recover the actual ploidy in the real data sets as slope heuristics do, even though all models are wrong. C. Keribin, in a discussion of S. Arlot's survey, gave some arguments to explain these behaviors [8].