A novel scan statistics approach for clustering identification and comparison in binary genomic data

In biomedical research a relevant issue is to identify time intervals or portions of a n-dimensional support where a particular event of interest is more likely to occur than expected. Algorithms that require to specify a-priori number/dimension/length of clusters assumed for the data suffer from a high degree of arbitrariness whenever no precise information are available, and this may strongly affect final estimation on parameters. Within this framework, spatial scan-statistics have been proposed in the literature, representing a valid non-parametric alternative.

We adapt the so called Bernoulli-model scan statistic to the genomic field and we propose a multivariate extension, named Relative Scan Statistics, for the comparison of two series of Bernoulli r.v. defined over a common support, with the final goal of highlighting unshared event rate variations. Using a probabilistic approach based on success probability estimates and comparison (likelihood based), we can exploit an hypothesis testing procedure to identify clusters and relative clusters. Both the univariate and the novel multivariate extension of the scan statistic confirm previously published findings.

The method described in the paper represents a challenging application of scan statistics framework to problem related to genomic data. From a biological perspective, these tools offer the possibility to clinicians and researcher to improve their knowledge on viral vectors integrations process, allowing to focus their attention to restricted over-targeted portion of the genome.