Methodology and software to detect viral integration site hot-spots

Background: Modern gene therapy methods have limited control over where a therapeutic viral vector inserts into the host genome. Vector integration can activate local gene expression, which can cause cancer if the vector inserts near an oncogene. Viral integration hot-spots or 'common insertion sites' (CIS) are scrutinized to evaluate and predict patient safety. CIS are typically defined by a minimum density of insertions (such as 2-4 within a 30-100 kb region), which unfortunately depends on the total number of observed VIS. This is problematic for comparing hot-spot distributions across data sets and patients, where the VIS numbers may vary. Results: We develop two new methods for defining hot-spots that are relatively independent of data set size. Both methods operate on distributions of VIS across consecutive 1 Mb 'bins' of the genome. The first method 'z-threshold' tallies the number of VIS per bin, converts these counts to z-scores, and applies a threshold to define high density bins. The second method 'BCP' applies a Bayesian change-point model to the z-scores to define hot-spots. The novel hot-spot methods are compared with a conventional CIS method using simulated data sets and data sets from five published human studies, including the X-linked ALD (adrenoleukodystrophy), CGD (chronic granulomatous disease) and SCID-X1 (X-linked severe combined immunodeficiency) trials. The BCP analysis of the human X-linked ALD data for two patients separately (774 and 1627 VIS) and combined (2401 VIS) resulted in 5-6 hot-spots covering 0.17-0.251% of the genome and containing 5.56-7.74% of the total VIS. In comparison, the CIS analysis resulted in 12-110 hot-spots covering 0.018-0.246% of the genome and containing 5.81-22.7% of the VIS, corresponding to a greater number of hot-spots as the data set size increased. Our hot-spot methods enable one to evaluate the extent of VIS clustering, and formally compare data sets in terms of hot-spot overlap. Finally, we show that the BCP hot-spots from the repopulating samples coincide with greater gene and CpG island density than the median genome density. Conclusions: The z-threshold and BCP methods are useful for comparing hot-spot patterns across data sets of disparate sizes. The methodology and software provided here should enable one to study hot-spot conservation across a variety of VIS data sets and evaluate vector safety for gene therapy trials.