Unlike previously described methods that use ‘smoothing’ techniques by borrowing information from signals of adjacent markers within each sample (4), our method borrows information on the GC content of the surrounding genomic region. The advantage of our approach is that each marker in the testing sample is independently adjusted regardless of neighboring markers in the same testing sample, so that the boundaries of a true CNV in a testing sample will not be affected by neighboring normal copy markers. In addition, unlike the ‘smoothing’ techniques that are trained on signal intensity only, our model is trained upon genome-wide GC distributions that are identical for each sample, effectively using more available information for improved model construction. Finally, Lowess regression is computationally intensive and may not scale up well for even a single chromosome with hundreds of thousands of markers. Unlike the smoothing approach, our method only uses a subset of autosome markers that are at least 1 Mb away from each other when building the regression model. While this simple approach circumvents the problem of dependency between neighboring markers, this also
