When performing VQSR, the data set has its variants overlapped with the training set, may I know if all the overlapped variants are used in the training or is it down sampled?
There isn't any downsampling of the training variants however there is some filtering that is done. By looking at the 1D annotation distributions, extreme outliers are removed from the training set. This behavior is controlled with the --stdThreshold argument to VariantRecalibrator.
I hope that helps,