Performance of DECoNT when polishing calls from unseen CNV callers. DECoNT learns a different set of weights and a different model for each WES-based CNV caller. To show the cross-model performance, we used DECoNT to correct CNV calls made by tools other than the ones used for training. We try every pair combination. Tools being pointed by an arrow are call-generating tools (i.e., being corrected). Tools at source of the arrow are the tools that are used to train the DECoNT model. Green arrows indicate improvement, and red arrows indicates deterioration in the corresponding performance metric. For each polished tool, we used 90% of the calls made on 802 1000 Genomes Project samples for training and the remaining 10% of the calls for testing. This roughly corresponds to a test set size of 80 samples.
