Background: DNA methylation classifiers (“episignatures”) help to determine the pathogenicity of variants of uncertain significance (VUS). However, their sensitivity is limited due to their training on unambiguous cases with strong-effect variants so that the classification of variants with reduced effect size or in mosaic state may fail. Moreover, episignature evaluation of mosaics as a function of their degree of mosaicism has not been developed so far. Results: We improved episignatures with respect to three categories. Applying (i) a minimum-redundancy-maximum-relevance feature selection we reduced their length by up to one order of magnitude without loss of accuracy. Performing (ii) repeated retraining of a support vector machine classifier by step-wise inclusion of cases in the training set that reached probability scores larger than 0.5, we increased the sensitivity of the episignature classifiers by 30%. In the newly diagnosed patients we confirmed the association between DNA methylation aberration and age at onset of KMT2B-deficient dystonia. Moreover, we found evidence for allelic series in KMT2B down to moderate effect variants associated with comparatively mild phenotypes such as late-onset focal dystonia. Retrained classifiers also can detect mosaics that previously remained below the 0.5-threshold, as we showed for KMT2D-associated Kabuki syndrome. Conversely, episignature classifiers are able to revoke erroneous exome calls of mosaicism, as we demonstrated by (iii) comparing presumed mosaic cases with a distribution of artificial in silico mosaics that represented all the possible variation in degree of mosaicism, variant read sampling, and methylation analysis.