Fig. 1

Schematic overview of SGAN method. a The SGAN method takes 23 prediction scores from available methods and 12 clinical evidence-based scores as input. Missing feature values were imputed with the mean of its 40 nearest neighboring variants. Discrete clinical evidence-based scores were converted into one-hot features adding Gaussian noise (mean is 0, standard deviation is 0.02) to make them continuous. Finally, the features were normalized by Minmax-scaling. b A total of 6498 labeled variants (1669 oncogenic and 4829 benign variants) and 60,000 unlabeled variants were used to train the model. The generative model generates fake (synthetic) samples by passing random noises into a linear model. The discriminative model distinguishes real samples from the fake, and classifies oncogenic variants and benign variants using labeled samples