Fig. 2

Predictive performance of quality features and machine learning (ML). a One-feature predictions. Predictive performances of each feature (no machine learning) as areas under receiver operating characteristics curves (auROC) ranging from 0.5 (random predictions) to 1 (perfect predictions). b Multi-feature predictions. Predictive performances of optimal generic and specialized machine learning models trained using given feature sets (y-axis) outperform one-feature predictions. Bars show mean value and error bars show standard deviations derived from 10-fold cross-validations within the grid search. Feature sets: RAW (raw data), MAP (genome mapping), LOC (genomic localization), TSS (transcription start sites profile), ALL (all features)