Accuracy or novelty: what can we gain from target-specific machine-learning-based scoring functions in virtual screening? | Synapse