Fig. 4From: Explaining compound activity predictions with a substructure-aware loss for graph neural networksPer-target comparison of global direction values. The two-dimensional kernel density plot shows the target-specific global direction values with \({\mathcal {L}}_{\texttt {MSE}}\) (x-axis) and \({\mathcal {L}}_{\texttt {MSE+UCN}}\) (y-axis) loss functions. The text-box reports the percentage of protein targets for which global direction (\(g_\text{dir}\)) was larger with \({\mathcal {L}}_{\texttt {MSE+UCN}}\) loss. Compound pairs considered at the minimum 50% MCS thresholdBack to article page