Fig. 5
From: Explaining compound activity predictions with a substructure-aware loss for graph neural networks
Protein targets with global direction improvements. Reported are the number of targets (y-axis) displaying a given improvement of the global direction metric \(g_{\text{dir}}\) using the proposed \(\mathcal {L_\mathrm {MSE+UCN}}\) loss compared to \({\mathcal {L}}_{MSE}\) (x-axis). Global direction improvements were binned into \(\le\)5%, between 5 and 10%, between 10 and 20%, and \(\ge\)20% thresholds. Colors indicate the loss function utilized during GNN training (\({\mathcal {L}}_{MSE}\), blue; \(\mathcal {L_\mathrm {MSE+UCN}}\), orange). A minimum threshold of 50% MCS was considered for this analysis