Fig. 9
From: Explaining compound activity predictions with a substructure-aware loss for graph neural networks
Mapping of feature attributions for visualizations after docking. Feature attribution values were mapped to two compound structures in the context of their binding receptors (PDB Ids. 2YDK and 1D3G). Attributions computed using Integrated Gradients (top row, PDB Id. 2YDK) and GradInput (bottom row, PDB Id. 1D3G), and using the \({\mathcal {L}}_{\text{MSE}}\) (left column) to \({\mathcal {L}}_{\mathrm {MSE+UCN}}\) (right column) losses, are reported