Fig. 2
From: Explaining compound activity predictions with a substructure-aware loss for graph neural networks
Schema of the proposed UCN loss. Two compounds sharing a scaffold are sampled from the training set, and their atom latent spaces computed via a forward pass of a GNN model. The uncommon latent nodes are used for the loss computation, targeting the activity difference between the compound pairs. In the illustrated example, the compound pair is composed by \(c_{i}\) and \(c_{j}\), with a large MCS and two substitution sites, highlighted in red for \(c_{i}\) and green for \(c_{j}\). Substituents (or decorations) differ for both compounds, and correspond to the uncommon nodes in the latent space