Fig. 8
From: Explaining compound activity predictions with a substructure-aware loss for graph neural networks
Exemplary explanations for test set molecules. a Ground-truth feature attributions from the benchmark, b Integrated Gradients with MSE loss, and c with MSE+UCN loss results are reported with a coloring scheme. In the first two examples (PDB Ids. 1D3G, 1F0R), compounds had a single substitution site. The model trained with the simpler MSE loss failed to correctly capture the direction of the activity change (indicated by the ground-truth). The third and fourth examples (PDB Ids. 4XT9, 5CF4) constitute compounds from pairs that differed in multiple substitution sites. Feature attribution methods are also be applicable. Both the UCN and the simple MSE loss provide similar colors for all but one site