Fig. 8From: Extended study on atomic featurization in graph neural networks for molecular property predictionComparison of exploitation of atom type features between models trained with representation F (full) and representation A (only atom types). Models that have access to additional information ignore atom types more oftenBack to article page