Fig. 3: Systematic evaluation of the ability of models based on different graph readout functions to learn from large-scale and complex datasets. | Nature Communications

Fig. 3: Systematic evaluation of the ability of models based on different graph readout functions to learn from large-scale and complex datasets.

From: Transfer learning with graph neural networks for improved molecular property prediction in the multi-fidelity setting

Fig. 3

A Train MAE for the PubChem low-fidelity models, with a model predicting the dataset mean for comparison. The MAE values are scaled to the range [0, 1] for each dataset. B The same but for the AstraZeneca and QMugs models. C 3D UMAP latent space visualisations for a selection of low-fidelity models using sum and neural readouts. Similar effects are observed for other datasets. The dataset sizes are: 1581928 (AZ-SD9), 1700745 (AZ-SD-1), 98472 (AID1949), 311910 (AID449762), and 647794 (QMugs after filtering). MAE denotes the mean absolute error and UMAP stands for uniform manifold approximation and projection. Source data are provided as a Source Data file.

Back to article page