Fig 1.
Summary of the Siamese graph convolutional network-based transfer learning workflow (SGT).
Fig 2.
Data distribution of the drug datasets.
The first graph in the first row is an overview of the proportions of positive drug samples of the targets in datasets a Tox21, b MUV, c PCBA, and d Toxcast, and other graphs show in detail the distribution of positive samples in each dataset.
Fig 3.
A graphical representation of the network described in this article.
a Siamese graph convolutional neural network with shared weights, b graph convolution operation, c graph pooling operation, and d graph gathering operation in the network.
Fig 4.
The transfer learning workflow for data-poor targets.
Two stages are used to transfer model parameters from data-rich targets to data-poor targets in specific datasets.
Fig 5.
Performance comparison between our model and the baseline model in multitask classification and regression tasks.
The area under the curve (AUC) of the ROC curve of various models in the a Tox21 and b Freesolv datasets.
Fig 6.
Performance comparison between our model and the baseline model in multitask classification tasks.
The area under the curve (AUC) of the ROC curve of various models in the a PCBA and b MUV datasets.
Fig 7.
Performance comparison between our model and the baseline model in multitask classification tasks.
The area under the curve (AUC) of the ROC curve of various models in the a HIV and b Toxcast datasets.
Fig 8.
Performance comparison between our model and the baseline model in the regression task.
a RMSE and b MAE (kcal/mol) are provided to perfectly reflect the performance of the models in the regression task of the Freesolv dataset.
Fig 9.
The correlation between predictions and observations of our model on different tasks in tox21 data set (index split).
Fig 10.
The correlation between predictions and observations of our model on different tasks in tox21 data set (random split).
Fig 11.
The correlation between predictions and observations of our model on different tasks in tox21 data set (scaffold split).
Table 1.
The area under curve (AUC) of the ROC curve of various models in Tox21, ToxCast, MUV and HIV data sets.
Table 2.
The Performances in FreeSolv data set.
R2, RMSE and MAE (kcal/mol) are provided to reflect the performance of the models.
Fig 12.
An example diagram showing the correlation between atomic and molecular toxicity using similarity maps.
Red represents a higher correlation, and green represents a lower correlation.