Fig. 4: Performance of the RareDDIE in drug synergy prediction based on transfer learning versus de novo learning.

a The ROC (Receiver Operating Characteristic) curves in drug synergy prediction based on three different settings: 1-shot w/o transfer, 1-shot w/ transfer, and 10-shot w/ transfer. 1-shot w/o transfer: directly training RareDDIE using the training set of the drug synergy dataset. 1-shot w/ transfer: using the pre-trained RareDDIE on the DDIE (Drug-Drug Interaction Event) dataset, followed by fine-tuning on the drug synergy test set in a 1-shot setting. 10-shot w/ transfer: using the pre-trained RareDDIE on the DDIE dataset, followed by fine-tuning on the drug synergy test set in a 10-shot setting. Each experiment is conducted three times, with a distinct set of randomly selected support samples used for training and prediction in each iteration. The shading represents the standard deviation. b The PR (Precision-Recall) curves for drug synergy prediction based on three different settings show the performance of 1-shot w/o transfer, 1-shot w/ transfer, and 10-shot w/ transfer. Each experiment is conducted three times, with a distinct set of randomly selected support samples used for training and prediction in each iteration. The shading represents the standard deviation. c The visualization of the embedding of the drug pair Topotecan with BEZ-235 alongside the embedding of its corresponding 8 cell lines. The model predicts synergy by assessing the distance between these embeddings. d The AUC (Area Under the Curve) on the training set during the training process for both transfer learning and de novo learning models. We recorded the average AUC values every 50 iterations over 12,000 iterations. This approach helps smooth out fluctuations caused by variations in data from different tasks in each iteration. e The AUC on the validation set during the training process for both transfer learning and de novo learning models. f The loss on the training set during the training process for both transfer learning and de novo learning models. Source data are provided as a Source Data File.