Prediction of Drug Sensitivity of HER2-Positive Breast Cancer Cell Line via Graph Neural Network
K. Kim, H.M. Ju, K.S. Kim, C.S. Kang, Chang‐Mo Kang, S.-K. Woo · 2023
In this study, the drug response prediction (DRP) model based on the graph neural network (GNN) was applied to predict the drug sensitivity of the HER2-positive breast cancer cell lines. GDSC dataset which contains the drug sensitivity of the enormous drug-cell line pairs was used for the training of the model. Drugs in the GDSC library was processed by RDKit software and transformed to the graph structure which composed of nodes, the elements of the molecules, and edges, the chemical bonds. The initial node embedding is set to the chemical properties of the each element. Multi-omics representation of the cancer cell lines were referred from the CCLE library and used as the input. The DRP model was trained with the 197,818 drug-cell line pairs. After training, the accuracy of the model was evaluated for the 7 kinds of HER2-positive breast cancer cell lines with the drugs within the training dataset. It is found that the Pearson correlation coefficient between the label and the predicted value was 0.83, while the R-squared was 0.68. The mean absolute error of log of IC50 was 1.1. In addition, we evaluated whether the HER2-positive cell lines shows similar tendency in drug sensitivity for the several drugs, and which drug is predicted to have high efficacy. It is found that docetaxel has highest efficacy among the drugs we selected. Besides, the boxplots of the predicted drug sensitivity for each drug shows that the HER2-positive breast cancer cell lines shows similar tendency. To sum up, we verified that the GNN-based drug response prediction model is capable of predicting the drug sensitivity of the HER2-positive breast cancer cell lines. We concluded that the GNN-based DRP model is appropriate to predict the efficacy of the unseen drugs according the subtype of the cancer cell lines. Validation of the sensitivity of the unseen drugs for the HER2-positive breast cancer cell lines which is not in the GDSC will be our future work.