Predicting Anticancer Drug Response With Interactive Graph Convolutional Networks Based On Multi-task Learnin

Cancer is a major disease that causes high mortality rates worldwide,and its treatment faces enormous challenges.Due to the genomic heterogeneity of cancer cells,the therapeutic effect of the same anti-cancer drug may vary significantly for different patients.Therefore,predicting the sensitivity and resistance of patients to drugs based on their genomic characteristics is important to achieve individualized treatment.Deep learning has shown great potential in the field of drug response prediction and is expected to provide more accurate tools for individualized treatment.Although there have been many machine learning-based methods for cancer drug response prediction,it remains a challenge to effectively integrate multiple information between cancer cell lines,drugs and their responses.In this paper,we aim to construct reliable anti-cancer drug response prediction models using deep learning methods,and accomplish the following two main tasks:(1)To address the heterogeneity between cell lines and drugs,this paper employs a neighborhood interaction(NI)-based heterogeneous graph convolution network method for end-to-end prediction of binary responses(sensitivity or resistance)or response concentrations(IC50 values)between cell lines and drugs.The method aggregates node-level neighbor features by graph convolution operations and considers element-level interactions with neighbors.The experimental results show that the method has significant advantages over state-of-the-art algorithms in predicting new cell lines and new drugs.In addition,a case study using two clinically approved drugs demonstrates that the model can detect drug responses in new cancer cell lines.(2)To better exploit the common training signal between the two tasks of binary response and response concentration,this paper employs a multi-task learning interactive graph convolutional network approach based on the previous method,introducing two auxiliary tasks in predicting binary responses: IC50 regression prediction and similarity network reconstruction.The method exploits the sharing of feature expression parameters between different tasks to enhance the generalization of the model on two tasks: cell line drug binary response and cell line drug response concentration,while maintaining the similarity structure in drug and cell line embedding.The experimental results show that the performance of predicting anticancer drug responses is significantly improved thanks to multi-task learning.In the dataset spanning experiments,the in vitro data can be well migrated to the in vivo patient data,demonstrating the potential of the approach employed in this paper to guide anti-cancer drug screening and personalized therapy.