Binary Classification of Drug Sensitivity Prediction Using Multimodal Deep Learning

Abstract

Abstract Precision medicine has revolutionized cancer treatment by tailoring cancer therapy to individual patients. The field of drug sensitivity prediction has witnessed significant growth, encompassing diverse contributions that range from multi-omics integration to novel computational network architectures. Many existing studies focused on predicting the continuous value of the half maximal inhibitory concentration (IC50), but few have focused on predicting the binary class of the response (sensitive or resistant). This study presents a Binary Multimodal Deep Learning classifier (BinaryMDL) trained on diverse data sources, including drug screening, gene expression, mutation, and clinical data, to predict binary drug responses. The training dataset comprises 166 drugs and 603 cell lines. Unlike prior studies, we addressed the issue of data imbalance by employing the class weight technique, which shows an improvement in the model’s consideration of the minority class. Furthermore, our model’s predictive ability improved when incorporating clinical data. The resulting model BinaryMDL showed a performance improvement compared to other studies, achieving an AUC of 0.890 and an auPR of 0.614. Our work advances binary drug sensitivity prediction, highlighting the potential of multimodal deep learning for precision cancer therapy.