BioXNet: a biologically inspired neural network for deciphering anti-cancer drug response in precision medicine

Abstract

Accurate prediction of anti-cancer drug responses in preclinical and clinical studies is crucial for drug discovery and personalized medicine. While machine learning models have demonstrated promising prediction accuracy in this task, their translational value in cancer therapy is constrained by the lack of model interpretability and insufficient patients data with genomic profiles to calibrate models. The rich cell line data has the potential to supplement patients data, but the difference between the drug response mechanisms in cell lines and human body needs to be characterized quantitatively. To address these challenges, we proposed the BioXNet, which captures drug response mechanisms by seamlessly integrating drug target information with genomic profiles (genetic and epigenetic modifications) into a single biologically inspired neural network. BioXNet exhibited superior performance in drug response prediction tasks in both preclinical and clinical settings. An analysis of BioXNets interpretability revealed its ability to identify significant differences in drug response mechanisms between cell lines and the human body. Notably, the key factor of drug response is the drug targeting genes in cell lines but methylation modifications in the human body. Furthermore, we developed an online human-readable interface of BioXNet for drug response exploration by medical professionals and laymen. BioXNet represents a step further towards unifying drug, cell line and patients data under a holistic interpretable machine learning framework for precision medicine in cancer therapy.