Abstract
The advancement of traditional Chinese medicine (TCM) faces challenges, due to the absence of a deep understanding of TCM mechanism at the perspective of modern biomedical practices. This results in how TCM selects herbs to treat diseases or symptoms prevailingly rely on clinicals' experience or TCM ancient books, at least in part lacking scientific basis. Herein, we present a novel deep learning-based approach, named Negative-Correlation-based TCM Architecture for Reversal (NeCTAR), to optimize the generation and combination of TCM formulas for guiding empiric therapy, by which we could, to some degree, narrow the gap between TCM and modern biomedical science. Our approach builds on a hypothesis that pathway alterations may serve as a proxy for the corresponding physiological changes induced by a certain disease, and 'inverse-fit' those alterations would provide a feasible therapeutic strategy to treat the disease. We leveraged ribonucleic acid sequencing (RNA-seq) data with Gene Set Enrichment Analysis to establish herb-pathway associations, integrating these insights into a multilayer perceptron model that incorporates top-k sparse projection and pathway reconstruction loss to predict the most therapeutically promising herbal components. NeCTAR demonstrated high concordance with experimental data across various disease models, including fatty liver disease, type 2 diabetes mellitus, and premature ovarian failure. Notably, NeCTAR could equally apply to single cell RNA-seq data. Overall, our study put forwards a novel interpretive framework underlying TCM mechanisms using modern biomedical foundation, by which we could prioritize herbal components based on existing TCM formulas treating diseases.