Predicting Herb Pairs against Triple-negative Breast Cancer by Integrating Graph Neural Network and Multiscale Interactome

초록

Triple-negative breast cancer (TNBC) is an aggressive subtype lacking targeted therapies, leading to high relapse rates and poor prognoses. Here, we aimed to identify a synergistic herb pair against TNBC by integrating graph neural networks (GNNs) and a multiscale interactome. We curated TNBC-related biomarkers and constructed an herb-compound-target-disease network by integrating multiple data sources. Using this dataset, we trained and evaluated three GNN architectures-graph convolutional network (GCN), graph attention network, and graph sample-and-aggregate (GraphSAGE)-on 6830 herb pairs annotated with compound and target information. We then applied a biased random walk algorithm to estimate the network effect of herb targets and TNBC-related proteins, identifying new herbal candidates with potential synergy. Among the tested GNNs, GraphSAGE showed the highest performance in distinguishing known versus unknown herb pairs, with significant accuracy gains (p < 0.001). We subsequently performed diffusion profile analysis on top-ranked herbal combinations, revealing key TNBC targets, such as AKT1 and TP53. This multiscale approach illuminated potential synergistic effects within herbal therapies for TNBC. Our findings demonstrate that integrating GNN-driven deep learning with network pharmacology can systematically uncover multi-target herbal therapies for TNBC. Moreover, the molecular network we present can guide the design of materials for the rapid screening of herb-target interactions, aligning this work with emerging sensing technologies.

키워드