Multi-target meridians classification based on the topological structure of anti-cancer phytochemicals using deep learning

Abstract

Ethnopharmacological relevance: Traditional Chinese medicine (TCM) meridian is the key theoretical guidance of prescription against tumor in clinical practice. However, there is no scientific and systematic verification of therapeutic action of herbs under meridians context. Several studies have determined the Chinese herbal medicine (CHM) phytochemicals for intrinsic attribute or meridians classification based on artificial intelligence (AI) tools. However, it is challenging to represent the complex molecular structures with large heterogeneity through the current technologies. In addition, the multiple correspondence between herbs and meridians has not been paid much attention.

Aim of the study: We aim to develop an AI framework to classify multi-target meridians through the topological structure of phytochemicals.

Materials and methods: A total of 354 anti-cancer herbs, their corresponding TCM meridians and 5471 ingredient compounds were collected from public databases of CancerHSP, ETCM, and Hit 2.0. The statistical analysis of herbal and compound datasets, clustering analysis of the associated cancers, and correlational analysis of meridian tropism were preliminary conducted. Then a deep learning (DL) hybrid model named GRMC consisting of graph convolutional network (GCN) and recurrent neural network (RNN) was employed to generate the meridian multi-label sequences based on molecular graph.

Results: The curing herbs against tumors have tight relationships to lung, liver, stomach, and spleen meridians. These herbs behave different properties in curing certain cancer. Certain cancer types have co-occurrence such as ovarian, bladder and cervical cancer. Compounds have multitarget meridians with characteristics of higher-order correlations. Compared with the other state-of-the-art algorithms on the datasets and previous methods dealing with conventional fixed fingerprints of herbal compounds, the proposed GRMC has superior overall performance on testing dataset with the one error of 0.183, hamming loss of 0.112, mean averaged accuracy (MAA) of 0.855, mean averaged precision (MAP) of 0.891, mean averaged recall (MAR) of 0.812, and mean averaged F1 score (MAF) of 0.849.

Conclusions: The proposed method can predict multi-targeted meridians through neural graph features in herbal compounds and outperforms several comparison methods. It could provide a basis for understanding the molecular scientific evidence of TCM meridians.

Keywords: Chinese herbal medicine; Deep learning; Graph convolutional neural network; Meridians classification; Recurrent neural network; Traditional Chinese medicine.