基于传统药性理论的抗肿瘤中药预测潜力探索研究OA

Objective To evaluate the feasibility and upper information limit of machine learning models using traditional medicinal property features as input variables for predicting the antitumor activity of traditional Chinese medicine(TCM),and to identify core medicinal property features with high statistical contribution.Methods An annotated dataset of antitumor TCM was constructed by integrating information from Chinese Materia Medica,the SymMap database,and multi-platform literature searches.Five categories of features including four natures,five flavors,meridian tropism,toxicity,and efficacy were extracted and subjected to one-hot encoding.Within a nested cross-validation framework,eight models were trained and hyperparameter-optimized using Bayesian optimization algorithm with the treestructured parzen estimator(TPE).Generalization performance was further validated on an independent test set,and model interpretability was analyzed via the shapley additive explanations(SHAP)method.Results The mean area under the curve(AUC)values of the eight models ranged from 0.594 to 0.629.On the independent test set,the AUC of all models fell between 0.510 and 0.565.SHAP analysis revealed that"bitter flavor","liver meridian",and"pungent flavor"were the core medicinal property features that were robust across models.Conclusion Traditional medicinal property features exhibit a certain statistical predictive power for the antitumor activity of TCM,yet a distinct upper information limit exists.This study provides a methodological baseline and hypothetical reference for subsequent multimodal fusion research.