目的 基于成分-基因双维度平行比较策略，结合机器学习与生物信息学分析中药寒热药性差异，为药性理论的现代阐释提供依据。方法 选取16味热性和18味寒性中药，在成分层面分析相对分子质量、辛醇-水分配系数、氢键供体数、氢键受体数及可旋转键数的全局分布与特异性成分；在基因层面构建中药-靶点二值化拓扑矩阵，采用基于同源药材隔离的组别留一交叉验证（leave-one-group-out cross-validation，Group-LOOCV）和置换检验评价分类模型，并结合χ²特征筛选、Shapley加性解释（Shapley additive explanations，SHAP）及功能富集分析解析关键靶点与相关通路。结果 寒热药成分的常规分子理化特征分布高度重叠。基因靶点分类模型中，脊回归分类器（ridge classifier，RidgeClass）的准确率为0.82，受试者工作特征曲线下面积（area under the receiver operating characteristic curve，AUC）为0.85，经检验P＜0.01。识别出再生家族成员1α（regenerating family member 1 alpha，REG1A）、核受体亚家族3 C组成员1（nuclear receptor subfamily 3 group C member 1，NR3C1）、分层蛋白（stratifin，SFN）及瞬时受体电位锚蛋白1（transient receptor potential ankyrin 1，TRPA1）等8个核心特征基因。热性药相关基因主要关联心血管疾病与细胞应激，寒性药相关基因主要关联肿瘤及铂类耐药，二者下游信号通路存在交叠。结论 在数据集及分析框架下，基因靶点网络特征较常规分子理化特征更有助于区分中药寒热药性，提示生物网络调控可能是药性差异的重要基础，为后续机制验证与药性研究提供线索。

Objective To investigate differences between cold and hot properties of traditional Chinese medicine (TCM) using a component-gene dual-dimensional parallel comparison strategy integrating machine learning and bioinformatics. Methods A total of 16 hot-property and 18 cold-property herbs were selected. At the component level, global distributions of molecular weight, octanol-water partition coefficient, hydrogen bond donor count, hydrogen bond acceptor count, and rotatable bond count, together with property-specific components, were analyzed. At the gene level, a binary herb-target topological matrix was constructed. Leave-one-group-out cross-validation (Group-LOOCV) with isolation of homologous herbs and permutation testing were used to evaluate classification models. Chi-square feature selection, Shapley additive explanations (SHAP), and functional enrichment analysis were combined to interpret key targets and associated pathways. Results Conventional molecular physicochemical features of components from cold- and hot-property herbs showed substantial overlap. Among target-based classifiers, the ridge classifier (RidgeClass) achieved an accuracy of 0.82 and an area under the receiver operating characteristic curve (AUC) of 0.85, with an empirical P < 0.01. A total of eight core feature genes were identified, including regenerating family member 1 alpha (REG1A), nuclear receptor subfamily 3 group C member 1 (NR3C1), stratifin (SFN), and transient receptor potential ankyrin 1 (TRPA1). Genes associated with hot-property herbs were mainly related to cardiovascular diseases and cellular stress, whereas those associated with cold-property herbs were mainly related to tumors and platinum drug resistance. Their downstream signaling pathways overlapped. Conclusion Within the present dataset and analytical framework, gene-target network features were more informative for distinguishing cold and hot medicinal properties than conventional molecular physicochemical features. These findings suggest that biological network regulation may contribute to medicinal property differences and provide leads for subsequent mechanistic validation and research.

R285;Q811.4

苏州市青年科技人才托举工程（QRC2507）；南京中医药大学2025年临床教学专项研究课题（NZYJY2025-L-45）；第三批吴门医派中医专家学术经验继承工作湿热治病专项课题（YSC2026024）；姑苏卫生人才计划（紧缺型人才）（SZWSRC2025078）