[关键词]
[摘要]
目的 面向免疫相关药食同源候选组方,构建融合轻量知识图谱、检索增强生成(retrieval-augmented generation,RAG)和规则核验的智能辅助原型。两阶段计算评价分别用于分析方法组件的相对贡献及考察该方案在另一生成模型条件下与常用基线的差异。方法 知识资源包括106种目录物质、73首免疫相关文献组方和386条来源组方正向共现关系。阶段I使用统一基座模型和固定参数,在30个经数据来源隔离核查的案例中比较无检索、语义RAG、知识图谱辅助RAG(GraphRAG)、量化低秩适配(quantized low-rank adaptation,QLoRA)和完整知识增强方案,并将参考组成Jaccard相似度(指标代码M-P01)作为主指标。阶段II另行纳入45个比较案例,在统一通义千问(Qwen)生成条件下比较纯Qwen、常规语义RAG、Microsoft GraphRAG和本研究方案。2个阶段均进行案例层配对分析,报告bootstrap置信区间并作Holm多重比较校正;规则响应一致性则以20组异常输入评价。结果 阶段I的540次预定生成全部完成。按预设规则,评价案例与39份评价文本及110条图谱路径之间未发现案例级记录重合。B2-B1的M-P01均值差为0.078 9(95% CI: 0.045 6~0.113 6;Holm P=0.000 369),各来源簇情景的差异方向均为正,QLoRA则未显示独立显著增益。阶段II中,Q3的Jaccard均值为0.178 5;与常规语义RAG相比,均值差为0.063 4(95% CI: 0.022 5~0.104 9;Holm P=0.014 98)。其与纯Qwen、Microsoft GraphRAG的差异均未达统计学意义。阶段I规则核验改变了2/90次输出的组成,20组异常输入的核验结果均与预设响应一致。结论 以各阶段预设的语义RAG为对照,阶段I的知识图谱辅助RAG和阶段II的本研究知识增强方案均获得更高的候选组成Jaccard相似度,差异具有统计学意义。这一结论仅适用于当前数据、模型和计算终点,不代表免疫功效、临床有效性或剂量适宜性,也不构成对法规合规性或安全性的一般判断。
[Key word]
[Abstract]
Objective To construct an intelligent auxiliary prototype integrating a lightweight knowledge graph, retrieval-augmented generation, and rule checking for candidate formulations of medicinal and edible homology related to immunity. The two-stage computational evaluation was respectively used to analyze the relative contributions of the method components and to examine the differences between this scheme and the common baseline under another generative model condition. Methods Knowledge resources included 106 catalog-listed substances, 73 immune-related literature-derived formulas, and 386 co-occurrence relations derived from source formulas. Stage I used a common base model and fixed parameters to compare no retrieval, semantic RAG, knowledge-graph-assisted RAG (GraphRAG), quantized low-rank adaptation (QLoRA), and the complete knowledge-enhanced approach in 30 cases after data-source overlap screening. Reference-composition Jaccard similarity (M-P01) was the primary endpoint. Stage II compared Qwen alone, conventional semantic RAG, Microsoft GraphRAG, and the proposed approach in 45 additional comparison cases under matched Qwen generation conditions. Both stages used paired case-level analyses, bootstrap confidence intervals, and Holm multiple comparison adjustment; 20 abnormal inputs assessed rule-response concordance. Results All 540 prespecified stage-I generations were completed. Under the prespecified rules, no case-level record overlap was found between the evaluation cases and the 39 evaluation documents or 110 graph paths. The mean difference of M-P01 of B2-B1 was 0.078 9 (95% CI: 0.045 6-0.113 6; Holm-adjusted P = 0.000 369), with a positive direction across source-cluster scenarios; QLoRA showed no independent significant increment. In stage II, the mean Jaccard similarity for Q3 was 0.178 5. Compared with the conventional semantic RAG, the mean difference was 0.063 4 (95% CI: 0.022 5—0.104 9; Holm-adjusted P = 0.014 98). Differences from Qwen alone and Microsoft GraphRAG were not statistically significant. In stage I, rule checking changed the output composition in 2 of 90 generations, and all 20 abnormal inputs produced the prespecified responses. Conclusion Taking the prespecified semantic RAG comparator for each stage as a control, candidate-composition Jaccard similarity was significantly higher for knowledge-graph-assisted RAG in stage I and the proposed knowledge-enhanced approach in stage II, and the difference was statistically significant. These findings apply only to the current data, models, and computational endpoints; they do not establish immune effects, clinical efficacy, dose appropriateness, or general judgment about regulatory compliance or safety.
[中图分类号]
TP18;R285
[基金项目]
中国食品科学技术学会食品科技专项基金(2025-M08); 国家重点研发计划(2018YFC1706500)