Abstract: The development of traditional Chinese medicine formulas involves key steps including screening of Chinese medicine formulas, determination of effective ingredients, identification of action targets, druggability studies, preclinical research, and clinical research.  To address the problems of low formula screening efficiency, difficulty in experimental research, and high time costs in traditional Chinese medicine formula development, an intelligent model was constructed for formula indication prediction to improve formula screening efficiency. A Formula Symptom Prediction based on Graph Convolution Network (FSPGCN) was proposed for formula indication prediction using a multi-attention mechanism and graph convolution. Firstly, herb features including properties, flavors, meridian tropism, and toxicity, as well as formula features including dosage form and category, were quantified. Secondly, multiple herb features were fused through a multi-attention mechanism to obtain initial formula features for modeling formula composition relationships. Thirdly, an indication attention encoder was constructed based on the known formula–indication association matrix to initialize indication features and distinguish primary and secondary indication features. Finally, a formula–indication heterogeneous network was constructed, and graph convolution was used to learn formula–indication relationship features. The probability was calculated by the inner product of the learned final formula and indication features, enabling formula indication prediction. On the traditional Chinese medicine formula dataset, FSPGCN achieves the best performance, with Accuracy, Recall, Precision, and F1-score of 0.902, 0.903, 0.901, and 0.902, respectively; compared with the best baseline PSGCN, these metrics increase by 0.083, 0.093, 0.079, and 0.086.  The experimental results indicate that integrating formula characteristics through a multi-attention mechanism can effectively characterize formulas and enable formula indication prediction.

Key words: graph convolution, attention mechanism, prediction of prescription indications, screening of traditional Chinese medicine formulas, heterogeneous network

摘要： 中药方药研发流程包括中药方剂筛选、确定有效成分、明确作用靶点、成药性研究、临床前研究和临床研究的关键步骤。针对中药方药研发中存在的方剂筛选效率低、实验研究难度大、时间成本高等问题，构建智能模型实现方剂适应症预测，提高方剂筛选效率，提出基于多注意力图卷积的方剂适应症预测模型(Formula Symptom Prediction based on Graph Convolution Network, FSPGCN)。首先，量化中药性味归经和毒性特征，以及方剂剂型和类别特征，其次基于多注意力机制融合多个中药特征获得方剂初始化特征以建模方剂组成关系，再次基于已知的方剂-适应症的关联关系矩阵构建适应症注意力编码器初始化适应症特征，区分主次适应症特征。最后构建方剂-适应症异构网络并使用图卷积学习方剂-适应症关系特征，通过学习到的最终方剂和适应症特征进行内积计算概率，实现方剂适应症预测。在中药方剂数据集上，FSPGCN获得了最优模型性能，准确率(Accuracy)、召回率(Recall)、精确率(Precision)和F1分数(F1-score)分别为0.902、0.903、0.901和0.902，与最佳基线模型PSGCN相比，上述指标分别提升0.083、0.093、0.079和0.086。实验结果表明，通过多注意力机制融合中药方剂特性能有效表征方剂实现方剂适应症预测。

关键词: 图卷积, 注意力机制, 方剂适应症预测, 中药方剂筛选, 异构网络