多分子指纹特征融合预测中药成分-靶点相互作用
本研究提出MMFi-DPBML深度学习框架,融合四种分子指纹与20种分子描述符预测中药成分-靶点相互作用。在自建TCMTS数据集上AUC达0.9873,加权F1为0.9860,优于现有模型;黄芩案例经分子对接验证。
要点
- MMFi-DPBML在TCMTS数据集AUC为0.9873,加权F1为0.9860。
- 模型融合MACCS、ECFP4、Torsion和2D药效团四种指纹及20种描述符。
- 黄芩案例预测的多个相互作用获分子对接强结合亲和力支持。
核心结论
- MMFi-DPBML能有效整合结构、空间与相互作用特征,准确预测中药成分-靶点相互作用,具药物发现计算工具潜力。
研究设计与样本
- 构建自定义TCMTS数据集,并在内部变体及公开基准Davis、KIBA上评估模型。
- 采用DPCNN处理分子指纹与描述符,BiLSTM编码蛋白序列,MLP整合特征空间进行预测。
主要发现
- 消融实验证实各组件有效,完整模型在TCMTS上AUC为0.9873,加权F1为0.9860。
- 对比实验显示MMFi-DPBML优于现有ITI模型,在多个数据集上展现稳健泛化能力。
- 黄芩案例中多个预测相互作用获分子对接强结合亲和力及TCMSP注释佐证。
关键数值与推荐
- TCMTS数据集上AUC为0.9873,加权F1为0.9860。
- 融合MACCS、ECFP4、Torsion、2D Pharmacophore四种指纹及20种分子描述符。
临床怎么用
- 可作为计算工具辅助识别中药复杂天然产物中的活性成分及其靶点。
- 预测结果需经分子对接等实验方法进一步验证。
局限性
- 未来工作需增强模型可解释性并支持多靶点药理学建模。
以上解读依据论文摘要(全文未开放获取或许可不允许翻译转载)。
原文
RESEARCH PURPOSE: This study proposes MMFi-DPBML, a deep learning framework that in-tegrates multi-molecular fingerprint features for predicting ingredient-target interactions (ITIs) in traditional Chinese medicine (TCM). By capturing di-verse structural and physicochemical features, the model aims to enhance prediction accuracy and support TCM modernization through data-driven approaches. METHOD: MMFi-DPBML fuses four types of molecular fingerprints-MACCS, ECFP4, Torsion, and 2D Pharmacophore-alongside 20 molecular descrip-tors, processed through a Deep Pyramid Convolutional Neural Network (DPCNN). Protein targets are represented by amino acid sequences and encoded using a stacked BiLSTM. A multilayer perceptron (MLP) inte-grates both feature spaces for final interaction prediction. A custom dataset (TCMTS) was constructed, and the model was evaluated against both inter-nal variants and public benchmarks (Davis, KIBA). RESULTS: Ablation studies confirmed the effectiveness of each component, with the full model achieving an AUC of 0.9873 and a weighted F1 score of 0.9860 on the TCMTS dataset. Comparative experiments demonstrated MMFi- DPBML's superior performance over state-of-the-art ITI models, showing robust generalization across diverse datasets. A case study on Scutellaria baicalensis further validated model predictions through molecular docking, with several predicted interactions corroborated by strong binding affinities and TCMSP annotations. CONCLUSION: MMFi-DPBML effectively integrates structural, spatial, and interaction-relevant features to accurately predict compound-target interactions in TCM. Its high performance and generalization capability underscore its po-tential as a computational tool for drug discovery, especially in identifying bioactive ingredients within complex natural compounds. Future work will focus on enhancing interpretability and supporting multi-target pharmaco-logical modeling.
引用
- Wang X, Li R, Zou Y, et al. MMFi-DPBML: Multi-molecular fingerprint feature fusion for predicting ingredient-target interactions in traditional Chinese medicine. J Ethnopharmacol. 2025. PMID: 40840725. doi:10.1016/j.jep.2025.120451
供执业医师工作参考,不替代临床判断;数值与分级以指南原文为准。