基于异质图动态特征学习的药物重定位预测

doi:10.12122/j.issn.1673-4254.2026.02.23

南方医科大学学报 ›› 2026, Vol. 46 ›› Issue (2): 456-465.doi: 10.12122/j.issn.1673-4254.2026.02.23

• • 上一篇

基于异质图动态特征学习的药物重定位预测

朱昊坤¹(), 郭延哺¹^,²(), 辛向军¹(), 李朝阳¹, 周冬明³^,⁴

^1.郑州轻工业大学软件学院，河南郑州 450001
^2.东南大学江苏省网络群体智能重点实验室，江苏南京 211189
^3.湖南信息学院电子科学与工程学院，湖南长沙 410151
^4.云南大学信息学院，云南昆明 650500

收稿日期:2025-07-09 出版日期:2026-02-20 发布日期:2026-03-10
通讯作者: 郭延哺,辛向军 E-mail:332316040996@zzuli.edu.cn;guoyanbu@ zzuli.edu.cn;2004006@zzuli.edu.cn
作者简介:朱昊坤，在读硕士研究生，E-mail: 332316040996@zzuli.edu.cn
基金资助:
国家自然科学基金(62403437);国家自然科学基金(62272090);河南省科技攻关项目(242102211039);河南省科技攻关项目(252102210182)

Drug repositioning prediction based on dynamic feature learning on heterogeneous graphs

Haokun ZHU¹(), Yanbu GUO¹^,²(), Xiangjun XIN¹(), Chaoyang LI¹, Dongming ZHOU³^,⁴

^1.School of Software Engineering, Zhengzhou University of Light Industry, Zhengzhou 450001, China
^2.Jiangsu Provincial Key Laboratory of Networked Collective Intelligence, Southeast University, Nanjing 211189, China
^3.School of Electronic Science and Engineering, Hunan University of Information Technology, Changsha 410151, China
^4.School of Information Science and Engineering, Yunnan University, Kunming 650500, China

Received:2025-07-09 Online:2026-02-20 Published:2026-03-10
Contact: Yanbu GUO, Xiangjun XIN E-mail:332316040996@zzuli.edu.cn;guoyanbu@ zzuli.edu.cn;2004006@zzuli.edu.cn
Supported by:
National Natural Science Foundation of China(62403437)

摘要/Abstract

摘要：

目的针对现有人工智能方法在复杂异质生物网络建模中难以挖掘网络节点间的协同关系、提取高阶拓扑语义特征等问题，本文提出一种异质图动态特征学习的药物重定位预测方法。方法该方法首先构建融合药物、疾病及其交互关系的异质生物图模型。设计动态门控注意力模块，结合动态图注意力机制动态提取药物与疾病的判别性拓扑特征。设计门控残差特征融合机制，精准融合多源相似性网络中的结构和语义信息，有效缓解特征冗余与信息缺失的问题，实现药物与疾病关联的精准预测。结果在多个数据集上的实验和案例分析表明，本文药物重定位预测方法的性能优于现有主流模型。结论所提方法可有效建模异质生物网络中的复杂关联关系，提升药物重定位预测的准确性，为复杂疾病的精准治疗和医学人工智能提供重要的技术支持。

关键词: 复杂生物网络, 图神经网络, 门控机制, 药物重定位

Abstract:

Objective To address the challenges faced by existing artificial intelligence methods in modeling complex heterogeneous biological networks, particularly their limitations in capturing collaborative relationships between nodes and in extracting high-order topological semantic features, we propose a novel drug repositioning prediction method based on dynamic representation learning on heterogeneous graphs. Methods A heterogeneous biological graph that integrates drugs, diseases, and their interaction relationships was constructed, based on which a dynamic gated attention module was designed to extract discriminative topological features of drugs and diseases by incorporating a dynamic graph attention mechanism. A gated residual feature fusion mechanism was developed to precisely integrate structural and semantic information from multiple similarity networks to reduce feature redundancy and information loss, thereby enabling accurate prediction of drug-disease associations. Results Experiments and case studies conducted on multiple drug datasets related to complex diseases demonstrated that the proposed method outperformed existing mainstream models in drug repositioning prediction. Conclusion The proposed method can effectively model complex associations in heterogeneous biological networks, enhance the accuracy of drug repositioning prediction, and provide important technical support for precision treatment of complex diseases and development of medical artificial intelligence.

Key words: complex biological networks, graph neural networks, gating mechanism, drug repositioning

朱昊坤, 郭延哺, 辛向军, 李朝阳, 周冬明. 基于异质图动态特征学习的药物重定位预测[J]. 南方医科大学学报, 2026, 46(2): 456-465.

Haokun ZHU, Yanbu GUO, Xiangjun XIN, Chaoyang LI, Dongming ZHOU. Drug repositioning prediction based on dynamic feature learning on heterogeneous graphs[J]. Journal of Southern Medical University, 2026, 46(2): 456-465.

图/表 11

图1 异质图动态特征学习方法总体结构图,异质图动态特征学习方法主要由药物-疾病二分图构建、异质子图构建、动态门控注意力和药物重定位预测组成

Fig.1 Overall architecture of the DRLHG, consisting of construction of a drug-disease bipartite graph, heterogeneous subgraphs, dynamic gated attention, and drug repositioning prediction.

图2 异质子图提取过程, 对于药物中心节点u₂和疾病中心节点v₂, 提取其二阶邻居节点, 并组合为异质子图

Fig. 2 Heterogeneous subgraph extraction process. For the drug central node u₂ and the disease central node v₂, their second-order neighbor nodes are extracted and combined into a heterogeneous subgraph.

图3 门控机制示意图, 通过引入SELU建模复杂非线性关系, 并应用Sigmoid门控选择性过滤噪声, 以获得关键特征表示

Fig.3 Schematic diagram of the gating mechanism. The key feature representations are obtained by introducing SELU to model complex nonlinear relationships and applying a Sigmoid gate to selectively filter noise.

表1 药物-疾病关联数据集信息

Tab.1 Drug-disease association dataset information

Dataset	Drugs	Diseases	Drug-disease associations
Gdataset	593	313	1933
Cdataset	663	409	2352
LRSSL	763	681	3051

表2 DRLHG和其他基线模型在10折交叉验证中的性能

Tab.2 Performance comparison of DRLHG and the baseline models in 10-fold cross-validation

Model	Gdataset		Cdataset		LRSSL
Model	AUROC	AUPR	AUROC	AUPR	AUROC	AUPR
NMFDR	0.9452±0.0089	0.9478±0.0104	0.9511±0.0053	0.9567±0.0124	0.9336±0.0049	0.9442±0.0115
DPNetinfer	0.9120±0.0055	0.9218±0.0071	0.9197±0.0036	0.9307±0.0055	0.9230±0.0044	0.9342±0.0093
FuHLDR	0.9063±0.0067	0.8970±0.0126	0.9423±0.0104	0.9235±0.0103	0.9311±0.0113	0.9196±0.0093
DRWBNCF	0.9121±0.0132	0.9314±0.0239	0.9276±0.0141	0.9395±0.0149	0.9253±0.0218	0.9389±0.0150
DRAGNN	0.9442±0.0086	0.9504±0.0093	0.9479±0.0077	0.9522±0.0102	0.9504±0.0103	0.9595±0.0083
PSGCN	0.9484±0.0101	0.9551±0.0094	0.9561±0.0120	0.9606±0.0077	0.9405±0.0137	0.9412±0.0103
HDGAT	0.9478±0.0102	0.9575±0.0117	0.9524±0.0097	0.9613±0.0103	0.9493±0.0133	0.9581±0.0106
DRLHG	0.9551±0.0079	0.9606±0.0058	0.9610±0.0039	0.9675±0.0027	0.9557±0.0075	0.9615±0.0054

图4 模型深度1、2、3、4的参数选择

Fig. 4 Parameter selection for model depths 1, 2, 3, and 4.

图5 80%~100%稀疏度水平的参数选择

Fig.5 Parameter selection across sparsity levels 80% to 100%.

图6 DRLHG 与其变体DRLHG-GM、DRLHG-DGAM和DRLHG-RCM之间的性能比较

Fig. 6 Performance comparison between DRLHG and its variants DRLHG-GM, DRLHG-DGAM and DRLHG-RCM.

表3 DRLHG预测的用于阿尔茨海默病的前10种候选药物

Tab.3 Top 10 candidate drugs for Alzheimer's disease predicted by DRLHG

Disease name	Drug	Drug ID	Evidence
Alzheimer's disease	Benazepril	DB00542	^[28]
	Lisinopril	DB00722	^[29]
	Rosuvastatin	DB01098	^[30]
	Oxaprozin	DB00991	^[31]
	Nizatidine	DB00585	N
	Mesoridazine	DB00933	N
	Simvastatin	DB00641	^[32]
	Dextromethorphan	DB00514	^[33]
	Diazepam	DB00829	^[34]
	Sertraline	DB01104	^[35]

表4 DRLHG预测的用于肺癌的前10种候选药物

Tab.4 Top 10 candidate drugs for lung cancer predicted by DRLHG

Disease name	Drug	Drug ID	Evidence
Lung cancer	Diethylstilbestrol	DB00255	^[36]
	Rosuvastatin	DB01098	^[37]
	Ofloxacin	DB01165	^[38]
	Heparin	DB01109	^[39]
	Levothyroxine	DB00451	N
	Celecoxib	DB00482	^[40]
	Valproic acid	DB00313	^[41]
	Carvedilol	DB01136	^[42]
	Orciprenaline	DB00816	N
	Sulfasalazine	DB00795	^[43]

图7 尼扎替丁与左旋甲状腺素的化学结构式

Fig.7 Chemical structures of nizatidine (A) and levothyroxine (B).

参考文献 43

[1]	Ullah A, Meng YH. Finding influential nodes via graph embedding and hybrid centrality in complex networks[J]. Chaos Solitons Fractals, 2025, 194: 116151-8. doi：10.1016/j.chaos.2025.116151
[2]	Aung HN, Ohsaki H. Node embedding accelerates randoms walk on a graph[C]//2024 IEEE 48th Annual Computers, Software, and Applications Conference, 2024: 537-545. doi：10.1109/compsac61105.2024.00079
[3]	Chen H, Cheng F, Li J. iDrug: Integration of drug repositioning and drug-target prediction via cross-network embedding[J]. PLoS Comput Biol, 2020, 16(7): e1008040-9. doi：10.1371/journal.pcbi.1008040
[4]	Sadeghi S, Lu J, Ngom A. A network-based drug repurposing method via non-negative matrix factorization[J]. Bioinformatics, 2022, 38(5): 1369-77. doi：10.1093/bioinformatics/btab826
[5]	Song SP, Liu B, Teng F, et al. Self-supervised contrastive learning for implicit collaborative filtering[J]. Eng Appl Artif Intell, 2025, 139: 109563. doi：10.1016/j.engappai.2024.109563
[6]	Zeng XX, Zhu SY, Lu WQ, et al. Target identification among known drugs by deep learning from heterogeneous networks[J]. Chem Sci, 2020, 11(7): 1775-97. doi：10.1039/c9sc04336e
[7]	Meng Y, Wang Y, Xu J, et al. Drug repositioning based on weighted local information augmented graph neural network[J]. Brief Bioinform, 2023, 25(1): bbad431. doi：10.1093/bib/bbad431
[8]	Sun X, Wang B, Zhang J, et al. Partner-specific drug repositioning approach based on graph convolutional network[J]. IEEE J Biomed Health Inform, 2022, 26(11): 5757-65. doi：10.1109/jbhi.2022.3194891
[9]	吴雪扬, 张煜, 张华, 等.基于注意力机制和多模态特征融合的猕猴脑磁共振图像全脑分割[J].南方医科大学学报,2023, 43 (12): 2118-25. doi：10.12122/j.issn.1673-4254.2023.12.17
[10]	Jia X, Sun X, Wang K, et al. DRGCL: drug repositioning via semantic-enriched graph contrastive learning[J]. IEEE J Biomed Health Inform, 2025, 29(3): 1656-67. doi：10.1109/jbhi.2024.3372527
[11]	Peng L, Yang C, Yang J, et al. Drug repositioning via multi-view representation learning with heterogeneous graph neural network[J]. IEEE J Biomed Health Inform, 2025, 29(3): 1668-79. doi：10.1109/jbhi.2024.3434439
[12]	Zhang F, Hu W, Liu Y. GCMM: graph convolution network based on multimodal attention mechanism for drug repurposing[J]. BMC Bioinformatics, 2022, 23(1): 372-83. doi：10.1186/s12859-022-04911-8
[13]	Zhang M, Chen Y.Inductive matrix completion based on graph neural networks[J]. arXiv, 2019: 1904. 12058.
[14]	洪永, 张鑫, 林铭俊, 等.基于深度可分离卷积与注意力机制的单导联心房颤动轻量级分类网络[J].南方医科大学学报, 2025, 45 (3): 650-60.
[15]	Maas AL, Hannun AY, Ng AY. Rectifier nonlinearities improve neural network acoustic models[C].Proc. International Conference on Machine Learning, 2013: 3.
[16]	Klambauer G, Unterthiner T, Mayr A, et al.Self-normalizing neural networks[J]. arXiv, 2017: 1706. 02515. doi：10.1016/j.toxlet.2017.07.175
[17]	王永贵, 于琦.结合图同构和混合阶残差门控图神经网络的会话推荐[J].计算机科学与探索, 2025, 19 (2): 502-12.
[18]	Ying R, You J, Morris C, et al.Hierarchical graph representation learning with differentiable pooling[J]. arXiv, 2018: 1806.08804.
[19]	Barakat A, Bianchi P. Convergence and dynamical behavior of the ADAM algorithm for nonconvex stochastic optimization[J]. SIAM J Optim, 2021, 31(1): 244-74. doi：10.1137/19m1263443
[20]	Poernomo A, Kang DK. Biased dropout and crossmap dropout: learning towards effective dropout regularization in convolutional neural network[J]. Neural Netw, 2018, 104: 60-7. doi：10.1016/j.neunet.2018.03.016
[21]	Knox C, Wilson M, Klinger CM, et al. DrugBank 6.0: the DrugBank knowledgebase for 2024[J]. Nucleic Acids Res, 2024, 52(d1): D1265-75. doi：10.1093/nar/gkad976
[22]	Hamosh A, Amberger JS, Bocchini C, et al. Online mendelian inheritance in man (OMIM®): victor McKusick's magnum opus[J]. Am J Med Genet A, 2021, 185(11): 3259-65. doi：10.1002/ajmg.a.62407
[23]	Yang X, Yang G, Chu J. The neural metric factorization for computational drug repositioning[J]. IEEE/ACM Trans Comput Biol Bioinform, 2023, 20(1): 731-41. doi：10.1109/tcbb.2022.3144429
[24]	Wang J, Wu Z, Peng Y, et al. Pathway-based drug repurposing with DPNetinfer: a method to predict drug-pathway associations via network-based approaches[J]. J Chem Inf Model, 2021, 61(5): 2475-85. doi：10.1021/acs.jcim.1c00009
[25]	Zhao BW, Wang L, Hu PW, et al. Fusing higher and lower-order biological information for drug repositioning via graph representation learning[J]. IEEE Trans Emerg Topics Comput, 2024, 12(1): 163-76. doi：10.1109/tetc.2023.3239949
[26]	Meng Y, Lu C, Jin M, et al. A weighted bilinear neural collaborative filtering approach for drug repositioning[J]. Brief Bioinform, 2022, 23(2): bbab581. doi：10.1093/bib/bbab581
[27]	Huang SH, Wang MH, Zheng X, et al. Hierarchical and dynamic graph attention network for drug-disease association prediction[J]. IEEE J Biomed Health Inform, 2024, 28(4): 2416-27. doi：10.1109/jbhi.2024.3363080
[28]	Yang MH, Ho TC, Chang CC, et al. Utilizing proteomic approaches to uncover the neuroprotective effects of ACE inhibitors: implications for Alzheimer's disease treatment[J]. Molecules, 2023, 28(16): 5938. doi：10.3390/molecules28165938
[29]	Thomas J, Smith H, Smith CA, et al. The angiotensin-converting enzyme inhibitor lisinopril mitigates memory and motor deficits in a Drosophila model of Alzheimer's disease[J]. Pathophysiology, 2021, 28(2): 307-19. doi：10.3390/pathophysiology28020020
[30]	Modiri A, Abdolmaleki Z, Paryani MR. The effect of rosuvastatin coated by nano-chitosan on developing hippocampus: association with hippocampal neurogenesis and memory in an Alzheimer's induced model of rats[J]. Anat Cell Biol, 2025, 58(1): 61-75. doi：10.5115/acb.24.250
[31]	Willems S, Busch R, Nawa F, et al. Structural optimization of oxaprozin for selective inverse Nurr1 agonism[J]. J Med Chem, 2024, 67(15): 13324-48. doi：10.1021/acs.jmedchem.4c01218
[32]	Salari A, Roghani M, Khalili M. HMG-CoA reductase inhibitor simvastatin ameliorates trimethyltin neurotoxicity and cognitive impairment through reversal of Alzheimer's-associated markers[J]. Metab Brain Dis, 2024, 40(1): 74. doi：10.1007/s11011-024-01515-4
[33]	Majhi PK, Sayyad S, Gaur M, et al. Tinospora cordifolia extract enhances dextromethorphan bioavailability: implications for Alzheimer's disease[J]. ACS Omega, 2024, 9(22): 23634-48. doi：10.1021/acsomega.4c01219
[34]	Chen JW, Zhang M, Shen ZY, et al. Low-dose diazepam improves cognitive function in APP/PS1 mouse models: Involvement of AMPA receptors[J]. Brain Res, 2024, 1845: 149207. doi：10.1016/j.brainres.2024.149207
[35]	Wang H, Li S, Zhang J, et al. Efficacy of selective serotonin reuptake inhibitors-related antidepressants in Alzheimer's disease: a meta-analysis[J]. Eur J Med Res, 2024, 29(1): 438. doi：10.1186/s40001-024-02006-z
[36]	Seo Y, Jeong SB, Woo JH, et al. Diethylstilbestrol, a novel ANO1 inhibitor, exerts an anticancer effect on non-small cell lung cancer via inhibition of ANO1[J]. Int J Mol Sci, 2021, 22(13): 7100. doi：10.3390/ijms22137100
[37]	Aksoy HN, Ceylan C. Comparison of the effects of statins on A549 nonsmall-cell lung cancer cell line lipids using Fourier transform infrared spectroscopy: rosuvastatin stands out[J]. Lipids, 2021, 56(3): 289-99. doi：10.1002/lipd.12296
[38]	Schuette W, Nagel S, von Weikersthal LF, et al. Randomized phase III trial of docetaxel plus carboplatin with or without levofloxacin prophylaxis in elderly patients with advanced non-small cell lung cancer: the APRONTA trial[J]. J Thorac Oncol, 2011, 6(12): 2090-6. doi：10.1097/jto.0b013e3182307e3c
[39]	Bendahl PO, Belting M, Gezelius E. Longitudinal assessment of circulating tumor cells and outcome in small cell lung cancer: a sub-study of RASTEN-a randomized trial with low molecular weight heparin[J]. Cancers: Basel, 2023, 15(12): 3176. doi：10.3390/cancers15123176
[40]	Qadir A, Khalid Z, Kashan Theba F, et al. Celecoxib and bevacizumab synergistically inhibit non-small cell lung cancer by inducing apoptosis and modulating VEGF and MMP-9 expression[J]. Pak J Pharm Sci, 2023, 36(2): 501-6.
[41]	Hu M, Cheng H, Yang Y, et al. Valproic acid increased the efficacy of EGFR TKIs on EGFR/TP53 co-mutated lung cancers and downregulated mutant-p53 levels[J]. Mol Carcinog, 2024, 63(2): 275-85. doi：10.1002/mc.23651
[42]	Ikhmais BA, Hammad AM, Abusara OH, et al. Investigating carvedilol's repurposing for the treatment of non-small cell lung cancer via aldehyde dehydrogenase activity modulation in the presence of β-adrenergic agonists[J]. Curr Issues Mol Biol, 2023, 45(10): 7996-8012. doi：10.3390/cimb45100505
[43]	Bagherpoor AJ, Shameem M, Luo X, et al. Inhibition of lung adenocarcinoma by combinations of sulfasalazine (SAS) and disulfiram-copper (DSF-Cu) in cell line models and mice[J]. Carcinogenesis, 2023, 44(4): 291-303. doi：10.1093/carcin/bgad020

基于异质图动态特征学习的药物重定位预测

Drug repositioning prediction based on dynamic feature learning on heterogeneous graphs

RichHTML

PDF

可视化

摘要/Abstract

引用本文

使用本文

图/表 11

参考文献 43

相关文章 1

编辑推荐

Metrics

本文评价