Prediction and verification of therapeutic drugs for triple-negative breast cancer using a knowledge graph-based drug repurposing model

doi:10.12122/j.issn.1673-4254.2026.08.22

Abstract

Abstract:

Objective To construct a knowledge graph-based drug repurposing model for predicting potential therapeutic drugs for triple-negative breast cancer (TNBC). Methods Drug-target interaction (DTI) affinity data were collected from the BindingDB database and filtered (including data of Kd, Ki, EC₅₀ and IC₅₀). The proposed KGNN model integrates graph convolutional network (GCN)‑extracted drug molecular graph features, ProtBERT-pretrained protein sequence representations, and STRING-derived protein-protein interaction (PPI) knowledge graphs. Multi-head attention mechanisms and gated fusion modules were used to model interaction dependencies. Model performance was evaluated using mean squared error (MSE), Pearson correlation coefficient (PCC), and concordance index (CI). Ablation studies were performed to assess module contributions, and cold-start experiments were conducted to test generalization ability of the model. Using data from TCGA, 1340 TNBC-associated pathogenic genes were screened by bioinformatics analyses and mapped to targets using UniProt. KGNN was applied to predict the candidate drugs, which were validated through molecular docking and molecular dynamics simulations. Results In the DTI affinity prediction task, KGNN outperformed the benchmark models including KronRLS, SimBoost, DeepDTA, FusionDTA, and GraphDTA (MSE=3.2697, PCC=0.8037, and CI=0.7862). Ablation studies confirmed the critical roles of the modules for enhancing model performance (multi-head attention increased MSE by 5.60%; PPI fusion increased MSE by 10.82%). In cold-start scenarios, KGNN maintained superior performance over the comparators in unseen drug/target settings, demonstrating robust generalization. The TNBC candidate drug predictions well aligned with docking affinities and dynamics simulations (Pearson correlation coefficient>0.85), while attention visualization highlighted the efficacy hotspots. Conclusion The KGNN model can effectively predict drug-target interactions to facilitate drug repurposing and the design of multi-target drugs while reducing the screening space and experimental validation costs.

Key words: drug repurposing, drug-target binding affinity, knowledge graph, attention mechanism, knowledge-aware graph neural network, triple-negative breast cancer

Diheng WU, Zhanfa XU, Yi LI, Mingxu ZHANG, Jiaze LIN, Yijun LÜ, Daogang GUAN, Genggeng QIN. Prediction and verification of therapeutic drugs for triple-negative breast cancer using a knowledge graph-based drug repurposing model[J]. Journal of Southern Medical University, 2026, 46(8): 1947-1955.

Figures/Tables 6

References 31

[1]	Bell CC, Gilan O. Principles and mechanisms of non-genetic resistance in cancer[J]. Br J Cancer, 2020, 122(4): 465-72. doi：10.1038/s41416-019-0648-6
[2]	Persi E, Wolf YI, Horn D, et al. Mutation–selection balance and compensatory mechanisms in tumour evolution[J]. Nat Rev Genet, 2021, 22(4): 251-62. doi：10.1038/s41576-020-00299-4
[3]	Lin A, Giuliano CJ, Palladino A, et al. Off-target toxicity is a common mechanism of action of cancer drugs undergoing clinical trials[J]. Sci Transl Med, 2019, 11(509): eaaw8412. doi：10.1126/scitranslmed.aaw8412
[4]	Vasan N, Baselga J, Hyman DM. A view on drug resistance in cancer[J]. Nature, 2019, 575(7782): 299-309. doi：10.1038/s41586-019-1730-1
[5]	Wang YP, Yin ZX. Drug-target interaction prediction through fine-grained selection and bidirectional random walk methodology[J]. Sci Rep, 2024, 14: 18104. doi：10.1038/s41598-024-69186-w
[6]	Peng Y, Zhao SW, Zeng ZL, et al. LGBMDF: a cascade forest framework with LightGBM for predicting drug-target interactions[J]. Front Microbiol, 2023, 13: 1092467. doi：10.3389/fmicb.2022.1092467
[7]	Hasan Mahmud SM, Chen WY, Liu YS, et al. PreDTIs: prediction of drug-target interactions based on multiple feature information using gradient boosting framework with data balancing and feature selection techniques[J]. Brief Bioinform, 2021, 22(5): bbab046. doi：10.1093/bib/bbab046
[8]	Zagami P, Carey LA. Triple negative breast cancer: Pitfalls and progress[J]. npj Breast Cancer, 2022, 8: 95. doi：10.1038/s41523-022-00468-0
[9]	Bai XP, Ni J, Beretov J, et al. Triple-negative breast cancer therapeutic resistance: Where is the Achilles' heel[J]? Cancer Lett, 2021, 497: 100-11. doi：10.1016/j.canlet.2020.10.016
[10]	Gupta R, Srivastava D, Sahu M, et al. Artificial intelligence to deep learning: machine intelligence approach for drug discovery[J]. Mol Divers, 2021, 25(3): 1315-60. doi：10.1007/s11030-021-10217-3
[11]	Liu TQ, Hwang L, Burley SK, et al. BindingDB in 2024: a FAIR knowledgebase of protein-small molecule binding data[J]. Nucleic Acids Res, 2025, 53(D1): D1633-44. doi：10.1093/nar/gkae1075
[12]	Szklarczyk D, Kirsch R, Koutrouli M, et al. The STRING database in 2023: protein-protein association networks and functional enrichment analyses for any sequenced genome of interest[J]. Nucleic Acids Res, 2023, 51(D1): D638-46. doi：10.1093/nar/gkac1000
[13]	Zheng QQ, Min SH, Zhou QH. Identification of potential diagnostic and prognostic biomarkers for LUAD based on TCGA and GEO databases[J]. Biosci Rep, 2021, 41(6): BSR20204370. doi：10.1042/bsr20204370
[14]	Zhang H, Liu XQ, Cheng WY, et al. Prediction of drug-target binding affinity based on deep learning models[J]. Comput Biol Med, 2024, 174: 108435. doi：10.1016/j.compbiomed.2024.108435
[15]	Zhu Y, Zhao LL, Wen NF, et al. DataDTA: a multi-feature and dual-interaction aggregation framework for drug-target binding affinity prediction[J]. Bioinformatics, 2023, 39(9): 560. doi：10.1093/bioinformatics/btad560
[16]	Luo WF, Deng J, He JC, et al. Integration of molecular docking, molecular dynamics and network pharmacology to explore the multi-target pharmacology of fenugreek against diabetes[J]. J Cellular Molecular Medi, 2023, 27(14): 1959-74. doi：10.1111/jcmm.17787
[17]	Zhao LH, Zhang ZX, Su HL, et al. Molecular docking-QSAR-Kronecker-regularized least squares-based multiple machine learning for assessment and prediction of PFAS-protein binding interactions[J]. J Hazard Mater, 2025, 492: 138069. doi：10.1016/j.jhazmat.2025.138069
[18]	Tang XW, Zhou YQ, Yang MY, et al. TC-DTA: predicting drug-target binding affinity with transformer and convolutional neural networks[J]. IEEE Transon Nanobioscience, 2024, 23(4): 572-8. doi：10.1109/tnb.2024.3441590
[19]	Nguyen T, Le H, Quinn TP, et al. GraphDTA: predicting drug-target binding affinity with graph neural networks[J]. Bioinformatics, 2021, 37(8): 1140-7. doi：10.1093/bioinformatics/btaa921
[20]	Öztürk H, Özgür A, Ozkirimli E. DeepDTA: deep drug-target binding affinity prediction[J]. Bioinformatics, 2018, 34(17): i821-9. doi：10.1093/bioinformatics/bty593
[21]	Yuan WN, Chen GX, Chen CY. FusionDTA: attention-based feature polymerizer and knowledge distillation for drug-target binding affinity prediction[J]. Brief Bioinform, 2022, 23: 506. doi：10.1093/bib/bbab506
[22]	Zhang YQ, Zhu GC, Li KW, et al. HLAB: learning the BiLSTM features from the ProtBert-encoded proteins for the class I HLA-peptide binding prediction[J]. Brief Bioinform, 2022, 23(5): 173. doi：10.1093/bib/bbac173
[23]	Mohamed SK, Nounu A, Nováček V. Biological applications of knowledge graph embedding models[J]. Brief Bioinform, 2021, 22(2): 1679-93. doi：10.1093/bib/bbaa012
[24]	He HH, Chen GX, Tang ZC, et al. Dual modality feature fused neural network integrating binding site information for drug target affinity prediction[J]. npj Digit Med, 2025, 8: 67. doi：10.1038/s41746-025-01464-x
[25]	Thafar MA, Alshahrani M, Albaradei S, et al. Affinity2Vec: drug-target binding affinity prediction through representation learning, graph mining, and machine learning[J]. Sci Rep, 2022, 12: 4751. doi：10.1038/s41598-022-08787-9
[26]	Pu YQ, Li JW, Tang JJ, et al. DeepFusionDTA: drug-target binding affinity prediction with information fusion and hybrid deep-learning ensemble model[J]. IEEE/ACM Trans Comput Biol And Bioinf, 2022, 19(5): 2760-9. doi：10.1109/tcbb.2021.3103966
[27]	Zeng X, Su GP, Li SJ, et al. Drug-Online: an online platform for drug-target interaction, affinity, and binding sites identification using deep learning[J]. BMC Bioinform, 2024, 25(1): 156. doi：10.1186/s12859-024-05783-w
[28]	Zhang L, Wang CC, Chen X. Predicting drug-target binding affinity through molecule representation block based on multi-head attention and skip connection[J]. Brief Bioinform, 2022, 23(6): 468. doi：10.1093/bib/bbac468
[29]	Wang JR, Xiao YH, Shang XQ, et al. Predicting drug-target binding affinity with cross-scale graph contrastive learning[J]. Brief Bioinform, 2023, 25: bbad516. doi：10.1093/bib/bbad516
[30]	Lu R, Wang J, Li P, et al. Improving drug-target affinity prediction via feature fusion and knowledge distillation[J]. Brief Bioinform, 2023, 24(3): 145. doi：10.1093/bib/bbad145
[31]	Zhu ZQ, Yao Z, Zheng X, et al. Drug-target affinity prediction method based on multi-scale information interaction and graph optimization[J]. Comput Biol Med, 2023, 167: 107621. doi：10.1016/j.compbiomed.2023.107621

Model	MSE↓	PCC↑	CI↑
KronRLS	5.2108±0.082	0.7237±0.021	0.7063±0.018
SimBoost	4.6471±0.065	0.7429±0.017	0.7277±0.017
DeepDTA	3.8972±0.053	0.7814±0.011	0.7426±0.014
FusionDTA	3.6054±0.047	0.7893±0.010	0.7598±0.014
GraphDTA	3.2845±0.041	0.8001±0.007	0.7959±0.008*
Ours	3.2697±0.038*	0.8037±0.006*	0.7862±0.008

Model	MSE↓	PCC↑	CI↑
KronRLS	5.2108±0.082	0.7237±0.021	0.7063±0.018
SimBoost	4.6471±0.065	0.7429±0.017	0.7277±0.017
DeepDTA	3.8972±0.053	0.7814±0.011	0.7426±0.014
FusionDTA	3.6054±0.047	0.7893±0.010	0.7598±0.014
GraphDTA	3.2845±0.041	0.8001±0.007	0.7959±0.008*
Ours	3.2697±0.038*	0.8037±0.006*	0.7862±0.008

Variants	MSE↓	PCC↑	CI↑
W/o MultiAttention	3.4528±0.049	0.7693±0.010	0.7620±0.011
W/o SelfAttention	3.3159±0.043	0.7762±0.009	0.7634±0.010
W/o KGNN	3.8467±0.058	0.7455±0.012	0.7589±0.013
W/o PPI	3.6234±0.051	0.7581±0.011	0.7611±0.012
All	3.2697±0.040	0.8037±0.006	0.7862±0.007

Variants	MSE↓	PCC↑	CI↑
W/o MultiAttention	3.4528±0.049	0.7693±0.010	0.7620±0.011
W/o SelfAttention	3.3159±0.043	0.7762±0.009	0.7634±0.010
W/o KGNN	3.8467±0.058	0.7455±0.012	0.7589±0.013
W/o PPI	3.6234±0.051	0.7581±0.011	0.7611±0.012
All	3.2697±0.040	0.8037±0.006	0.7862±0.007

Scenario	Model	MSE↓	PCC↑	CI↑
Unseen drug	KronRLS	6.4234±0.094	0.6819±0.013	0.5337±0.016
	SimBoost	5.7003±0.081	0.7026±0.012	0.5874±0.014
	DeepDTA	4.6016±0.069	0.7238±0.011	0.6288±0.013
	FusionDTA	4.2188±0.063	0.7433±0.010	0.6523±0.012
	GraphDTA	3.9764±0.058	0.7687±0.009	0.6811±0.011
	Ours	3.8842±0.052	0.7754±0.008	0.7112±0.008
Unseen target	KronRLS	6.8579±0.102	0.6217±0.021	0.4826±0.017
	SimBoost	6.0720±0.089	0.6576±0.018	0.5017±0.016
	DeepDTA	5.8666±0.084	0.6811±0.015	0.5473±0.014
	FusionDTA	5.4801±0.078	0.7002±0.013	0.5799±0.010
	GraphDTA	5.1212±0.071	0.7137±0.012	0.5943±0.007
	Ours	4.9795±0.065	0.7206±0.010	0.6033±0.007
All unseen	KronRLS	7.6275±0.115	0.5828±0.015	0.4517±0.018
	SimBoost	6.7518±0.101	0.6160±0.014	0.4833±0.015
	DeepDTA	5.9264±0.088	0.6412±0.013	0.5225±0.013
	FusionDTA	5.7730±0.083	0.6705±0.011	0.5496±0.012
	GraphDTA	5.6472±0.079	0.6826±0.011	0.5631±0.011
	Ours	5.5977±0.072	0.7015±0.009	0.5763±0.010