An Explainable Machine Learning Framework for Network-Based Drug Repurposing Across Neurological Disorders

Öz

Neurodegenerative diseases affect an estimated 57 million people worldwide, a number projected to double every 20 years, and treatment options remain limited by disease heterogeneity and the absence of curative therapies. Computational drug repurposing is one route to shorter development timelines, but published network-based frameworks for neurodegeneration are typically restricted to a single disease, report opaque rankings, and are evaluated under splitting procedures that let information about held-out associations enter the features. This study measures how much of the apparent accuracy of such a model survives once those weaknesses are removed. Gene-drug-disease associations from DGIdb and CTD covering 206 genes, 757 drugs and eight neurodegenerative diseases yielded 3,959 gene-drug-disease triplets, 2,225 known drug-disease associations and 6,056 possible drug-disease pairs, a bipartite density of 0.37. Thirty features per pair were derived from the gene-drug network, and a Random Forest was compared against logistic regression and three unfitted rankers under six evaluation protocols. Reproducing the original protocol gave a ROC-AUC of 0.980 and an average precision of 0.958, but an unfitted Jaccard threshold on the same data reached 0.975, and once held-out positives were excluded from the negative pool every model reached exactly 1.000, showing that the label was encoded in the features rather than learned. Removing the scored pair's own edges lowered the Random Forest to 0.916 on a single split and 0.933 (SD 0.008) under repeated cross-validation; drug-grouped cross-validation gave 0.805 (SD 0.026) and leave-one-disease-out validation 0.706 (SD 0.054), against 0.721 (SD 0.025) for ranking drugs by network degree alone. Held-out associations were nonetheless recovered with a ROC-AUC of 0.969 against 0.644 for the degree ranker, which is the one task on which the model clearly exceeds a connectivity baseline. SHAP attributions on the corrected model are led by drug and disease connectivity rather than by gene-set overlap, reversing the conclusion drawn under the leaky protocol; these are descriptions of model behaviour and not evidence of biological mechanism. Highly ranked unlabelled candidates are dominated by high-degree cytotoxic agents, and the top candidate for Lewy body disease is a typical antipsychotic contraindicated in that condition, so candidates are reported separately as known, rediscovered and unlabelled and require pharmacological and experimental triage before any therapeutic claim.

Anahtar Kelimeler

• Neurodegenerative disorders
• Drugs
• Targets
• Machine learning
• Interactions.

Nörolojik Hastalıklarda Ağ Tabanlı İlaç Yeniden Kullanımı için Açıklanabilir Bir Makine Öğrenimi Çerçevesi

Öz

Nörodejeneratif hastalıklar dünya çapında tahmini 57 milyon kişiyi etkilemektedir ve bu sayının her 20 yılda bir ikiye katlanacağı öngörülmektedir. Hastalıkların heterojenliği ve iyileştirici tedavilerin bulunmaması nedeniyle mevcut tedavi yaklaşımları sınırlı kalmaktadır. Nörodejeneratif hastalıklar için mevcut ilaç yeniden kullanım çerçeveleri genellikle “kara kutu” tahminleri, tek bir hastalığa odaklanma ve bilgi sızıntısını önleyen sistematik değerlendirme stratejilerinin eksikliğinden muzdariptir. Ağ tabanlı özellik mühendisliğinin yorumlanabilir makine öğrenimi ile entegre edilmesinin, tahminler için mekanik bir gerekçe sağlarken, birden fazla nörodejeneratif durumdaki yeniden kullanım adaylarının sistematik olarak tanımlanmasını mümkün kılabileceği hipotezi ortaya atılmıştır. Gen-ilaç-hastalık ilişkileri, 221 gen, 1.501 ilaç ve sekiz nörodejeneratif hastalık için derlenmiş veritabanlarından (DGIdb, CTD) çıkarılmıştır. İki parçalı bir ağ oluşturuldu ve merkeziyet metrikleri açısından analiz edildi; bu analiz, Jaccard benzerliği, kesirli gen kümesi kapsamı ve ortalama toplama ağ özellikleri dahil olmak üzere ilaç-hastalık çifti başına 30 özellik üretti. Kenar maskeleme, gerçekçi bir değerlendirme simülasyonu sağlamak amacıyla özellik hesaplamasından önce bilinen ilişkilerin %20'sini gizledi. 2.828 çift üzerinde eğitilen bir Rastgele Orman sınıflandırıcısı, bekletilen veriler üzerinde 0,975 ROC-AUC ve 0,958 ortalama hassasiyet elde etti. SHAP analizi, Jaccard benzerliğini (ortalama mutlak değer: 0,089) baskın öngörücü olarak belirledi ve gen kümesi örtüşmesini terapötik ilişkilerin birincil itici gücü olarak doğruladı. En yüksek puanlı adaylar büyük ölçüde bilinen ilaçları yeniden keşfetti (Alzheimer için takrin, Huntington için klozapin), ara sıra yeni tahminler de (izokarboksazid) ortaya çıktı. Gelecekteki çalışmalar, yapısal kenetlenme verilerini entegre etmeli, hastalığa özgü eğitim kullanmalı ve adayları prospektif deneysel tarama yoluyla doğrulamalıdır; böylece ağdan türetilen tahminlerin ötesinde klinik alaka kurulabilir.

Anahtar Kelimeler

• Neurodegenerative disorders
• Drugs
• Targets
• Machine learning
• Interactions.

Kaynakça

1. [1] C. Ricci, “Neurodegenerative Disease: From Molecular Basis to Therapy,” Int J Mol Sci, vol. 25, no. 2, Jan 12, 2024.
2. [2] C. Ricci, “Neurodegenerative Disease: From Molecular Basis to Therapy, 2nd Edition,” Int J Mol Sci, vol. 26, no. 5, Feb 24, 2025.
3. [3] F. Imam, R. Saloner, J. W. Vogel, V. Krish, G. Abdel-Azim, M. Ali, L. An, F. Anastasi, D. Bennett, A. Pichet Binette, A. L. Boxer, M. Bringmann, J. M. Burns, C. Cruchaga, J. L. Dage, A. Farinas, L. Ferrucci, C. A. Finney, M. Frasier, O. Hansson, T. J. Hohman, E. C. B. Johnson, M. Kivimaki, R. Korologou-Linden, A. Ruiz Laza, A. I. Levey, I. Liepelt-Scarfone, L. Lu, N. Mattsson-Carlgren, L. T. Middleton, K. Nho, H. S.-H. Oh, R. C. Petersen, E. M. Reiman, O. Robinson, J. D. Rothstein, A. J. Saykin, A. Shvetcov, C. Slawson, B. Smets, M. Suárez-Calvet, B. M. Tijms, M. Timmers, F. Vieira, N. Vilor-Tejedor, P. J. Visser, K. A. Walker, L. M. Winchester, T. Wyss-Coray, C. Yang, N. Bose, S. Lovestone, and C. The Global Neurodegeneration Proteomics, “The Global Neurodegeneration Proteomics Consortium: biomarker and drug target discovery for common neurodegenerative diseases and aging,” Nature Medicine, vol. 31, no. 8, pp. 2556-2566, 2025/08/01, 2025.
4. [4] V. O. K. Li, Y. Han, T. Kaistha, Q. Zhang, J. Downey, I. Gozes, and J. C. K. Lam, “DeepDrug as an expert guided and AI driven drug repurposing methodology for selecting the lead combination of drugs for Alzheimer’s disease,” Scientific Reports, vol. 15, no. 1, pp. 2093, 2025/01/15, 2025.
5. [5] K. Temiz, A. Gul, and E. Gov, “5-Repurposed Drug Candidates Identified in Motor Neurons and Muscle Tissues with Amyotrophic Lateral Sclerosis by Network Biology and Machine Learning Based on Gene Expression,” Neuromolecular Med, vol. 27, no. 1, pp. 24, Apr 3, 2025.
6. [6] Ç. Oğuztüzün, Z. Gao, H. Li, and R. Xu, “Precision Drug Repurposing (PDR): Patient-level modeling and prediction combining foundational knowledge graph with biobank data,” J Biomed Inform, vol. 163, pp. 104786, Mar, 2025.
7. [7] J. Fang, P. Zhang, Q. Wang, C. W. Chiang, Y. Zhou, Y. Hou, J. Xu, R. Chen, B. Zhang, S. J. Lewis, J. B. Leverenz, A. A. Pieper, B. Li, L. Li, J. Cummings, and F. Cheng, “Artificial intelligence framework identifies candidate targets for drug repurposing in Alzheimer's disease,” Alzheimers Res Ther, vol. 14, no. 1, pp. 7, Jan 10, 2022.
8. [8] C. Zang, H. Zhang, J. Xu, H. Zhang, S. Fouladvand, S. Havaldar, F. Cheng, K. Chen, Y. Chen, B. S. Glicksberg, J. Chen, J. Bian, and F. Wang, “High-throughput target trial emulation for Alzheimer’s disease drug repurposing with real-world data,” Nature Communications, vol. 14, no. 1, pp. 8180, 2023/12/11, 2023.

9. [9] J. Xu, C. Mao, Y. Hou, Y. Luo, J. L. Binder, Y. Zhou, L. M. Bekris, J. Shin, M. Hu, F. Wang, C. Eng, T. I. Oprea, M. E. Flanagan, A. A. Pieper, J. Cummings, J. B. Leverenz, and F. Cheng, “Interpretable deep learning translation of GWAS and multi-omics findings to identify pathobiology and drug repurposing in Alzheimer's disease,” Cell Rep, vol. 41, no. 9, pp. 111717, Nov 29, 2022.
10. [10] C. Su, Y. Hou, J. Xu, Z. Xu, M. Zhou, A. Ke, H. Li, J. Xu, M. Brendel, J. R. M. A. Maasch, Z. Bai, H. Zhang, Y. Zhu, M. C. Cincotta, X. Shi, C. Henchcliffe, J. B. Leverenz, J. Cummings, M. S. Okun, J. Bian, F. Cheng, and F. Wang, “Identification of Parkinson’s disease PACE subtypes and repurposing treatments through integrative analyses of multimodal data,” npj Digital Medicine, vol. 7, no. 1, pp. 184, 2024/07/09, 2024.
11. [11]Y. A. Qadri, S. Shaikh, K. Ahmad, I. Choi, S. W. Kim, and A. V. Vasilakos, "Explainable Artificial Intelligence: A Perspective on Drug Discovery," Pharmaceutics, vol. 17, 2025.
12. [12] Z. P. Wang, X. Li, N. Matsumoto, M. Venkatesan, J. H. Chang, J. Moran, H. Choi, B. Li, Y. Meng, M. E. Hernandez, and J. H. Moore, “Drug repurposing for Alzheimer's disease using a graph-of-thoughts based large language model to infer drug-disease relationships in a comprehensive knowledge graph,” BioData Min, vol. 18, no. 1, pp. 51, Aug 5, 2025.
13. [13] S. S. Bhat, V. Niranjan, T. Chandran, S. R. Kulkarni, S. Makkar, V. Dixit, C. K. Subhasa, and A. Vishal, “Multi-omics integration and machine learning reveal biomarker networks and therapeutic targets in Alzheimer's disease,” Advances in Biomarker Sciences and Technology, vol. 8, pp. 81-101, 2026/11/01/, 2026.
14. [14] G. Riddick, H. Song, S. Ahn, J. Walling, D. Borges-Rivera, W. Zhang, and H. A. Fine, “Predicting in vitro drug sensitivity using Random Forests,” Bioinformatics, vol. 27, no. 2, pp. 220-4, Jan 15, 2011.
15. [15] C. Yan, M. E. Grabowska, A. L. Dickson, B. Li, Z. Wen, D. M. Roden, C. Michael Stein, P. J. Embí, J. F. Peterson, Q. Feng, B. A. Malin, and W.-Q. Wei, “Leveraging generative AI to prioritize drug repurposing candidates for Alzheimer’s disease with real-world clinical validation,” npj Digital Medicine, vol. 7, no. 1, pp. 46, 2024/02/26, 2024.
16. [16] I. G. McKeith, B. F. Boeve, D. W. Dickson, G. Halliday, J. P. Taylor, D. Weintraub, D. Aarsland, J. Galvin, J. Attems, C. G. Ballard, A. Bayston, T. G. Beach, F. Blanc, N. Bohnen, L. Bonanni, J. Bras, P. Brundin, D. Burn, A. Chen-Plotkin, J. E. Duda, O. El-Agnaf, H. Feldman, T. J. Ferman, D. Ffytche, H. Fujishiro, D. Galasko, and J. B. Toledo, “Diagnosis and management of dementia with Lewy bodies: Fourth consensus report of the DLB Consortium,” Neurology, vol. 89, no. 1, pp. 88-100, Jul 4, 2017.
17. [17] J. Ooi, M. R. Hayden, and M. A. Pouladi, “Inhibition of Excessive Monoamine Oxidase A/B Activity Protects Against Stress-induced Neuronal Death in Huntington Disease,” Mol Neurobiol, vol. 52, no. 3, pp. 1850-1861, Dec, 2015.
18. [18] C. Elkan, and K. Noto, “Learning classifiers from only positive and unlabeled data,” in Proceedings of the 14th ACM SIGKDD International Conference on Knowledge Discovery and Data Mining, Las Vegas, NV, USA, 2008, pp. 213-220.
19. [19] M. Zietz, D. S. Himmelstein, K. Kloster, C. Williams, M. W. Nagle, and C. S. Greene, “The probability of edge existence due to node degree: a baseline for network-based predictions,” GigaScience, vol. 13, pp. giae001, Jan 2, 2024.