Patent-augmented knowledge graph drug repurposing for duchenne muscular dystrophy: a time-sliced evaluation
BIODATA MINING, cilt.19, sa.1, ss.1-18, 2026 (SCI-Expanded, Scopus)
- Yayın Türü: Makale / Tam Makale
- Cilt numarası: 19 Sayı: 1
- Basım Tarihi: 2026
- Doi Numarası: 10.1186/s13040-026-00605-6
- Dergi Adı: BIODATA MINING
- Derginin Tarandığı İndeksler: Academic Search Ultimate (EBSCO), Natural Science Collection (ProQuest), Biological Science Database (ProQuest), Biomedical Reference Collection: Corporate Edition (EBSCO), Health Research Premium Collection (ProQuest), Scopus, Technology Collection (ProQuest), Aerospace Database, Science Citation Index Expanded (SCI-EXPANDED), BIOSIS, EMBASE, INSPEC, Directory of Open Access Journals
- Sayfa Sayıları: ss.1-18
- Açık Arşiv Koleksiyonu: AVESİS Açık Erişim Koleksiyonu
- Van Yüzüncü Yıl Üniversitesi Adresli: Evet
Özet
Background
Patent documents may complement curated biomedical resources, but adding patent-derived relations to a knowledge graph does not itself establish improved drug-repurposing prediction or mechanism. We evaluated patent augmentation in a historical, time-sliced Duchenne muscular dystrophy (DMD) knowledge-graph completion study with explicit temporal control, target-leakage removal, source ablation, label-blind multi-seed scoring, extraction-quality auditing, non-embedding baselines, and cutoff sensitivity.
Methods
The leakage-safe UMLS 2018AA graph contained 47 816 edges. Adding 345 patent edges produced an Integrated graph with 48 161 edges, 41 335 global nodes, and 91 namespaced relations. The frozen universe comprised 3 684 candidate CUIs and 3 489 ranking units. TransE, DistMult, ComplEx, and RotatE were optimized on UMLS-only data, confirmed with three seeds, and refit with five seeds per scorable graph view. Four strict post-cutoff DMD clinical-development positives were evaluated under a positive-unlabeled framework. Four graph-topological rules and four prespecified cutoff dates were evaluated, and a domain expert in health sciences audited a 200-instance random sample of 541 model-included patent predications.
Results
UMLS-only TransE achieved MRR 0.00556 versus 0.00295 for Integrated; Integrated improved one of four strict-positive ranks. The UMLS-only shortest-path baseline achieved MRR 0.0260 and recovered three positives in the Top 250, although ties were extensive. The domain-expert audit classified 106 predications as valid, 89 as invalid, and 5 as unresolved; decidable validity was 54.4% (Wilson 95% CI 47.4–61.2%). Integrated MRR increased from 0.00217 to 0.00295 across cutoffs, but remained below the time-aligned UMLS anchor at every cutoff. Whole-ranking Spearman correlations were 0.930–0.935, whereas Top-25 Jaccard overlap was 0.282–0.389. All ten mechanistic Top-10 units were classified as nonspecific graph-derived hypotheses; none had a candidate-specific patent mechanism or candidate-specific DMD evidence.
Conclusions
Patent-derived relations materially perturbed KGE rankings but did not establish uniform retrospective improvement, causal patent contribution, or therapeutic efficacy. Extraction quality, temporal vocabulary alignment, model architecture, simple topology, and candidate identity materially affected the conclusions. Patent-augmented knowledge graphs should therefore be used as provenance-rich hypothesis-generation systems rather than treatment-validation tools.