Patent Data Added to a Drug Repurposing Knowledge Graph Failed to Beat Simpler Baselines for Duchenne Muscular Dystrophy

Adding patent-derived relations to a biomedical knowledge graph did not reliably improve computational drug repurposing predictions for Duchenne muscular dystrophy, according to a time-sliced evaluation published open access in BioData Mining on 22 September 2026.

Their baseline graph was drawn from the Unified Medical Language System, version 2018AA, in a leakage-safe configuration containing 47,816 edges. On top of that they layered 345 relations extracted from patent documents, producing an integrated graph of 48,161 edges spanning 41,335 nodes and 91 namespaced relation types. The frozen candidate universe comprised 3,684 concept identifiers and 3,489 ranking units.

Methodological rigor was the study’s defining feature. The team used four knowledge graph embedding models—TransE, DistMult, ComplEx, and RotatE—optimized on the curated UMLS-only data, confirmed with three random seeds, and then refit with five seeds on each scorable graph view. Scoring was label-blind across seeds to avoid optimistic bias. Four strict post-cutoff positive drug-development events for DMD were evaluated under a positive-unlabeled framework, and four prespecified cutoff dates were tested to check whether conclusions depended on where the temporal boundary was drawn.

The headline result was negative for patent augmentation. The UMLS-only TransE model achieved a mean reciprocal rank of 0.00556, while the integrated graph with patent edges reached only 0.00295—roughly half. Patent augmentation improved the rank of just one of the four strict post-cutoff positives. Even more striking, a simple shortest-path baseline built on graph topology alone, with no embedding model at all, achieved a mean reciprocal rank of 0.0260 and recovered three of the positives within its top 250 candidates, although the authors note that ties in this ranking were extensive.

Ranking perturbation, however, was real. Whole-ranking Spearman correlations between the UMLS-only and integrated views ranged from 0.930 to 0.935, indicating broadly similar global orderings, yet the Jaccard overlap of the top 25 candidates was only 0.282 to 0.389. In other words, patent edges materially reshuffled the highest-priority shortlist without making it demonstrably better. Across cutoff dates, integrated mean reciprocal rank rose from 0.00217 to 0.00295 but stayed below the time-aligned UMLS-only anchor at every cutoff tested.

The study also audited the quality of the patent-derived data itself. A domain expert in health sciences reviewed a random sample of 200 instances drawn from the 541 patent predications that the model actually consumed. Of these, 106 were classified as valid, 89 as invalid, and 5 as unresolved, yielding a decidable validity of 54.4 percent, with a Wilson 95 percent confidence interval of 47.4 to 61.2 percent. In other words, in this audited sample, nearly half of the machine-extracted patent relations feeding the model could not be defended as accurate.

Mechanistic claims fared no better. All ten top-ranked units surfaced by the integrated model as mechanistic hypotheses were classified as nonspecific graph-derived artifacts: none carried a candidate-specific patent mechanism or candidate-specific DMD evidence.

The authors are careful about what their results do and do not establish. Patent-derived relations materially perturbed the embedding rankings, but the study did not demonstrate uniform retrospective improvement, a causal contribution from the patent data, or any therapeutic efficacy. Extraction quality, temporal vocabulary alignment between patents and curated vocabularies, model architecture, simple graph topology, and candidate identity all materially affected the conclusions. Their recommendation is measured: patent-augmented knowledge graphs should be treated as provenance-rich hypothesis-generation systems, not as treatment-validation tools.

Subject of Research: A time-sliced knowledge graph evaluation of patent-augmented drug repurposing predictions for Duchenne muscular dystrophy

Article Title: Patent-augmented knowledge graph drug repurposing for duchenne muscular dystrophy: a time-sliced evaluation

Article References: Uludağ¹, F., & CELİK¹, H. E. (2026). Patent-augmented knowledge graph drug repurposing for duchenne muscular dystrophy: a time-sliced evaluation. BioData Mining. https://doi.org/10.1186/s13040-026-00605-6

Image Credits: AI Generated

DOI: 10.1186/s13040-026-00605-6

Keywords: drug repurposing, knowledge graph, Duchenne muscular dystrophy, patents, knowledge graph embeddings, TransE, machine learning, biomedical informatics, UMLS, hypothesis generation, Patent-augmented, knowledge