Pharmacology knowledge graphs enable drug repurposing without chemical structure information

Background Drug repurposing offers a cost-effective alternative to de novo discovery, yet the relative contributions of model complexity, data volume, and feature modalities to knowledge graph–based repurposing remain poorly quantified under rigorous temporal validation. Methods We constructed a pharmacology knowledge graph from ChEMBL 36 comprising 5,348 entities (3,127 drugs, 1,156 proteins, 1,065 indications) and 20,015 edges across 4 relation types. We enforced a strict temporal split (training: ≤ 2022; testing: 2023–2025) with biologically verified hard negatives mined from failed assays and clinical trials. We benchmarked five Knowledge Graph Embedding models (TransE, TransR, RotatE, ComplEx, DistMult; 0.78–0.81M parameters) and a Standard GNN (3.44M parameters) that incorporates drug chemical structure using a GAT encoder and ESM-2 embeddings evaluated by PR-AUC and Hits@ k on drug–protein and drug–indication link prediction. Scaling (0.78M–9.75M parameters; 25–100% data) and feature ablation studies isolated contributions of model capacity, graph density, and node feature modalities. Results Feature ablation revealed a counter-intuitive performance hierarchy. Removing the GAT-based drug structure encoder entirely from the GNN and retaining only topological embeddings combined with ESM-2 protein features improved drug–protein PR-AUC from 0.5631 to 0.5785, while simultaneously reducing VRAM usage from 5.30 GB to 353 MB. Additionally, replacing the GAT encoder with Morgan fingerprints further degraded performance (PR-AUC = 0.5286), indicating that explicit chemical structure representations can be not only redundant but detrimental for predicting pharmacological network interactions. Scaling the GNN beyond 2.44 M parameters yielded diminishing performance gains, whereas increasing training data consistently improved model performance with no observable ceiling. External validation confirmed 6 of the top 14 novel predictions (42.9%) as established therapeutic indications. Conclusions Drug pharmacological behavior can be accurately predicted using target-centric information and drug–network topology alone, without requiring explicit drug chemical structure representations. Model performance is substantially more sensitive to data volume and node density than to architectural complexity, with scaling in model size yielding limited returns relative to improvements in graph coverage. Consequently, state-of-the-art performance is achievable on budget hardware, with a model using only 352 MB VRAM on a consumer GPU.