Pharmacology Knowledge Graphs: Do We Need Chemical Structure for Drug Repurposing?

Youssef Abo-Dahab; Ruby Hernandez; Ismael Caleb Arechiga Duran

cs.AIMar 2, 2026

Pharmacology Knowledge Graphs: Do We Need Chemical Structure for Drug Repurposing?

Authors: Youssef Abo-Dahab, Ruby Hernandez, Ismael Caleb Arechiga Duran

Organizations: University of California, San Francisco, San Francisco, CA, USA. · Stanford University, Stanford, CA, USA.

Abstract

The contributions of model complexity, data volume, and feature modalities to knowledge graph-based drug repurposing remain poorly quantified under rigorous temporal validation. We constructed a pharmacology knowledge graph from ChEMBL 36 comprising 5,348 entities including 3,127 drugs, 1,156 proteins, and 1,065 indications. A strict temporal split was enforced with training data up to 2022 and testing data from 2023 to 2025, together with biologically verified hard negatives mined from failed assays and clinical trials. We benchmarked five knowledge graph embedding models and a standard graph neural network with 3.44 million parameters that incorporates drug chemical structure using a graph attention encoder and ESM-2 protein embeddings. Scaling experiments ranging from 0.78 to 9.75 million parameters and from 25 to 100 percent of the data, together with feature ablation studies, were used to isolate the contributions of model capacity, graph density, and node feature modalities. Removing the graph attention based drug structure encoder and retaining only topological embeddings combined with ESM-2 protein features improved drug protein PR-AUC from 0.5631 to 0.5785 while reducing VRAM usage from 5.30 GB to 353 MB. Replacing the drug encoder with Morgan fingerprints further degraded performance, indicating that explicit chemical structure representations can be detrimental for predicting pharmacological network interactions. Increasing model size beyond 2.44 million parameters yielded diminishing returns, whereas increasing training data consistently improved performance. External validation confirmed 6 of the top 14 novel predictions as established therapeutic indications. These results show that drug pharmacological behavior can be accurately predicted using target-centric information and drug network topology alone, without requiring explicit chemical structure representations.

Figures & tables

Metric	De Novo Drug Discovery	Drug Repurposing
Average Cost	$2.0–$ 3.0 Billion	$300 Million
Time to Market	10–17 Years	3–12 Years
Success Rate	$\sim$ 10% (Phase I to Approval)	$\sim$ 30%
Primary Risk	Safety & Efficacy	Efficacy
Key Advantage	Novelty & IP Exclusivity	Speed, Cost, & Safety

Table 1: Comparative Analysis of De Novo Drug Discovery vs. Drug Repurposing Pinzi et al. (2024)

Figure 1: The scatter plot illustrates the trade-off between predictive performance (Drug-Protein PR-AUC) and computational cost (Peak VRAM, log scale). While highly expressive architectures like the Blackwell GNN (purple diamond) establish the predictive ceiling (0.5910 PR-AUC) at a massive memory cost, scaling structural model complexity exhibits severe diminishing returns. Crucially, the Efficient Topological Model (green star)—which ablates explicit 2D/3D drug structure encoders in favor of pure topological embeddings and ESM-2 protein features—achieves 95% of state-of-the-art performance (0.5785 PR-AUC) utilizing less than 1% of the memory footprint. This demonstrates that for macro-scale repurposing, combining relational topology with high-fidelity target representations offers a vastly superior efficiency trade-off compared to computationally heavy small-molecule graph convolutions.

Model	Params	VRAM	Drug–Protein		Drug–Indication
Model	Params	VRAM	PR-AUC	Hits@10	PR-AUC	Hits@10
TransE	0.78M	1.0 GB	0.1871	0.1594	0.2954	0.4268
TransR	0.81M	1.0 GB	0.2017	0.1620	0.2402	0.3028
RotatE	0.78M	1.5 GB	0.1855	0.1783	0.0720	0.2718
ComplEx	0.78M	0.9 GB	0.2262	0.2188	0.2989	0.4028
DistMult	0.78M	1.0 GB	0.2314	0.2241	0.3246	0.3803

Table 2: Benchmark Comparison of Knowledge Graph Embeddings vs. GNNs. The highly expressive Blackwell GNN sets the absolute predictive ceiling when hardware constraints are lifted, but KGEs (especially ComplEx and DistMult) provide highly memory-efficient alternatives for budget GPUs.

Figure 2: Benchmark comparison of Knowledge Graph Embeddings (KGEs) versus the Standard GNN. Performance is measured by Area Under the Precision-Recall Curve (PR-AUC) for both Drug-Protein (blue) and Drug-Indication (orange) link prediction tasks. While shallow KGEs like ComplEx and DistMult offer highly memory-efficient baselines, the Standard GNN establishes a significantly higher predictive ceiling due to the integration of ESM-2 protein representations.

Variant	Params	VRAM	Drug–Protein		Drug–Indication
Variant	Params	VRAM	PR-AUC	Hits@10	PR-AUC	Hits@10
Standard GNN (GAT + ESM-2)	3.44M	5.30 GB	0.5631	0.5134	0.8175	0.8620
Ablation 1: No ESM-2	3.29M	5.26 GB	0.5186	0.4419	0.7937	0.8310
Ablation 2: No Drug Graphs	3.29M	353 MB	0.5785	0.5234	0.8060	0.8042
Ablation 3: Both Ablated	3.14M	345 MB	0.4658	0.4340	0.7952	0.8183
Morgan FP MLP + ESM-2	3.69M	$\sim$ 350 MB	0.5286	0.4813	0.7937	0.8254

Table 3: Feature Ablation Study. Removing the GAT drug encoder (Ablation 2) and relying purely on topological embeddings increases PR-AUC for drug-protein interactions while slashing VRAM usage by 93%, demonstrating that static 2D/3D drug structures add sparse noise and massive computational overhead compared to pure relational topology.

Track 1: Data Scaling (Fixed Params: 3.44M, sd=256)
Experiment	Params	Drug–Protein		Drug–Indication
Experiment	Params	PR-AUC	Hits@10	PR-AUC	Hits@10
25% Data	3.44M	0.4320	0.3782	0.5749	0.6338
50% Data	3.44M	0.5140	0.4603	0.6882	0.7634
75% Data	3.44M	0.5381	0.4850	0.7919	0.8155
100% Data (Standard GNN)	3.44M	0.5631	0.5134	0.8175	0.8620

Table 4: Scaling Laws: Data vs. Parameters. A model with half the parameters (1.66M) trained on 100% of the data is slighly better than a 3.44M parameter model trained on 50% of the data. Furthermore, scaling parameters beyond 2.44M yields diminishing returns, which does not happen with adding more data points proving that data volume is a tighter bottleneck than parameter count.

Rank	Drug	Predicted Indication	Validation Status	Evidence
4	Diltiazem Hydrochloride	Stroke	Clinical Efficacy	NORDIL Trial Hansson et al. (2000)
5	Dabigatran Etexilate Mesylate	Pulmonary Embolism	FDA Approved	FDA Label U.S. Food and Drug Administration (2014)
7	Cortisone Acetate	Serum Sickness	FDA Approved	FDA Label Drugs.com (2025)
8	Captopril	Stroke	Class Effect Efficacy	PROGRESS/HOPE Trials Gorelick (2002)
9	Methotrexate Sodium	Immune System Diseases	Standard Practice	Clinical Guidelines Weinblatt et al. (1985)
14	Pantoprazole Sodium	Stomach Ulcer	Standard Practice	Clinical Guidelines Savarino et al. (2018)

Table 5: External validation of top-ranked novel drug–indication predictions from the Blackwell GNN. Predictions were restricted to edges absent from the training graph (

\leq

2022). Six of the top 14 predictions (42.9%) are confirmed as true therapeutic indications by FDA labels or primary clinical literature. Etrasimod for atopic dermatitis (Rank 22) represents an investigational prediction independently corroborated by an ongoing Phase 2 clinical trial.

Figure 3: Feature ablation study isolating the impact of structural versus topological features on Drug-Protein PR-AUC. Removing the explicit 2D/3D drug graphs (Ablation 2) yields the highest predictive performance, outperforming both the full GAT-based model and the Morgan fingerprint (MLP) baseline. This demonstrates that for macro-scale repurposing of approved drugs, pure relational topology provides a superior signal compared to static chemical structures. Conversely, removing the ESM-2 protein sequence representations (Ablation 1 and 3) severely degrades performance.

Figure 4: Scaling laws contrasting data volume versus model complexity. The plot tracks Drug-Protein PR-AUC as resources are scaled relative to the 3.44M parameter baseline on the full graph. The model is highly sensitive to graph density (blue line), showing a rapid decline as training data is reduced. In contrast, reducing the network’s parameter capacity by over 50% (red line) results in negligible performance loss. Notably, a lightweight model (1.66M parameters) trained on the complete knowledge graph outperforms a massive model (3.44M parameters) constrained to 50% of the data.

Configuration	GPU	Params	Peak VRAM	Train Time	GPU TFLOPS	DP PR-AUC
KGE (TransR)	RTX 3090	0.81M	1.0 GB	$\sim$ 12 min	23.68	0.2017
Standard GNN	RTX 3090	3.44M	5.30 GB	44.8 min	23.68	0.5631
Ablation 2 (No Drug Graphs)	RTX 3090	3.29M	353 MB	42.1 min	23.68	0.5785
Scaled Model (sd=512)	RTX 3090	9.75M	5.40 GB	45.2 min	23.68	0.5349
Blackwell GNN	RTX PRO 6000	9.73M	36.8 GB	60.2 min	66.75	0.5910

Table 6: Comprehensive computational cost comparison across model configurations. GPU TFLOPS values are empirically measured single-precision (FP32) throughput. Despite the RTX PRO 6000 delivering

2.8\times

higher throughput, the Blackwell GNN’s 36.8 GB memory footprint and attention-based message passing offset throughput gains, resulting in only a modest absolute runtime increase over budget models. The ablated 353 MB model achieves the best efficiency–performance trade-off across all dimensions.

Figure 5: Computational footprint scaling. (Left) Total training time scales linearly with graph density (data scaling) due to the computational overhead of message-passing over edges. Altering the parameter count (hidden dimension) has virtually no impact on training speed. (Right) Peak GPU VRAM remains static at $\sim$ 5.2 GB across all scales, indicating that memory bottlenecks are driven by static node features (e.g., ESM-2 embeddings) and graph topology, rather than the model’s learnable weight matrices.

Method	Task	Metric	Reported result
TxGNN Huang et al. (2024)	Drug–disease	AUPRC / rel. gain	0.913; +49.2% ind., +35.1% contra.
NetMedGPT Firoozbakht et al. (2026)	Drug–disease	AUPRC	0.96 (indication), 0.91 (contraindication)
DTGHAT Jiang et al. (2025)	Drug–target	AUC / AUPRC	0.96 $\pm$ 0.01 / 0.95 $\pm$ 0.01 (5-fold CV)
LLaDR Xiang et al. (2025)	DRKG link prediction	AUC	0.850 (best reported variant)
Our 353 MB Model	Drug–Protein	ROC-AUC / PR-AUC	0.976 / 0.578
(Ablation 2)	Drug–Indication	ROC-AUC / PR-AUC	0.988 / 0.806

Table 7: Comparison of recent literature alongside our highly efficient 353 MB model.

Explore similar work

DrugReason: Dynamic Multi-View Reasoning over Knowledge Graph and Language Evidence for Drug Repurposing
Sep 6, 2026Zijie Liu, Hongxuan Li, Zhen Tan +5LLM Reasoning with Graphs Multi-Hop Reasoning
University of North Carolina at Chapel Hill · Stevens Institute of Technology · Emory University +1
Pharmacogenomic Knowledge Graph Augmentation for Graph Neural Network-Based Drug-Drug Interaction Prediction
Jun 5, 2026Juergen DietrichGraph Neural Networks Knowledge Graphs
AI Solutions Berlin
Large Language Models Meet Biomedical Knowledge Graphs for Mechanistically Grounded Therapeutic Prioritization
Apr 17, 2026Chih-Hsuan Wei, Chi-Ping Day, Zhizheng Wang +8LLM Grounding Large Language Model-Guided Scientific Discovery
Division of Intramural Research (DIR), National Library of Medicine (NLM), National Institutes of Health (NIH); Bethesda, MD 20894, USA · Cancer Data Science Laboratory, Center for Cancer Research, National Cancer Institute (NCI), National Institutes of Health (NIH), Bethesda, MD 20894, USA · Dartmouth Health, Dartmouth Hitchcock Medical Center, Lebanon, NH, USA +2

Sep 6, 2026J/K move · Enter open · S save

DrugReason: Dynamic Multi-View Reasoning over Knowledge Graph and Language Evidence for Drug Repurposing

Zijie Liu, Hongxuan Li, Zhen Tan, Jinhao Duan+4

University of North Carolina at Chapel Hill · Stevens Institute of Technology · Emory University+1

LLM Reasoning with Graphs Multi-Hop Reasoning

Open paper AskPDF arXiv

Drug repurposing aims to identify new therapeutic uses for existing compounds and, compared with de novo drug discovery, offers a faster and more cost-effective path to clinical translation. However, the space of candidate drug-disease pairs is enormous and their underlying relationships often depend on complex multi-hop biological mechanisms, making it difficult to reliably predict which pairs represent true therapeutic relationships. Existing approaches tackle this from two directions: knowledge graph-based methods organize curated biomedical evidence into structured relational networks for grounded multi-hop reasoning, while LLM-based methods leverage pretrained knowledge to generate flexible mechanistic rationales. Yet neither is sufficient alone - KGs are confined to observed graph structure while LLMs lack factual grounding and risk hallucination. To address this gap, we propose DrugReason, a multi-view reasoning framework that integrates grounded KG reasoning with LLM-generated mechanistic inference for drug repurposing. DrugReason adaptively routes diverse reasoning paths to specialized experts conditioned on the query context, while a cross-expert distillation objective enables knowledge sharing without sacrificing expert specialization. Experiments on PharmaDB, DDInter, and DrugBank show that DrugReason improves average performance over strong single-view reasoning baselines and achieves competitive or superior results compared with graph-based alternatives, while providing interpretable routing-based predictions.

Abstract

Explore similar work

DrugReason: Dynamic Multi-View Reasoning over Knowledge Graph and Language Evidence for Drug Repurposing

Pharmacogenomic Knowledge Graph Augmentation for Graph Neural Network-Based Drug-Drug Interaction Prediction

Large Language Models Meet Biomedical Knowledge Graphs for Mechanistically Grounded Therapeutic Prioritization