Exploring Drug Safety Through Knowledge Graphs: Protein Kinase Inhibitors as a Case Study

David Jackson; Michael Gertz; Jürgen Hesser

q-bio.BMFeb 17, 2026

Exploring Drug Safety Through Knowledge Graphs: Protein Kinase Inhibitors as a Case Study

Authors: David Jackson, Michael Gertz, Jürgen Hesser

Organizations: Institute of Informatics, University of Amsterdam, Amsterdam, Netherlands · Institute of Computer Science, Heidelberg University, Heidelberg, Germany · Data Analysis and Modeling in Medicine, Mannheim Institute for Intelligent Systems in Medicine (MIISM), Medical Faculty Mannheim, Heidelberg University, Mannheim, Germany

Abstract

Adverse Drug Reactions (ADRs) are a leading cause of morbidity and mortality. Existing prediction methods rely mainly on chemical similarity, machine learning on structured databases, or isolated target profiles, but often fail to integrate heterogeneous, partly unstructured evidence effectively. We present a knowledge graph-based framework that unifies diverse sources, drug-target data (ChEMBL), clinical trial literature (PubMed), trial metadata (ClinicalTrials.gov), and post-marketing safety reports (FAERS) into a single evidence-weighted bipartite network of drugs and medical conditions. Applied to 400 protein kinase inhibitors, the resulting network enables contextual comparison of efficacy (HR, PFS, OS), phenotypic and target similarity, and ADR prediction via target-to-adverse-event correlations. A non-small cell lung cancer case study correctly highlights established and candidate drugs, target communities (ERbB, ALK, VEGF), and tolerability differences. Designed as an orthogonal, extensible analysis and search tool rather than a replacement for current models, the framework excels at revealing complex patterns, supporting hypothesis generation, and enhancing pharmacovigilance. Code and data are publicly available at https://github.com/davidjackson99/PKI_KG.

Figures & tables

Panel of Figure 1 — Figure 1: Overview of the $N_{Giant}$ component properties and research coverage of PKIs by clinical trial phase.

Drug	Condition	Weight
Erlotinib	Non-small cell lung carcinoma (C0007131)	273
Gefitinib	Non-small cell lung carcinoma (C0007131)	212
Imatinib	Chronic myelogenous leukemia (C0023473)	173
Imatinib	Gastrointestinal stromal tumor (C0238198)	126
Sorafenib	Advanced Adult Hepatoma (C1706732)	118

Table 1: Top 5 most weighted edges in

N_{DC}

Figure 2: Chord graph linking PKI pairings by the number of common related conditions (left) and chord graph linking PKI pairings by the number of common targets (right). A cutoff was introduced (left: > 10 conditions, right: > 1 targets and weight > 0.4) for clarity.

Drug $u$	Drug $v$	$\|\Gamma(u)\cap\Gamma(v)\|$	Drug $u$	Drug $v$	$\|\Gamma_{w}(u)\cap\Gamma_{w}(v)\|$
Sirolimus	Everolimus	103	Sirolimus	Everolimus	281
Sunitinib	Everolimus	75	Sorafenib	Everolimus	260
Sunitinib	Sorafenib	74	Sunitinib	Everolimus	232
Erlotinib	Gefitinib	71	Erlotinib	Gefitinib	221
Erlotinib	Sorafenib	70	Sorafenib	Sunitinib	216

Table 2: Top 5 drug node pairs with most common neighbors in

N_{DC}

where

\Gamma(u)

denotes the set of neighbors of node u and

\Gamma_{w}(u)

denotes the set of neighbors of node u including the weight between the endpoints.

Drug $u$	Drug $v$	$\mathcal{S}_{uv}$
Brigatinib	Lorlatinib	0.184
Sirolimus	Everolimus	0.181
Ripasudil Hydrochloride Dihydrate	Ripasudil	0.165
Itacitinib	Parsaclisib	0.154
Ulixertinib	Rabusertib	0.154

Table 3: Top 5 drug pairs with highest normalized similarity.

Figure 3: Scatter plot displaying the HR scores collected from all papers for eight exemplary PKIs. Dots represent single papers and lines represent median HR values. Blue dots indicate a favorable effect and red dots represent a negative effect.

Target	Adverse Events	Exemplary PKI
B-RAF	Skin Reaction, Rash Maculo-Papular, Cholestasis, Hand-foot Syndrome, Blood Lactate Dehydrogenase Increase	Sorafenib
ABL	White Blood Cell Count Decreased, Cytopenia, Blood Test Abnormal, Fluid Retention, Pulmonary Edema	Imatinib
FKBP1A	Blood Cholesterol and Urea Increase, Proteinuria, Osteonecrosis, Interstitial Lung Disease, Mouth Ulcer, Aphthous Ulcer	Everolimus
VEGFR	Blood Pressure Increase, Hypertension Ageusia, Proteinuria, Glossodynia Flatulence, Oral Pain	Regorafenib

Table 4: Four exemplary targets with the adverse events of highest correlation

Drug	Weight	Effect	HR	PFS/OS	Approv.	AE	T	Receptor
Erlotinib	273	0.49	0.78	5.8/10.9	2004	43	132	ERbB
Gefitinib	212	0.50	0.68	8.5/15.9	2003	36	42	ERbB
Osimertinib	70	0.80	0.59	12.1/26.9	2015	46	7	ERbB
Crizotinib	52	0.66	0.44	11.4/33.9	2011	59	10	EML4-ALK
Afatinib	37	0.67	0.66	11.9/15.4	2013	53	22	ERbB
Alectinib	29	0.75	0.37	9.3/41.2	2015	58	5	EML4-ALK

Table 5: Drug characteristics including weight, effect, hazard ratio, survival metrics, approval year, adverse events, and receptor targets

Figure 4: Directed graph linking NSCLC drugs to the genes they bind to. Node colors are divided into drugs, VEGFR, ALK, ERbB, and other genes.

Figure 5: Directed graph linking drugs to the genes they bind to, having two types of drugs and targets, respectively. Green nodes and large red nodes represent the drugs/targets from the original set in Figure 4 . Purple nodes are the drugs with an overlapping target profile towards the drugs from the original set and small red nodes are the gene targets only hit by ’purple’ drugs. Sorafenib was excluded here for clarity as it shares its targets with many other drugs.

Table 11

PMID	Punchline Text	Population	Part.	Bias
27755796	Median progression-free survival (PFS) was 30 months (gefitinib vs. erlotinib: 31 vs. 24 months; P > 0.05).	Patient Drug Resistance Japanese	42	0.08
22806307	The CSF concentration and penetration rate of erlotinib were significantly higher than those of gefitinib (P = 0.0008 and <0.0001, respectively).	c-erbB-1 Genes Cerebrospinal Fluid	15	0.19
21684626	Exploratory analyses showed that there was no significant difference in RR and PFS in the gefitinib arm compared to the erlotinib arm (RR (%) 47.9 vs. 39.6, p=0.269; median survival (months) 4.9 vs. 3.1, p=0.336).	Female Smoker Adenocarcinoma Mutation	96	0.09
19680652	There was an association between the disease control with gefitinib and erlotinib (p = 0.031).	Female Non-Smokers Adenocarcinoma	21	0.06
16467097	After treatment with erlotinib or gefitinib, patients with EGFR exon 19 deletions had significantly longer median survival than patients with EGFR L858R (34 versus 8 months; log-rank P = 0.01).	Patient EGFR Receptors	70	0.12
19765296	Erlotinib was slightly superior to gefitinib in all measures of response, although the differences were not statistically significant.	Patient Population	N/A	0.07

Table 8: Partial review of six RCT’s directly comparing Erlotinib and Gefitinib for the treatment of NSCLC.

Abbreviation	Meaning
ADR	Adverse Drug Reaction
FAERS	FDA Adverse Event Reporting System
HR	Hazard Ratio
KG	Knowledge Graph
NCT	National Clinical Trial
NIH	National Institutes of Health

Table 9: List of abbreviations and their meanings

Explore similar work

Pharmacology Knowledge Graphs: Do We Need Chemical Structure for Drug Repurposing?
Mar 2, 2026Youssef Abo-Dahab, Ruby Hernandez, Ismael Caleb Arechiga DuranBiomedical KGs KG Embedding
University of California, San Francisco, San Francisco, CA, USA. · Stanford University, Stanford, CA, USA.
CrossADR: enhancing adverse drug reactions prediction for combination pharmacotherapy with cross-layer feature integration and cross-level associative learning
Mar 16, 2026Y. CheungClinical Risk Prediction Graph Neural Networks
School of Mathematical Sciences and Shanghai Key Laboratory of PMMP, East China Normal University, 200241, Shanghai, China
Knowledge-augmented Agentic AI for Mental Health Medication Information Seeking
Jun 24, 2026Huizi Yu, Jian Liu, Wenkong Wang +10KG Construction Healthcare
Department of Medicine and Therapeutics, The Chinese University of Hong Kong, Shatin, Hong Kong SAR, China · Department of Psychiatry, The Chinese University of Hong Kong, Shatin, Hong Kong SAR, China · School of Control Science and Engineering, Shandong University, Ji’nan, Shandong, China +5

Mar 2, 2026J/K move · Enter open · S save

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

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

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

Biomedical KGs KG Embedding Molecular Representation Learning

Open paper AskPDF arXiv

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.

Abstract

Explore similar work

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

CrossADR: enhancing adverse drug reactions prediction for combination pharmacotherapy with cross-layer feature integration and cross-level associative learning

Knowledge-augmented Agentic AI for Mental Health Medication Information Seeking