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

Y. Cheung

cs.LGMar 16, 2026

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

Authors: Y. Cheung

Organizations: School of Mathematical Sciences and Shanghai Key Laboratory of PMMP, East China Normal University, 200241, Shanghai, China

Abstract

Combination pharmacotherapy offers substantial therapeutic advantages but also poses substantial risks of adverse drug reactions (ADRs). The accurate prediction of ADRs with interpretable computational methods is crucial for clinical safety management, drug development, and precision medicine. However, managing ADRs remains a challenge due to the vast search space of drug combinations and the complexity of physiological responses. Current graph-based architectures often struggle to effectively integrate multi-scale biological information and frequently rely on fixed association matrices, which limits their ability to capture dynamic organ-level dependencies and generalize across diverse datasets. Here we propose CrossADR, a hierarchical framework for organ-level ADR prediction through cross-layer feature integration and cross-level associative learning. It incorporates a gated-residual-flow graph neural network to fuse multi-scale molecular features and utilizes a learnable ADR embedding space to dynamically capture latent biological correlations across 15 organ systems. Systematic evaluation on the newly constructed CrossADR-Dataset-covering 1,376 drugs and 946,000 unique combinations-demonstrates that CrossADR consistently achieves state-of-the-art performance across 80 distinct experimental scenarios and provides high-resolution insights into drug-related protein protein interactions and pathways. Overall, CrossADR represents a robust tool for cross-scale biomedical information integration, cross-layer feature integration as well as cross-level associative learning, and can be effectively utilized to prevent ADRs in clinical decision-making.

Figures & tables

Figure 1: Overview of the CrossADR framework for organ-level adverse drug reaction (ADR) prediction. (A) Input: drug pairs, as well as the 15 ADR labels at organ level. (B) The training and application workflow of CrossADR for predicting ADRs and performing downstream PPI network and pathway analysis. (C) Architectural comparison between baseline methods (GNN-KG-based, OrganADR) and the proposed CrossADR model, highlighting two proposed important module: (1) the cross-layer feature integration module, (2) the learnable ADR embeddings as well as cross-level gated module and attention mechanisms.

Figure 2: Detailed architecture of CrossADR. (A) Task setting for ”emerging drug -emerging drug“ ADR prediction task. (B) The cross-layer GNN on KG module. (C) The workflow of CrossADR. (D) Molecular feature. (E) The cross-level gated module and attention module with learnable ADR embedding space.

Figure 3: Knowledge Graph (KG) construction and ablated KGs. (A) Integration of TWOSIDES and ADReCS, as well as the topology of the basic KG. (B,C,D) Three ablated KGs for comparison. B, C and D for ablated KG 1, 2 and 3. Based on the basic KG, different type of information is removed accordingly.

Figure 4: Data annotation and sample generation for the CrossADR-Dataset. (A) Mapping drug combinations from TWOSIDES and ADRECS to 15 specific organs to create multi organ ADR labels. (B) Integration of synergistic and ADR-related information from DrugBank to define positive and negative samples.

Figure 5: Construction and statistics of the CrossADR-Dataset. (A) Determine the drugs to be analyzed for CrossADR. (B) Construction of one sample (one drug combination and its annotations); definition of positive sample and negative sample. (C) CrossADR database, CrossADR-Dataset D and CrossADR-Dataset R. (D, E) Comparison of the number of drugs, total combination space and other metrics between OrganADR-Dataset and the proposed CrossADR-Dataset.

Type	Seed	Train Drugs	Valid Drugs	Test Drugs	Triplets-Trainset	Triplets-Validset	Triplets-Testset
\multirow 10*CrossADR-Dataset D	10	1022	119	121	64190	1118	1240
	15	1017	120	122	64826	1160	1142
	20	1021	117	119	64224	1040	1234
	25	1023	119	123	65070	1260	1032
	30	1028	115	119	62922	1234	1318
	35	1020	120	122	68636	956	798

Table 1: Statistic of the CrossADR-Dataset. The symbols

|\mathcal{D}_{\text{train}}|

|\mathcal{D}_{\text{valid}}|

, and

|\mathcal{D}_{\text{test}}|

in the main manuscript represent Train/Valid/Test drug sets (in the table header), respectively. The symbols

|\mathcal{C}_{\text{train}}|

|\mathcal{C}_{\text{valid}}|

, and

|\mathcal{C}_{\text{test}}|

in the main manuscript represent the Triplets in the Trainset/Validset/Testset (in the table header). Each triplet refers to a single sample, with a balanced

1:1

ratio between positive and negative samples.

Relation	Type of Source Node	Type of Target Node	Number of Edges	KG: Basic	KG: Ablation 1	KG: Ablation 3
ppi	gene/protein	gene/protein	642,150	$\checkmark$	$\checkmark$	$\checkmark$
associated with	effect/phenotype	gene/protein	3,330	$\checkmark$	$\checkmark$
associated with	gene/protein	effect/phenotype	3,330	$\checkmark$	$\checkmark$
parent-child	effect/phenotype	effect/phenotype	37,472	$\checkmark$
target	drug	gene/protein	16,380	$\checkmark$	$\checkmark$	$\checkmark$
target	gene/protein	drug	16,380	$\checkmark$	$\checkmark$	$\checkmark$

Table 2: Overview of the KGs’ edges used in this manuscript. The semantic relations, including source and target node categories, and the total edge count for each triplet type are shown. The columns ”KG: Basic” and ”KG: Ablation 1–3” define the corresponding edge subsets for the 4 KGs.

Figure 6: Performance comparison of CrossADR against baseline models. The figure is organized into a $4\times 4$ panel matrix: rows 1 to 4 correspond to different Knowledge Graph configurations (KG: Basic, KG: Ablation 1, KG: Ablation 2, and KG: Ablation 3, respectively); the first two columns present results for CrossADR-Dataset D (ROC-AUC and F1-score), while the last two columns show results for CrossADR-Dataset R (ROC-AUC and F1-score). In each boxplot, individual scatter points represent results from independent experimental runs, the central horizontal line indicates the median value, the box boundaries denote the upper and lower quartiles, and the whiskers extend to illustrate the overall data range excluding outliers. For each model in each subplot, N = 60.

Datasource	Model	PR-AUC	ROC-AUC	ACC	F1	HL
\multirow 7*CrossADR-Dataset D	CrossADR	83.83% $\pm$ 3.99%	83.75% $\pm$ 4.31%	76.44% $\pm$ 3.91%	75.32% $\pm$ 4.76%	24.25% $\pm$ 3.93%
	OrganADR	83.17% $\pm$ 4.25%	82.76% $\pm$ 4.04%	75.34% $\pm$ 3.87%	74.52% $\pm$ 5.32%	24.66% $\pm$ 3.87%
	EmerGNN (Modified)	81.57% $\pm$ 4.32%	79.86% $\pm$ 4.51%	72.71% $\pm$ 4.68%	73.69% $\pm$ 4.13%	27.29% $\pm$ 4.68%
	RandomForest	61.94% $\pm$ 4.68%	61.31% $\pm$ 3.98%	56.14% $\pm$ 2.59%	38.37% $\pm$ 5.35%	43.86% $\pm$ 2.59%
	KNN	58.20% $\pm$ 4.05%	60.35% $\pm$ 5.03%	56.91% $\pm$ 3.04%	48.28% $\pm$ 5.23%	43.09% $\pm$ 3.04%
	GaussianNB	57.76% $\pm$ 5.26%	59.14% $\pm$ 5.96%	56.98% $\pm$ 5.19%	49.20% $\pm$ 7.34%	43.02% $\pm$ 5.19%

Table 3: Performance of CrossADR with baseline models across 5 metrics. CrossADR, OrganADR and EmerGNN are run three times with random seeds 10, 20, 30. Four ML models are run one time with fixed random state 1. The process above is repeated on ten datasets, which are generated with the random seeds 10, 15, 20, 25, 30, 35, 40, 45, 50 and 55. Given that the ratio of validation set to test set is 1:1, the two sets serve as mutual validation and test sets, and are tested twice. The reported values in table are the mean value and standard deviation calculated from a total of 60 runs for deep-learning models and 10 runs for machine-learning models. All evaluation metrics are shown as percentages (%), where higher values are better for ROC-AUC, PR-AUC, ACC and F1, and lower values indicate better performance for HL. Best results in bold , second best underlined .

		CrossADR (Model 1) vs OrganADR (Model 2)					CrossADR (Model 1) vs EmerGNN (Model 2)
Dataset	KG	Model 1	Model 2	P-Value	Sig	Effect Size	Model 1	Model 2	P-Value	Sig	Effect Size
\multirow 4*CrossADR Dataset-D	Basic	83.7%	82.8%	$2.24\times 10^{-3}$	**	0.413	83.7%	79.9%	$8.96\times 10^{-15}$		1.327
	Ablation 1	68.8%	66.9%	$2.74\times 10^{-3}$	**	0.404	68.8%	66.9%	$1.91\times 10^{-3}$	**	0.420
	Ablation 2	69.6%	64.7%	$1.18\times 10^{-11}$		1.084	69.6%	64.8%	$8.99\times 10^{-11}$		1.016
	Ablation 3	65.5%	62.4%	$1.11\times 10^{-4}$		0.535	65.5%	62.2%	$6.01\times 10^{-4}$		0.468
\multirow 4*CrossADR Dataset-R	Basic	83.6%	80.7%	$9.75\times 10^{-11}$		1.690	83.6%	79.4%	$1.22\times 10^{-18}$		1.650

Table 4: Significance test of CrossADR with baseline model OrganADR and EmerGNN. Consistent with the previous annotation, the reported values in table are calculated from a total of 60 runs. The values reported for Model 1 and Model 2 represent the mean ROC-AUC across experimental runs. ** (

P<0.01

), *** (

P<0.001

). Note that EmerGNN implemented in this comparison is the Modified EmerGNN adapted in this manuscript.

Figure 7: Evaluation of CrossADR verses OrganADR across 15 organs. (A-D) Comprehensive evaluation across four KG configurations (Basic and Ablations 1–3) using five metrics: PR-AUC, ROC-AUC, Accuracy, F1-Score, and Hamming Loss (HM-loss). Each of the 15 axes represents a specific organ category.

		CrossADR (Model 1) vs Ablated 1 (Model 2)					CrossADR (Model 1) vs Ablated 2 (Model 2)
Dataset	KG	Model 1	Model 2	P-Value	Sig	Effect Size	Model 1	Model 2	P-Value	Sig	Effect Size
\multirow 4*CrossADR Dataset-D	Basic	83.7%	82.4%	$9.29\times 10^{-6}$		0.626	83.7%	81.8%	$1.74\times 10^{-6}$		0.686
	Ablation 1	68.8%	67.4%	$1.22\times 10^{-4}$		0.450	68.8%	67.3%	$8.56\times 10^{-3}$	**	0.351
	Ablation 2	69.6%	67.6%	$1.00\times 10^{-3}$		0.451	69.6%	65.4%	$7.23\times 10^{-9}$		0.871
	Ablation 3	65.5%	64.0%	$1.31\times 10^{-2}$		0.330	65.5%	62.1%	$9.81\times 10^{-5}$		0.540
\multirow 4*CrossADR Dataset-R	Basic	83.6%	81.7%	$7.47\times 10^{-12}$		1.099	83.6%	80.9%	$1.24\times 10^{-10}$		1.006

Table 5: Significance test of CrossADR with two ablation studies. Consistent with the previous annotation, the reported values in table are calculated from a total of 60 runs. The values reported for Model 1 and Model 2 represent the mean ROC-AUC across experimental runs. * (

P<0.05

), ** (

P<0.01

), *** (

P<0.001

). Ablation 1 and Ablation 2 refer to ablated variants of CrossADR, with detailed information in the main text.

Figure 8: Case study and interpretability analysis of CrossADR. (A) Drug combination to be analyzed: Drug 1, Donepezil and Drug 2, Benzatropine. (B) Predicted ADRs for the drug combination. (C) Visualization of the PPI network. The proteins in the PPI network are identified as important proteins by CrossADR. (D) KEGG pathway enrichment analysis of the key proteins. (E) Gene Ontology (GO) Biological Process enrichment analysis of the key proteins.

Explore similar work

Exploring Drug Safety Through Knowledge Graphs: Protein Kinase Inhibitors as a Case Study
Feb 17, 2026David Jackson, Michael Gertz, Jürgen HesserClinical Risk Prediction Biomedical KGs
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
CoFEND: A Cross-Modal Fusion End-to-End Network for Cold-Start Drug-Drug Interaction Prediction
Jul 3, 2026Di Wu, Hongyi Sun, Haichao Xu +3Multimodal Graph Learning Drug Discovery
School of Physics and Electronic Science, Zunyi Normal University
Enhanced Drug-drug Interaction Prediction Using Adaptive Knowledge Integration
Mar 13, 2026Pengfei Liu, Jun Tao, Zhixiang RenKnowledge Augmentation for Language Models
Pengcheng Laboratory, Shenzhen 518000 · School of Computer Science and Engineering, Sun Yat-sen University, Guangzhou 510006, China

Feb 17, 2026J/K move · Enter open · S save

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

David Jackson, Michael Gertz, Jürgen Hesser

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

Clinical Risk Prediction Biomedical KGs KG Construction Drug Discovery

Open paper AskPDF arXiv

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.

Abstract

Explore similar work

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

CoFEND: A Cross-Modal Fusion End-to-End Network for Cold-Start Drug-Drug Interaction Prediction

Enhanced Drug-drug Interaction Prediction Using Adaptive Knowledge Integration