OPBench: A Graph Benchmark to Combat the Opioid Crisis

Tianyi Ma; Yiyang Li; Yiyue Qian; Zheyuan Zhang; Zehong Wang; Chuxu Zhang; Yanfang Ye

cs.LGFeb 16, 2026

OPBench: A Graph Benchmark to Combat the Opioid Crisis

Authors: Tianyi Ma, Yiyang Li, Yiyue Qian, Zheyuan Zhang, Zehong Wang, Chuxu Zhang, Yanfang Ye

Organizations: University of Notre Dame · Amazon · University of Connecticut

Abstract

The opioid epidemic continues to ravage communities worldwide, straining healthcare systems, disrupting families, and demanding urgent computational solutions. To combat this lethal opioid crisis, graph learning methods have emerged as a promising paradigm for modeling complex drug-related phenomena. However, a significant gap remains: there is no comprehensive benchmark for systematically evaluating these methods across real-world opioid crisis scenarios. To bridge this gap, we introduce OPBench, the first comprehensive opioid benchmark comprising five datasets across three critical application domains: opioid overdose detection from healthcare claims, illicit drug trafficking detection from digital platforms, and drug misuse prediction from dietary patterns. Specifically, OPBench incorporates diverse graph structures, including heterogeneous graphs and hypergraphs, to preserve the rich and complex relational information among drug-related data. To address data scarcity, we collaborate with domain experts and authoritative institutions to curate and annotate datasets while adhering to privacy and ethical guidelines. Furthermore, we establish a unified evaluation framework with standardized protocols, predefined data splits, and reproducible baselines to facilitate fair and systematic comparison among graph learning methods. Through extensive experiments, we analyze the strengths and limitations of existing graph learning methods, thereby providing actionable insights for future research in combating the opioid crisis. Our source code and datasets are available at https://github.com/Tianyi-Billy-Ma/OPBench.

Figures & tables

Figure 1. Dataset construction process for applications in Opbench . (a) Opioid Overdose Detection : We leverage the PDMP dataset from the State of Ohio to construct a heterogeneous graph with four node types to detect patients at risk of opioid overdose. (b) Illicit Online Drug Trafficking : We collect metadata from X/Twitter and construct a hypergraph and a multi-relation graph to capture group-wise interactions and diverse relation types among users for community detection and role classification tasks. (c) Opioid Misuse Detection : We employ the NHANES data to construct a heterogeneous graph over five node types to predict opioid misuse through dietary patterns.

Dataset	Domain	Type	$\|{\mathcal{V}}\|$	$\|{\mathcal{E}}\|$	Task	$\|{\mathcal{Y}}\|$	Class Ratio
Pdmp-OD-Det	Healthcare	Het.	54,318	321,437	BC	2	1.0 / 1.0
X-HyDrug-Comm	Social Network	Hp.	2,936	33,893	ML	6	1.0/1.4/2.1/1.1/0.3/4.1
X-HyDrug-Role	Social Network	Hp.	2,936	33,893	MC	4	1.0 / 0.8 / 4.0 / 1.2
X-MRDrug-Role	Social Network	MR.	27,945	436,385	MC	4	1.0 / 1.2 / 4.4 / 56.3
NHANES-Diet	Dietary	Het.	13,742	445,468	BC	2	1.0/1.0

Table 1. Statistics of benchmark Opbench . BC, ML, and MC denote binary, multi-label, and multi-class classification, respectively. Het., MR. and Hp. denote heterogeneous graph, multi-relation graph, and hypergraph, respectively.

Model	Train-10% Valid-10% Test-80%			Train-20% Valid-10% Test-70%			Train-50% Valid-10% Test-40%
Model	AUC	F1-Macro	F1-Micro	AUC	F1-Macro	F1-Micro	AUC	F1-Macro	F1-Micro
MLP	72.57 $\pm$ 6.86	68.77 $\pm$ 4.97	67.06 $\pm$ 6.54	77.02 $\pm$ 0.69	71.12 $\pm$ 6.32	72.40 $\pm$ 5.79	75.41 $\pm$ 2.67	60.52 $\pm$ 8.59	66.18 $\pm$ 8.59
GCN ( Kipf and Welling, 2017 )	60.96 $\pm$ 0.35	52.95 $\pm$ 6.58	57.86 $\pm$ 0.27	68.58 $\pm$ 0.26	57.58 $\pm$ 8.47	60.20 $\pm$ 4.35	70.91 $\pm$ 0.55	66.99 $\pm$ 8.01	67.13 $\pm$ 3.35
GAT ( Veličković et al., 2018 )	61.76 $\pm$ 4.48	53.94 $\pm$ 7.67	56.70 $\pm$ 3.11	70.75 $\pm$ 2.49	59.10 $\pm$ 10.36	61.77 $\pm$ 5.60	75.78 $\pm$ 2.85	69.64 $\pm$ 6.43	70.82 $\pm$ 2.20
R-GCN ( Schlichtkrull et al., 2018 )	63.37 $\pm$ 3.95	55.85 $\pm$ 10.78	58.99 $\pm$ 4.51	70.54 $\pm$ 3.99	59.79 $\pm$ 12.91	63.14 $\pm$ 6.37	73.94 $\pm$ 0.25	67.93 $\pm$ 0.33	68.03 $\pm$ 0.34
HAN ( Wang et al., 2019 )	79.14 $\pm$ 0.45	75.78 $\pm$ 4.92	74.82 $\pm$ 4.73	84.11 $\pm$ 4.82	76.45 $\pm$ 11.09	79.44 $\pm$ 3.05	87.06 $\pm$ 4.11	78.02 $\pm$ 9.22	82.41 $\pm$ 2.54

Table 2. The performance comparison among Pdmp-OD-Det . The best performance is bolded and runner-ups are underlined.

Model	Train-10% Valid-10% Test-80%			Train-20% Valid-10% Test-70%			Train-50% Valid-10% Test-40%
Model	Acc.	F1-Macro	F1-Micro	Acc.	F1-Macro	F1-Micro	Acc.	F1-Macro	F1-Micro
MLP	82.63 $\pm$ 0.68	47.55 $\pm$ 4.27	70.55 $\pm$ 2.11	84.05 $\pm$ 0.43	58.48 $\pm$ 4.29	74.60 $\pm$ 1.07	84.59 $\pm$ 0.38	56.13 $\pm$ 3.44	74.96 $\pm$ 0.97
GCN ( Kipf and Welling, 2017 )	78.68 $\pm$ 0.23	39.09 $\pm$ 0.97	63.58 $\pm$ 0.76	80.14 $\pm$ 0.24	41.37 $\pm$ 1.84	65.36 $\pm$ 0.28	80.66 $\pm$ 0.33	45.39 $\pm$ 2.32	66.67 $\pm$ 1.19
GAT ( Veličković et al., 2018 )	79.42 $\pm$ 0.51	42.21 $\pm$ 2.28	64.72 $\pm$ 0.78	80.44 $\pm$ 0.59	47.54 $\pm$ 2.75	66.68 $\pm$ 1.75	82.10 $\pm$ 0.54	55.12 $\pm$ 1.68	71.08 $\pm$ 1.19
HGNN ( Feng et al., 2019 )	80.70 $\pm$ 0.16	45.40 $\pm$ 1.40	65.84 $\pm$ 0.20	82.31 $\pm$ 0.32	49.96 $\pm$ 1.66	69.96 $\pm$ 0.93	82.49 $\pm$ 2.88	50.97 $\pm$ 2.88	69.77 $\pm$ 0.77
HNHN ( Dong et al., 2020 )	82.55 $\pm$ 0.65	54.64 $\pm$ 1.78	71.80 $\pm$ 1.37	83.89 $\pm$ 0.48	57.60 $\pm$ 4.97	74.26 $\pm$ 1.02	85.05 $\pm$ 0.86	58.07 $\pm$ 3.28	76.01 $\pm$ 1.71

Table 3. The performance comparison among X-HyDrug-Comm . The best performance is bolded, and runner-ups are underlined.

Model	Train-10% Valid-10% Test-80%			Train-20% Valid-10% Test-70%			Train-50% Valid-10% Test-40%
Model	Acc.	F1-Macro	F1-Micro	Acc.	F1-Macro	F1-Micro	Acc.	F1-Macro	F1-Micro
MLP	67.46 $\pm$ 0.53	45.25 $\pm$ 2.37	67.46 $\pm$ 0.53	68.35 $\pm$ 1.03	47.60 $\pm$ 1.40	68.35 $\pm$ 1.03	69.58 $\pm$ 1.75	48.89 $\pm$ 2.25	69.58 $\pm$ 1.75
GCN ( Kipf and Welling, 2017 )	64.87 $\pm$ 0.51	40.55 $\pm$ 0.66	64.87 $\pm$ 0.51	65.20 $\pm$ 0.78	39.36 $\pm$ 0.76	65.20 $\pm$ 0.78	66.08 $\pm$ 1.12	44.49 $\pm$ 2.06	66.08 $\pm$ 1.12
GAT ( Veličković et al., 2018 )	65.57 $\pm$ 0.81	40.37 $\pm$ 1.44	65.57 $\pm$ 0.81	66.42 $\pm$ 1.28	41.26 $\pm$ 1.55	66.42 $\pm$ 1.28	67.35 $\pm$ 2.04	41.49 $\pm$ 1.65	67.35 $\pm$ 2.04
HGNN ( Feng et al., 2019 )	67.69 $\pm$ 0.21	43.45 $\pm$ 2.00	67.69 $\pm$ 0.21	68.64 $\pm$ 0.90	46.06 $\pm$ 1.86	68.64 $\pm$ 0.90	69.38 $\pm$ 0.92	49.02 $\pm$ 2.78	69.38 $\pm$ 0.92
HNHN ( Dong et al., 2020 )	66.89 $\pm$ 0.63	40.83 $\pm$ 1.43	66.89 $\pm$ 0.63	67.90 $\pm$ 0.86	41.11 $\pm$ 1.49	67.90 $\pm$ 0.86	68.94 $\pm$ 1.55	45.40 $\pm$ 3.88	68.94 $\pm$ 1.55

Table 4. The performance comparison among X-HyDrug-Role . The best performance is bolded and runner-ups are underlined.

Model	Train-10% Valid-10% Test-80%			Train-20% Valid-10% Test-70%			Train-50% Valid-10% Test-40%
Model	GMean	F1-Macro	F1-Micro	GMean	F1-Macro	F1-Micro	GMean	F1-Macro	F1-Micro
MLP	40.53 $\pm$ 12.80	62.38 $\pm$ 2.40	94.43 $\pm$ 0.11	52.60 $\pm$ 2.61	66.43 $\pm$ 1.20	94.94 $\pm$ 0.17	59.74 $\pm$ 1.57	70.30 $\pm$ 1.09	95.56 $\pm$ 0.29
Oversampling ( Zheng et al., 2015 )	57.61 $\pm$ 2.11	62.66 $\pm$ 0.77	92.40 $\pm$ 0.49	60.83 $\pm$ 1.35	64.33 $\pm$ 0.67	92.59 $\pm$ 0.52	65.53 $\pm$ 2.53	68.45 $\pm$ 1.95	94.10 $\pm$ 0.41
SMOTE ( Chawla et al., 2002 )	57.98 $\pm$ 3.68	62.68 $\pm$ 1.69	92.50 $\pm$ 1.09	60.92 $\pm$ 0.99	65.52 $\pm$ 0.63	93.49 $\pm$ 0.09	65.52 $\pm$ 2.78	69.10 $\pm$ 1.53	94.31 $\pm$ 0.22
GCN ( Kipf and Welling, 2017 )	33.84 $\pm$ 20.01	60.31 $\pm$ 3.59	94.48 $\pm$ 0.10	52.03 $\pm$ 4.45	64.38 $\pm$ 1.64	94.75 $\pm$ 0.18	53.55 $\pm$ 3.69	66.08 $\pm$ 1.38	95.22 $\pm$ 0.26
GAT ( Veličković et al., 2018 )	26.87 $\pm$ 16.89	53.43 $\pm$ 3.39	91.35 $\pm$ 0.41	44.79 $\pm$ 8.41	58.63 $\pm$ 3.33	92.78 $\pm$ 0.52	55.67 $\pm$ 3.44	65.55 $\pm$ 1.73	94.40 $\pm$ 0.37

Table 5. The performance comparison among X-MrDrug-Role . The best performance is bolded and runner-ups are underlined.

Model	Train-10% Valid-10% Test-80%			Train-20% Valid-10% Test-70%			Train-50% Valid-10% Test-40%
Model	Acc.	F1-Macro	F1-Micro	Acc.	F1-Macro	F1-Micro	Acc.	F1-Macro	F1-Micro
MLP	68.87 $\pm$ 1.27	69.27 $\pm$ 2.30	68.87 $\pm$ 1.27	70.62 $\pm$ 0.62	71.49 $\pm$ 0.90	70.62 $\pm$ 0.62	72.76 $\pm$ 0.89	73.90 $\pm$ 0.97	72.76 $\pm$ 0.89
GCN ( Kipf and Welling, 2017 )	71.51 $\pm$ 1.20	73.01 $\pm$ 1.48	71.51 $\pm$ 1.20	73.46 $\pm$ 0.85	75.53 $\pm$ 0.90	73.46 $\pm$ 0.85	74.90 $\pm$ 0.60	77.15 $\pm$ 0.67	74.90 $\pm$ 0.60
GAT ( Veličković et al., 2018 )	73.74 $\pm$ 0.74	76.12 $\pm$ 1.01	73.74 $\pm$ 0.74	74.97 $\pm$ 0.64	77.36 $\pm$ 0.66	74.97 $\pm$ 0.64	75.32 $\pm$ 0.90	77.80 $\pm$ 0.90	75.32 $\pm$ 0.90
R-GCN ( Schlichtkrull et al., 2018 )	72.74 $\pm$ 1.06	74.73 $\pm$ 1.25	72.74 $\pm$ 1.06	74.26 $\pm$ 0.70	76.62 $\pm$ 0.60	74.26 $\pm$ 0.70	75.42 $\pm$ 0.66	77.96 $\pm$ 0.58	75.42 $\pm$ 0.66
HAN ( Wang et al., 2019 )	74.06 $\pm$ 0.82	76.27 $\pm$ 1.07	74.06 $\pm$ 0.82	75.10 $\pm$ 0.55	77.69 $\pm$ 0.37	75.10 $\pm$ 0.55	75.90 $\pm$ 0.89	78.46 $\pm$ 0.84	75.90 $\pm$ 0.89

Table 6. The performance comparison among NHANES-Diet . The best performance is bolded and runner-ups are underlined.

Appendix figures & tables5 assets

Supplementary material from the paper’s appendix.

Appendix

Community Type	Representative Drugs (Keywords)
Cannabis	Cannabis and cannabis-infused products
Opioid	Oxycodone, Codeine, Morphine, Fentanyl, Hydrocodone
Hallucinogen	LSD, MDMA, Shroom, Mescaline, DMT, Ketamine
Stimulant	Cocaine, Amphetamine, Methamphetamine
Depressant	Xanax, Farmapram, Valium, Halcion, Ativan, Klonopin
Others	Drugs not listed above

Appendix

Table 7. Drug community types and drug keywords.

Node Type	#Nodes	Feature Dim
Patient ( P. )	30,574	768
Prescriber ( Pr. )	21,159	768
Pharmacy ( Ph. )	2,517	768
Drug ( D. )	68	768
Edge Type	#Edges
P. -take- D.	76,172

Appendix

Table 8. Statistics for the benchmark PDMP-OD-Det .

Node Type	#Nodes	Feature Dim
User ( U. )	4,826	37
Food ( F. )	5,896	814
Ingredient ( I. )	2,792	768
Category ( C. )	174	768
Habit ( H. )	54	768
Edge Type	#Edges

Appendix

Table 9. Statistics for the benchmark NHANES-Diet .

Figure 2. Inference time (ms) across hidden dimensions (64, 128, 256, 512). Left and middle panels show hypergraph neural networks on X-HyDrug-Comm and X-HyDrug-Role ; right panel shows heterogeneous graph neural networks on NHANES-Diet .

		Train-10%Valid-10%Test-80%				Train-20%Valid-10%Test-70%				Train-50%Valid-10%Test-40%
	Model	LR	WD	Drop	Dim	LR	WD	Drop	Dim	LR	WD	Drop	Dim
Pdmp-OD-Det	MLP	0.001	0.0001	0.0	512	0.01	0.0	0.5	512	0.01	0.0001	0.5	256
	GCN	0.0001	0.0001	0.3	256	0.0001	0.0	0.5	512	0.0001	0.0	0.5	128
	GAT	0.0001	0.0	0.5	256	0.0001	0.0001	0.3	256	0.0001	0.0	0.5	256
	R-GCN	0.01	0.0	0.0	512	0.0001	0.0	0.5	64	0.0001	0.0	0.5	512
	HAN	0.001	0.0001	0.3	512	0.001	0.0	0.3	256	0.001	0.0001	0.3	512

Appendix

Table 10. Best hyperparameters for models on Opbench . LR, WD, Drop, and Dim denote learning rate, weight decay, dropout, and hidden dimension, respectively. OOM indicates out-of-memory error during training.

Benchmark Design

Benchmark design creates tasks, datasets, metrics, and protocols that make AI capabilities measurable and comparisons credible.

See all 391 papers

OgBench: A Framework for Evaluating Graph Neural Networks on Omics Data
May 15, 2026Louisa Cornelis, Johan Mathe, Louis Van Langendonck +2Graph Neural Networks Graph Representation Learning
UC Santa Barbara · Atmo, Inc. · Universitat Politècnica de Catalunya
A Rubric-Guided Large Language Model Solution for Opioid Use Disorder Computable Phenotyping
Sep 4, 2026Mengxian Lyu, Paredes Pardo, Cheng Peng +7Medical LLMs Clinical Information Extraction
Department of Health Outcomes and Biomedical Informatics, College of Medicine, University of Florida, Gainesville, FL, USA · Department of Engineering Education, Herbert Wertheim College of Engineering, University of Florida, Gainesville, FL, USA · Department of Psychiatry, College of Medicine, University of Florida, Gainesville, FL, USA +2
Benchmarking graph-based models for in-silico toxicity prediction in drug discovery
Sep 29, 2026Noel Suarez-Barro, Manuel Lama, Juan C. VidalDrug Discovery Graph Representation Learning

May 15, 2026J/K move · Enter open · S save

OgBench: A Framework for Evaluating Graph Neural Networks on Omics Data

Louisa Cornelis, Johan Mathe, Louis Van Langendonck, Guillermo Bernárdez+1

UC Santa Barbara · Atmo, Inc. · Universitat Politècnica de Catalunya

Graph Neural Networks Graph Representation Learning Graph Classification

Open paper AskPDF arXiv

Graph Neural Networks (GNNs) have become the dominant framework for inductive graph-level learning. Yet most benchmarks focus on the regime

n \gg p

, where the number of graphs

n

greatly exceeds the number of nodes per graph

p

. This overlooks biological domains such as omics, which operate in the opposite

n \ll p

regime, characterized by large graphs of genes, transcripts, or proteins across few patient samples. This raises the question: how do GNNs perform in this low-sample, high-node omics setting? We introduce OgBench (Omics-Graph Bench), the first benchmarking platform for graph-level prediction in the

n \ll p

regime characteristic of omics data. We provide a standardized, end-to-end modular infrastructure from raw omics data to families of featured graphs with varied structural properties. We benchmark classical GNNs, as well as GNNs designed for large graphs and omics applications, alongside MLPs and machine learning baselines to establish reference performances. Our results show that widely used GNNs often do not outperform simple MLPs and classical baselines. These findings challenge the prevailing assumption that graph structure inherently adds value in this domain, fostering a critical reassessment of current learning paradigms. Ultimately, by exposing these limitations, OgBench provides the open-source ecosystem necessary for the community to develop and validate novel architectures explicitly tailored for biological graphs. The code is available at https://github.com/geometric-intelligence/ogbench.

Abstract

Benchmark Design

Explore similar work

OgBench: A Framework for Evaluating Graph Neural Networks on Omics Data

A Rubric-Guided Large Language Model Solution for Opioid Use Disorder Computable Phenotyping

Benchmarking graph-based models for in-silico toxicity prediction in drug discovery