Graph Neural Networks Research Articles

Graph Neural Networks and Multimodal DTI Features for Schizophrenia Classification: Insights from Brain Network Analysis and Gene Expression.

Schizophrenia (SZ) stands as a severe psychiatric disorder. This study applied diffusion tensor imaging (DTI) data in conjunction with graph neural networks to distinguish SZ patients from normal controls (NCs) and showcases the superior performance of a graph neural network integrating combined fractional anisotropy and fiber number brain network features, achieving an accuracy of 73.79% in distinguishing SZ patients from NCs. Beyond mere discrimination, our study delved deeper into the advantages of utilizing white matter brain network features for identifying SZ patients through interpretable model analysis and gene expression analysis. These analyses uncovered intricate interrelationships between brain imaging markers and genetic biomarkers, providing novel insights into the neuropathological basis of SZ. In summary, our findings underscore the potential of graph neural networks applied to multimodal DTI data for enhancing SZ detection through an integrated analysis of neuroimaging and genetic features.

A graph neural network that combines scRNA-seq and protein-protein interaction data.

A graph neural network that combines scRNA-seq and protein-protein interaction data.

Local-Global Structure-Aware Geometric Equivariant Graph Representation Learning for Predicting Protein-Ligand Binding Affinity.

Predicting protein-ligand binding affinities is a critical problem in drug discovery and design. A majority of existing methods fail to accurately characterize and exploit the geometrically invariant structures of protein-ligand complexes for predicting binding affinities. In this study, we propose Geo-protein-ligand binding affinity (PLA), a geometric equivariant graph representation learning framework with local-global structure awareness, to predict binding affinity by capturing the geometric information of protein-ligand complexes. Specifically, the local structural information of 3-D protein-ligand complexes is extracted by using an equivariant graph neural network (EGNN), which iteratively updates node representations while preserving the equivariance of coordinate transformations. Meanwhile, a graph transformer is utilized to capture long-range interactions among atoms, offering a global view that adaptively focuses on complex regions with a significant impact on binding affinities. Furthermore, the multiscale information from the two channels is integrated to enhance the predictive capability of the model. Extensive experimental studies on two benchmark datasets confirm the superior performance of Geo-PLA. Moreover, the visual interpretation of the learned protein-ligand complexes further indicates that our model offers valuable biological insights for virtual screening and drug repositioning.

Evading control flow graph based GNN malware detectors via active opcode insertion method with maliciousness preserving

With the continuous advancement of machine learning, numerous malware detection methods that leverage this technology have emerged, presenting new challenges to the generation of adversarial malware. Existing function-preserving adversarial attacks fall short of effectively modifying portable executable (PE) malware control flow graphs (CFGs), thereby failing to bypass the graph neural network (GNN) models that utilize CFGs for detection. To solve this issue, we introduce a novel base modification method called active opcode insertion, which modifies PE CFGs while preserving functionality by inserting a processed sequence of benign and jump opcodes to connect with the original base block. Using reinforcement learning, MalAOI identifies optimal insertion points and benign opcode sequences to autonomously generate adversarial malware that evades GNN model detection. We tested our approach on the BODMAS and SOREL-20M datasets, and the results demonstrate that MalAOI-generated adversarial malware achieves an average evasion rate of 93.73% against the GNN detection model, with only 12.87% increase in byte size.

A review of hyperspectral image classification based on graph neural networks

Hyperspectral images provide rich spectral-spatial information but pose significant classification challenges due to high dimensionality, noise, mixed pixels, and limited labeled samples. Graph Neural Networks (GNNs) have emerged as a promising solution, offering a semi-supervised framework that can capture complex spatial-spectral relationships inherent in non-Euclidean hyperspectral image data. However, existing reviews often concentrate on specific aspects, thus limiting a comprehensive understanding of GNN-based hyperspectral image classification. This review systematically outlines the fundamental concepts of hyperspectral image classification and GNNs, and summarizes leading approaches from both traditional machine learning and deep learning. Then, it categorizes GNN-based methods into four paradigms: graph recurrent neural networks, graph convolutional networks, graph autoencoders, and hybrid graph neural networks, discussing their theoretical underpinnings, architectures, and representative applications. Finally, five key directions are further highlighted: adaptive graph construction, dynamic graph processing, deeper architectures, self-supervised strategies, and robustness enhancement. These insights aim to facilitate continued innovation in GNN-based hyperspectral imaging, guiding researchers toward more efficient and accurate classification frameworks.

H2GnnDTI: hierarchical heterogeneous graph neural networks for drug target interaction prediction.

Identifying drug target interactions is a crucial step in drug repurposing and drug discovery. The significant increase in demand and the expensive nature for experimentally identifying drug target interactions necessitate computational tools for automated prediction and comprehension of drug target interactions. Despite recent advancements, current methods fail to fully leverage the hierarchical information in drug target interactions. Here we introduce H2GnnDTI, a novel two-level hierarchical heterogeneous graph learning model to predict drug target interactions, by integrating the structures of drugs and proteins via a low-level view GNN (LGNN) and a high-level view GNN (HGNN). The hierarchical graph consists of high-level heterogeneous nodes representing drugs and proteins, connected by edges representing known DTIs. Each drug or protein node is further detailed in a low-level graph, where nodes represent molecules within each drug or amino acids within each protein, accompanied by their respective chemical descriptors. Two distinct low-level graph neural networks are first deployed to capture structural and chemical features specific to drugs and proteins from these low-level graphs. Subsequently, a high-level graph encoder is employed to comprehensively capture and merge interactive features pertaining to drugs and proteins from the high-level graph. The high-level encoder incorporates a structure and attribute information fusion module designed to explicitly integrate representations acquired from both a feature encoder and a graph encoder, facilitating consensus representation learning. Extensive experiments conducted on three benchmark datasets have shown that our proposed H2GnnDTI model consistently outperforms state-of-the-art deep learning methods. The codes are freely available at https://github.com/LiminLi-xjtu/H2GnnDTI. Supplementary data are available at Bioinformatics online.

RDDGN: minimizing the total resistance distance and diameter using graph neural networks

RDDGN: minimizing the total resistance distance and diameter using graph neural networks

ScNET: learning context-specific gene and cell embeddings by integrating single-cell gene expression data with protein-protein interactions.

Recent advances in single-cell RNA sequencing (scRNA-seq) techniques have provided unprecedented insights into the heterogeneity of various tissues. However, gene expression data alone often fails to capture and identify changes in cellular pathways and complexes, as they are more discernible at the protein level. Moreover, analyzing scRNA-seq data presents further challenges due to inherent characteristics such as high noise levels and zero inflation. In this study, we propose an approach to address these limitations by integrating scRNA-seq datasets with a protein-protein interaction network. Our method utilizes a unique dual-view architecture based on graph neural networks, enabling joint representation of gene expression and protein-protein interaction network data. This approach models gene-to-gene relationships under specific biological contexts and refines cell-cell relations using an attention mechanism. Next, through comprehensive evaluations, we demonstrate that scNET better captures gene annotation, pathway characterization and gene-gene relationship identification, while improving cell clustering and pathway analysis across diverse cell types and biological conditions.

TwitterTagNet: an extensive graph dataset for node classification in co-occurring hashtag networks

This study introduces a comprehensive dataset of co-occurring Twitter hashtags, specifically designed to benchmark the performance of graph neural network (GNN) models in node classification tasks, with a focus on hashtag virality. The dataset, collected over 10 days and comprising 4655 nodes and 5901 edges, incorporates a diverse set of 12 features for each hashtag. These features include standard metrics, such as structural characteristics within co-occurrence networks, and novel attributes, including semantic similarity, polarity, subjectivity, and a unique user activity measure based on activity days and tweet volume. A new metric for determining user activity levels on social networks is introduced and used as a feature in the dataset. The dataset is divided into two parts: a feature matrix representing each hashtag’s attributes and a connection matrix detailing the co-occurrence relationships between hashtags. Labels indicating viral or non-viral behavior were assigned during data collection, providing a valuable resource for evaluating the effectiveness of GNN models in classifying hashtag virality. The high F1-scores achieved using these models underscore the significance of integrating structural, content-based, and user-related features, as well as the relationships between hashtags, in accurately classifying hashtag virality on Twitter. This work also demonstrates the significant influence of the selected features and relationships within the dataset, highlighting the importance of a comprehensive graph-based approach in understanding and predicting hashtag virality.

Spatio-Temporal Graph Neural Networks for Streamflow Prediction in the Upper Colorado Basin

Spatio-Temporal Graph Neural Networks for Streamflow Prediction in the Upper Colorado Basin

Predicting orthognathic surgery results as postoperative lateral cephalograms using graph neural networks and diffusion models

Orthognathic surgery, or corrective jaw surgery, is performed to correct severe dentofacial deformities and is increasingly sought for cosmetic purposes. Accurate prediction of surgical outcomes is essential for selecting the optimal treatment plan and ensuring patient satisfaction. Here, we present GPOSC-Net, a generative prediction model for orthognathic surgery that synthesizes post-operative lateral cephalograms from pre-operative data. GPOSC-Net consists of two key components: a landmark prediction model that estimates post-surgical cephalometric changes and a latent diffusion model that generates realistic synthesizes post-operative lateral cephalograms images based on predicted landmarks and segmented profile lines. We validated our model using diverse patient datasets, a visual Turing test, and a simulation study. Our results demonstrate that GPOSC-Net can accurately predict cephalometric landmark positions and generate high-fidelity synthesized post-operative lateral cephalogram images, providing a valuable tool for surgical planning. By enhancing predictive accuracy and visualization, our model has the potential to improve clinical decision-making and patient communication.

Scalable training of trustworthy and energy-efficient predictive graph foundation models for atomistic materials modeling: a case study with HydraGNN

We present our work on developing and training scalable, trustworthy, and energy-efficient predictive graph foundation models (GFMs) using HydraGNN, a multi-headed graph convolutional neural network architecture. HydraGNN expands the boundaries of graph neural network (GNN) computations in both training scale and data diversity. It abstracts over message passing algorithms, allowing both reproduction of and comparison across algorithmic innovations that define nearest-neighbor convolution in GNNs. This work discusses a series of optimizations that have allowed scaling up the GFMs training to tens of thousands of GPUs on datasets consisting of hundreds of millions of graphs. Our GFMs use multitask learning (MTL) to simultaneously learn graph-level and node-level properties of atomistic structures, such as energy and atomic forces. Using over 154 million atomistic structures for training, we illustrate the performance of our approach along with the lessons learned on two state-of-the-art US Department of Energy (US-DOE) supercomputers, namely the Perlmutter petascale system at the National Energy Research Scientific Computing Center and the Frontier exascale system at Oak Ridge Leadership Computing Facility. The HydraGNN architecture enables the GFM to achieve near-linear strong scaling performance using more than 2000 GPUs on Perlmutter and 16,000 GPUs on Frontier.

Impact of crystal structure symmetry in training datasets on GNN-based energy assessments for chemically disordered CsPbI3

Robust solutions combining computational chemistry and data-driven approaches are in high demand in various areas of materials science. For instance, such methods can use machine learning models trained on a limited dataset to make structure-to-property predictions over large search spaces. This paper examines the impact of data selection mechanisms on thermodynamic property assessments for chemically modified lead halide perovskite γ-CsPbI3 and non-perovskite δ-CsPbI3. For disordered states of these phases, complete composition/configuration spaces are built by adding Cd or Zn substitutions of Pb and Br substitutions of I and comprise 2,946,709 and 2,995,462 inequivalent spatial arrangements of substituents, respectively. Using the properties of 1162 entries of the built spaces evaluated by means of density functional theory, we implement independent procedures for training graph neural networks (GNNs). In each of them, a training dataset is constructed depending on the defect contents and presence of low- and high-symmetry structures. The results show that symmetries of training structures can significantly influence quality of the subsequent GNNs’ predictions and can result in twofold increase in errors due to the preferential selection of high-symmetry structures.

CatTSunami: Accelerating Transition State Energy Calculations with Pretrained Graph Neural Networks

CatTSunami: Accelerating Transition State Energy Calculations with Pretrained Graph Neural Networks

CiRLExplainer: Causality-Inspired Explainer for Graph Neural Networks via Reinforcement Learning.

In this article, we propose a new graph neural network (GNN) explainability model, CiRLExplainer, which elucidates GNN predictions from a causal attribution perspective. Initially, a causal graph is constructed to analyze the causal relationships between the graph structure and GNN predicted values, identifying node attributes as confounding factors between the two. Subsequently, a backdoor adjustment strategy is employed to circumvent these confounders. Additionally, since the edges within the graph structure are not independent, reinforcement learning is incorporated. Through a sequential selection process, each step evaluates the combined effects of an edge and the previous structure to generate an explanatory subgraph. Specifically, a policy network predicts the probability of each candidate edge being selected and adds a new edge through sampling. The causal effect of this action is quantified as a reward, reflecting the interactivity among edges. By maximizing the policy gradient during training, the reward stream of the edge sequence is optimized. The CiRLExplainer is versatile and can be applied to any GNN model. A series of experiments was conducted, including accuracy (ACC) analysis of the explanation results, visualization of the explanatory subgraph, and ablation studies considering node attributes as confounding factors. The experimental results demonstrate that our model not only outperforms current state-of-the-art explanation techniques, but also provides precise semantic explanations from a causal perspective. Additionally, the experiments validate the rationale for considering node attributes as confounding factors, thereby enhancing the explanatory power and ACC of the model. Notably, across different datasets, our explainer achieved improvements over the best baseline models in the ACC-area under the curve (AUC) metrics by 5.89%, 5.69%, and 4.87%, respectively.

A Graph Neural Network Model for Financial Fraud Prevention

Financial fraud prevention is a critical challenge for banks, payment processors, and online financial services. Traditional fraud detection models, including rule-based systems and machine learning classifiers, often struggle with adaptive fraud tactics, requiring frequent retraining to remain effective. Recent advancements in graph neural networks (GNNs) have enabled fraud detection models to leverage relational transaction data, capturing multi-hop fraud patterns and collusive fraud schemes that are difficult to detect with conventional approaches. This study proposes a GNN-based fraud prevention framework that models financial transactions as a heterogeneous graph, where nodes represent users and transactions, while edges encode financial relationships such as payment frequency, transaction amount similarity, and shared device usage. The GNN model learns fraud indicators by aggregating information from neighboring transactions, allowing it to detect complex fraud networks and coordinated money laundering activities. The proposed system was evaluated on large-scale transaction datasets, demonstrating higher fraud detection accuracy and lower false positive rates compared to traditional fraud detection models. The results confirm that graph-based fraud detection improves scalability and adaptability, making it a more effective approach for modern financial institutions seeking real-time fraud prevention solutions.

Graph attention and Kolmogorov–Arnold network based smart grids intrusion detection

The digital revolution in power systems has increased their complexity and interconnectivity, thereby exacerbating the risk of cyberattacks. To protect critical power infrastructure, there is an urgent need for an advanced intrusion detection system capable of capturing the intricate interactions within smart grids. Although traditional graph neural network (GNN)-based methods have exhibited substantial potential, they primarily rely on network data (e.g., IP addresses and ports) to construct the graph structure, failing to adequately integrate physical data from power grid devices. Moreover, these methods typically employ fixed activation functions in downstream deep networks, which limits the accurate representation of complex nonlinear attack patterns, thereby reducing detection accuracy. To address these challenges, this paper introduces GraphKAN, a novel intrusion detection framework that leverages graph attention network (GAT) and Kolmogorov–Arnold network (KAN) to enhance detection precision in smart grids. GraphKAN firstly constructs a graph-structured representation with power devices, information technology devices, and communication network devices as nodes, and integrates the physical connections and logical dependencies among infrastructure elements as edges, providing a comprehensive view of device interactions. Furthermore, the GAT module utilizes multi-head attention mechanisms to dynamically allocate node weights, extracting global features that encompass both feature information and interaction patterns. The KAN introduces learnable activation functions based on parameterized B-splines, enhancing the nonlinear expression of the global features extracted by GAT and significantly improving the detection accuracy of complex attack patterns. Experiments conducted on datasets obtained from Mississippi State University and Oak Ridge National Laboratory demonstrate that GraphKAN achieves detection accuracies of 97.63%, 98.66%, and 99.04% for binary, ternary, and 37-class intrusion detection tasks, respectively. These results represent substantial improvements over state-of-the-art models, including GA-RBF-SVM, BGWO-EC, and Net_Stack, with accuracy gains of 5.73%, 0.89%, and 3.52%, respectively. The findings underscore the efficacy of GraphKAN in enhancing intrusion detection accuracy in smart grids and its robust performance in complex attack scenarios.

Scalable Blockchain Fraud Detection Using Spatial-Temporal Graph Neural Networks

The increasing adoption of blockchain technology has led to a surge in financial fraud, including money laundering, Ponzi schemes, and illicit fund transfers. Traditional fraud detection techniques, such as rule-based systems and supervised machine learning models, struggle to handle the high-volume, high-velocity, and dynamically evolving nature of blockchain transactions. These limitations necessitate a scalable and adaptive approach to detect fraudulent activities efficiently. This study introduces a Spatial-Temporal Graph Neural Network (STGNN)-based fraud detection framework, specifically designed for scalable anomaly detection in large-scale blockchain networks. By modeling blockchain transactions as a spatial-temporal graph, the proposed system captures structural dependencies between wallets and temporal patterns of fund movements. The STGNN model employs graph convolutional networks (GCN) or graph attention networks (GAT) for spatial feature extraction and gated recurrent units (GRU) or temporal convolutional networks (TCN) for sequential fraud pattern recognition. Additionally, to ensure scalability, the framework incorporates graph partitioning techniques, parallelized mini-batch training, and distributed processing, enabling real-time fraud detection across high-throughput blockchain networks. Extensive experiments conducted on Bitcoin and Ethereum transaction datasets demonstrate that the STGNN model achieves higher accuracy, lower false positive rates, and improved computational efficiency compared to rule-based fraud detection systems, supervised ML models, and static GNNs. Case studies further confirm the model’s effectiveness in detecting large-scale fraud schemes, such as DeFi exploits, cross-chain laundering, and coordinated illicit transactions. This research highlights the potential of graph-based deep learning techniques in blockchain security, providing a foundation for future advancements in scalable fraud detection, cross-chain anomaly detection, and decentralized financial security monitoring.

Image Object Detection Technology Based on Graph Neural Network

Image Object Detection Technology Based on Graph Neural Network

A deep learning architecture for aligning cross-domain geographic knowledge graph

Geographic knowledge graph (GeoKG) alignment is important for the integration and knowledge discovery of multisource geographic information and the generation of large-scale and high-quality knowledge graphs (KGs). However, the existing models/technologies face many challenges when dealing with large-scale multisource complex GeoKG alignment tasks, including the inconsistency of attribute and relationship values caused by domain differences, the inability to perceive relationships and entities, and missing geographic domain training data. To address these issues, we propose a GeoKG alignment model based on depth relationships and neighborhood awareness (named DRNA-GCNE). The DRNA-GCNE model adopts a graph neural network as the infrastructure and uses the graph attention technique to evaluate and weight the entity’s relationship attributes dynamically, thus enhancing the ability to perceive structural and semantic information in the GeoKG; concurrently, the relationship information and the multihop neighbor characteristics of the entity are effectively integrated, and the representation of the entity is further enriched. Finally, the training technique of normalized loss mining for multiple negative samples is shown. This approach increases the model’s capacity for generalization. The DRNA-GCNE model, as evaluated on two public datasets and our GeoEA2024 Chinese dataset, significantly outperforms current GeoKG entity alignment methods across key metrics.