Source-linked AI summary
Graph Neural Networks in Network Neuroscience
Alaa Bessadok, Mohamed Ali Mahjoub, Islem Rekik
TL;DR
The review addresses how GNNs can support network neuroscience despite incomplete connectomic data and heterogeneous brain-graph representations. It surveys and taxonomizes GNN methods for prediction, classification, and integration, while identifying synthesis applications and scalability and data limitations. The review covers 30 papers, with about 36% studying brain-graph synthesis and cross-domain, cross-resolution, and cross-time synthesis as prominent applications.
Problem
Connectomic datasets often lack modalities or follow-up scans because medical imaging is costly, while limited training graphs hinder learning-based models.
Method
The paper systematically reviews and taxonomizes GNN-based network-neuroscience methods across brain-graph prediction, integration, and disease classification.
Results
About 36% of the 30 reviewed papers studied brain-graph synthesis, with cross-domain, cross-resolution, and cross-time synthesis identified as important applications.
Takeaways & Limitations
GNN applications in network neuroscience include generating missing brain graphs, classifying disease-related brain states, and integrating population brain graphs.
Takeaways & Limitations
Existing longitudinal prediction methods face scalability constraints, including costly edge convolutions on large graphs and limited joint prediction of follow-up graphs.
Abstract
from arXiv · showhide
Noninvasive medical neuroimaging has yielded many discoveries about the brain connectivity. Several substantial techniques mapping morphological, structural and functional brain connectivities were developed to create a comprehensive road map of neuronal activities in the human brain -namely brain graph. Relying on its non-Euclidean data type, graph neural network (GNN) provides a clever way of learning the deep graph structure and it is rapidly becoming the state-of-the-art leading to enhanced performance in various network neuroscience tasks. Here we review current GNN-based methods, highlighting the ways that they have been used in several applications related to brain graphs such as missing brain graph synthesis and disease classification. We conclude by charting a path toward a better application of GNN models in network neuroscience field for neurological disorder diagnosis and population graph integration. The list of papers cited in our work is available at https://github.com/basiralab/GNNs-in-Network-Neuroscience.
1 INTRODUCTION
Network neuroscience models the brain as a graph of anatomical regions linked by morphological, structural, or functional connectivities. The review motivates GNNs for learning these non-Euclidean structures and addressing incomplete, multimodal, multiscale, and temporal brain-graph data.
- Brain graphs represent anatomical regions as nodes and pairwise morphological, structural, or functional connectivities as edges.
- Morphological, structural, and functional brain graphs derive respectively from T1-weighted, diffusion-weighted, and resting-state functional MRI.
- Connectomic representations span domain, resolution, and time axes, enabling different connectivity types, higher-resolution graphs, and predicted topological changes.
- Incomplete observations arise from costly clinical scans and time-consuming neuroimaging preprocessing.
- Graph generative models address missing data by predicting multimodal, high-resolution, and temporal brain networks from single-domain, single-scale, or single-timepoint graphs.
- The review analyzes GNN-based methods for neuroscience problems because prior reviews generally treated GNNs or neuroscience topics separately rather than their specific applications.
2.1 GNN overview
GNNs apply geometric deep learning to graph-structured data by propagating and transforming node information. In brain graphs, the framework accommodates both individual connectomes and population graphs, including cases without natural node features.
- GNNs are a geometric-deep-learning approach for learning on directed and undirected non-Euclidean graphs.
- A brain connectome is represented as G = (N, E, A, F), with A encoding pairwise relationships between ROIs and F encoding node features.
- At the population level, a graph represents relationships among connectomes, with an affinity matrix capturing similarity between subjects.
- The GCN propagation rule produces learned graph representations using an activation function, graph-convolution weights, self-loops, and degree normalization.
2.2 Brain graph overview
Brain graphs are scalable connectome representations built by parcellating neuroimaging data into brain regions and connecting them with weighted edges. The review distinguishes graph modalities and organizes GNN applications into prediction, classification, and integration.
- Brain graphs use parcellated brain regions as nodes because modeling individual neurons and synapses is computationally expensive.
- Morphological brain graphs estimate interregional morphology from measures such as sulcal depth and cortical thickness in T1-weighted images.
- GNN-based network-neuroscience methods comprise brain graph prediction, disease classification, and brain graph integration.
- Functional brain graphs derive interregional relationships from correlations between ROI-averaged BOLD signals in functional MRI.
- Structural brain graphs derive from diffusion tensor or diffusion spectrum imaging, which measures water diffusion to characterize white-matter connectivity.
2.3 Literature search and taxonomy definition
The review systematically surveys GNN applications in network neuroscience and defines a taxonomy spanning prediction, integration, and classification. It further separates graph-based models from population-based models according to their input graphs.
- The review searched multiple electronic databases for GNN and connectome studies published from 1 January 2017 through 31 December 2020.
- Non-GNN brain-graph studies, including CNN-based and other machine-learning architectures, were excluded from the review.
- The taxonomy contains three application flavors: brain graph prediction, integration, and classification.
- Graph-based models learn from graphs whose nodes are anatomical regions and whose edges encode morphological, functional, or structural connectivity.
- Population-based models instead use subject graphs, where nodes represent individual brain graphs and edges quantify pairwise subject similarity.
2.4 Brain graph prediction
Brain graph prediction methods address incomplete connectomic data by mapping source graphs across domains, resolutions, or time, with GNN architectures supporting synthesis of missing brain-network information. The review highlights progress alongside persistent concerns about end-to-end training, topology preservation, scalability, and evaluation scope.
- Motivation: High medical-scan costs and incomplete follow-up or multimodal imaging leave real-world connectomic datasets incomplete, motivating brain graph synthesis.The review also notes that small training sets can produce poor results for learning-based models.
- Cross-domain graph prediction: Cross-domain prediction maps a source brain graph to a target graph, including single-target and multi-target prediction.Existing approaches include GANs with GCN-based generators and discriminators, autoencoders, and adversarially regularized multi-graph models.
- Challenges and insights: Reported challenges include accumulated errors, GAN mode collapse, limited topology preservation, costly large-graph convolutions, and evaluation of multigraph prediction on morphological data only.The review calls for scalable, topology-aware, multimodal, and broader connectomic evaluations.
- Cross-resolution graph prediction: Super-resolution prediction generates a high-resolution brain graph with more nodes and connectivities from a low-resolution graph.A graph U-autoencoder used GCN layers and a superresolving propagation rule based on graph eigendecomposition, with evaluation on a real-world connectomic dataset.
- Cross-time graph prediction: Brain graph evolution prediction maps a baseline graph to a trajectory of follow-up brain graphs across future timepoints.A cascaded GAN framework generates each timepoint graph using the previously synthesized graph, enabling end-to-end trajectory prediction.
2.5 Brain graph integration
Brain graph integration learns a representative connectional brain template from multimodal graphs across a population. Existing GNN-based integration can capture shared connectivity patterns, but current methods remain limited by fixed graph sizes, non-end-to-end training, and limited interpretability.
- Problem statement: The integration goal is to map multimodal brain graphs from all subjects in a population into one holistic fingerprint graph representing shared connectivity patterns.This representative connectional brain template is intended to help identify biological patterns differing between typical and atypical populations.
- Existing works: MGINet clusters subjects, extracts representative edges within clusters, and averages cluster-specific graphs to produce a population brain template.Its clustering, cluster-specific template generation, and population-template estimation blocks are learned separately.
- Challenges and insights: Current architectures generate templates only from multimodal fixed-size brain graphs, limiting integration of non-isomorphic multi-resolution graphs.They also do not explain which brain connectivity types contribute most to template learning.
2.6 Brain graph classification
Brain graph classification asks whether connectomic representations can distinguish brain states and identify disease-related biomarkers. The reviewed methods include whole-graph, embedding-based, and population-graph learning, with biomarker identification targeting clinically relevant connectivities and treatment response.
- Brain state classification and biomarkers: Brain graph classification distinguishes healthy and unhealthy subjects while seeking discriminative connectivities that fingerprint disease.The review divides this area into brain state classification and biomarker identification.
- Brain state classification: Brain state classification can learn from a whole-brain graph, its learned embedding, or a graph population whose nodes represent subjects’ brain graphs.These alternatives correspond to different levels of representation for predicting a subject’s brain state.
- Biomarker identification: Biomarker identification seeks discriminative biomarkers that may characterize disease-related brain connectivities and treatment response in clinical settings.The review summarizes disease-classification studies in Table 3.
A- Graph embedding-based classification
Graph embedding-based classification compresses brain graphs into low-dimensional embeddings and maps those embeddings to diagnostic labels. Reviewed approaches include graph-based and population-based learning, with attention mechanisms supporting interpretation of abnormal functional connections.
- A- Graph embedding-based classification: Brain graph embeddings provide compressed representations that can be mapped to labels such as normal control or autistic.
- A- Graph embedding-based classification: Graph-based models learn from anatomical brain regions as nodes and their morphological, functional, or structural connectivities as edges.
- A- Graph embedding-based classification: Population-based models instead represent subjects’ brain graphs as nodes and similarities between subjects as edges.
- A- Graph embedding-based classification: A GAN-based approach used GCN layers to learn graph representations before applying a linear SVM to classify connectomes.
- A- Graph embedding-based classification: An edge-weighted GAT followed by DiffPool learned embeddings and used attention maps to identify abnormal functional brain associations in bipolar disorder.
- A- Graph embedding-based classification: A remaining challenge is explaining why embedding-based models make their predictions while integrating different brain graph types.
B-Graph-based classification
Graph-based classification uses the whole brain graph rather than only learned graph embeddings to predict subject states. Existing models combine GNNs with recurrent components, but neuroscientifically interpretable explanations remain difficult.
- B-Graph-based classification: Graph-based classification leverages the whole brain graph to make the target prediction.
- B-Graph-based classification: GCN-LSTM models classified functional connectivity in demented and healthy subjects while regularizing with gender and age prediction.
- B-Graph-based classification: Other approaches combined GCN and recurrent neural network models for brain disease classification.
- B-Graph-based classification: A common limitation is that these models struggle to explain classification results in a neuroscientifically interpretable way.
C-Population-based classification
Population-based classification constructs graphs of subjects to combine imaging and non-imaging information for brain-state prediction. Connectome-level population modeling remains limited, while scarce datasets increase overfitting concerns.
- C-Population-based classification: Population-based classification constructs a subject graph whose nodes are brain graphs and whose edges encode relationships between subjects.
- C-Population-based classification: The population graph is mapped to labels such as normal control or autistic through a learned classification function.
- C-Population-based classification: Most reviewed population graphs used imaging and non-imaging data, representing subjects with neuroimaging and linking them through phenotypic similarity.
- C-Population-based classification: One reported model integrated functional and structural brain graphs to predict Alzheimer’s disease using calibrated adjacency and modality-specific GCNs.
- C-Population-based classification: Scarce connectomic datasets and missing instances make classification difficult because models tend to overfit.
- C-Population-based classification: Biomarker-identification models seek discriminative brain subgraphs in addition to predicting labels.
- C-Population-based classification: Message-passing and pooling architectures generate ROI subgraphs interpreted as candidate biomarkers.
- C-Population-based classification: Biomarker methods face reproducibility and explainability challenges, including instability across training strategies and limited insight into model decisions.
3 DISCUSSION AND OUTLOOK
The review organizes GNN network-neuroscience work around prediction, integration, and classification, and identifies clinical translation as constrained by data scarcity, model selection, reproducibility, and explainability. It highlights multimodal synthesis, few-shot learning, and topology-aware modeling as future directions.
- 3 DISCUSSION AND OUTLOOK: The review identified 30 papers published from 2017 through 2020 across brain graph prediction, integration, and disease classification.
- 3 DISCUSSION AND OUTLOOK: About 36% of reviewed papers studied brain graph synthesis, especially cross-domain, cross-resolution, and cross-time prediction.
- 3 DISCUSSION AND OUTLOOK: Morphological brain graphs were the most common modality, with synthesis models mainly trained on T1-weighted MRI-derived graphs.
- 3 DISCUSSION AND OUTLOOK: Existing synthesis studies generally do not combine multiple modalities to predict missing brain graphs from structural, functional, and diffusion MRI together.
- 3 DISCUSSION AND OUTLOOK: Among disease-classification studies, only four papers addressed interpretability, leaving reproducibility and explainability underdeveloped.
- 3 DISCUSSION AND OUTLOOK: Cortical parcellation effects on GNN learning and predictive performance remain largely unexamined outside cross-resolution work.
- 3 DISCUSSION AND OUTLOOK: 50% of frameworks used GCNs and 50% used other graph convolution operations, leaving model selection task-dependent and unresolved.