[Paper Review] PTGB: Pre-Train Graph Neural Networks for Brain Network Analysis

Yi Yang, Hejie Cui+1 more|arXiv (Cornell University)|May 20, 2023

Functional Brain Connectivity Studies4 citations

TL;DR

PTGB is a self-supervised graph neural network pre-training framework for brain network analysis that addresses data scarcity and inconsistent parcellation atlases through two-level contrastive learning and a data-driven atlas mapping pipeline. It achieves state-of-the-art performance across multiple brain disorder prediction tasks, with a 4.27% relative improvement over baselines on key metrics.

ABSTRACT

The human brain is the central hub of the neurobiological system, controlling behavior and cognition in complex ways. Recent advances in neuroscience and neuroimaging analysis have shown a growing interest in the interactions between brain regions of interest (ROIs) and their impact on neural development and disorder diagnosis. As a powerful deep model for analyzing graph-structured data, Graph Neural Networks (GNNs) have been applied for brain network analysis. However, training deep models requires large amounts of labeled data, which is often scarce in brain network datasets due to the complexities of data acquisition and sharing restrictions. To make the most out of available training data, we propose PTGB, a GNN pre-training framework that captures intrinsic brain network structures, regardless of clinical outcomes, and is easily adaptable to various downstream tasks. PTGB comprises two key components: (1) an unsupervised pre-training technique designed specifically for brain networks, which enables learning from large-scale datasets without task-specific labels; (2) a data-driven parcellation atlas mapping pipeline that facilitates knowledge transfer across datasets with different ROI systems. Extensive evaluations using various GNN models have demonstrated the robust and superior performance of PTGB compared to baseline methods.

Motivation & Objective

• To address the challenge of limited labeled data in brain network analysis, which hampers deep learning performance.
• To overcome inconsistencies in region-of-interest (ROI) parcellation systems across different neuroimaging datasets, which hinder knowledge transfer.
• To develop a pre-training framework that learns structural inductive biases from unlabeled brain networks without task-specific labels.
• To enable effective cross-dataset adaptation by learning a shared, interpretable mapping between different ROI atlases.
• To improve downstream disease prediction performance through joint optimization of graph-level and node-level representations.

Proposed method

• Proposes a two-level node contrastive sampling strategy that simultaneously optimizes local neighborhood and global graph-level representations.
• Introduces a multi-dataset, self-supervised pre-training objective based on mutual information maximization across multiple brain network graphs.
• Employs a data-driven atlas mapping pipeline using three specialized regularizers: sparsity, locality, and community regularizers to align ROIs across different parcellation templates.
• Uses variance-based sorting to identify and prioritize the most informative virtual ROIs for cross-atlas alignment.
• Applies MAML-inspired meta-learning principles to enable fast adaptation of pre-trained models to downstream tasks.
• Integrates a graph-level contrastive learning component that summarizes global graph structure into local node representations.

Figure 1: Overview of the proposed framework PTGB . The initial features of the source datasets are projected to a fixed dimension through atlas transformation followed by variance-based feature alignment, which facilitates self-supervised GNN pre-training on multiple datasets via the novel two-leve

Experimental results

Research questions

• RQ1 Can self-supervised pre-training on large-scale unlabeled brain networks improve downstream disease prediction performance?
• RQ2 How effective is two-level contrastive learning in capturing both local and global structural patterns in brain networks?
• RQ3 To what extent can a data-driven atlas mapping pipeline align ROIs across different parcellation systems without relying on manual annotations?
• RQ4 What is the contribution of each component in the PTGB framework—specifically, the regularizers and contrastive objectives—to overall performance?
• RQ5 How does PTGB generalize across diverse datasets with varying ROI definitions and clinical outcomes?

Key findings

• The full PTGB framework achieves a relative improvement of 4.27% in both accuracy and AUC over the strongest baseline, demonstrating the value of jointly learning shared substructure knowledge across multiple graphs.
• Removing the sparsity regularizer leads to the largest performance drop, confirming its critical role in robustness and overfitting prevention.
• The inclusion of locality and community regularizers significantly improves performance, validating the importance of spatial and modular brain structure awareness.
• Two-level contrastive learning (node and graph-level) converges faster and achieves better downstream adaptation than single-level alternatives.
• The variance-based ROI alignment method successfully identifies overlapping brain regions across PPMI, BP, and HIV datasets, with visualizations confirming regional correspondence.
• The framework maintains strong performance across three distinct datasets—PPMI, BP, and HIV—demonstrating its robustness and generalizability to diverse clinical and imaging contexts.

Figure 2: Visual demonstration of the sample types where $X_{i,p}$ is the anchor and $\mathbf{S_{1}}$ / $\mathbf{S_{4}}$ are sampled as 1-hop neighbors.