A Brain Graph Foundation Model: Pre-Training and Prompt-Tuning across Broad Atlases and Disorders

TL;DR

BrainGFM introduces a graph-based brain foundation model pre-trained on a large, multi-atlas fMRI corpus to address data heterogeneity and limited cross-parcellation transfer. It fuses Graph Contrastive Learning and Graph Masked Autoencoders within a Graph Transformer backbone, augmented by atlas/parcellation and task/disorder prompts, plus meta-learning for few-shot and language prompts for zero-shot transfer. Across 25 disorders and 8 parcellations, BrainGFM demonstrates superior generalization and efficiency compared with time-series and Connectome/FC-based FMs, while enabling flexible downstream adaptation through graph and language prompts. This work offers a scalable, generalizable framework for multi-atlas neuroimaging analysis with broad translational potential in clinical neuroscience and brain-computer interfacing.

Abstract

As large language models (LLMs) continue to revolutionize AI research, there is a growing interest in building large-scale brain foundation models to advance neuroscience. While most existing brain foundation models are pre-trained on time-series signals or connectome features, we propose a novel graph-based pre-training paradigm for constructing a brain graph foundation model. In this paper, we introduce the Brain Graph Foundation Model, termed BrainGFM, a unified framework that leverages graph contrastive learning and graph masked autoencoders for large-scale fMRI-based pre-training. BrainGFM is pre-trained on a diverse mixture of brain atlases with varying parcellations, significantly expanding the pre-training corpus and enhancing the model's ability to generalize across heterogeneous fMRI-derived brain representations. To support efficient and versatile downstream transfer, we integrate both graph prompts and language prompts into the model design, enabling BrainGFM to flexibly adapt to a wide range of atlases, neurological and psychiatric disorders, and task settings. Furthermore, we employ meta-learning to optimize the graph prompts, facilitating strong generalization to previously unseen disorders under both few-shot and zero-shot learning conditions via language-guided prompting. BrainGFM is pre-trained on 27 neuroimaging datasets spanning 25 common neurological and psychiatric disorders, encompassing 2 types of brain atlases (functional and anatomical) across 8 widely-used parcellations, and covering over 25,000 subjects, 60,000 fMRI scans, and a total of 400,000 graph samples aggregated across all atlases and parcellations.

Figures (10)

• Figure 1: The pipeline of our proposed BrainGFM. (a) A large-scale brain fMRI graph dataset is constructed for pre-training. (b) BrainGFM is pre-trained using graph contrastive and masked autoencoder strategies, with atlas/parcellation tokens [A/P] to encode atlas-specific information. (c) We introduce graph prompts and use meta-learning to optimize them for few-shot adaptation, keeping the graph FM backbone frozen. (d) Finally, we freeze both the model and graph prompts, and use language prompts to enable zero-shot transfer to new tasks. Note that "Schf." means Schaefer atlas.
• Figure 2: Performance comparison across different settings (Full-Shot, Few-Shot, Zero-Shot) on three datasets: ABIDE II, ADHD 200, and ADNI 2. The results demonstrate the progressive performance gains achieved by incorporating graph prompts (G-Prompt), meta-learning (Meta L.), and language prompts (Lan. Prompt) into the FM (BrainGFM), especially in few-shot and zero-shot settings.
• Figure 3: The performance of models pre-trained on different atlases varies across downstream atlases. The experiments are conducted on ABIDE II dataset for ASD classification.
• Figure 4: Comparison of performance and efficiency across different FMs.
• Figure 5: Performance of different graph pre-training methods.