BrainGFM

BrainGFM is a graph foundation model for functional brain networks derived from resting-state and task fMRI. Most existing brain foundation models are pretrained on raw time-series signals or region-of-interest (ROI) feature vectors; BrainGFM instead treats each subject's functional connectome as a graph and learns transferable representations directly over that graph structure. This shift lets a single pretrained model adapt across many neurological and psychiatric conditions without bespoke, task-specific retraining for every new dataset or brain atlas.

The model was introduced by researchers at Lehigh University and Stanford University and published as a conference paper at ICLR 2026 (preprint posted to arXiv in June 2025). It addresses a long-standing fragmentation problem in fMRI-based machine learning: studies typically use different atlases, parcellations, and disorder labels, so models rarely transfer across cohorts. BrainGFM is pretrained on a deliberately heterogeneous mixture of atlases and parcellations to learn representations that generalize across these differences.

By coupling large-scale self-supervised pretraining with prompt-based adaptation, BrainGFM targets the small-sample regime that characterizes most clinical neuroimaging studies, where individual disorder cohorts often contain only tens to hundreds of subjects.

#Key Features

• Graph-native pretraining: Combines graph contrastive learning and a graph masked autoencoder to pretrain over functional connectomes, rather than over time-series or flat ROI features as in prior brain foundation models.
• Atlas- and parcellation-agnostic: Trained across 2 atlas types (functional and anatomical) and 8 widely used parcellations, with atlas/parcellation tokens so one model serves heterogeneous fMRI representations.
• Graph and language prompts: Integrates both graph prompts and language prompts, letting the model be steered toward a specific disorder or task via lightweight prompt tuning instead of full fine-tuning.
• Meta-learned prompts for generalization: Uses meta-learning to optimize the graph prompts, enabling few-shot and zero-shot transfer to previously unseen disorders through language-guided prompting.

#Technical Details

BrainGFM uses a Graph Transformer backbone with positional encoding and specialized tokens that encode atlas/parcellation identity and task/disorder context. Pretraining spans 27 neuroimaging datasets covering 25 neurological and psychiatric disorders, more than 25,000 subjects, roughly 60,000 fMRI scans, and about 400,000 graph samples aggregated across atlases and parcellations. The self-supervised objective mixes graph contrastive learning with a graph masked autoencoder, and downstream adaptation is performed through graph and language prompt tuning rather than retraining the full network. On the Schaefer-100 atlas, reported results include AUCs of roughly 70.3 on ADHD-200 (ADHD), 71.2 on ABIDE II (ASD), 80.3 on ADNI 2 (Alzheimer's disease), and 79.9-80.4 on Healthy Brain Network depression and OCD tasks, with the authors reporting state-of-the-art performance over time-series and connectome baselines across the evaluated disorders. A pretrained checkpoint is released alongside the code.

#Applications

BrainGFM is aimed at computational neuroscience and neuroimaging researchers who build fMRI-based classifiers for conditions such as ADHD, autism spectrum disorder, Alzheimer's disease, depression, and OCD. Because adaptation is prompt-based and supports few- and zero-shot settings, the model is most useful in the typical clinical-cohort regime where labeled subjects are scarce and where datasets were collected with different atlases or parcellations. It also provides a reusable pretrained backbone for transfer-learning studies that would otherwise train a separate model per cohort.

#Impact

BrainGFM extends the foundation-model paradigm into graph-structured functional neuroimaging, an area where transfer across atlases and disorders has been a persistent obstacle. Its combination of graph-native pretraining with meta-learned graph and language prompts offers a route to unified, atlas-agnostic brain-network models and a shared backbone for the neuroimaging community. As a recent ICLR 2026 contribution with a public code release and pretrained weights, its long-term adoption is still emerging, and reported gains are based on the authors' own benchmarks rather than independent replication; broader validation across external cohorts and clinical settings remains future work.