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

Researchers have introduced BrainGFM, a graph-based foundation model for neuroscience trained on a massive dataset of over 25,000 subjects and 60,000 fMRI scans. The model, pre-trained on 27 datasets covering 25+ neurological and psychiatric disorders, uses graph contrastive learning and masked autoencoders. It integrates graph and language prompts for flexible adaptation, enabling few-shot and zero-shot learning to generalize to new, unseen disorders, marking a significant step toward large-scale brain analysis.