BrainSymphony: A parameter-efficient multimodal foundation model for brain dynamics with limited data

Researchers have unveiled BrainSymphony, a new parameter-efficient multimodal foundation model for analyzing brain dynamics that requires substantially less data than current state-of-the-art models. The lightweight architecture combines fMRI time series and structural connectivity data through parallel transformer streams and a novel signed graph transformer. Despite its compact design, it consistently outperforms much larger models on benchmarks for prediction, classification, and network discovery, and has demonstrated interpretable results showing drug-induced cortical reorganization.