BrainSymphony: A parameter-efficient multimodal foundation model for brain dynamics with limited data
arXiv Preprint Archive June 23, 2025 Moein Khajehnejad, Forough Habibollahi, Devon Stoliker et al.
A lightweight foundation model called BrainSymphony integrates fMRI time series and diffusion-derived structural connectivity, enabling unimodal or multimodal training without architectural changes and requiring less data than larger models. It processes fMRI data through parallel spatial and temporal transformer streams, distills embeddings via a Perceiver module, and encodes anatomical connectivity with a signed graph transformer. The model outperforms larger counterparts on benchmarks for prediction, classification, and network discovery. Attention maps from an independent psilocybin dataset reveal drug-induced reorganization of cortical hierarchies, demonstrating interpretability and generalizability. The work shows that architecturally informed multimodal models can surpass much larger models, advancing AI applications in neuroscience.