BrainSymphony Reveals Psilocybin-Induced Network Reorganization
Introduction fMRI foundation models have grown rapidly in size [1-2], but scale may be suboptimal for neuroimaging given domain constraints and interpretability needs. BrainSymphony instead embeds neurobiological priors to build a lightweight, parameter-efficient, multimodal model capturing spatiotemporal BOLD dynamics and diffusion-MRI connectivity. We test out-of-distribution generalization on PsiConnect (62 participants; pre/post psilocybin; rest, meditation, music, movie) and relate model-derived dynamics to MEQ-30 mystical-type experience ratings. BrainSymphony reconstructs unseen fMRI time series with high fidelity and exposes fine-grained psychedelic network reorganization beyond classical functional connectivity. Methods BrainSymphony uses three fMRI encoders: an ROI-wise Spatial Transformer with mixed positional embeddings (including cortex-gradient priors), a Temporal Transformer with sinusoidal time embeddings, and a 1D convolutional stream for short-range transients. Encoder outputs are fused by a Perceiver module. We applied a frozen model pretrained on HCP/HCP-Aging to PsiConnect (baseline and psilocybin; rest, meditation, music, movie). Reconstruction was quantified by R², Pearson r, and MAE against ROI-shuffled controls. Directed influence matrices were computed from Perceiver attention (incoming/outgoing) and used to decode context and to characterize psilocybin–baseline reorganization, including modulation by experience intensity. Results BrainSymphony reconstructed unseen psychedelic fMRI with high fidelity; R² and Pearson r exceeded shuffled controls (Fig. 1). Directed influence matrices decoded rest/meditation/music/movie at baseline (~0.64) and remained above chance under psilocybin (~0.46; chance 0.25), consistent with reduced modular boundaries and increased integration. Psilocybin increased outgoing influence in DMN, Control, and Visual networks, suggesting reduced DMN autonomy. Incoming influence highlighted Visual cortex as a dominant driver even in eyes-closed states, with limbic/salience drivers strongest in affective contexts. These effects scaled with Mystical Experience Questionnaire intensity with conventional functional connectivity failing to identify them. Discussion Compact, domain-informed foundation models can be predictive and mechanistically useful. BrainSymphony's attention-derived directed influence reveals psilocybin-driven redistribution of large-scale communication that tracks behavioral context and subjective intensity of the experience. Visual territories emerge as consistent drivers even with closed eyes, aligning with internally generated imagery; limbic circuits amplify during emotionally rich contexts; DMN/Control/ systems show condition-specific increases in being influenced, reflecting decreased segregation and increased integration. BrainSymphony demonstrates that efficiency and interpretability can coexist with strong generalization. Acknowledgments A.R. is affiliated with The Wellcome Centre for Human Neuroimaging, supported by core funding from Wellcome [203147/Z/16/Z]. A.R. is a CIFAR Azrieli Global Scholar in the Brain, Mind & Consciousness Program. Figure Figure 1: BrainSymphony reconstruction and attention-based reorganization. (a) Paired dots: real vs permuted ROI series across conditions; higher R², r, lower MAE. (b) Circos: Admin–Baseline attention Δ (top 500 edges) colored by source network; inner track = total outgoing. (c) Network-mean incoming attention Δ. (d) High vs Low MEQ receptive-attention maps plus inter-network Δ matrices. References [1] Caro, J. O., Fonseca, A. H. D. O., Averill, C., Rizvi, S. A., Rosati, M., Cross, J. L., & van Dijk, D. (2023). BrainLM: A foundation model for brain activity recordings. BioRxiv, 2023-09[2] Dong, Z., Li, R., Wu, Y., Nguyen, T. T., Chong, J., Ji, F., & Zhou, J. H. (2024). Brain-jepa: Brain dynamics foundation model with gradient positioning and spatiotemporal masking. Advances in Neural Information Processing Systems, 37, 86048-86073