#3 · Systems biology

Pretrained generative models extend to fMRI forecasting, diagnosis, and imagery

The common thread is reuse of large-scale pretrained or generative models for neural data. Brain-DiT pretrains a diffusion transformer with metadata conditioning on 349,898 fMRI sessions across 24 datasets. A structure-aware VAE generates functional connectivity matrices for generative replay in continual fMRI diagnosis of brain disorders. Latent functional alignment adapts the DynaDiff perception decoder to reconstruct mental imagery from NSD-Imagery. OmniMouse applies the scaling question to mouse visual cortex, training multi-modal, multi-task models on 150B neural tokens from 3.1M neurons and reporting that performance scales with data rather than model size. The zeta-law framework addresses when added biomedical data improves cross-modal discoverability, based on covariance spectral decay. What is new is the extension of these strategies beyond perception decoding into diagnosis and imagery, together with explicit scaling analysis. Quantitative comparisons against baselines are not established by these summaries.