The conventional way to predict behavioural phenotypes from functional MRI is to apply Kernel Ridge Regression (KRR) on functional connectivity (FC) matrices. This approach requires a separate model to be trained and tuned for each behavioural score and each cohort, does not transfer across datasets, and its accuracy drops sharply when only a few subjects are available – a common situation in clinical and cognitive studies.
We introduce BrainPFN (Prior‑Data Fitted Network), a meta‑learning framework for brain‑behaviour regression. A domain‑specific prior is built from roughly 7 000 real FC matrices; from this prior we sample one billion synthetic training examples for meta‑training. At inference time a labelled context is fed through the network in a single forward pass, eliminating the need for per‑score or per‑cohort refitting.
Across four independent public cohorts, BrainPFN outperforms KRR in low‑sample regimes and surpasses recent foundation models that were pre‑trained on datasets four to nine times larger. Importantly, our meta‑training pool is assembled from open repositories and represents only a fraction of the data used by competing models that rely on restrictive agreements such as the UK Biobank. We release the pool, code, and checkpoints publicly, showing that a modest, openly available dataset can match or beat models trained at Biobank scale.
Review: BrainPFN leverages massive synthetic data and amortized inference to deliver fast, robust behavioural predictions, highlighting the strong potential of open data for building high‑performing brain science models.