henhaobio
Sign in

bioRxivNeurosciencePreprint

BrainLM: A foundation model for brain activity recordings

Josue Ortega CaroiD, Antonio H. de O. Fonseca, Christopher L. AverilliD, Syed Asad RizviiD, Matteo Rosati, James M. CrossiD, Prateek MittaliD, Emanuele ZappalaiD, Daniel Levine, Rahul M. DhodapkariD, Insu HaniD, Amin KarbasiiD, Chadi G. AbdallahiD, David van DijkiD

PDFDOI
Structured summary available · Jump to summary ↓

Abstract

AO_SCPLOWBSTRACTC_SCPLOWWe introduce the Brain Language Model (BrainLM), a foundation model for brain activity dynamics trained on 6,700 hours of fMRI recordings. Utilizing self-supervised masked-prediction training, BrainLM demonstrates proficiency in both fine-tuning and zero-shot inference tasks. Fine-tuning allows for the accurate prediction of clinical variables like age, anxiety, and PTSD as well as forecasting of future brain states. Critically, the model generalizes well to entirely new external cohorts not seen during training. In zero-shot inference mode, BrainLM can identify intrinsic functional networks directly from raw fMRI data without any network-based supervision during training. The model also generates interpretable latent representations that reveal relationships between brain activity patterns and cognitive states. Overall, BrainLM offers a versatile and interpretable framework for elucidating the complex spatiotemporal dynamics of human brain activity. It serves as a powerful "lens" through which massive repositories of fMRI data can be analyzed in new ways, enabling more effective interpretation and utilization at scale. The work demonstrates the potential of foundation models to advance computational neuroscience research.

Abstract summary

Claude Opus 5.5 · xhigh effort · 18 Oct 2026 · from the abstract · CC BY-NC-ND 4.0

A masked-prediction model pretrained on 6,700 hours of fMRI that fine-tunes to predict clinical variables, forecasts future brain states, and recovers functional networks without supervision.

Massive repositories of functional MRI (fMRI) recordings exist but, as the paper frames it, need new ways of being analysed at scale. BrainLM applies self-supervised masked-prediction pretraining to brain activity dynamics so that a single model can be fine-tuned for clinical prediction or queried without task-specific training.

Key results

  1. ClaimedAfter self-supervised masked-prediction pretraining, the model can be fine-tuned to predict participant-level clinical variables including age, anxiety and post-traumatic stress disorder (PTSD), and to forecast future brain states.

    Pretraining corpus is 6,700 hours of fMRI recordings; the abstract reports no accuracy, correlation or error figures for any of these predictions, and no baseline comparison. · Abstract

  2. ClaimedThe fine-tuned model is reported to transfer to external cohorts that were not part of training.

    Generalisation to new cohorts is asserted; the cohorts are not named and no held-out performance numbers are given in the abstract. · Abstract

  3. ClaimedWithout any network-level labels during training, the model is said to pick out intrinsic functional networks from raw recordings, and its latent space is described as relating activity patterns to cognitive states.

    Presented as zero-shot behaviour and as qualitative interpretability of the representation; no network-recovery metric or comparison with established parcellations appears in the abstract. · Abstract

Methods, data and reproducibility

A foundation model for brain activity dynamics trained by self-supervised masked prediction on a large corpus of functional MRI recordings, then evaluated in two regimes: fine-tuning for clinical-variable prediction and brain-state forecasting, and zero-shot inference for functional-network identification and inspection of latent representations.

  • fMRI recordings used for self-supervised pretraining · human · n = 6,700 hours
  • External cohorts not seen during training, used to test generalisation; not named in the abstract · human
Methods & tools
BrainLM, functional MRI
Limitations (2)
  • Not one quantitative result appears in the abstract: no prediction accuracies, forecasting errors or comparisons with task-specific models or standard fMRI analyses. · noted in this summary
  • Neither the pretraining corpus nor the external validation cohorts is identified, and no code, weights or data location is given. · noted in this summary

Read this if you work with large fMRI archives and want to know what a pretrained sequence model claims to offer over per-study pipelines for prediction and network discovery.

Citations
50

Reproducibility ledger

Nobody has reported trying to reproduce this paper yet. Did you replicate an experiment or run the code? Your report saves the next lab a week.

Sign in to file a reproducibility report.

Related papers

Similarity Map
SelectedSame fieldCross-field
THISWang '26Nakagi '24Donoso '25Bhattacharj '24Wang '23Jain '20
68% semantic matchNeuroscience
Yifan Wang, Shaonan Wang, Wenhao Cai+13 more·bioRxiv·Apr 8, 2026· 2 citations