Research map / Probabilistic and scientific modeling / Molecules and proteins

BrainWorld: A Structural-Prior-Conditioned Generative Model for Whole-Brain 4D fMRI Dynamics

NeurIPS 2026 · Probabilistic and scientific modeling › Molecules and proteins · approach: Generative protein structure design · 0 citations (OpenAlex)

Read the paperCheck your idea against 108,967 papers

AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Proposes BrainWorld, a latent diffusion transformer that generates long-horizon whole-brain 4D fMRI conditioned on past FC, sMRI (as structural prior inside the DiT denoiser), and optional stimulus embeddings, and reuses DiT features for downstream transfer.

Problem

Existing fMRI foundation models focus on representation learning and downstream prediction rather than condition-aware long-horizon generation of voxel-level whole-brain 4D fMRI; structure-function fusion is usually parallel or post-hoc.

Approach

Use a 4D fMRI VAE with a conditional DiT that injects sMRI as subject-level anatomical context throughout the denoising process (cross-attention plus adaptive normalization), autoregressively rolls out 40-frame latent blocks with FC-based context, and extracts intermediate DiT features for downstream tasks.

Key results

• Full fine-tuning: BrainWorld F+S-prior reaches R=0.962 on NKI age regression and 80.75% ACC on ABCD-style gender classification versus best baseline values around 0.946 R and 76.51% ACC.
• On 400-frame SALD generation, BrainWorld obtains MSE 0.218 and R 0.596 versus MONAI's 0.295 and 0.508 (R gain reported as favorable).
• Generated-example augmentation raises HCP-400 gender accuracy from 81.45% to 85.39% at GE=1.0 and SALD-400 Pearson from 0.7613 to 0.8104 at GE=0.75.

Surprising findings

• Concatenating sMRI features post hoc (F+S-concat) does not consistently improve over fMRI-only features, while structural-prior injection does.
• Most baselines degrade as generated-example ratio increases during augmentation.

Limitations

• Autoregressive diffusion rollout may accumulate errors and increase latency over very long horizons.
• Evaluates condition-aware prediction rather than explicit counterfactual or fine-grained stimulus control.
• Generated examples are label-preserving augmentations; structural benefit is computational evidence, not biological causality.

Experimental setup

Datasets: HCP SALD CineBrain HCP Movie ABIDE NKI ABCD BHRC ADNI CHCP HCP Task NSD

Benchmarks: HCP LR1 SALD CineBrain HCP Movie ABIDE NKI ABCD BHRC ADNI (MCI)

Metrics: MSE RMSE Pearson correlation ACC F1 Cohen's d

Baselines: BrainMAE BrainLM NeuroSTORM MONAI latent diffusion Brain-JEPA BrainMass Brain-DiT SWIFT SlimBrain Omni-fMRI Brain Harmony BrainIAC

Builds on: WF-VAE v-prediction diffusion Brain Harmony BrainLM MONAI latent diffusion

Closest papers by meaning

• BrainVista: Modeling Naturalistic Brain Dynamics as Multimodal Next-Token Prediction (NeurIPS 2026 · similarity 0.87)
• Brain Harmony: A Multimodal Foundation Model Unifying Morphology and Function into 1D Tokens (NeurIPS 2025 · similarity 0.87)
• BrainLM: A foundation model for brain activity recordings (ICLR 2024 · similarity 0.87)
• Synthesizing Realistic fMRI: A Physiological Dynamics-Driven Hierarchical Diffusion Model for Efficient fMRI Acquisition (ICLR 2025 · similarity 0.87)
• A scalable generative model for dynamical system reconstruction from neuroimaging data (NeurIPS 2024 · similarity 0.86)
• Brain-OF: An Omnifunctional Foundation Model for fMRI, EEG and MEG (NeurIPS 2026 · similarity 0.85)