Research map / Probabilistic and scientific modeling / Molecules and proteins
Learning High-Order Relationships of Brain Regions
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 HyBRiD, which learns weighted hyperedges of brain regions end-to-end via learnable node masks and an information-bottleneck objective (multi-head drop-bottleneck) to predict cognition from fMRI correlation features.
Problem
Identifying informative high-order (hyperedge) relationships among brain regions from fMRI for phenotypic prediction, since pairwise connectivity methods miss them and enumerating high-order relations is exponential.
Approach
A Constructor learns K multi-head binary node masks as hyperedges and a Weighter computes a weight per hyperedge; training uses an information-bottleneck bound (multi-head drop-bottleneck) trading off informativeness about the target against redundancy with inputs, avoiding exponential search.
Key results
- On ABCD, HyBRiD's CPM r outperforms state-of-the-art predictive models on 7 of 8 datasets, average improvement 12.1%.
- On ABIDE, HyBRiD consistently outperforms all baselines on FIQ, VIQ, PIQ, average improvement 8.8% over state of the art.
- On synthetic data, HyBRiD achieves average 28.3% F1 improvement over strongest baselines.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| use learned hyperedges instead of the strongest predictive baseline | HyBRiD vs BrainNetTF · ABCD (7 of 8 tasks; Rest 1 is the exception) | CPM r: average improvement of 12.1% | Mixed |
| use learned hyperedges instead of the strongest predictive baseline | HyBRiD vs BrainNetGNN · ABIDE (FIQ, VIQ, PIQ) | CPM r: average improvement of 8.8% | Helped |
| add learned masking rather than expose all nodes to each head (ablation) | HyBRiD · ABCD SST 1, EN-back 1, MID 1, Rest 1, SST 2, EN-back 2, MID 2, Rest 2 | CPM r: 0.297_{±0.035} -> 0.361_{±0.058}; 0.274_{±0.057} -> 0.348_{±0.061}; 0.323_{±0.059} -> 0.386_{±0.060}; 0.221_{±0.034} -> | Helped |
| use learnable masks rather than random masks (ablation) | HyBRiD · ABCD SST 1, EN-back 1, MID 1, Rest 1, SST 2, EN-back 2, MID 2, Rest 2 | CPM r: 0.256_{±0.069} -> 0.361_{±0.058}; 0.191_{±0.046} -> 0.348_{±0.061}; 0.255_{±0.080} -> 0.386_{±0.060}; 0.190_{±0.051} -> | Helped |
| use thresholded binary masks rather than soft masks (ablation) | HyBRiD · ABCD (SST 1, EN-back 1, MID 1, Rest 1, SST 2, EN-back 2, MID 2, Rest 2) | CPM r: 0.343_{±0.042} -> 0.361_{±0.058}; 0.314_{±0.040} -> 0.348_{±0.061}; 0.320_{±0.055} -> 0.386_{±0.060}; 0.245_{±0.061} -> | Mixed |
| increase the number of hyperedge heads K (ablation) | HyBRiD · three datasets in the K ablation | CPM r | Mixed |
| increase the information-bottleneck trade-off coefficient β (ablation) | HyBRiD · ABCD datasets | CPM r | Mixed |
Surprising findings
- Higher-degree hyperedges are more significant, with a strong positive correlation between degree and significance.
- Even at a strict threshold of 1e-8, about 60% of learned hyperedges remain significant.
Limitations
- Only considers static high-order relations, not dynamic fMRI temporal evolution.
- Does not provide a direct method for interpreting complex high-order relationships.
- Rest 1 dataset is an exception where HyBRiD is second-best, hypothesized to stem from sparse resting-state connectivity.
Experimental setup
Datasets: ABIDE ABCD synthetic hypergraph dataset
Benchmarks: ABIDE ABCD
Metrics: CPM r MSE precision recall F1 runtime
Baselines: pairwise CPM kNN hypergraph l1 hypergraph l2 hypergraph BrainNetGNN BrainGNN BrainGB BrainNetTF HGCN
Builds on: information bottleneck drop-bottleneck connectome-based predictive modeling stop-gradient straight-through estimator
Closest papers by meaning
- Beyond Pairwise Connections: Extracting High-Order Functional Brain Network Structures under Global Constraints (NeurIPS 2025 · similarity 0.88)
- Information Bottleneck-Guided Adaptive Hypergraph Transformer for Brain Disease Diagnosis (NeurIPS 2026 · similarity 0.85)
- Brain Harmony: A Multimodal Foundation Model Unifying Morphology and Function into 1D Tokens (NeurIPS 2025 · similarity 0.85)
- In Silico Mapping of Visual Categorical Selectivity Across the Whole Brain (NeurIPS 2025 · similarity 0.85)
- FAST-Brain: A Flow-Aligned Spatio-Temporal Surrogate Brain Model (NeurIPS 2026 · similarity 0.84)
- $\textit{NeuroPath}$: A Neural Pathway Transformer for Joining the Dots of Human Connectomes (NeurIPS 2024 · similarity 0.84)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.