Research map / Graph learning / Graph self-supervised learning
Biologically Plausible Brain Graph Transformer
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 BioBGT, a graph transformer for brain graphs that adds network-entanglement node importance encoding and a community-contrastive, functional-module-aware self-attention, reporting gains on ABIDE, ADNI and ADHD-200 disorder detection.
Problem
Existing graph transformers for brain graphs treat nodes uniformly and ignore small-world properties (hubs and functional modules), so learned representations lack biological plausibility and perform poorly on brain disorder detection.
Approach
Measure node importance via network entanglement (change in density-matrix spectral entropy when a node's local connections are perturbed) and encode it as a learnable input embedding; obtain functional-module-aware node representations via Louvain communities and a contrastive objective over two edge-dropped views, then use them in a kernel-style self-attention.
Key results
- BioBGT reaches 74.00% ACC on ABIDE and 71.06% ACC on ADHD-200, versus the strongest baseline BRAINNETTF at 68.24% ACC on ABIDE and 70.80% on ADHD-200.
- On ADNI (three-class), BioBGT achieves 52.08% ACC and 62.33% AUC, and the paper reports BioBGT's ACC as best across all three datasets.
- Ablation: replacing NE with degree, betweenness, closeness or eigenvector centrality, or Laplacian PE, lowers performance relative to full BioBGT on all datasets.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| use BioBGT instead of the strongest baseline | brain graph classification · ADHD-200 | accuracy: 70.80±2.70 -> 71.06±0.08 | Helped |
| use BioBGT instead of the strongest baseline | brain graph classification · ABIDE | accuracy: 68.24±2.24 -> 74.00±2.01 | Helped |
| use BioBGT instead of the strongest baseline | brain graph classification · ADNI | accuracy: 49.03±1.27 -> 52.08±2.08 | Helped |
| use BioBGT instead of the strongest baseline | brain graph classification · ADNI | AUC: 62.25±2.93 -> 62.33±5.98 | Helped |
| remove network entanglement-based node importance encoding (ablation) | BioBGT · ADHD-200, ABIDE, and ADNI | classification performance | Hurt |
| replace functional module-aware attention and its extractor with normal self-attention (ablation) | BioBGT · ADHD-200, ABIDE, and ADNI | classification performance | Hurt |
| replace NE encoding with Laplacian positional encoding (ablation) | BioBGT · ABIDE, ADNI, and ADHD-200 | accuracy: ABIDE 60.60±2.25 vs 74.00±2.01; ADNI 49.43±2.42 vs 52.08±2.08; ADHD-200 67.56±3.24 vs 71.06±0.08 | Mixed |
| replace NE encoding with degree centrality encoding (ablation) | BioBGT · ABIDE, ADNI, and ADHD-200 | accuracy: ABIDE 61.20±1.88 vs 74.00±2.01; ADNI 50.57±1.64 vs 52.08±2.08; ADHD-200 69.78±2.33 vs 71.06±0.08 | Mixed |
| replace NE encoding with Laplacian positional and degree centrality encodings (ablation) | BioBGT · ABIDE, ADNI, and ADHD-200 | accuracy: ABIDE 63.00±1.63 vs 74.00±2.01; ADNI 50.94±1.25 vs 52.08±2.08; ADHD-200 70.67±4.00 vs 71.06±0.08 | Mixed |
Surprising findings
- On ABIDE, BRAINNETTF has higher AUC (78.38%) than BioBGT (73.33%), and on ADHD-200 BRAINNETTF's AUC (79.36%) also exceeds BioBGT's (71.64%).
Limitations
- Computational cost of network entanglement and quadratic self-attention limits applicability, as acknowledged by the authors.
- Functional module labels for ADNI and ABIDE are unavailable, so Louvain communities are used, and empirical ADHD-200 labels are imprecise.
- Results are averaged over 10 random runs on small datasets with large standard deviations.
Experimental setup
Datasets: ABIDE ADNI ADHD-200
Benchmarks: ABIDE ADNI ADHD-200
Metrics: ACC AUC F1 sensitivity specificity
Baselines: SVM Random Forest SAN Graph Transformer Graphormer SAT BRAINNETTF Polynormer Gradformer GTSP GAT BrainGNN
Builds on: Graphormer SAN SAT BrainGB Louvain algorithm
Closest papers by meaning
- BrainHGT: A Hierarchical Graph Transformer for Interpretable Brain Network Analysis (AAAI 2026 · similarity 0.91)
- Long-range Brain Graph Transformer (NeurIPS 2024 · similarity 0.87)
- $\textit{NeuroPath}$: A Neural Pathway Transformer for Joining the Dots of Human Connectomes (NeurIPS 2024 · similarity 0.86)
- Do We Really Need Message Passing in Brain Network Modeling? (ICML 2025 · similarity 0.85)
- Information Bottleneck-Guided Adaptive Hypergraph Transformer for Brain Disease Diagnosis (NeurIPS 2026 · similarity 0.85)
- MAGNET: Manifold-Aware Graph Diffusion Network for Connectome Generation (NeurIPS 2026 · 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.