菜单
Use this model doc whenever the user wants to run BrainGNN for fMRI phenotype prediction, including graph construction, training, and evaluation. This document focuses on model-level usage and delegates upstream preprocessing to fmri-skill (and optionally hcpya-skill for HCP data).
用 Codex 或 Claude 帮你安装 复制这段 Prompt，粘贴到 Codex、Claude 或其他助手里，让它检查 Skill 页面并帮你完成安装。
基于 SOC 职业分类
Use this skill when users need to build, populate, or extend a domain-specific knowledge graph from literature and structured databases. Triggers include: 'build knowledge graph', 'extract claims from papers', 'ingest data into graph', 'batch extract claims', 'knowledge graph construction', 'populate graph from PubMed', 'extract structured claims', 'ingest atlas data', or any request involving knowledge graph population from scientific literature or biomedical databases. Covers both structured data ingestion (Phase 1) and LLM-based claim extraction from papers (Phase 2).
Use this skill whenever the user wants to formalize a network architecture and derive theoretical components from a research idea. Triggers include: 'method design', 'design method', 'network architecture', 'formula derivation', 'method-design', 'theoretical framework', 'derive equations', 'compare alternatives', 'architecture tournament', 'rank designs', 'design space exploration', or any request to transform IDEA.md into a detailed METHOD.md. This skill is the **mandatory interface-layer method formalizer** in NeuroClaw: it reads IDEA.md, optionally drafts N alternative architectures and ranks them via a Bradley-Terry pairwise tournament (Tournament Mode), designs concrete network structures (layers, modules, connections), performs mathematical derivations (equations, loss functions, proofs), and always outputs a structured METHOD.md.
Use this skill whenever the user wants to remove site/scanner/batch effects from neuroimaging features before running downstream models, run mega-analysis across multiple datasets, or evaluate models with leave-site-out / site-stratified protocols. Triggers include: 'harmonize', 'ComBat', 'CovBat', 'site effect', 'scanner effect', 'batch effect', 'leave-site-out', 'mega-analysis', 'multi-site', 'cross-site', 'neuroHarmonize'. This is a horizontal cross-cutting layer between dataset skills and model skills.
Use this skill whenever the user wants to run phenotype-prediction models, browse model cards, map model inputs/outputs, or choose an execution route for fMRI/sMRI based models. This is a model-entry orchestration skill: it routes requests to model-specific docs and delegates preprocessing to modality skills.
Use this model doc whenever the user wants to run Com-BrainTF (Community-aware Brain Transformer) for fMRI phenotype prediction. Com-BrainTF uses dense FC matrices with a two-level Transformer (per-community local + global) and DEC pooling. NeuroClaw auto-derives community partitions from atlas naming conventions (Yeo 7-net for Schaefer, lobe-based for AAL).
Use this model doc whenever the user wants to run IBGNN (Interpretable Brain Graph Neural Network) for fMRI phenotype prediction. IBGNN is a PyG-based GNN with a learnable MLP message function over [x_i, x_j, edge_attr], designed for connectome-based brain disorder analysis with post-hoc edge-mask explainer support.
| name | braingnn |
| description | Use this model doc whenever the user wants to run BrainGNN for fMRI phenotype prediction, including graph construction, training, and evaluation. This document focuses on model-level usage and delegates upstream preprocessing to fmri-skill (and optionally hcpya-skill for HCP data). |
| license | MIT License (NeuroClaw custom skill - freely modifiable within the project) |
| layer | base |
| skill_type | model |
| dependencies | ["fmri-skill","run_models"] |
BrainGNN is an interpretable graph neural network for fMRI analysis and phenotype prediction.
models/braingnn/ (Windows-compatible, 无需 torch_sparse)Research use only.
NeuroClaw 版本对原始 BrainGNN 做了以下关键改动：
augment_adj 不再使用 spspmm 做邻接矩阵平方，改用 add_self_loops + remove_self_loops，在 Windows 上可直接运行。torch.tanh() 还原为 Pearson r，对角线置零。--task 参数切换，regression 时 nclass=1，输出 raw scalar，用 MSELoss。pool.select.weight，forward 中做了兼容处理。--include-t1 将 z-scored GM volume 作为额外 1 维 node feature 拼接。neuroclaw (Python 3.11)data/braingnn_input/<atlas>/sub-*.pt 文件（由 fmri-skill 生成）data/t1_volume/<atlas>/sub-*.npz（GM volume）python models/braingnn/scripts/train.py \
--atlas schaefer_100_7net \
--labels-csv data/hcp_gender_labels.csv \
--subjects-file data/ready_subjects.txt \
--fold 0 --kfold 5 \
--n-epochs 50 --batch-size 16 --lr 0.005 \
--include-t1
python models/braingnn/scripts/train.py \
--atlas aal_116 \
--labels-csv data/hcp_age_labels.csv \
--subject-col subject_id --label-col age \
--task regression --label-scaling standardization \
--fold 0 --n-epochs 50
python models/braingnn/scripts/sweep_atlases.py
对所有可用 atlas 跑 fold 0，输出 CSV 对比表。
python models/braingnn/scripts/train.py --atlas aal_116 --dry-run
| 文件 | 作用 |
|---|---|
models/braingnn/net/braingnn.py | 模型定义：MyNNConv + TopKPooling + FC head + loss functions |
models/braingnn/scripts/data_adapter.py | 数据加载：NeuroClaw .pt -> PyG InMemoryDataset |
models/braingnn/scripts/train.py | 训练入口：K-fold CV, classification/regression |
models/braingnn/scripts/sweep_atlases.py | Atlas 对比扫描脚本 |
data/braingnn_input/<atlas>/sub-<id>.pt){
"subject_id": str,
"atlas": str,
"n_rois": int,
"time_series": Tensor[T, n_roi],
"fc_matrix": Tensor[n_roi, n_roi], # Fisher-z transformed
"node_features": Tensor[n_roi, n_roi],
"edge_index": Tensor[2, n_edge],
"edge_attr": Tensor[n_edge, 1],
"roi_names": list[str],
}
data/t1_volume/<atlas>/sub-<id>.npz){
"subject_id": str,
"atlas": str,
"n_rois": int,
"roi_names": list[str],
"gm_volume_mm3": ndarray[n_roi],
"gm_fraction": ndarray[n_roi],
}
models/braingnn/checkpoints/<atlas>/fold<N>.pt){
"state_dict": OrderedDict,
"args": dict,
"best_val_acc": float, # classification
"test_acc": float,
"n_roi": int,
"indim": int,
}
| 参数 | 默认值 | 说明 |
|---|---|---|
--atlas | (必填) | atlas 名，需匹配 data/braingnn_input/ 下子目录 |
--fold | 0 | CV fold index |
--kfold | 5 | 总 fold 数（自动根据最小类别数调整） |
--n-epochs | 50 | 训练轮数 |
--batch-size | 16 | batch size |
--lr | 0.01 | 学习率 |
--ratio | 0.5 | TopKPooling 保留比例 |
--n-communities | 8 | MyNNConv 中间层 community 数 |
--include-t1 | False | 是否融合 T1 GM volume |
--task | classification | classification / regression |
--lamb3/4/5 | 0.1 | topk_loss 和 consist_loss 权重 |
n_roi 和 indim，无需手动指定。data/braingnn_cache/<atlas>/ 目录，否则会加载旧缓存。sweep_atlases.py 已自动处理。roi_mask 参数做子图选择，用于消融实验。fmri-skillhcpya-skill / hcpa-skill / hcpd-skill / hcpep-skilldependency-planner + conda-env-managerclaw-shell执行前需明确计划确认。
Created At: 2026-03-28 19:53 HKT Last Updated At: 2026-05-15 22:50 HKT Author: chengwang96