How yanghanwu/graphtgi works
GraphTGI is a research model for predicting transcription-factor (TF) to target-gene interactions. The principal workflow samples known and negative pairs, assembles similarity features and interaction edges into a graph, trains and evaluates a graph-attention model across five folds, then reports classification metrics and plots ROC curves. Similarity matrices and association pairs are visible in the data tree; the listed TRRUST file’s role is not established by the sampled code. The graph reflects the shown calls; the data-loading code uses hard-coded paths, so runtime path resolution is uncertain.
Main entry points: train.py Fold training · main.py Run and plot · model.py GraphTGI model · utils.py Load similarities
Components 13
Workflow
- Run and plot
main.py - Fold training
train.py
Data inputs
- Input datasets
data - Load similarities
utils.py - Sample pairs
utils.py - Build interaction graph
utils.py
Graph learning
- GAT encoder
model.py - Node embeddings
model.py - Bilinear decoder
model.py - GraphTGI model
model.py
Evaluation
- Fold metrics
train.py - ROC visualization
main.py
Other components
- Research user
Connections 16
- Research user → Run and plot: runs
- Run and plot → Fold training: invokes
- Fold training → Build interaction graph: builds graph
- Fold training → Sample pairs: samples pairs
- Build interaction graph → Load similarities: loads matrices
- Load similarities → Input datasets: reads similarities
- Sample pairs → Input datasets: reads pairs
- Fold training → GraphTGI model: trains and scores
- GraphTGI model → GAT encoder: calls encoder
- GAT encoder → Node embeddings: produces
- GraphTGI model → Bilinear decoder: scores pairs
- Fold training → GAT encoder: encodes folds
- Fold training → Bilinear decoder: scores test pairs
- Fold training → Fold metrics: computes metrics
- Fold metrics → Run and plot: returns scores
- Run and plot → ROC visualization: plots curves