TabNado

Predicts binding from epigenomic cofactors (ChIP-seq, CUT&TAG, CUT&RUN) over tiled TSS windows or user-defined BED regions, with support for both GANDALF (neural tabular) and XGBoost backends.

Uses datasets prepared with QuantNado via SeqNado.

What It Does

1. Builds train/eval/test tabular datasets from genomic signal.
2. Runs hyperparameter sweeps (GANDALF or XGBoost backend).
3. Trains a final model with best hyperparameters.
4. Evaluates predictions and exports metrics/figures.
5. Computes SHAP feature importance outputs.

Pipeline at a Glance

1. Data prep and split: chr8 for eval, chr9 for test.
2. Sweep stage: identifies best hyperparameters.
3. Train stage: produces final_model/.
4. Evaluate stage: metrics, scatter, UMAP artifacts.
5. SHAP stage: feature importance and spatial SHAP artifacts.

Main Entry Points

• Initialise the parameters: tabnado-init
• Full pipeline: tabnado-run --params params.yaml
• Data only: tabnado-data --params params.yaml
• Sweep only: tabnado-sweep --params params.yaml
• Train only: tabnado-train --params params.yaml
• Evaluate only: tabnado-evaluate --params params.yaml
• SHAP only: tabnado-shap --params params.yaml

See the navigation for detailed setup and configuration.