Scripts and notebooks to reproduce results for cellina, including counterfactual LOO benchmarking, ablation studies, disentanglement benchmarking, and application analyses on the CRC spatial transcriptomics dataset.
| Path | Description |
|---|---|
scripts/ |
Core training, evaluation, benchmarking, and utility scripts |
notebooks/ablations/ |
Lambda-sweep ablation studies for model components |
notebooks/application/ |
End-to-end CRC case study with pathway and counterfactual analysis |
notebooks/disentanglement/ |
Latent disentanglement benchmark |
notebooks/loo_benchmarks/ |
Leave-one-celltype-out benchmark suite |
notebooks/neighb_perts/ |
Gene-range neighborhood perturbation studies |
notebooks/spatial_ablations/ |
Spatial loss and link-prediction-weight ablations |
environments/ |
Conda environment files |
Each baseline or model variant has its own conda environment:
conda env create -f environments/<env>.yml| Environment | Used for |
|---|---|
cellina.yml |
Main Cellina training and evaluation |
cellina_graph.yml |
CellinaGraph (+ GCN) experiments |
cpa_env.yml |
CPA |
spatialprop_env.yml |
SpatialProp |
mintflow_env.yml |
MintFlow |
cellina.yml |
scGEN (via pertpy) |
- Ablation study of each component of
cellina— classifier, discriminator and edge loss - Marginal log likelihood for unseen region/cell type combination
- Benchmark measuring disentanglement on 3 datasets
- Counterfactual prediction of cell states in unseen niches