Pretrained checkpoints, preprocessed cohort data, and the curated pathway graph for ProtoPathway, an interpretable-by-design multimodal framework for cancer survival prediction.

Layout

pathways/pathways_base_*.pkl           curated Reactome + Hallmark pathway graph
raw_inputs/                            raw files for re-running preprocessing from scratch
    Reactome/                          Reactome hierarchy files (GMT, relations, names)
    Hallmark/                          MSigDB Hallmark gene sets
    {cohort}/                          rna_clean.csv, clinical CSV, SurvPath splits
cohorts/{cohort}/                      preprocessed cohort data and trained models
    gene_expression.csv                preprocessed expression matrix
    bipartite_graph.pt                 cohort-specific gene-pathway graph
    labels.csv                         survival times, events, and bins
    data_splits.pkl                    5-fold CV splits (SurvPath-compatible)
    checkpoints/best_fold_{0..4}.pt    trained model weights

Cohorts

Five TCGA cohorts: BRCA (N=714), BLCA (N=359), COADREAD (N=227), HNSC (N=392), STAD (N=318). Gene expression is the preprocessed SurvPath release. WSI patch features (UNI2-h) are not redistributed here and should be obtained from the Mahmood Lab directly.

Quick load

python

from huggingface_hub import snapshot_download

# Everything for one cohort plus the shared pathway file
snapshot_download(
    repo_id="AmayaGS/ProtoPathway",
    local_dir="./assets",
    allow_patterns=["cohorts/TCGA-BLCA/*", "pathways/*"],
)

Citation

bibtex

@article{protopathway2026,
  title   = {ProtoPathway: Biologically Structured Prototype-Pathway Fusion for Multimodal Cancer Survival Prediction},
  author  = {Amaya Gallagher-Syed, Costantino Pitzalis, Myles J. Lewis, Michael R. 
  Barnes, Gregory Slabaugh},
  journal = {arXiv preprint arXiv:2605.21454},
  year    = {2026},
}