GNN-mpIF — Processed Graph Datasets

Preprocessed PyTorch Geometric cell-graph datasets for the capstone project "Enhancing Immunotherapy Predictions with Graph Neural Networks" — predicting breast-cancer receptor status (ER / PR / HER2) from multiplex immunofluorescence (mpIF) imaging using GNNs (a GCN baseline and a SPACE-GM/GIN model).

Code: github.com/hussenmi/capstone-mpif

Private dataset. These are derived artifacts from unpublished capstone work — do not redistribute without permission.

What's here

Each file is a serialized PyG InMemoryDataset (CellGraphDataset) — one graph per tissue region, with cells as nodes (biomarker expression as node features) and spatial neighborhood edges. Files follow PyG's root/processed/data.pt convention, one dataset per receptor target.

Path	Target	md5
training/processed/data.pt	base graphs (unlabeled / default)	b8ac4025…
training/ER_status/processed/data.pt	ER status	b8ac4025… (same graphs as base)
training/HR_status/processed/data.pt	HR status	753c2c8e…
training/PR_status/processed/data.pt	PR status	4ea63c2d…
training/HER2_status/processed/data.pt	HER2 status	a1c07e30…

Each file is 570 MB (2.85 GB total). The raw source data is not included, so these processed tensors are the only copy — they cannot be regenerated without the original mpIF inputs.

Restoring / loading

Download a target's data.pt back into a PyG root/processed/ layout, then load with the project's CellGraphDataset:

# fetch one target (e.g. ER) into a local root dir
hf download hussenmi/gnn-mpif --repo-type dataset \
  --include "training/ER_status/processed/data.pt" \
  --local-dir ./gnn-mpif-data

# or fetch everything
hf download hussenmi/gnn-mpif --repo-type dataset --local-dir ./gnn-mpif-data

# from the capstone-mpif repo (training/)
from create_graphs_for_classification import CellGraphDataset

# root must be the dir that CONTAINS processed/data.pt
dataset = CellGraphDataset(root="./gnn-mpif-data/training/ER_status", response_label_dict=None)
print(dataset, len(dataset), dataset.num_node_features, dataset.num_classes)

Plain PyG load (without the project class):

import torch
data, slices = torch.load("training/ER_status/processed/data.pt", weights_only=False)

Source & citation

Derived from multiplex immunofluorescence (mpIF) tumor-microenvironment data. The GIN backbone follows SPACE-GM (Wu et al., Nature Reviews Cancer 2023, s41568-023-00582-6). Please cite the original work when using these graphs.