PathLUPI Checkpoints

Official model checkpoints for PathLUPI, introduced in Genome-Anchored Foundation Model Embeddings Improve Molecular Prediction from Histology Images.

PathLUPI uses transcriptomic profiles as privileged information during training to learn genome-anchored representations from histology. At inference time, the model requires only WSI features extracted with CONCH; transcriptomic data are not required.

For model code, data preparation, and inference instructions, see the PathLUPI GitHub repository.

Available Checkpoints

This repository provides the five cross-validation folds for 37 TCGA-trained tasks reported in the paper:

• 24 biomarker and molecular prediction tasks
• 13 survival prognosis tasks
• 185 checkpoint files in total

External cohorts in the paper were evaluated with the checkpoint for the corresponding TCGA-trained task and therefore do not require separate checkpoint files.

Biomarker Prediction

Cancer type	Task
BLCA	FGFR3 mutation
BRCA	ER, HER2, PIK3CA, PR, TNBC, TP53
CRC	BRAF, KRAS, TP53
GBMLGG	IDH1 mutation
LIHC	TP53 mutation
LUAD	EGFR, KRAS, TP53
NSCLC	Tumor mutation burden (TMB)
SKCM	BRAF mutation

Molecular Subtyping

Directory	Cohort
BLCA_MolSub	Bladder cancer
BRCA_MolSub	Breast cancer
CRC_MolSub	Colorectal cancer
GBMLGG_MolSub	Glioma
HNSC_MolSub	Head and neck cancer
PanGI_MolSub	Pan-gastrointestinal cancers
UCEC_MolSub	Endometrial cancer

Survival Prognosis

BLCA, BRCA, CRC, GBM, HNSC, KIRC, LGG, LIHC, LUAD, LUSC, SKCM, STAD, and UCEC.

GBM and LGG are provided as separate survival models, consistent with the final analysis in the paper.

Repository Structure

PathLUPI/
├── subtyping/
│   ├── BRCA_ERSub/
│   │   ├── fold0.pth.tar
│   │   ├── fold1.pth.tar
│   │   ├── fold2.pth.tar
│   │   ├── fold3.pth.tar
│   │   └── fold4.pth.tar
│   └── ...
└── survival/
    ├── GBM/
    │   ├── fold0.pth.tar
    │   └── ...
    ├── LGG/
    │   ├── fold0.pth.tar
    │   └── ...
    └── ...

Download

Install the Hugging Face CLI:

pip install -U huggingface_hub

Download all checkpoints:

hf download peterjin0703/PathLUPI \
  --local-dir ./checkpoints/PathLUPI

Download one task only, for example BRCA ER prediction:

hf download peterjin0703/PathLUPI \
  --include "subtyping/BRCA_ERSub/*" \
  --local-dir ./checkpoints/PathLUPI

Download one survival model only:

hf download peterjin0703/PathLUPI \
  --include "survival/GBM/*" \
  --local-dir ./checkpoints/PathLUPI

Checkpoint Format

Each file is a PyTorch state_dict containing model weights only. Validation metrics, training epochs, optimizer states, and other training metadata are not stored in the checkpoint.

import torch

state_dict = torch.load(
    "checkpoints/PathLUPI/subtyping/BRCA_ERSub/fold0.pth.tar",
    map_location="cpu",
    weights_only=True,
)
model.load_state_dict(state_dict)
model.eval()

The model architecture and task-specific construction code are provided in the GitHub repository.

CONCH Dependency

These checkpoints operate on WSI features extracted using CONCH. The original CONCH weights are not redistributed here. Please obtain them directly from the MahmoodLab CONCH model page and follow its license and access requirements.

Citation

@article{jin2025pathlupi,
  title={Genome-Anchored Foundation Model Embeddings Improve Molecular Prediction from Histology Images},
  author={Jin, Cheng and Zhou, Fengtao and Yu, Yunfang and Ma, Jiabo and Wang, Yihui and Xu, Yingxue and others},
  journal={arXiv preprint arXiv:2506.19681},
  year={2025}
}

License

The PathLUPI checkpoints are released under the CC BY-NC-ND 4.0 license. CONCH is distributed separately under its own terms.