Digital Spatial Pathway Mapping Reveals Prognostic Tumor States in Head and Neck Cancer

Digital Spatial Pathway Mapping Reveals Prognostic Tumor States in Head and Neck Cancer.
Julius Hense*, Mina Jamshidi Idaji*, Laure Ciernik, Jonas Dippel, Fatma Ersan, Maximilian Knebel, Ada Pusztai, Andrea Sendelhofert, Oliver Buchstab, Stefan Fröhling, Sven Otto, Jochen Hess, Paris Liokatis, Frederick Klauschen, Klaus-Robert Müller, Andreas Mock
* Equal contribution

https://github.com/bifold-pathomics/xMIL-Pathways

📄 Preprint: https://www.biorxiv.org/content/10.1101/2025.11.24.689710v1.abstract

@article {Hense2025.11.24.689710,
	author = {Hense, Julius and Idaji, Mina Jamshidi and Ciernik, Laure and Dippel, Jonas and Ersan, Fatma and Knebel, Maximilian and Pusztai, Ada and Sendelhofert, Andrea and Buchstab, Oliver and Fr{\"o}hling, Stefan and Otto, Sven and Hess, Jochen and Liokatis, Paris and Klauschen, Frederick and M{\"u}ller, Klaus-Robert and Mock, Andreas},
	title = {Digital Spatial Pathway Mapping Reveals Prognostic Tumor States in Head and Neck Cancer},
	year = {2025},
	doi = {10.1101/2025.11.24.689710},
	publisher = {Cold Spring Harbor Laboratory},
	journal = {bioRxiv}
}

Summary: In this work, we infer transcriptome-derived signaling pathway activities directly from routine H&E slides. We propose to use the MIL heatmaps for stratifying the patients. In this regard, we propose a spatial tumor pathway activity score (TAPAS) for quantifying spatially resolved pathway activity based on XAI heatmaps. In this repository, we share the codes for this manuscript. We have also shared data of an exemplar patient at Zenodo. We have prepared the pipelines in a way that they work smoothly with the shared data.

Usage

The workflow of this work is as follows:

1. Train a (Transformer-based) MIL model to predict the biomarker from H&E slide and create heatmaps for the test set. For this step, we used a codebase developed by us, publicly available at xMIL. Please also see our NeurIPS 2024 publication about explaining MIL models.
2. If you have multiple models (e.g., from your cross-validation training), aggregate the heatmaps. For this step, you can use the code in the folder heatmap_aggregation.
3. Perform tissue segmentation: for our subsequent analyses, we segment the H&E slide into tumor, non-tumor, and border. For this step, you can use the code in the folder tissue_segmentation.
4. You can perform the IHC-H&E analyses using the code at folder ihc_he_analyses. This includes IHC-H&E registration, aggregating the IHC activations within the patches of the H&E slide and overlapping the heatmap and IHC activations.
5. From the tumor area and the generated (aggregated) heatmap, you can compute TAPAS score. For this step, you can use the code in the folder tapas.
6. You can find the code for analyses doing patient stratification using TAPAS score and clinical metadata at patient_stratification.

Contact us

If you face issues using our codes, you can open an issue in this repository, or contact us:

📧 Julius Hense and Mina Jamshidi

License and citation

If you find our codes useful in your work, please cite us:

@article {Hense2025.11.24.689710,
	author = {Hense, Julius and Idaji, Mina Jamshidi and Ciernik, Laure and Dippel, Jonas and Ersan, Fatma and Knebel, Maximilian and Pusztai, Ada and Sendelhofert, Andrea and Buchstab, Oliver and Fr{\"o}hling, Stefan and Otto, Sven and Hess, Jochen and Liokatis, Paris and Klauschen, Frederick and M{\"u}ller, Klaus-Robert and Mock, Andreas},
	title = {Digital Spatial Pathway Mapping Reveals Prognostic Tumor States in Head and Neck Cancer},
	year = {2025},
	doi = {10.1101/2025.11.24.689710},
	publisher = {Cold Spring Harbor Laboratory},
	journal = {bioRxiv}
}

©️ This code is provided under CC BY-NC-ND 4.0. Please refer to the license file for details.