Session 1: Artificial Intelligence & Cancer

Cellohood: Discovery of immunological cellular neighborhoods from protein markers in spatial tumor data

Marcin MOZEJKO¹, Bodenmiller BERND^2,3, Ewa SZCZUREK ^1,6, Krzysztof GOGOLEWSKI¹, Daniel SCHULZ^2,3, Nils ELING^2,3, Joanna KRAWCZYK¹, Michelle DANIEL^2,3, Eikie STAUB⁴, Henoch HONG⁴, Marie MORFOUACE⁵

¹University of Warsaw, Warsaw, Poland
² Department of Quantitative Biomedicine, University of Zurich, Zurich, Switzerland
³ Institute for Molecular Health Sciences, ETH Zurich, Zurich, Switzerland
⁴Merck KGaA, Darmstadt, Germany
⁵Gustav Roussy Institute, Paris, France
⁶AI for Health, Helmholtz, Munich, Germany

Introduction

Spatial imaging of single cells and their protein markers in tumor tissues provides crucial insights into tumor-microenvironment interactions. While cellular neighborhood analysis is key to understanding these mechanisms, current approaches rely on predefined cell types or neighborhood-wide marker aggregations, which sacrifice valuable single-cell resolution data.

Method

We developed Cellohood, an innovative AI tool utilizing a permutation-invariant, transformer-based autoencoder designed for cellular neighborhood modeling. The system compresses information about individual cells and their marker expression within local environments, delivering both neighborhood-level representations and single-cell profiles. From these representations, we derived novel cellular neighborhood prototypes characterized by cell types, protein markers, and spatial arrangements.

The model's architecture enables flexible multi-resolution analysis, allowing researchers to investigate both broad tissue organization as well as detailed cellular interactions. As a result, patients can be effectively described through the abundance and arrangements of cellular neighborhood prototypes.

Results

We demonstrated Cellohood's capabilities across multiple spatial imaging technology datasets:

- Mouse spleen lupus CODEX data,

- DLPFC Prefrontal Cortex Visium 10x data,

- Breast cancer IMC data from Jackson et al.,

- NSCLC and TNBC IMC data from the Immucan consortium.

Our analysis revealed that patient representations generated by Cellohood correlate meaningfully with disease stages, types/histologies, and patient prognosis. In cancer datasets, coarse-resolution analysis uncovered distinct whole-slide tumor infiltration patterns, while high-resolution examination revealed specific neighborhood interactions significantly linked to clinical outcomes. Spatial analysis further demonstrated that the neighborhood representation successfully encodes diverse tissue architectures.

Conclusions

Thanks to AI-powered transformer-based generative modeling, Cellohood is the first model to utilize complete cell marker information during training without resorting to coarse, neighborhood-wide approximation. Results on multiple datasets obtained using different technologies demonstrated that Cellohood enables marker-driven discovery of cellular-microenvironment interactions and their clinical implications.

?