202P - Discovery of immunological cellular neighborhoods from protein markers in spatial tumor data

Background

Spatial imaging of single cells and their protein markers in multiple tumor tissues offers detailed insights into the mechanisms of tumor-microenvironment interactions. The identification and characterization of cellular neighborhoods are vital to elucidating these mechanisms. However, existing approaches for cellular neighborhood analyses resort to predefined cell types or coarse, neighborhood-wide cell marker aggregations, failing to preserve the marker information at the single-cell level resolution.

Methods

We developed Cellohood - the first permutation-invariant, transformer-based autoencoder designed to model cellular neighborhoods explicitly. Cellohood compresses information about each cell and its marker expression for a given neighborhood, providing representations of the cellular neighborhoods. Based on these representations, we derived novel cellular neighborhood prototypes. Cell types, protein markers, and spatial arrangement patterns describe each identified prototype. Consequently, patients can be described by the abundance of cellular neighborhood prototypes and their mutual arrangements.

Results

We showcased the performance of Cellohood by applying it to the Imaging Mass Cytometry data from a non-small cell lung cancer (NSCLC) patient cohort (N=192) from the IMMUcan (Integrated iMMUnoprofiling of large adaptive CANcer patient cohorts) consortium. Prototypes discovered by Cellohood, including tumor neighborhoods characterized by i) low B2M expression, ii) IDO1+ tumor cells, iii) plasma cell enrichment, and iv) T-reg infiltration, defined patient clustering that surpassed the TNM cancer staging in terms of survival prediction. Analysis of macrophage-enriched neighborhoods indicated that CD206+ macrophages are a key differentiating factor between cancer histologies.

Conclusions

Thanks to 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 the NSCLC cohort from the IMMUcan consortium demonstrated that Cellohood enables marker-driven discovery of tumor-microenvironment interactions and their clinical implications.

Legal entity responsible for the study

IMMUcan (Integrated iMMUnoprofiling of large adaptive CANcer patient cohorts).

Funding

IMI2 JU grant agreement 821558, supported by EU’s Horizon 2020 and EFPIA.

Disclosure

All authors have declared no conflicts of interest.