Abstract 2416: CELLama: cell embedding leveraging language model abilities for analyzing cell types and spatial relationship in tumor

Abstract Background: Single-cell RNA sequencing (scRNA-seq) and spatial transcriptomics (ST) have opened up new avenues that will expand our knowledge on cellular heterogeneity and spatial tissue architecture in the tumor microenvironment. However, the mere volume of data as well as different error profiles pose significant challenges in comprehensive data integration and subsequent analysis, particularly for robust cell type identification and spatial context analysis. Here, we developed CELLama (Cell Embeddings Leverage Language Model Abilities), a foundation model that relies on natural language processing (NLP) methods to embed the information from the cell and space. Methods: CELLama converts scRNA-seq data into natural language-like sentences of input, with major focus on gene expression and including metadata on tissue of origin and disease status. These will be projected to high-dimensional vectors for each cell by a pre-trained sentence transformer. This embedding was also fine-tuned with sentence pairs of the generated sets labeled for cosine similarity that utilized gene expression profiles. In testing the performances of CELLama and the other existing models on multitissue and disease datasets, zero-shot cell typing and spatial context analysis were performed. To analyze the spatial architecture of the tumor microenvironment, neighborhood cell type patterns were generated, and cell subtypes within the tumor microenvironment were clustered based on these patterns. Results: CELLama demonstrated better integration and annotation of cells across multiple datasets. It showed better performance than other foundation cell embedding models including scGPT in terms of accuracy, precision, and recall, especially after the addition of metadata like tissue of origin. As an example of leveraging CELLama for spatial analysis of tumor microenvironment architecture, it has been applied to identify five distinct fibroblast subpopulations in lung cancer, differentiated by their spatial context. Conclusions: CELLama represents a unique cellular data analysis leveraging NLP, which changes more complicated transcriptomic data into highly interpretable language-like formats, integrates various metadata and spatial context, and will further extend to atlas-level integrations and disease-specific tissue analyses. It simplifies complex analyses by providing a thorough understanding of intricate cell types within the tumor microenvironment and accelerating the study of their spatial architectural patterns. Citation Format: Jiwon Kim, Jeongbin Park, Sumin Kim, Dongjoo Lee, Sungwoo Bae, Haenara Shin, Daeseung Lee, Hongyoon Choi. CELLama: cell embedding leveraging language model abilities for analyzing cell types and spatial relationship in tumor [abstract]. In: Proceedings of the American Association for Cancer Research Annual Meeting 2025; Part 1 (Regular Abstracts); 2025 Apr 25-30; Chicago, IL. Philadelphia (PA): AACR; Cancer Res 2025;85(8_Suppl_1):Abstract nr 2416.