Is Deciphering Cell Types in Complex Tissues the Next Frontier for Protein Language Models?

Spatial proteomics allows for the analysis of protein expression within tissue contexts. While cellular expression and spatial location are commonly utilized, the properties of protein markers can offer additional valuable information. Therefore, we propose a deep learning framework that utilizes embeddings from the ESM2 protein language model to incorporate prior biological knowledge of protein markers. Our framework integrates these marker priors with single-cell expression and spatial information through a cross-Attention mechanism to generate informative cell representations. We demonstrate that this integration significantly improves cell type annotation performance across multiple spatial proteomics datasets, highlighting the value of leveraging PLM-derived marker knowledge.