Crosstalk In: crosstalk-aware inference of tumor microenvironment cell infiltration

The tumor microenvironment (TME) is a complex ecosystem whose cellular composition and interactions shape tumor progression, therapeutic response, and patient prognosis. However, most bulk deconvolution methods infer cell-type composition while treating cell types independently.

In this paper, we develop CrosstalkIn, a crosstalk-aware bulk deconvolution framework that constructs patient-specific cell-cell networks by integrating Gene Ontology (GO)-based functional similarity and protein-protein interaction (PPI)-based molecular relationships. A modified random walk with restart algorithm is then applied to calculate cell Infiltration Scores (InScores).

Benchmark results on two flow-cytometry-validated cohorts demonstrate that CrosstalkIn achieves superior deconvolution performance, with the largest or second-largest Spearman correlations for most evaluated cell types. In lower-grade glioma, CrosstalkIn-derived InScores identify survival-associated cell types, generate accurate prognostic risk scores, and stratify patients into distinct survival groups. Across multiple adenocarcinoma cohorts, the risk score shows consistent prognostic value, particularly in advanced-stage patients. In melanoma, CrosstalkIn improves immunotherapy response prediction and identifies biologically interpretable cell-type biomarkers.

CrosstalkIn provides a robust framework for crosstalk-aware inference of TME cell infiltration from bulk transcriptomic data and supports cancer prognosis and immunotherapy response prediction.