BIOCHEMICAL AND SPATIAL MODELING OF TUMOR MICROENVIRONMENT FROM BIOPSY PROTEOMICS TO PREDICT METASTATIC PROGRESSION

Abstract

Metastasis causes about 90% of cancer deaths, but the anticipation of its occurrence from primary tumor biopsies remains an intractable clinical dilemma. Spatial organization and biochemical heterogeneity of TME are regulators of metastatic formation but are usually not interrogated by typical diagnostic approaches. To overcome this deficiency, we established a hybrid modeling approach that unites biopsy-derived proteomics, spatial statistical analysis, and machine learning to predict metastatic potential. Employing multiplexed proteomic imaging technologies like imaging mass cytometry, we profiled in situ protein expression patterns in the patient's biopsy tissue. These data were used to calculate spatial features like cell–cell interaction graphs, spatial proximity scores, and neighborhood-based immune signatures. Machine learning classifiers—graph neural networks and random forests—were trained to predict these spatial-biochemical signatures from recognized metastatic endpoints. Our model performed with excellent prediction capability (AUC ≈ 0.90) in a retrovalidation set across various solid cancers. Notably, we also found specific TME signatures that correlate with metastasis: dense foci of PD-L1⁺ macrophages and FOXP3⁺ regulatory T cells at the invasive edge were correlated with a higher risk of metastasis, and abundant CD8⁺ T cell infiltration at the invasive edge with lower progression. This integrated strategy provides a clinically feasible method for predicting metastasis from standard biopsy samples, which could potentially lead to earlier treatment and therapy stratification. Our results highlight the translational value of the combination of spatial proteomics and computational modeling towards improved prognostic precision in oncology.