Machine learning-driven proteomics classifier deciphers tumor origins of primary and metastatic squamous cell carcinomas

Squamous cell carcinoma (SCC) occurs across multiple organs with highly similar histology, making the diagnosis of SCC of unknown primary (SCCUP) particularly challenging. To address this, we established a machine learning–based 39-protein biomarker classifier (39PBC) trained on proteomic profiles from 387 SCC samples collected at seven tertiary hospitals. The classifier accurately predicted the origin of primary and metastatic SCCs from cervical, esophageal, lung, nasopharyngeal, and head and neck sites, with validation in internal (n = 324) and external (n = 63) cohorts yielding AUCs of 0.924–0.961 and 0.971 and accuracies above 87%. Immunohistochemistry of 509 cases further identified a simplified five-marker panel (four robust site-specific markers CCDC6, LGALS7, LGALS9, and P16, together with EBER) suitable for routine screening. Importantly, 39PBC demonstrated reliable performance in real-world SCCUP and dual-primary cases. Proteomic profiling also uncovered distinct prognostic and molecular landscapes, implicating metabolic activation as a driver of progression and immune modulation as a site-specific feature. Together, these findings establish a clinically applicable workflow that integrates high-resolution proteomics with practical IHC validation, offering a public resource to improve SCCUP diagnosis, enable cost-effective clinical translation, and provide mechanistic insights into SCC metastasis.