ADC target profiling in NSCLC: Generalizable AI separates TROP-2 and cMET phenotypes

BACKGROUND: Antibody drug conjugates (ADCs) targeting TROP-2 and cMET are entering clinical trials in NSCLC. Their translation depends on reliable biomarker assessment, a task still dominated by subjective visual scoring and inconsistent reproducibility. METHODS: We built a modular AI pipeline that detects cells, classifies carcinoma cells, and quantifies membranous and cytoplasmic expression. A membranous scorer trained on TROP-2 was applied zero-shot to cMET, HER2, and PD-L1. The analysis covered 1,142 resected NSCLCs, integrating expression maps with clinicopathologic, molecular, and tumor microenvironment (TME) features. RESULTS: The AI scorer recapitulated pathologist annotations with near-perfect correlation (r = 0.98-0.99, ρ = 97-98, τ = 0.88-0.89, CCC = 0.97-0.98) for TROP-2 and generalized to other markers (cMET r = 0.99, ρ = 0.96, τ = 0.91, CCC = 0.96; HER2 r = 0.93, ρ = 0.72, τ = 0.60, CCC = 0.92; and PD-L1 [TPS] r = 0.85, ρ = 0.84, τ = 0.68, CCC = 0.82 / [H-score] r = 0.87, ρ = 0.86, τ = 0.70, CCC = 0.85). Its agreement with six pathologists matched interobserver variability (0.86-0.96). Expression maps revealed contrasting spatial and cellular patterns: TROP-2 dominated LUSC (mean H-Scores 141.3/103.2 vs. 74.5/45.7 in LUAD for membrane/cytoplasm) and marked immune-deserted tumors. cMET prevailed in LUAD (mean 50.4 vs. 20.6 in LUSC), co-localized with fibroblast-rich, immune-active TME and KRAS mutations. CONCLUSIONS: Foundation model-based scoring produces expert-level, scalable biomarker quantification. The resulting TME phenotypes - TROP-2-high immune-deserted vs cMET-high, immune-active - reveal therapeutic implications for combining ADCs with immunotherapies or kinase inhibitors.