AI-powered image-based spatial profiling of MET-mutated non-small cell lung cancer identifies immune-active MET exon 14 skipping subtypes as potential immunotherapy targets

Background MET alterations are oncogenic drivers in non-small cell lung cancer (NSCLC), but their impact on the tumor microenvironment (TME) remains unclear. Spatial analysis of whole slide images (WSIs) enables high-resolution TME characterization, overcoming limitations of bulk sequencing. This study used AI-powered spatial analysis to profile immune phenotypes and cellular composition across MET mutations in NSCLC.

Methods We retrospectively analyzed 371 H&E-stained WSIs from NSCLC biopsies collected during routine clinical care using the AI-based SCOPE IO algorithm (Lunit) to quantify tumor cellular densities and immune phenotypes (inflamed, immune-excluded, immune-desert). Cases were stratified by MET status: exon 14 skipping (METex14, N=241), amplification (METamp, N=31), and wildtype (METwt, N=99). SCOPE IO metrics and targeted sequencing-based immune gene expression (iGEX) were compared across groups. METex14 tumors were further grouped into inflamed (N=63) and non-inflamed (N=158) subtypes. A machine learning (ML) model trained on iGEX features associated with METex14 subtypes was used to impute subtypes in a second NSCLC cohort with immunotherapy outcomes (N=205).

Results METex14 tumors had more inflamed phenotypes than METamp and METwt (29% vs 10% and 15%, P=2E-3), with higher densities of endothelial cells, fibroblasts, and lymphocytes in cancer areas (P≤2E-3), elevated inflamed scores (P=0.01), and lower immune-desert scores (P=0.02). METamp tumors showed more immune-desert phenotypes (79% vs 52% and 63%, P=0.01), increased presence of mitotic cells (P=4E-8), fewer non-tumor cells (P<0.05), and higher immune-desert scores (P=0.047). iGEX confirmed these findings: METex14 tumors had higher overall iGEX (192 genes), while METamp tumors showed elevated proliferation-associated genes (18 genes) (P<0.05). Inflamed METex14 tumors had more lymphocytes, macrophages, and other non-tumor cells in cancer and stromal areas (P≤2E-3), with increased iGEX in immune activation pathways (166 genes, P<0.05). ML-based feature selection identified 46 differentially expressed genes distinguishing METex14 subtypes with high accuracy (ROC-AUC=0.94). In a second cohort, tumors classified as inflamed were associated with improved survival under immunotherapy (HR=0.5, P=0.004) (figure 1).

Conclusions AI-powered spatial analysis and iGEX profiling revealed distinct TME profiles across MET mutations in NSCLC. METex14 tumors exhibited immune-active TMEs, while METamp tumors were immune-deficient and proliferative. A subset of METex14 tumors showed high immune activity, suggesting potential responsiveness to immunotherapy (figure 2). These findings highlight MET-driven NSCLC heterogeneity and the utility of spatial AI tools for immunotherapy stratification and biomarker development.