Artificial intelligence-powered spatial analysis of tumor infiltrating lymphocytes and tertiary lymphoid structures in non-small cell lung cancer patients treated with immune-checkpoint inhibitors±chemotherapy.

Abstract

8572 Background: This study evaluates the predictive utility of an artificial intelligence (AI)-powered whole-slide image (WSI) analyzer for assessing Tumor-infiltrating lymphocytes (TILs) and tertiary lymphoid structures (TLSs) in patients (pts) treated with ICIs, either as monotherapy or in combination with chemotherapy. Methods: An AI-powered WSI analyzer (Lunit SCOPE IO, Lunit, Seoul, Korea) was utilized to segment cancer area (CA) and cancer stroma (CS), and identification of tumor infiltrating cells (TILs) and tertiary lymphoid structure (TLS) on tumor tissues. Pre-treatment H&E-stained WSIs were obtained from Samsung Medical Center (n = 533), and other multi-center cohorts (Shen et al, 2024, n=634). After quality control, 1,144 samples (98.0%) were used included in the final analysis. Pts were stratified into risk groups; good risk (high TILs in CA and high TLS area per CA), poor risk group (low TILs in CA and low TLS area per CA), and intermediate risk group (others). Among them, 988 pts had available PD-L1 expression data, 435 underwent whole transcriptome sequencing, and 292 underwent whole exome sequencing. Results: TILs in CA correlated significantly with the interferon gamma pathway (ρ=0.49, P<0.001), and the T-cell inflamed score (ρ=0.56, P<0.001). Similarly, TLS area per CA was significantly correlated with TLS imprinting (ρ=0.48, P<0.001), and B cell receptor signature (ρ=0.42, P<0.001). Among 1,144 pts, 1,044 received ICI monotherapy, and 100 underwent combination therapy with ICI and chemotherapy. ICIs were administered as first-line therapy in 245 pts (24.1%), and second line in 524 (45.8%). The risk groups were distributed as follows: good risk (n=279, 24.4%), intermediate risk (n=437, 38.2%), and poor risk (n=428, 37.4%). Pts with PD-L1 tumor proportion score ≥50% were more frequent in the good-risk group (47.3%) than in intermediate (37.5%) or poor-risk groups (30.6%, P=0.001). Smoking history showed no significant association with risk groups (P=0.958). Pts receiving ICI monotherapy showed significant differences in overall response rate (ORR: 28.9% vs. 19.7% vs. 16.3%, P<0.001), median progression-free survival (mPFS: 6.1 vs. 3.5 vs. 2.4 months, P<0.001), and median overall survival (mOS: 26.4 vs. 14.6 vs. 11.3 months, P<0.001) among the good, intermediate, and poor-risk groups, respectively. Similar trends were observed in pts receiving ICI plus chemotherapy: mPFS (10.3 vs. 8.4 vs. 4.7 months, P=0.005), and mOS (27.9 vs. 22.4 vs. 17.6 months, P=0.047). Notably, KEAP1 mutations were significantly more frequent in the poor-risk group (17.4% vs. 7.9%, P=0.020). Conclusions: AI-powered analysis of TILs and TLSs effectively stratifies NSCLC pts into risk groups, predicting efficacy outcomes of ICIs with or without chemotherapy.