Artificial Intelligence-based tumor microenvironment and PD-L1 analysis using digital pathology to predict pembrolizumab response in metastatic triple-negative breast cancer.

1110 Background: The combination of pembrolizumab and chemotherapy improves survival in programmed death ligand 1 (PD-L1) positive metastatic triple-negative breast cancer (mTNBC). However, responses vary even among PD-L1 positive, and predictive biomarkers remain undefined. This study investigates the predictive biomarkers to pembrolizumab through digital pathology and artificial intelligence (AI)-based tumor microenvironment (TME) and PD-L1 analysis. Methods: We retrospectively analyzed 53 PD-L1 positive, mTNBC patients treated with pembrolizumab at Gangnam Severance Hospital (2017–2024). PD-L1 positivity was defined as a combined positive score (CPS) ≥ 10. Immune phenotypes and immune cell density in both tumor and stroma were analyzed in 67 H&E images using Lunit SCOPE IO, an AI-powered whole slide image analyzer. PD-L1 CPS was analyzed in paired PD-L1 staining images by both Lunit uIHCv2 analyzer and pathologists. Samples were categorized as pre-(pre) or post-treatment (post). Pre-samples were collected before any therapy exposure, while post-samples were obtained after recurrence following neo/adjuvant therapy. These features were analyzed for their association with pembrolizumab response and clinical outcome. Results: With a median follow-up of 13.2 months, the median age was 53 years, and 16 patients (22.5%) were de novo stage IV TNBC. AI-assessed PD-L1 positivity was seen in 52.2% (35/67) of cases, compared to 74.6% (50/67) by pathologist. Overall, AI-based PD-L1 positive cases had a median progression-free survival (mPFS) of 8.8 months (mo) vs 6.7 mo in PD-L1 negative (p = 0.028), while pathologist-reported cases showed 7.9 mo vs 6.3 mo, respectively (p = 0.17). AI-based PD-L1 positivity in pre-samples was associated with better PFS with pembrolizumab (mPFS 7.7 mo vs 4.4mo, HR 0.32, p = 0.014), while post-samples showed no significant association (mPFS 7.3 mo vs 6.4mo, HR 0.69, p = 0.4). Notably, post-samples (50.0%) had a higher proportion of cases with AI-based CPS ≥ 10 compared to pre-samples (36.4%), primarily driven by increased PD-L1-expressing macrophages, as revealed by AI-based cell composition analysis (22.7% vs 7.7%, p = 0.0007). When categorized by AI-based immune phenotype, notable differences were seen between pre/post-samples despite PD-L1 positivity. Post-samples showed a higher prevalence of the immune-desert phenotype, reflecting significant changes in the TME following prior therapy exposure (40.0% vs 20.0%, p = 0.06). Conclusions: This study highlights the role of the TME and PD-L1 assessed by AI in predicting pembrolizumab response in mTNBC. While PD-L1 positivity in pre-samples was associated with outcome, PD-L1 positivity in post-samples showed limited association with PFS, potentially influenced by immune desert phenotype and increased PD-L1-expressing macrophages.