Artificial intelligence-based tumor microenvironment and PD-L1 analysis to predict pembrolizumab response in metastatic triple-negative breast cancer

Pembrolizumab plus chemotherapy improves survival in PD-L1–positive metastatic triple-negative breast cancer (mTNBC), but responses remain heterogeneous and robust predictive biomarkers are lacking. We evaluated artificial intelligence (AI)–based digital pathology to identify predictors of pembrolizumab benefit through quantitative assessment of PD-L1 expression and the tumor microenvironment (TME). We retrospectively analyzed 49 PD-L1–positive mTNBC patients treated with pembrolizumab between 2021 and 2024. PD-L1 positivity was defined as a combined positive score (CPS) ≥ 10. Tumor samples were classified as pre- or post-treatment based on exposure to prior systemic therapy. Of the 49 patients, 23 contributed a pre-treatment specimen only, 24 a post-treatment specimen only, and 2 contributed both; specimens were therefore not serially paired. Immune phenotypes and immune cell densities were quantified from hematoxylin–eosin (H&E) and PD-L1–stained whole slide images using the AI-powered platform Lunit SCOPE IO (H&E) and Lunit SCOPE uIHC (PD-L1 immunohistochemistry). PD-L1 status was evaluated by both AI and pathologists. Progression-free survival (PFS) was compared using Kaplan–Meier and Cox regression analyses. Median follow-up was 13.2 months. AI-based PD-L1 positivity was observed in 51.5% of samples versus 74.2% by pathologists. AI-positive cases had significantly longer median PFS (8.8 vs. 6.7 months; HR 0.50, 95% CI 0.27–0.94; p = 0.031), while pathologist-assessed PD-L1 was not (median 7.2 vs. 7.9 months; HR 0.63, 95% CI 0.38–1.04; p = 0.072). In pre-treatment samples, AI-based PD-L1 positivity was associated with longer PFS (7.2 vs. 4.4 months; HR 0.29, 95% CI 0.12–0.71; p = 0.006), but no such association was observed in post-treatment samples. The benefit associated with AI-based positivity emerged after the first 6 months of treatment (HR 0.30 beyond six months vs. 0.73 before), consistent with the delayed separation typical of immune checkpoint blockade. Among PD-L1–positive immune cells, macrophages accounted for a higher proportion in post-treatment than in pre-treatment CPS ≥ 10 samples (34% vs. 16%; p = 0.0044). AI-based PD-L1 and TME profiling identified patients with longer PFS on pembrolizumab in mTNBC, particularly in pre-treatment samples, while post-treatment PD-L1 positivity, which was accompanied by an increased proportion of PD-L1–positive macrophages, showed no such association. These findings are hypothesis-generating and require validation in an independent cohort before AI-based scoring can be considered for patient selection in mTNBC.