Digital Versus Manual PD-L1 Scoring in Advanced NSCLC From the IMpower110 and IMpower150 Trials

Treatment selection in patients with advanced NSCLC is based on programmed death-ligand 1 (PD-L1) expression, which is usually scored manually and is subject to intra- and inter-pathologist variability. A PD-L1 clone-agnostic artificial intelligence (AI) model for AI-based measurement of PD-L1 (AIM-PD-L1) was developed and assessed in advanced NSCLC using clinical samples from two phase 3 trials.

IMpower110 evaluated atezolizumab versus chemotherapy in PD-L1-positive metastatic, stage IV, squamous or nonsquamous NSCLC. IMpower150 evaluated atezolizumab, carboplatin, and paclitaxel, with or without bevacizumab, versus carboplatin, paclitaxel, and bevacizumab in patients with metastatic nonsquamous NSCLC. AIM-PD-L1 was developed and deployed on SP263-stained whole slide images (IMpower110, n = 509; IMpower150, n = 766) for digital scoring of tumor cell (TC) PD-L1 expression and identification of human-interpretable features (HIFs) associated with survival outcomes.

Overall percentage agreements between scoring methods for TC more than or equal to 50% and more than or equal to 1% cutoffs were high. Survival analyses were similar for PD-L1 subgroups between scoring methods at both TC cutoffs. A nonsignificant improvement in survival outcomes was observed in patients treated with atezolizumab-containing regimens and classified as positive by digital scoring but missed by manual scoring. Two HIFs in the cancer epithelium-density of all PD-L1-positive TC and immune cells-were nominally associated with overall survival. Many HIFs were identified to be predictive of significantly improved progression-free survival with atezolizumab-containing regimens versus control.

AIM-PD-L1 digital SP263 PD-L1 scoring is concordant with manual scoring, revealing similar predictivity for benefit, and could potentially be used as a predictive marker for patient stratification and selection for anti-PD-(L)1 therapy.