Heterogeneity Assessment of PD-L1 Expression in Biopsy Versus Resected Specimens of Lung Adenocarcinoma - Machine Learning Approach

Abstract

Introduction PD-L1 is a biomarker that is used to predict the response to immunotherapy of patients with lung cancer. A pathologist manually assesses the PD-L1 expression mainly on biopsy specimens. Even though PD-L1 is an established biomarkers in the routine practice, still some challenges remain. This thesis aims to address two issues: 1) whether an automated scoring system can objectively and reliably reproduce the manual scores, and 2) how the PD-L1 score on biopsy is comparable with surgically resected samples.

Methods Paired biopsy and resection tumor tissues from 26 patients with adenocarcinoma of the non-small cell lung cancer (NSCLC) were included in this thesis. Immunohistochemistry was used to detect the expression of PD-L1. The slides were digitalized, and the clinical data was retrieved from the patient’s journals. A supervised machine learning model (ML) was developed to assess the tumor proportion score (TPS) of PD-L1 expression in the whole slide images. Sensitivity and specificity of ML-derived PD-L1 scores and the intra-class correlation coefficient (ICC) of biopsy vs resected PD-L1 scores were computed.

Results There was a moderate correlation (r=0.59, P<0.001) between digital and manual scores using PD-L1 TPS as continuous variable. The ML model showed high performance in PD-L1 scoring with a sensitivity/specificity of 0.88/0.92 and 0.85/0.96 at both <1 vs ≥1 % and <50 vs ≥50 % TPS cutoffs, respectively. No correlation was observed between biopsy vs resected PD-L1 scores at 1% cutoff using either digital or manual scores. However, at 50% cutoff, both digital and manual scores show high level of consistency (manual TPS ICC, 0.82, P<0.001; digital TPS ICC: 0.7, P=0.01) across paired biopsy and resection tissues.

Conclusion The biopsies were found to be equivalent to corresponding resected specimens for determining PD-L1 expression at 50% cutoff in adenocarcinoma of NSCLC. Our machine learning algorithm is found to be robust in detecting PD-L1 positive tumor cells with an accuracy like pathologists. Validation of the findings and algorithms is warranted in large-scale cohort.