579 Novel digital image approach of multiplex immunofluorescence based PD-L1 expression enables the stratification of advanced NSCLC patients treated with durvalumab

Abstract

Background Pathologist-based scoring of PD-L1 expression on tumor cells using IHC 1 has shown clinical utility in predicting favorable overall survival in advanced non-small cell lung cancer (NSCLC) patients treated with anti-PD-(L)1 therapies including durvalumab. [2][3] Quantitative Continuous Scoring (QCS) 4 enables the continuous measurement of the PD-L1 expression on single cells and the selection of the PD-L1 expression cutoff that best stratifies anti-PD-L1-treated patients with respect to prevalence and log-rank test p-value. 5 We present here the extension of QCS to PD-L1 measured by multiplex immunofluorescence (mIF) 6 to evaluate its ability to optimize patient stratification. Methods Pre-treatment tumor samples from advanced NSCLC patients enrolled in durvalumab nonrandomized phase 1/2 trial (CP1108/NCT01693562) 2 , were stained by mIF panel containing PD-L1. 6 Similarly to IHC PD-L1 QCS, mIF PD-L1 QCS consists of two deep-learning models, first to segment epithelium regions and second to detect membrane, cytoplasm and nuclei of each epithelium cell, transferring for the second model annotations from IHC to mIF domain. 7 The mIF images are normalized based on batch statistics prior to image analysis. PD-L1 expression is measured for each epithelium cell as the average of the PD-L1 signal in the segmented membrane. Cells with expression higher than an expression threshold (T PDL1 ) are considered positive. A slide is considered QCS-positive if it comprises a greater percentage of PD-L1 positive cells (QCS-score) than a cutoff value (CoV). ResultsThe QCS-scores are computed on 119 NSCLC patients treated with durvalumab. As a first proof of concept that QCS-scoring can replicate tumor proportion scoring (TPS), we optimize T PDL1 as to maximize the correlation between QCS and TPS scores (figure 1). Second, we estimate for different combinations of (T PDL1 , CoV) the log rank p-value associated with the stratification between patients with low and high QCS scores. A subregion of the parameter space was identified for which the stratification is significant (p<0.01) with more than 50% prevalence in the positive subgroup (figure 2). The p-value is minimized (p=7.2 10 -5 ) for (T PD-L1 =37, CoV=0.75%), yielding a median OS of 5.58 months and 13.44 months in the QCS negative and positive subgroups respectively, similar to those of IHC PD-L1 manual scoring with 25% cutoff. Conclusions The extension of QCS to mIF imaging provides opportunities to evaluate continuous PD-L1 expression of single tumor cells in relation to spatial distribution of other cells (e.g. PD1+ CD8+ T cells) and identify predictive biomarkers of tumor-immune cell interactions of anti-PD-(L)1 therapies. Trial Registration CP1108/NCT01693562 Abstract 579 Figure 1 Correlation to pathologist-based TPS score Top: Lineplots of Pearson and Spearman correlations as well as of Lin correlation coefficient between the pathologist-based TPS scores and the QCS-based scores, computed for increasing expression threshold values (TPD-L1). The QCS shows maxim...