Digital quantification of lymphocytic infiltration on routine H&E images and immunotherapy response in non–small cell lung cancer.

Abstract

9066 Background: Current biomarker(s) for immuno-oncology (IO) therapy response prediction in lung cancer are limited. Additional predictive biomarkers are useful to help refine patient selection and guide precision therapy. Methods: Biopsy and surgical specimens stained with hematoxylin-eosin (H&E) were subjected to whole-slide scanning for 446 advanced stage non-small cell lung cancer (NSCLC) treated with single agent immune check point inhibitors (ICI). A machine learning model was trained on H&E images for classification of tumor infiltrating lymphocytes (TILs), tumor cells, and stromal cells in specific tissue types. Results: TIL levels were found to be highly variable, with a range of 12 to 4270 cells/mm2, and median of 319 (Q1 = 159, Q3 = 681). TIL levels were assessed on tissue samples from multiple organs which had shown primary or metastatic NSCLC, and were similar across all specimen sites except the liver, for which median TIL levels were significantly lower, at 90 cells/mm2. There was no correlation between tumor mutational burden (TMB) and TIL levels, while high TIL levels were correlated with high PD-L1 (≥ 50%) expression. Patients who experienced a partial/complete response to ICI therapy had a trend to higher median TILs compared to those who had progressive/stable disease (350 versus 310 cells/mm2, P = 0.09). In a multivariable analysis after controlling for covariates (incl. sex, age, cigarette smoking, ECOG, PD-L1, TMB & treatment line), a higher TIL level (≥ 250 cells/mm2) was an independent predictor of IO response for both progression-free survival (PFS; HRadj 0.70; 95% CI, 0.55 - 0.89; P = 0.003) and overall survival (HRadj 0.73; 95% CI, 0.56 - 0.95; P = 0.02). In a ROC analysis considering single biomarkers, PD-L1 had the highest AUC (0.68, P < 0.001), while TIL (AUC = 0.53, P = 0.08) and TMB (AUC = 0.55, P = 0.05) had similar AUC values for classifying responders from non-responders based on objective response rate. Using weighted linear regression approach to combine the biomarkers, paired PD-L1/TMB had the greatest AUC (0.70, P < 0.001) compared to PD-L1 single assay. In the PD-L1 negative (< 1%, N = 50) subgroup, TIL levels had superior predictive performance for classification of IO responders (AUC = 0.77, P = 0.02) compared to TMB (AUC = 0.57, P = 0.3), such that patients with a high TIL level (≥ 250 cells/mm2) had an improved PFS (median PFS: 2.7 vs 2.2 months; HR = 0.48; 95% CI, 0.26 - 0.87; P = 0.02). Conclusions: Digital TIL quantification with use of machine learning is feasible. TIL levels appear to be a robust and independent biomarker of likelihood of response to IO treatment in NSCLC, especially in the PD-L1 negative subgroup. The findings of this study are under validation in additional lung cancer cohorts.