Background: Current biomarker(s) for immune-oncology (IO) therapy response prediction in lung cancer are limited. Additional predictive/prognostic biomarkers are useful to help refine patient selection and guide precision therapy.

Methods: A machine learning (ML) automated method was developed to segment tumor, stroma and lymphocytes (TILs) in whole slide hematoxylin and eosin (H&E)–stained images of NSCLC patients. A primary set (n = 446, stage III-IV, DFCI cohort) and a secondary set (n = 523, stage I-III, Norway cohort) were used to evaluate the predictive and prognostic impact of TILs, respectively. Targeted DNA sequencing (Oncopanel) and PD-L1 protein expression data were used to evaluate independence and interaction between tumor mutational burden (TMB), PD-L1 and TIL levels with clinical outcomes after anti-PD-(L)1 monotherapy in the primary set.

Results: In the primary set, higher TIL levels was an independent predictor of IO response for both progression-free (HRadj 0.70; P = 0.003) and overall survival (HRadj 0.73; P = 0.02). In the secondary set using the identical ML algorithm, high TIL levels was an independent favorable prognostic factor regardless of endpoint (DSS, TTR), pathological stage and histology. In the PD-L1 negative subgroup of the primary set, TIL levels had superior predictive accuracy for IO response (AUC = 0.77) compared to TMB (AUC = 0.57).

Conclusion: TIL levels appear to be a robust and independent biomarker of likelihood of response to IO treatment in NSCLC. Furthermore, this study shows the potential of ML–based image analysis for extracting quantitative immune information from histopathological slides.