Machine learning reveals a PD-L1–independent prediction of response to immunotherapy of non-small cell lung cancer by gene expression context. (November 2020)

• Record Type: Journal Article
• Authors: Wiesweg, Marcel; Mairinger, Fabian; Reis, Henning; Goetz, Moritz; Kollmeier, Jens; Misch, Daniel; Stephan-Falkenau, Susann; Mairinger, Thomas; Walter, Robert F.H.; Hager, Thomas; Metzenmacher, Martin; Eberhardt, Wilfried E.E.; Zaun, Gregor; Köster, Johannes; Stuschke, Martin; Aigner, Clemens; Darwiche, Kaid; Schmid, Kurt W.; Rahmann, Sven; Schuler, Martin
• Is Part Of: European journal of cancer. Volume 140(2020)
• Page Start: 76
• Page End: 85
• Publication Date: 2020-11
• Subjects: Lung cancer -- PD-L1 -- Immunotherapy -- Predictive factors -- Machine learning; Cancer -- Periodicals; Neoplasms -- Periodicals; Cancer -- Périodiques; Cancer; Tumors; Electronic journals; Periodicals; Electronic journals; 616.994
• DOI: 10.1016/j.ejca.2020.09.015
• Languages: English
• ISSNs: 0959-8049

Abstract

Abstract: Objective: Current predictive biomarkers for PD-1 (programmed cell death protein 1)/PD-L1 (programmed death-ligand 1)-directed immunotherapy in non-small cell lung cancer (NSCLC) mostly focus on features of tumour cells. However, the tumour microenvironment and immune context are expected to play major roles in governing therapy response. Against this background, we set out to apply context-sensitive feature selection and machine learning approaches on expression profiles of immune-related genes in diagnostic biopsies of patients with stage IV NSCLC. Methods: RNA expression levels were determined using the NanoString nCounter platform in formalin-fixed paraffin-embedded tumour biopsies obtained during the diagnostic workup of stage IV NSCLC from two thoracic oncology centres. A 770-gene panel covering immune-related genes and control genes was used. We applied supervised machine learning methods for feature selection and generation of predictive models. Results: Feature selection and model creation were based on a training cohort of 55 patients with recurrent NSCLC treated with PD-1/PD-L1 antibody therapy. Resulting models identified patients with superior outcomes to immunotherapy, as validated in two subsequently recruited, separate patient cohorts (n = 67, hazard ratio = 0.46, p = 0.035). The predictive information obtained from these models was orthogonal to PD-L1 expression as per immunohistochemistry: Selecting by PD-L1 positivity at immunohistochemistry plus model prediction identified patients with highly favourable outcomes. Independence of PD-L1 positivity and model predictions were confirmed in multivariate analysis. Visualisation of the models revealed the predictive superiority of the entire 7-gene context over any single gene. Conclusion: Using context-sensitive assays and bioinformatics capturing the tumour immune context allows precise prediction of response to PD-1/PD-L1-directed immunotherapy in NSCLC. Highlights: Establishment of a novel predictor of immunotherapy response in lung cancer. Machine-learning approach increases precision by integrating tumor immune context. Gene expression context is more powerful than single markers.