Machine learning-based predictors for immune checkpoint inhibitor therapy of non-small-cell lung cancer

Abstract

Immunotherapy targeting programmed cell death protein 1/programmed death-ligand 1 (PD-1/PD-L1) is a standard of care in the treatment of stage IV non-small-cell lung cancer (NSCLC). However, only a minority of patients responds to anti-PD-1/PD-L1 monotherapy. Tumor-centric predictive biomarkers applicable to small diagnostic specimens such as PD-L1 expression and tumor mutational burden [1.Prelaj A. Tay R. Ferrara R. et al.Predictive biomarkers of response for immune checkpoint inhibitors in non–small-cell lung cancer.Eur J Cancer. 2019; 106: 144-159Abstract Full Text Full Text PDF PubMed Scopus (130) Google Scholar] only allow enrichment of cohorts with higher probability of treatment response. Efficacy of immunotherapy is governed by a complex interplay of tumor-intrinsic properties (genomic and epigenomic), the tumor microenvironment, the systemic state of the immune system, and de novo or acquired resistance [2.Cogdill A.P. Andrews M.C. Wargo J.A. Hallmarks of response to immune checkpoint blockade.Br J Cancer. 2017; 117: 1-7Crossref PubMed Scopus (153) Google Scholar]. Capturing this ‘cancer-immune set point’ [3.Chen D.S. Mellman I. Elements of cancer immunity and the cancer–immune set point.Nature. 2017; 541: 321-330Crossref PubMed Scopus (2577) Google Scholar] requires insight in both tumor biology and the tumor microenvironment. Recent technological advances have allowed studying hundreds of genes in small, diagnostic biopsies, which are the clinically feasible biosample format in most patients with stage IV cancer. Specific signatures have been validated to carry predictive information across cancer types [4.Ayers M. Lunceford J. Nebozhyn M. et al.IFN-γ–related mRNA profile predicts clinical response to PD-1 blockade.J Clin Invest. 2017; 127: 2930-2940Crossref PubMed Scopus (1755) Google Scholar, 5.Danaher P. Warren S. Lu R. et al.Pan-cancer adaptive immune resistance as defined by the Tumor Inflammation Signature (TIS): results from The Cancer Genome Atlas (TCGA).J Immunother Cancer. 2018; 6: 63.Crossref PubMed Scopus (211) Google Scholar]. We set out to explore and validate the predictive value of a machine-learning approach based on archival, formalin-fixed paraffin-embedded tumor biopsies. Patients with advanced or metastatic NSCLC and available surplus routine biopsy specimens were sequentially enrolled to a training (n = 55) and validation cohort (n = 36; supplementary Table S1, available at Annals of Oncology online). All patients had received anti-PD-1 antibodies in second- or further line. Expression analysis of 770 immune-related genes was performed on the NanoString nCounter platform (NanoString Technologies, Inc., Seattle, USA). Clinical end points were best response, time-to-treatment-failure and overall survival following immunotherapy. From the expression data, predictive feature sets were selected by ensemble-based penalized regression techniques and by utilizing previously published expression signatures of immune cell subtypes. Best performing machine-learning techniques and associated hyperparameters were selected by cross-validation from a set of state-of-the-art algorithms. The feature selection process allowed identifying a subset of approximately 20 of 770 genes that associated with clinical outcome (Figure 1A). We utilized the training cohort to derive prediction models based on the end point of best response following immunotherapy. In the validation cohort, these models successfully identified all ‘top responders’. Concordant prediction of clinical benefit by our models identified a subgroup of patients that benefits from immunotherapy (P = 0.035, hazard ratio = 0.32, Figure 1B). PD-L1 immunohistochemistry appeared to confer an orthogonal layer of information: Incorporating PD-L1 tumor proportion score (PD-L1 TPS) provided a combined prediction with an even stronger predictive value (favorable/intermediate/unfavorable, P = 0.006, Figure 1C). Among patients with PD-L1 positive tumors 10 of 13 were correctly classified (77%); in particular, all patients in this group who did not benefit were correctly identified (3/3 patients, Figure 1D). Our findings show that machine-learning techniques based on nCounter RNA expression data can be applied to achieve immunotherapy response prediction. This approach appeared to provide information in addition to PD-L1 expression. Limitations of our study include the limited sample size of both training and validation cohorts and missing comparison with the predictive information of tumor mutational burden. The employed platform allows analysis of RNA extracted even from small formalin-fixed paraffin-embedded biopsies. Integration into a standard diagnostic workflow relying on small biopsy specimens is feasible. Thus, nCounter analysis can be cost-effective and integrated into the standard molecular pathology workup to enable rapid clinical decision making in precision immunotherapy of NSCLC. This work was supported by a research grant from Bristol-Myers Squibb to the Department of Medical Oncology and by a grant of the Medical Faculty of the University Duisburg-Essen (IFORES fellowship to MW). The West German Cancer Center receives grant support from the Oncology Center of Excellence Program of the Deutsche Krebshilfe (110534); and as partner site of the German Cancer Consortium (DKTK) by the German Federal and State governments. The funding sources had no influence on the analysis and interpretation of data.