Bioinformatics Model of Serum Biomarkers to Prognosticate the Response to Programmed Death-1/Ligand-1 Targeted Immunotherapy in Metastatic Non–Small Cell Lung Cancer. (11th September 2019)

• Record Type: Journal Article
• Authors: Tarhoni, Imad; Multani, Maneet; Fughhi, Ibtihaj; Gerard, David; Fidler, Mary; Batus, Marta; Bonomi, Philip; Borgia, Jeffrey
• Is Part Of: American journal of clinical pathology. Volume 152(2019)Supplement 1
• Page Start: S137
• Page End: S137
• Publication Date: 2019-09-11
• Subjects: Diagnosis, Laboratory -- Periodicals; Pathology -- Periodicals; 616.07
• DOI: 10.1093/ajcp/aqz126.008
• Languages: English
• ISSNs: 0002-9173

Abstract

Abstract: Introduction: Immune-checkpoint inhibitors revolutionized the therapeutic paradigm for metastatic non–small cell lung cancer (NSCLC). The average response, however, still hovers at 20%, demonstrating the urgent need for biomarkers predictive of response. High-throughput laboratory technology promises to serve as an insightful and robust tool to recognize and select patterns of biomarkers in serum. We applied machine learning on serum immune-checkpoint biomarkers for prognostication of response to immunotherapy in advanced NSCLC. Method: Pretreatment sera from 106 advanced NSCLC cases who failed frontline chemotherapy were evaluated for 16 soluble immune-checkpoint molecules using the Human Immuno-Oncology Checkpoint Protein Panel (MilliporeSigma). This panel constituted BTLA, CD27, CD28, TIM-3, HVEM, CD40, GITR, GITRL, LAG-3, TLR-2, PD-1, PD-L1, CTLA-4, CD80, CD86, and ICOS. Primary data points were collected and calculated via a Luminex FLEXMAP 3D system (xPONENT v4.0.3 Luminex Corp). The minimum follow-up after treatment was 12 months. Response patterns were categorized based on their overall survival (OS) as long-term responders (>12 months) or short-term responders (<12 months). Values were analyzed with the clinical outcomes using "Survminer" and "survival" R packages to determine the log-rank-based cutoff values associated with overall survival. Finally, machine learning methods were implemented using "caret" and "rpart" R packages to fit a classification model to predict the response pattern. The model was trained and tested on random fractions of the cohort. Results: BTLA4, HVEM, CD40, GITRL, LAG-3, PD-1, CD80, and CD86 serum levels significantly correlated with OS (all P values ≤.02 and HR of 0.27, 0.5, 4.59, 0.17, 0.12, 0.48, 3.64, and 0.37, respectively). The algorithm composing PD-1, LAG-3, CD86, and CTLA4 predicted the response pattern with PPV of 81%, specificity of 87%, and accuracy of 75%. Conclusion: The serum immune-checkpoint predictive model might assist in the tissue and gene-based profiling of immune-checkpoints to predict the benefit from immunotherapy.