BackgroundImmune checkpoint inhibitors (ICIs) have revolutionized the cancer therapy landscape due to long-term benefits in patients with advanced metastatic disease. However, robust predictive biomarkers for response are still lacking and treatment resistance is not fully understood.MethodsWe profiled approximately 800 pre-treatment and on-treatment plasma proteins from 143 ICI-treated patients with non-small cell lung cancer (NSCLC) using ELISA-based arrays. Different clinical parameters were collected from the patients including specific mutations, smoking habits, and body mass index, among others. Machine learning algorithms were used to identify a predictive signature for response. Bioinformatics tools were used for the identification of patient subtypes and analysis of differentially expressed proteins and pathways in each response group.ResultsWe identified a predictive signature for response to treatment comprizing two proteins (CXCL8 and CXCL10) and two clinical parameters (age and sex). Bioinformatic analysis of the proteomic profiles identified three distinct patient clusters that correlated with multiple parameters such as response, sex and TNM (tumors, nodes, and metastasis) staging. Patients who did not benefit from ICI therapy exhibited significantly higher plasma levels of several proteins on-treatment, and enrichment in neutrophil-related proteins.ConclusionsOur study reveals potential biomarkers in blood plasma for predicting response to ICI therapy in patients with NSCLC and sheds light on mechanisms underlying therapy resistance.

Importance: Advanced stage non-small cell lung cancer (NSCLC) patients with no driver mutations are typically treated with immune checkpoint inhibitor (ICI)-based therapy, either in the form of monotherapy or concurrently with chemotherapy, while treatment modality selection is based mainly on programmed death ligand 1 (PD-L1) expression levels in the tumor. However, PD-L1 assays are only moderately predictive of therapeutic benefit. Objective: To develop a novel decision-making tool for physicians treating NSCLC patients on whether to administer immune checkpoint inhibitor (ICI) therapy alone or in combination with chemotherapy. Design, setting, and participants: This multicenter observational study includes patients from an ongoing clinical trial (PROPHETIC;NCT04056247). Patients were recruited from 13 different centers (total n=425; 58 patients were excluded) from June 2016 and June 2021. Plasma samples were obtained prior to treatment initiation, and deep proteomic profiling was conducted. PROphet computational model for predicting clinical benefit (CB) probability at 12 months was developed based on the plasma proteomic profile. The model performance was validated in a blinded manner. Following validation, training and prediction was performed over the entire cohort using cross-validation methodology. The patients were divided into four groups based on their PD-L1 expression level combined with their CB probability, and the survival outcome was examined for each group. The data were analyzed from July to October 2022. Main outcome and measures: Clinical benefit from ICI-based treatment, overall survival (OS) and progression-free survival (PFS). Results: The model displayed strong predictive capability with an AUC of 0.78 (p-value = 5.00e-05), outperforming a PD-L1-based predictive model (AUC = 0.62; p-value 2.76e-01), and exhibited a significant difference in OS and PFS between patients with low and high CB probabilities. When combining CB probability with PD-L1 expression levels, four patient subgroups were identified; (i) patients with PD-L1>=50% and a negative PROphet result who significantly benefit from ICI-chemotherapy combination therapy compared to ICI monotherapy; (ii) patients with PD-L1≥50% and a positive PROphet result who benefit similarly from either treatment modalities; (iii) patients with PD-L1<50% and a negative PROphet result who do not benefit from either treatment modalities; (iv) patients with PD-L1<50% and a positive PROphet score who benefit from combination therapy. Conclusions and relevance: The PROphet model displayed good performance for prediction of CB at 12 months based on a plasma sample obtained prior to treatment. Our findings further demonstrate a potential clinical utility for informing treatment decisions for NSCLC patients treated with ICIs by adding resolution to the PD-L1 biomarker currently used to guide treatment selection, thereby enabling to select the most suitable treatment modality for each patient.

The PROphet test provides a decision-making tool for first-line treatment of non-small cell lung cancer (NSCLC) patients without driver mutations. The test is based on plasma proteomics profiling using the SomaScan platform, followed by a machine learning-based analysis. During the PROphet predictor development, a set of predictive proteins was identified, termed resistance-associated proteins (RAPs). Here we set to examine the analytical validity of the SomaScan assay as the biomarker measurement method in the PROphet test, focusing either on all proteins measured using this platform, or specifically on the RAPs. Experimental precision analysis displayed a median coefficient of variation (CV) of 3.9% and 4.7% for intra-plate and inter-plate examination, respectively, when studying all proteins, with no difference between technicians. Notably, the RAPs displayed lower median CV values and lower CV range. The median accuracy rate of measurements obtained in sites was 88% and 94% for all proteins and for the RAPs, respectively, with no significant difference between technicians. Computational precision examination displayed high precision for 13 out of 14 examined samples, with median standard error of 0.1689. Cross-platform comparison between SomaScan platform and other proteomics platforms, including immunoassays, proximity extension assay (PEA) and mass spectrometry, displayed a median Spearman coefficient of 0.51 and 0.53 for all examined proteins and RAPs, with no significant difference between the two protein sets. Last, the effect of the cross-platform correlations on the prediction capabilities was examined, displaying only a minor effect of the correlation on the prediction capabilities of the PROphet predictor, with a median of 80% agreement between predictions obtained between the two examined proteomic platforms. Taken together, these results demonstrate strong analytical performance for the SomaScan technology, while the RAPs displayed improved capabilities in most of the analytical performance analyses.