Patients with irresectable stage III/IV advanced melanoma or stage IV NSCLC were included if they had received intravenous monotherapy with either ipilimumab (anti-CTLA-4; 3 mg/kg every 3 weeks for four courses), nivolumab (anti-PD-1; 3 mg/kg every 2 weeks until a maximum of 2 years) or pembrolizumab (anti-PD-1; 2 mg/kg every 3 weeks until a maximum of 2 years) as standard of care at the Erasmus University Medical Center (Rotterdam, The Netherlands), Leiden University Medical Center (Leiden, The Netherlands) or University Hospital Zürich (Zürich, Switzerland), all being referral hospitals. Patients who received ICI combination therapy or who were treated with a prior line of any form of immunotherapy were excluded, pretreatment with corticosteroids was not considered. Following written informed consent of the patients, blood samples were collected before the first administration of ICIs and after the last administration of any previous treatment if applicable.

Binary grouping was performed according to patients with (responders) or without (non-responders) clinical benefit based on Response Evaluation Criteria in Solid Tumors V.1.1. For determination of best overall response (BOR), confirmation of a complete or partial response (CR/PR) was not required, but a minimum duration of 90 days was required for stable disease (SD). Patients with a CR, PR or SD as BOR were considered to have obtained clinical benefit after ICIs and were defined as responders. Patients with progressive disease (PD) were defined as non-responders. Additionally, binary grouping was performed according to progression-free survival, measured from start of treatment to death or the first evaluation time point that PD is detected; the two groups included late (>140 days) or no progression (responders) versus patients with early progression within 140 days (non-responders). Clinical parameters and chemistry or blood parameters were evaluated at baseline and included age, gender, WHO performance score, pathological tumor type, presence of brain metastases and serum lactate dehydrogenase (LDH) levels.

Venous blood of patients was collected at baseline using either sodium-heparin or EDTA as anticoagulant, and isolation of PBMCs was done within 4 hours or within 24 hours depending on the local study protocol (online supplemental table 1). PBMC were isolated by density gradient centrifugation and cryopreserved until further use. Erythrocyte lysis was only performed in the Mel-CTLA4-B cohort if PBMC still contained considerable erythrocyte contamination after isolation. Importantly, kinase activity was affected by erythrocyte lysis during PBMC isolation (online supplemental figure 1a). Moreover, overall reduced kinase activity was observed in PBMCs that were isolated after 24 hours from blood collected in EDTA anticoagulated tubes (online supplemental figure 1b).

Cryopreserved cells were thawed, washed with phosphate-buffered saline and lysed using ice-cold M-PER lysis buffer (Mammalian Protein Extraction Reagent, Thermo Fisher Scientific, Massachusetts, USA) containing 1:100 Protease Inhibitor Cocktail (Thermo Fisher Scientific). Cell lysate was obtained by centrifugation at 14.000 ×g for 10 min at 4°C. After centrifugation, the supernatant was snap-frozen in aliquots and stored at −80°C. The protein concentration was determined using the Bradford assay (Thermo Fisher Scientific) with bovine serum albumin (BSA) as the standard.

PBMC kinomic activity was measured using protein tyrosine kinase (PTK) PamChip-96 microarrays (catalog # 86311, PamGene International BV, ’s-Hertogenbosch, The Netherlands) using standard manufacturer protocol. In short, the microarrays were blocked with 2% BSA (Calbiochem # 126609) to prevent non-specific binding. After blocking, the arrays were washed three times with 1×PK buffer (50 mM Tris-HCl pH 7.5, 10 mM MgCl₂, 0.01% Brij35, 2 mM DTT). For kinomic profiling 2 µg PBMC lysate protein was used in 40 µL PTK assay buffer (1×PK buffer, 10mM DTT, 400 µM ATP, 1×PTK additive (PamGene International BV), 1:400 Halt Phosphatase Inhibitors (Thermo Fisher Scientific), 0.01% BSA (Calbiochem), and fluorescein isothiocyanate-labeled antiphosphotyrosine antibody (PamGene). During incubation, the reaction mixture was pumped up and down through the porous membrane for 60 cycles at 2 cycles/min. Incubation and read-out of the microarrays was performed with 96 arrays in parallel on a PamStation-96. Typically, three to four technical replicates of each sample were measured in the same run. Time courses of peptide phosphorylation were followed by recording fluorescent images which were quantified by automated image analysis in Bionavigator 6.3.67 (PamGene). Analysis of signals was performed after local background subtraction in Bionavigator 6.3.67 interfaced to the open source statistical program R 3.3.1 (R-project, www.rproject.org).

Analysis was specifically performed per cohort and comprised a quality check, removal of outlier replicates, data transformation and averaging of the kinase signal replicates. To check data quality, the first step was exclusion of a low portion of arrays that showed clear visual defects (eg, broken membrane and large stains) or technical replicates clearly deviating from the other replicates of the same samples. In the second step, data was log-transformed (CTLA-4 cohorts: Mel-CTLA4-A, Mel-CTLA4-B) or normalized using the variance-stabilizing normalization (VSN) method (PD-1 cohorts: Mel-PD1-A, Mel-PD1-B, NSCLC-PD1). Prior to log-transformation, a small fraction of negative values in the (background corrected) signals was handled by setting all signals<1 or equal to 1. Signal-positive spots were required to show a positive trend in the recorded phosphorylation time course. Peptides for which such a trend could not be detected in >75% of the samples were excluded from further analysis. Effectively, 88 to 113 peptides were included for further analysis. Technical replicates of each patient sample were averaged resulting in a single kinase activity profile per included patient for subsequent analysis.

Principal component analysis (PCA) was performed on the transformed and filtered data and PCA scores were used to identify systematic variation, for example, as a result of different sub-cohorts and/or the use of different PamStation-96 microarray plates. Systematic variation was handled by applying ComBat29 batch correction, and outlier patient samples were removed. For the small Mel-CTLA4-A cohort, no such correction was necessary. For the Mel-CTLA4-B cohort, a batch effect was observed and corrected for the batches in which samples were lysed together. A small number (ranging from one to eight) of outlier samples according to PCA scores or heat map visualizations of the transformed profiles were removed. In most cases, these were samples showing low overall kinase activity. Other reasons for removing measured patient samples from the analysis was unavailability of clinical data or if, after re-evaluation, patients were observed not to fulfill the selection criteria.

Kinase activity profiles were correlated to the clinical response to ICIs. Univariate (per peptide) analysis was performed using a two-sided two sample t-test with binary grouping as a covariate (responders vs non-responders). Peptides with p<0.05 were regarded as significant; in addition, the proportion of false discoveries in a set of significant peptides was estimated using the false discovery rate (FDR) method of Benjamini and Hochberg. Classification analysis was performed using partial least square discriminant analysis (PLS-DA) with patients divided in binary groups as previously described.13 In short, a classification model is trained using all tested peptides, that is, without prior peptide selection based on responder and non-responder differences. The resulting model is a set of coefficients (one for each peptide+an offset) that can be applied to new observations to obtain a score that predicts classification in either groups. For each cohort, the correct classification rate (CCR) was estimated using cross validation. Approximately 90% binomial CIs (CI₉₀) for the CCR estimates were obtained using the exact method (note that a CI₉₀ implies with 95% confidence that the CCR is higher than the lower limit). Prediction of kinases responsible for the changes in peptide phosphorylation in the kinomic profiles were obtained using the Upstream Kinase Analysis tool in Bionavigator 6.3.67 (PamGene). Results were visualized by annotation to a kinase phylogenetic tree using the web-based Coral tool (http://phanstiel-lab.med.unc.edu).30 Additionally, Ingenuity Pathway Analysis (Qiagen, Hilden, Germany) was used to perform gene set enrichment analyzes using the delta and unadjusted p value of each peptide (two-sided two sample t-tests). Pathway activation or inhibition were predicted by the z-score statistic, and further explored by the MAP tool (Qiagen, Hilden, Germany). Full access of the data sets generated during the current study is provided in the supplementary materials (online supplemental tables 2-6).

The correlation of kinase activity profiles with treatment response is visualized for the discovery cohort Mel-CTLA4-A using a heat map (figure 2a). A profound difference in kinase activity was observed between the responder and non-responders. Generally, the phosphorylation signal of peptides was higher in responders compared with non-responders. For 83% of the target peptides, a significantly higher signal was found in responders compared with non-responders (two-sided two sample t-test, p value<0.05; FDR<5%). This overall increase in kinase activity was confirmed in the cross-validation cohort Mel-CTLA4-B (figure 2b), although less pronounced since only 23% of the target peptides displayed a significantly higher signal in responders compared with non-responders (two-sided two sample t-test, p value<0.05, FDR=18%). In addition, the relative increase in signal was higher in the Mel-CTLA4-A cohort (median Log2 fold change=0.84; SD=0.15; ~80% increase) compared with the Mel-CTLA4-B cohort (median Log2 fold change=0.40; SD=0.17; ~30% increase). This increase in signal appears to reflect systemic, non-specific signaling in the responders compared with the non-responders.

Discovery cohort Mel-PD1-A included melanoma patients who were treated with anti-PD-1. The profound higher kinase activities as observed in the responders in the two anti-CTLA-4 cohorts was not observed in the three anti-PD-1 cohorts. Therefore, the data was normalized for overall kinase activity using the VSN method. As a consequence, the data reflects differences in the ratio between peptides on the array rather than the overall differences. Responder patients showed a different kinase activity profile when compared with non-responders (figure 3a). For 17% of the peptides in cohort Mel-PD1-A, a significant different signal was found in responders compared with non-responders (two-sided two sample t-test, p<0.05; FDR=29%). These differentially phosphorylated peptides represented both higher and lower signals in responders compared with non-responders. Likewise, differential peptide phosphorylation was observed in the cross-validation cohorts Mel-PD1-B and NSCLC-PD1 (figure 3b–c). In Mel-PD1-B, 16 peptides (18%) displayed significantly differential signals for response (two-sided two sample t-test, p<0.05, FDR=25%) whereas NSCLC-PD1, 18 peptides (19%) were significantly differently phosphorylated in responders compared with non-responders (two-sided two sample t-test, p<0.05, FDR=24%). The specific differences in kinase activities differentiating responders and non-responders varied between cohorts.

To investigate the potential use of kinase activity profiling as a biomarker for response to ICI therapy, classification analysis was performed using the binary grouping of responders and non-responders. Because of the kinase profile variation between the CTLA-4 and PD-1 ICIs cohorts, a separate PLS-DA classification model was trained for each cohort and predictive scores for each patient were obtained using cross-validation. This resulted in a CCR of 100% (90% CI 74% to 100%) in the discovery cohort Mel-CTLA4-A and 83% (64% to 93%) in cross-validation cohort Mel-CTLA4-B (figure 2c and d). The CCR was 93% (80% to 99%) in the discovery cohort Mel-PD1-A, 78% (63% to 88%) in cross-validation cohort Mel-PD1-B, and 68% (56% to 78%) in the second cross-validation cohort NSCLC-PD1 (figure 3d–f).

Bioinformatics was applied to interrogate the biological processes underlying response or resistance to ICIs. Here, we have zoomed in on the anti-PD-1 cohorts as the overall increase of kinase activity in anti-CTLA-4 responders hampered proper identification of relevant and recognized peptide targets. Identified kinases were annotated to a phylogenetic tree for protein tyrosine kinases (figure 4). In the discovery cohort Mel-PD1-A, predictions revealed that the vascular endothelial growth factor (VEGF) family kinases and feline sarcoma (FES)/FES-related (FER) have relatively higher activity in responders compared with non-responders. Similarly, but less pronounced, the activity of TYRO-3, AXL and MER kinases of the TAM-family and the tropomyosin receptor kinase (TRK)-family were positively correlated with response in this cohort. In the cross-validation cohort Mel-PD1-B; however, the activity of several kinases, including the SRC family kinases, were negatively correlated with response. Interestingly, the above observations were corroborated in the second cross-validation cohort NSCLC-PD1, where VEGF kinases were found to have higher activity and the SRC family kinases to have lower kinase activity in responders compared with non-responders. Canonical pathway analysis using Mel-PD1-A and NSCLC-PD1 cohorts revealed the importance of peptide targets involved in immune cell migration/leukocyte extravasation and co-stimulation of T helper cells. Finally, increased kinase activity in the STAT3, ERBB, VEGF and EGF signaling pathways were found to be related to response to anti-PD-1, whereas the PTEN activation was associated with resistance to anti-PD-1 (figure 5 and online supplemental figure 3).