Circulating microRNAs associated with prolonged overall survival in lung cancer patients treated with nivolumab

Abstract

Background

The introduction of immune check-point inhibition in non-small cell lung cancer (NSCLC) therapy represents improved prospects for the patients. The response rates to check-point inhibitors are approximately 20% in unselected NSCLC patients. Increasing levels of tumor PD-L1 expression are associated with higher response rates. However, patients with low PD-L1 levels may also have durable responses, and improved strategies for patient stratification are needed.

Material and methods

In this study, we investigated circulating microRNAs aiming to identify circulating predictive biomarkers associated with increased overall survival after immune check-point treatment. Using next generation sequencing, we performed microRNA profiling in serum from NSCLC patients (n = 20) treated with nivolumab. Serum samples from 31 patients were used for validation using qPCR assays. Serum samples were collected prior to immune therapy initiation.

Results

Based on multivariate regression analysis, we identified a signature of seven microRNAs (miR-215-5p, miR-411-3p, miR-493-5p, miR-494-3p, miR-495-3p, miR-548j-5p and miR-93-3p) significantly associated with overall survival (OS) > 6 months in discovery cohort (p = .0003). We further validated this in another similar set of samples (n = 31) and the model was significantly associated with overall survival (OS) > 6 months (p = .001) with sensitivity and specificity of 71% and 90%, respectively.

Conclusions

In this study of circulating microRNAs, we have identified a 7-miR signature associated with survival in nivolumab-treated NSCLC patients. This signature may lead to better treatment options for patients with NSCLC, but a validation in an independent cohort is needed to confirm the predicted potential.

Background

Worldwide, lung cancer has for several years been the leading cause of cancer related deaths [1], and despite improvements in treatment regimens, the majority of patients diagnosed in advanced stage will progress during or shortly after treatment [2]. Immunotherapy has emerged as a promising treatment modality and the blockade of immune check-points has revolutionized the field of cancer treatment. Immune check-points are immune inhibitory pathways which normally mediate immune tolerance and mitigate tissue damage during an inflammation. A number of check-points are so far identified, representing promising targets for blockade [3]. Programmed death protein 1 (PD-1) and its ligands, programmed death ligands 1 and 2 (PD-L1 and PD-L2), are immune check-point proteins that primarily function to limit the activation of T cells in tissues during inflammatory responses. Tumor cells upregulate PD-L1 expression and block antitumor responses in the tumor microenvironment by inhibiting T-cell activation. A number of antibodies against PD-1 and its ligands have been developed and are in clinical use today [3]. Several studies have shown significant response rates and survival in non-small cell lung cancer (NSCLC) patients treated with the PD-1-inhibitor nivolumab [4,5], leading to an approval both by US Food and Drug Administration (FDA) and European Medicines Agency (EMA). Antibodies against PD-1 induce antitumor immune response by re-activating the patient’s own immune system [3].

As shown in several clinical studies, approximately 20% of unselected NSCLC patients respond to immune check-point inhibitors, with an additional 30% experiencing a stabilization of the disease, some of these experiences durable disease control with few side-effects [4,5]. Unfortunately, the treatment is extremely expensive, urging for robust predictive biomarkers to identify patients most likely to respond to the treatment. The PD-L1 as a biomarker has proven to indicate an increased likelihood of response in clinical studies with different PD-1 and PD-L1 inhibitors [6]. However, some patients with PD-L1-negative tumors may experience durable responses, while some patients with PD-L1-positive tumors progress quickly [7]. Different assays as well as different cut-off values for determination of PD-L1 expression have been used in the different studies [8,9]. Standardization of the staining and scoring methods of PD-L1 expression is required [10]. In a recent project, four different assays were tested on 39 tumor samples concluding that although quite consistent, interchanging of the assays and cut-offs would lead to different treatment for some of the patients [11]. This discrepancy calls for additional robust predictive biomarkers which are easy to evaluate, in order to offer patients personalized and effective treatment. A high mutational burden has been shown to correlate with durable clinical benefit of immunotherapy and a prolonged PFS in NSCLC patients. This can be explained by the increased number of T-cell epitopes as a consequence of tumor specific mutations, resulting in an improved T-cell activation against the tumor cells. This analysis has so far mainly been performed by tumor biopsy sequencing [12].

MicroRNAs are small oligonucleotides with regulatory capacity in the cell. By complementary binding of microRNA to mRNA, translation can be inhibited. The expression patterns of microRNAs in tumor cells are very different from that of normal cells, explaining many of the abnormal processes in tumors [13]. Recent studies have also identified microRNAs in biofluids such as serum, plasma, saliva and urine [14], suggesting that tumor specific microRNAs are secreted into the microenvironment. Studies have shown that microRNAs in circulation exhibit important features required for biomarkers. They are stable, seem to be protected from degradation and can easily be analyzed, and hence are attractive biomarker candidates [15]. It has also been shown that tumor-derived microRNAs can modulate cells in the tumor microenvironment [16], and may be essential in the cross talk between tumor cells and immune cells. Specific microRNAs are associated with immune cells active in tumor stroma and suggested to mediate changes in the tumor microenvironment [17].

Blood-based biomarkers are of great interest since acquiring blood samples is less invasive than tissue biopsies and less stressful for the patients. In this study, we investigate circulating microRNAs in serum as predictive and prognostic markers for response to nivolumab in lung cancer patients. To our knowledge, this is the first study to report on circulating microRNAs in serum of nivolumab-treated lung cancer patients.

Material and methods

Fifty-eight patients with advanced NSCLC were included in a nivolumab named-patient-use-program (Bristol Meyer Squibbs, BMS) from June to August 2015 at Oslo University Hospital – The Norwegian Radium Hospital [18]. Of these, 28 were diagnosed with adenocarcinomas and 23 with squamous cell carcinomas. All patients had previously been treated with one or more lines of standard chemotherapy for metastatic disease as previously described [18]. Blood samples were collected from the patients before treatment with nivolumab. Whole blood was collected in tubes with serum clot activator and centrifuged at 2410 rpm for 12 minutes within one hour after collection. Serum was stored at −80 °C until use. Blood samples from 51 of the 58 patients were available at baseline, and were used in this study (Table 1). All the patients have been followed during therapy until death, and clinical data have been recorded. One year after the last inclusion, 14 patients (24%) were still on treatment, and 23 patients (40%) were still alive. Out of the 51 samples, 20 samples were selected for the discovery set. Due to few samples we selected the discovery set based on survival data, and first priority was to start with a balanced number of patients in the two survival groups. Nine patients died within six months of inclusion, and were categorized as poor responders, and 11 patients exhibited an overall survival (OS) longer than six months, categorized as good responders. Samples from the discovery set were analyzed using next generation sequencing (NGS). The remaining 31 samples used in the validation set, were not inspected until the analyses were performed. As seen in Table 1, some of the parameters within this group are more skewed. Samples from the validation set were analyzed using qPCR. This study was approved by the Regional Ethical Committee (2015/1587).

Table 1. Clinical table shows the number of patients in the discovery set and validation set.

	Discovery set		Validation set
	OS <6m	OS >6m	OS <6m	OS >6m
Number of patients	9	11	10	21
Sex
Men	4	7	9	8
Women	5	4	1	13
Histology
Adenocarcinomas	5	7	3	13
Squamous cell carcinomas	4	4	7	8
Number of treatment lines before start of nivolumab
1	1	5	6	7
2	5	5	2	10
3	1	1	2	4
4	2	0	0	0
ECOG status
0	0	1	0	8
1	7	9	5	10
2	2	1	3	2
3	0	0	2	1

The response groups are divided into overall survival >6 months (6m) and <6 months (6m). Sex, histology, and number of treatment lines before start of nivolumab and ECOG status are shown for the different groups.

RNA isolation

RNA was extracted from 500 µl serum for NGS analysis, and from 250 µl for qPCR analysis, using miRCURY™ RNA Isolation Kit. Quality control procedure to monitor RNA isolation efficiency, inhibition, detection of outliers and degree of hemolysis was applied prior to profiling. Spike-ins were added to estimate RNA quality and to monitor reproducibility and linearity of the sequencing reaction. The quality control was performed using qPCR with monitoring of the added spike-ins. RNA isolation, quality control monitoring, NGS, RT-qPCR and normalization were performed at Exiqon Services (Vedbaek, Denmark) and according to the protocols.

The level of hemolysis was assessed using the ratio of miR-451 and miR-23a. Ratios above Cq = 7 indicate affection of hemolysis [19], and the sample should be interpreted with caution.

Next generation sequencing

Profiling of microRNAs was performed using NGS with Illumina NextSeq500, in the serum of 20 nivolumab treated lung cancer patients. RNA molecules were ligated to adaptors and converted into cDNA. The NGS library was prepared using a protocol optimized for biofluids based on NEBNext^® Small RNA Library Prep Kit (New England Biolabs), and amplified with PCR. Quality control of the library was performed by Agilent 2100 Bioanalyzer. The library pool was sequenced on Illumina NextSeq 500. After sequencing, the adaptors were trimmed off and mapped.

Normalization

Reads not aligned to the reference genome were discarded. The microRNAs were mapped to miRBase version 20. After mapping, the reads are divided by the total number of mapped reads in the sample and then multiplied with one million, denoted tags per million (TPM. Before any statistical analysis, the data were a normalized by the trimmed mean of M-values (TMM normalization). This is performed to correct for variation in library size or sequencing depth between the samples, and reduce the skewness and the false positive rate.

Validation of microRNAs

For validation of the results, qPCR panels from Exiqon were utilized (miRCURY LNA™ Universal RT microRNA PCR System, Exiqon) preloaded with primers specific to the microRNAs. First the microRNAs are polyadenylated and reverse transcribed into cDNA, then loaded onto the panels. The amplification was performed on a Roche Lightcycler 480. Normalization of the data was performed using two stable expressed microRNAs (miR-93-5p and miR-222-3p) identified by the NormFinder software [20].

Statistics

The Generalized Linear Model likelihood ratio test (GLM-RT) was performed to identify significantly differentially expressed microRNAs in the discovery set. The microRNAs were ranked in order of evidence for differential expression at p < .05. Leave one out cross validation was performed to check if any of the identified microRNAs were driven by one or two of the samples. Furthermore, the generalized logistic regression model was built with elastic net penalty using GLMNET package in R [21]. The five-fold cross validation was performed to find the lambda parameter that returns the smallest possible root mean squared error statistics for a selected alpha parameter. Using this lambda parameter and elastic net penalty, the model was built on discovery cohort using differentially expressed microRNAs from the GLM-RT test. The predictive performance of the model was tested using leave one out cross-validation in which one observation was omitted at each step and model was built using all the remaining observations and tested on omitted observation. Overall survival below/over six months was set as outcome and the seven identified microRNAs were set as predictors. The predictive model was further tested on the validation set. A receiver operating characteristics (ROC) curve was generated from the prediction analyses. To further assess the relationship between the predictors and survival, Kaplan Meier analyses and multivariate Cox regression were conducted. All the statistical calculations were performed in R version 3.4.3 using the following R-packages, caret (version 6.0-78), glmnet (version 2.0-13), pROC (version 1.10.0), verification (version 1.42), ROCR (version 1.0-7), survcomp (version 1.28.4), and survival (version 2.41-3) [21–26]. The R-script used for the model building and prediction analyses is included in Supplementary Table 1.

Results

The samples analyzed with NGS and qPCR showed overall good quality. All the samples passed the quality control, and they were not affected by hemolysis. The highest number of reads was mapped to small RNAs or aligned to locations where no known microRNA or small RNA is located. Reads mapped to miRBase accounted for less than 15% of the total reads on average, see Supplementary Figure 1. For the NGS profiling of microRNAs, an average of 15.2 reads per sample was obtained. With a TPM ≥1 TPM, 309 microRNAs were identified, and for TPM ≥10 TPM, 154 microRNAs were identified according to entries in miRBase release 20 (Supplementary Figure 2).

In the discovery set, using the GLM-RT test, levels of 25 microRNA were significantly different in the two response groups (p < .05). After correcting for multiple testing, only two microRNAs were significantly distributed in the two response groups (FDR < 0.05). The test indicated that some of the microRNAs were driven by one or two of the samples. To test the robustness of the identified microRNAs, leave one out tests were performed. Only microRNAs consistently significantly differentially expressed in all the tests were considered for validation. This resulted in seven microRNAs (miR-215-5p, miR-411-3p, miR-493-5p, miR-494-3p, miR-495-3p, miR-548j-5p and miR-93-3p) with significant different levels (p < .05) in serum of the two response groups. Some of the microRNAs were detected with reads below 10 TPM. Since we considered that low abundant microRNAs are biological relevant, these were not filtered out. Multivariate regression confirmed a significant association with OS for all the seven microRNAs (p = .005).

To estimate how well the seven identified microRNAs combined could predict survival, we fitted the generalized logistic regression model with elastic net penalty using GLMNET package in R and tested the predictive accuracy using leave one out cross validation (Figure 1). This was performed on the discovery set based on the two groups; OS >6 months (n = 11) and OS <6 months (n = 9). The model could predict OS with sensitivity of 100% and specificity of 77% (Figure 1). Furthermore, the model was tested on the validation cohort. The validation set was divided into group 1 (n = 10) with OS < 6 months, and group 2 (n = 21) with OS >6 months. The model was able to predict the OS of patients with sensitivity of 71%, specificity of 90% and area under the curve (AUC) of 81% (Figure 1). To visualize how the seven microRNAs were correlated, a hierarchical clustering was performed using spearman correlation as distance measure and average linkage (Supplementary Figure 3). The seven microRNAs showed a significant relationship with survival time when using fitted values at optimal threshold (p = .002). This is visualized with a Kaplan Meier curve for the validation set in Figure 2. A multivariate Cox regression analysis was performed to examine the relationship of independent variables with the outcome. Gender, ECOG status, number of treatment lines before nivolumab, histology and a calculated score for the seven microRNAs were included as variables. In the discovery set, only the calculated 7-miRs model score was significantly (p = .04) associated with outcome. In the validation set, ECOG status and the 7-miR model score contributed significantly (p = .025) to the outcome (Supplementary Table 2).

Figure 1. ROC curve shows the sensitivity and specificity based on the seven microRNAs detected in the discovery set. Results from the validation set showed highly concordance between the two data sets, and the predictors were able to identify the responders to nivolumab with sensitivity of 71%, and specificity of 90%.

Figure 2. High calculated 7-miRs model score (blue line) was significantly associated with poor overall survival. Here visualized with Kaplan Meier survival curve.

To evaluate if the signature could be applied to other cut-offs than six months, we included 9 months, 12 months and 18 months as cut-offs. The model was significantly linked with survival at all the cut-off at p < .05, except in the validation set at 18 months (p = .06). A slight decrease in sensitivity and specificity was shown using cut-offs at longer OS, as shown in Supplementary Table 3.

To test if our microRNA-signature was predictive of therapy response or barely prognostic, we inspected the same microRNAs analyzed in a previous study on NSCLC [27]. Profiling of microRNAs was performed in serum of 38 surgical resected NSCLC patients, 16 COPD patients and 16 healthy volunteers. Serum from COPD and healthy volunteers were used as controls. We extracted data from six of the seven microRNAs (miR-215-5p, miR-411-3p, miR-494-3p, miR-495-3p, miR-548j-5p and miR-93-3p), miR-493-5p was not on the platform. A t-test revealed that all the microRNAs except miR-93-3p had a higher abundance in serum of controls as compared to serum of NSCLC patients, of which miR-494-3p, miR-495-3p, miR-548j-5p were significant. The level of miR-93-3p was significantly higher in NSCLC patients as compared to the controls. Multivariate Cox regressions did not show any significant association between OS and the levels of the six microRNAs in serum (data not shown).

Discussion

We profiled microRNAs in serum from NSCLC patients and identified seven microRNAs (miR-215-5p, miR-411-3p, miR-493-5p, miR-494-3p, miR-495-3p, miR-548j-5p, miR-93-3p) associated with OS after treatment with the immune check-point inhibitor nivolumab. Based on the levels of the seven identified microRNAs, the nivolumab-treated lung cancer patients with an OS longer than six months were identified. This microRNA signature was significantly validated, resulting in sensitivity of 71% and specificity of 90%. A blood-based microRNA signature can be of great value to determine lung cancer patients who most likely will benefit from treatment with immune check-point inhibitors.

To further investigate if the identified microRNAs only reflected a prognostic signature, six of the microRNAs profiled in serum of a different cohort of NSCLC patients and controls were examined (cohort 2). None of the six microRNAs were associated with overall survival. This finding strengthens the value of the identified microRNAs as predictive biomarkers. Nevertheless, estimation of prognostic impact requires a large number of patients which underlines the need for further investigation. Furthermore, significantly higher levels of circulating miR-494-3p, miR-495-3p, miR-548j-5p, and lower levels of miR-93-3p were found in the controls as compared to NSCLC patients in cohort 2. Interestingly, high levels of miR-494-3p, miR-495-3p, miR-548j-5p and low levels of miR-93-3p were associated with prolonged survival in present study; when levels of our signature microRNAs are similar to the controls without cancer, the patients are more likely to respond. This also strengthens our signature as a predictive signature of response to checkpoint inhibitors.

Studies have shown that miR-215 can induce cell-cycle arrest in a p53 dependent manner, indicating a tumor suppressor role for the molecule. Tumor suppressor genes are often found repressed in cancer [28]. The expression of miR-215 has been reported to be significantly decreased in colon tumors compared to normal tissues. However, in contrast to our study, high levels of miR-215 were associated with poor OS [29]. Decreased levels of miR-215-5p measured both in tissue and serum, have been reported in metastatic breast cancer patients as compared to non-cancerous controls [30]. In a study of NSCLC, miR-215-5p was down-regulated in lung tumor tissue as compared to adjacent lung tissue, and over-expression of miR-215-5p inhibited progression of lung cancer [31]. This is in line with our study, as low levels of miR-215-5p were associated with decreased OS, which may be explained by the hampered tumor suppressor effect of miR-215-5p.

The microRNAs miR-548j-5p, miR-411-3p and miR-93-3p are not well described in the literature which may be explained by low abundance measures in serum. Interestingly, a previous study has shown that the miR-548 family (including miR-548j) is involved in regulation of interferon- λ1 (IFN-λ1/IL-29), an essential molecule for host response against viral infection [32]. More recently, inhibition of miR-93-3p by circular RNA has been associated with development of hepatocellular carcinoma [33]. In our dataset, low levels of miR-548j-5p and high levels of miR-93-3p were associated with short OS. Upregulation of miR-494-3p has been linked to lung cancer progression and worse survival in lung cancer patients [34], whereas miR-493-5p has been described as a suppressor of invasiveness in breast cancer [35]. In a study of microRNAs associated with resistance to radiotherapy in NSCLC patients, miR-495-3p demonstrated a higher abundance in plasma of those with complete response than in those with less response to the treatment [36]. In our study, increased levels of miR-494-3p, miR-493-5p and miR-495-3p were detected in lung cancer patients with OS >6 months.

Several of the microRNAs identified (miR-411-3p, miR-493-5p, miR-494-3p, miR-495-3p and miR-548j-5p) were detected at very low levels. Since, low abundant microRNAs may be important biomarkers; we did not exclude any of them. It is not expected that tumor-associated microRNAs are detected at high levels in blood [37]. Furthermore, microRNA profiling of red and white blood cells have demonstrated that the majority of circulating microRNAs detected in blood are externalized from the blood cells [38]. Low abundant markers can be difficult to measure in blood, and repetitions of the analysis may be needed using a clinical approach. In addition, standardization of the method is crucial in order to increase the accuracy of the method.

The results obtained in this study are not correlated to PD-L1 expression in tumor due to lack of tumor material. The distribution of patients with different histology, gender and ECOG status’ was not equal in the two response groups, which in non-optimal. These are variables known to influence survival. However, a multivariate Cox regression analysis with the abovementioned variables included, confirmed the predictive potential of the seven microRNAs.

A blood-based predictive test for distinguishing responders from non-responders has many beneficial factors. A blood test can be repeated often for monitoring the disease and is cost effective. Furthermore, the analyses allow for quickly obtained results. On the other hand, factors such as hemolysis, sample preparation, RNA isolation, normalization and method for microRNA detection are previously shown to affect the results [39]. In order to bring microRNA analysis into the clinic, a standardization of these parameters is required.

Heterogeneity is another obstacle in the field of biomarker research. The big diversity of mutations and aberrant cell signaling between tumors, even within tumors at the same stage and histological subgroup, can result in different microRNA profiles. Therefore, it is crucial to combine predictors in order to capture the heterogeneity of the patients. Some of the microRNAs identified in this project were detected only in a few patients. This may be explained by the heterogeneity of the tumors of the patients participating in this study, comprising both adenocarcinomas and squamous cell carcinomas. Despite few samples and low abundance of the markers, the 7-miR signature obtained sensitivity of 71% and specificity of 90%. A drop in both sensitivity and specificity is expected when validated in new cohorts. To further emphasize the predictive potential, a validation in a large independent cohort is needed.

In this study, we chose OS as clinical endpoint. This is in accordance with other studies on check-point inhibitors. Recently, in a large phase 3 multicenter trial, significant improved OS was seen in the NSCLC patients when treated with atezolizumab (anti-PD-L1) compared to docetaxel, albeit the progression free survival was similar in the two groups [40]. Growing evidence indicates that many patients experience off-therapy effect and long-term survival even if the response is partial, stable and the disease progresses [41]. The major goal with this therapy was to increase the lifetime for the patients, and this signature was developed to identify patients with no response to the treatment. These patients have a short expected lifetime and may benefit from other types of treatment. We used a cut-off on six months since it’s not likely that microRNAs detected at baseline can reflect both those with no response to the treatment and those with response followed by progression as one group. We tested different cut-offs and it was a clear drop in sensitivity and specificity at later time-points. However, the signature performed well with cut-offs on both six and nine months.

Conclusions

In this explorative study, we identified seven circulating microRNAs associated with increased OS after treatment with nivolumab. With sensitivity of 71% and specificity of 90% in the validation cohort, the predictive potential of the seven identified microRNAs is a promising tool for identification of responders to nivolumab. Due to few samples in the cohort, the seven microRNAs should be included in a larger independent study to externally validate the predictive potential of the seven microRNAs.

Disclosure statement

No potential conflict of interest was reported by the authors.

Additional information

Funding

This project was supported with grant from Helse Sør-Øst and the Norwegian Cancer Society.