Blood platelet RNA profiles do not enable for nivolumab response prediction at baseline in patients with non-small cell lung cancer

1 Introduction

Lung cancer, especially non-small cell lung cancer (NSCLC), is the most common cause of cancer-related mortality in the world, causing over one million deaths annually [1]. Interestingly, PD-(L)1 checkpoint inhibitors have shown remarkable effectiveness, durability of response and tolerability [2], although overall response rates remain low (∼20%) [3, 4]. Numerous studies have identified correlations between smoking, tumor tissue mutational load, infiltration of CD8-positive T-cells and response to anti-PD-(L)1 immunotherapy [5, 6]. At the moment, current stratification of patients for anti-PD-(L)1 targeted therapy is hampered by limited accuracy and concordance of available biomarkers, including the currently used tumor tissue PD-L1-protein expression marker. Identification of patients with a low likelihood of response to anti-PD-(L)1 immunotherapy, while still correctly identifying individuals who most likely benefit from this therapy, may prevent unnecessary treatment and concomitant costs, and potential exposure of patients to serious immunological adverse events.

Blood platelets are cell fragments involved in hemostasis, initiation of wound healing and metastasis of cancer [7, 8]. They originate from the megakaryocytes that reside in the bone marrow and lung parenchyma [7, 9], and are loaded with pre-mature RNA molecules [10–12]. Upon environmental cues, platelets splice their mRNA molecules resulting in platelet RNA repertoires of both spliced and unspliced mRNAs [10, 14]. In addition, platelets have the capability to sequester RNAs [15] and load platelet microparticles with RNAs that can be transferred to other cells, such as endothelial and tumor cells [16, 17]. Functionally, platelet-derived TGFß and direct physical platelet-tumor cell interactions can induce tumor cell epithelial-mesenchymal transition [18], a process that has previously also been associated with resistance to anti-PD1 immunotherapy [19]. In addition, platelets can behave as immunomodulators in inflammatory conditions [20, 21], regulate immune cell extravasation [22, 23], and cross-communicate with lymphocytes [24] and neutrophils [25]. It was shown that platelets armed with immunotherapeutic antibodies home to lung tumors in mice [26], exhibiting profound anti-tumor activity. Also, previously pretreatment increased platelet-to-lymphocyte ratio in patients with metastatic NSCLC treated with nivolumab was associated with worse overall survival and a reduced response rate [27]. Platelets are thus potentially involved in the immune response towards tumor cells, and might thereby alter their RNA composition.

We recently provided evidence that platelets may serve as blood-based liquid biopsies for the detection and subclassification of cancer [14, 28], as well as for therapy response monitoring [29, 30]. In this prospective study, we investigated whether platelet RNA signatures may provide classification power for nivolumab immunotherapy response prediction before start of treatment.

2 Methods

3 Results

We prospectively collected 300 blood samples from patients with advanced NSCLC that were selected for treatment with the PD-1 monoclonal antibody nivolumab. These were subjected to the thromboSeq protocol (Fig. 1A). 14 samples (4%) were excluded from follow-up analyses according to previously introduced thromboSeq quality measures (Supplemental Figure 1) [37, 41]. The remaining 286 samples were included for the analyses (Table 1). We observed no statistically significant differences between responders and non-responders regarding age, gender, and hospital of patient inclusion (Table 1). We also observed no statistically significant difference in total platelet counts between responders and non-responders of whom this data was available (p = 0.660, n = 185, unpaired Student’s t-test, Table 1). Tumor tissue was partially available for correlative PD-L1 immunohistochemistry analysis. Additional demographic variables are presented in Table 1 and Table S1.

OPEN IN VIEWER

Fig. 1

Platelet RNA-based nivolumab response prediction. (A) Schematic overview of the experimental setup. Blood of patients eligible for treatment with PD-1 inhibitor nivolumab was included from one month before till start of treatment (baseline. t = 0). Tumor response read-out based on CT-imaging and according to the RECIST 1.1 criteria were performed at 6–8 weeks. 3 months. and 6 months. after start of nivolumab therapy. A responder was defined as PR. CR at three months and SD at three and six months. A non-responder was defined as PD and SD after three. but PD after six months. (B) Heatmap and unsupervised hierarchical clustering analysis of RNAs with differential RNA levels between nivolumab responders (n = 88) and non-responders (n = 198). Columns indicate samples. rows indicate RNAs. and color intensity represents the Z-score transformed RNA expression values. Clustering of samples showed non-random partitioning (p < 0.003. Fisher’s exact test). The clinical and technical variables ‘hospital of sample location’. ‘library size’. ‘sex’. ‘age’. ‘smoking status’. ‘tumor stage’. and ‘PD-L1 tumor tissue status’ are indicated on top of the heatmap with each owns color coding (legend on the right). (C) Receiver operating characteristics curve of the thromboSeq nivolumab response prediction algorithm at moment of treatment baseline. Indicated are the training series (dashed gray line). evaluation series (gray line). and validation series (red line). Also indicated are sample numbers. AUC values. and the 95% confidence intervals (95% -CI). (D) 2×2 tables summarizing the classification results for the evaluation series (upper table) and validation series (lower table) of the thromboSeq nivolumab response prediction algorithm at moment of treatment baseline geared towards a rule-out setting. Indicated are sample numbers and percentage of total samples.

The full dataset was subdivided into training (n = 88), evaluation (n = 91), and validation series (n = 107; Supplemental Figure 1C), as previously performed according to our thromboSeq dry-lab protocol [37]. For comparative analysis, we forced the patients from whom the tumor tissue PD-L1 status was available into the validation series. The hospital of origin (i.e. sample collection) was equally distributed among the three series (Table S2). The separation of samples among the three series was defined in the study protocol and based on preliminary data and power calculations (see Methods).

We first performed an ANOVA comparison on all platelet RNA samples, identifying platelet RNAs with discriminative power between nivolumab responders and non-responders. From the 3750 platelet RNAs that were detected in the dataset in total, 27 had statistically significant differential RNA levels (False Discovery Rate (FDR)< 0.05, Table S3). From these RNAs, only one RNA from the GSEA inflammatory response gene data set [42] had a statistically significant differential expression level (NAMPT, FDR = 0.0002).

Heatmap dendrogram clustering of the samples applying these 27 RNAs resulted in non-significant clustering between responders and non-responders (p-value: 0.24, Fisher’s exact test, data not shown). Adjusting the ANOVA-statistics significance threshold employing a PSO-function (see Methods) resulted in dendrogram clustering discriminating responders from non-responders (p-value< 0.003, Fisher’s exact test, used FDR-threshold: 0.419696, Fig. 1B). Improving discriminating power by adjusting the ANOVAs FDR-threshold implies that also statistically non-significant RNAs can contribute to the clustering accuracy. This provides the rationale for the PSO-enhanced classification algorithm development selecting an RNA panel based on most contributive power to a discriminative biomarker panel.

We next performed PSO-enhanced training of an SVM-classification algorithm, forcing the algorithm to identify optimal algorithm settings including the RNA biomarker panel for a rule-out nivolumab response prediction classifier. This would enable accurate identification of patients who do respond to nivolumab, while providing the opportunity to exclude patients that will not respond. This training process resulted into an area under the curve (AUC) for the training series of 0.73 (95% -CI: 0.63–0.84, n = 88 samples, as determined by LOOCV), and for the evaluation series of 0.64 (95% -CI: 0.51–0.76, n = 91 samples) for the evaluation series. Subsequent validation of the 107-samples validation series resulted into an AUC of 0.58 (95% -CI: 0.45–0.70, n = 5 spliced RNA biomarker panel, Fig. 1C). The five-gene biomarker panel as unbiasedly selected by the thromboSeq software was composed of the RNAs HPSE, HBD, PRSS50, CRYM, and LARP1 (SupplementalFigure 2A). Using the classification score readout towards a ‘rule-out’ application, in a sensitivity of 96% (95% -CI: 82–100%, n = 28) with a specificity of 24% (95% -CI: 14–36%, n = 63) in the evaluation series, resulted in a concomitant sensitivity in the validation series of 76% (95% -CI: 59–89%, n = 34) and a specificity of 18% (95% -CI: 10–29%, n = 73, Fig. 1D), with which our predefined threshold of a successful new biomarker is not reached.

Subsequent analysis of classification accuracies according to PD-L1 tumor tissue status of whom the data was available showed on average increased classification scores for patients responding to nivolumab as compared to patients not responding to nivolumab (Supplemental Figure 2B).

Anti-PD-(L)1 therapies have revolutionized cancer treatment, although, (non-invasive) predictive biomarkers are highly desirable [43]. In this proof-of-concept study, we aimed to associate spliced platelet RNA profiles to nivolumab response prediction measures at moment of therapy baseline.

Platelets act as local systemic responders during tumorigenesis [8], thereby suffering from tumor-mediated platelet education, and altering its behavior [18, 44]. In previous studies, we have demonstrated that platelet RNA can function as a biomarker trove to detect and classify cancer from blood [14, 28]. Using our previously developed thromboSeq platform, we could not find solid RNA repertoires associated with tumor response at moment of start of therapy. Therefore, we were in this setting and study setup unable to successfully validate a nivolumab response prediction thromboSeq algorithm. Therefore at this point, interrogation of platelet RNA signatures before start of nivolumab treatment seems not to be useful for nivolumab therapy selection.

Because platelets are regarded as local immuno-like responders, the role of platelets in anti-tumor activity prior to and during anti-PD-(L)1 therapy, including direct and indirect physical interactions with lymphocytes, macrophages, and tumor cells should not be underestimated. In our five-gene platelet biomarker panel, both HPSE (heparanase) and LARP1 (La RibonucleoProtein 1) have lower expression values in nivolumab responders. HPSE is active in tumor invasion and inflammatory processes [45, 46], where-as LARP1 is associated with dengue virus and coronavirus disease-19 (COVID-19) [47, 48]. Higher expression levels in nivolumab non-responders may suggest either a different inflammatory process required for anti-tumor response, or a highly progressive tumor, which may explain the insensitivity to nivolumab treatment.

In this light, the current results were unexpected. We have several explanations for our findings; 1) platelet RNA is only very subtly altered in the presence of immunotherapy-sensitive as opposed to immunotherapy-resistant NSCLC tumors for which insufficient sample numbers were available to uncover such differences. Alternatively, the employed machine learning algorithms and gene panel selection approach had insufficient discriminatory power; 2) the alterations are masked by the presence of multiple platelet subpopulations; 3) the alterations may be best measured after start of treatment, e.g. a couple of weeks following the first nivolumab exposure; 4) insufficient samples were included into the training process; 5) despite matching of several confounding factors, remaining (unmeasured) factors and/or pre-analytical variables may increase the gene expression noise as opposed to the (low-level) signal and 6) in the power calculation the estimated RNA expression difference was overestimated, resulting into a too small anticipated cohorts sample size and therefore to a reduced algorithm’s performance. Also, we cannot exclude that platelet RNA is of value as a predictive biomarker platform for other (immuno) therapeutics. Here, re-analysis of the data using other machine learning approaches is of interest and the data is publicly available for such purposes.