Genomics of NSCLC patients both affirm PD-L1 expression and predict their clinical responses to anti-PD-1 immunotherapy

In clinical trials of PD-1 or PD-L1 checkpoint immunotherapies, patients with NSCLC separate into groups that respond or do not respond to immunotherapy treatment [1‐4]. Objective responses for NSCLC in these studies range from 19.0–23.0%. Patients are selected for immunotherapy based on immunohistochemistry (IHC) detection of PD-L1 reactivity. Positive PD-L1 reactivity in tumors is considered to be important to predicting the success of PD-1 and PD-L1 immunotherapy treatments [2, 5]. However, IHC results to detect PD-L1 reactivity can vary depending upon different IHC platforms, differences in anti-PD-L1 antibodies, differences in scoring systems, and differences in positivity cut-off values [4, 6‐9]. In the Blueprint PD-L1 IHC Assay Comparison Project [10], similar antibody-specific differences were seen. In all, this variability presents challenges for using PD-L1 reactivity as a sole marker for diagnosis and as a marker to predict the success of PD-1 and PD-L1 immunotherapy treatments.

More likely, there is a complex profile of molecules that contributes to the regulation of PD-L1 and to the subsequent immunosuppressive effects that NSCLC cells have on immune cells [11]. Chae et al. suggested that reliable predictive molecules need to be identified that can be used to select patients who would benefit from immunotherapy, yet limit the exposure of patients who would not benefit or have adverse reactions [12]. Multifaceted predictive biomarker systems have also been proposed that contain input on PD-L1 expression, tumor mutations, and the roles of inflammatory cells to identify patients that would respond or not respond to immunotherapy treatment [13, 14]. For better treatment outcomes, there is a recognized need to develop additional methods that can identify a profile of molecules that contributes to the regulation of PD-L1 expression and affirms IHC PD-L1 positive reactivity. One approach is to use the influence of patient cell genomics on tumor cell signaling to identify the downstream effects on PD-L1 expression.

In this study, we hypothesized that patient tumor cell genomics influences cell signaling and the expression of PD-L1, chemokines, and immunosuppressive molecules. We also hypothesized that these profiles can be used to predict patient clinical responses. Rizvi et al. assessed the mutational profiles that determined sensitivity to PD-L1 blockade from patients with NSCLC treated with pembrolizumab [15] and we used the Rizvi et al. dataset to test our hypothesis. We first assessed the effect of patient genomics on the expression profile of 24 molecules: PD-L1, 9 chemokines, and 14 immunosuppressive molecules. Differences among patient-specific models reflected the input that their deleterious gene mutation profiles had on modeled signaling pathways and the expression of PD-L1, chemokines, and immunosuppressive molecules. Second, we used the expression profiles of these 24 chemokines and immunosuppressive molecules to sort patients into those that would or would not respond to PD-1 immunotherapy. The 9 chemokines were used to generate an index to predict dendritic cell infiltration and PD-L1 and the 14 immunosuppressive molecules were selected as tumor-derived molecules with a long list of reported immunosuppressive functions (Additional file 1: Table S1). Our results suggest that patient-specific chemokine and immunosuppressive molecule expression profiles can be used to accurately predict clinical responses thus differentiate among patients who would or would not respond to PD-1 immunotherapy.

In this retrospective study, we used a recent dataset from NSCLC patients treated with pembrolizumab and identified deleterious gene mutational profiles in patient exomes. We annotated the deleterious gene mutational profiles into a cancer network to create NSCLC patient-specific predictive computational simulation models. We used these models as a tool to identify and validate a profile of 24 chemokines and immunosuppressive molecules that could accurately affirm expression of PD-L1 and predict patient clinical responses to PD-1 immunotherapy. We found that patient tumor cell genomics influenced cell signaling and altered the expression of PD-L1, 9 chemokines, and 14 immunosuppressive molecules. We also found that expression profiles of these 24 chemokines and immunosuppressive molecules could be used to identify patients who would or would not respond to PD-1 immunotherapy. Adding chemokine and immunosuppressive molecule expression profiles to a predicted PD-L1 profile allowed models to achieve a greater than 85.0% predictive correlation among predicted and reported patient clinical responses. This differentiated patients who would and would not benefit from PD-1 or PD-L1 immunotherapies. To validate our results, we used retrospective correlation of our simulation models against patient genomic signatures and clinical outcome data that was available in the NSCLC cohort of the Rizvi et al. study [15]. We also used Weka 3 to validate predictions determined using the predictive computational simulation models. The Weka 3 results were similar to that generated via machine-learning methods and the Chi-square test was used to show no differences among the match rate results in these datasets. It is important to note that expanding PD-L1 expression profiles to include 23 additional chemokine and immunosuppressive molecule expression responses allowed models to achieve a greater than 85.0% correlation among predicted and reported patient clinical responses.

The 24 molecules used in this study have immunosuppressive properties. The role of PD-L1 in tumor pathogenesis is well known. Increased expression of PD-L1 on tumor cells inhibits T-cell proliferation, reduces T-cell survival, inhibits cytokine release, and promotes T-cell apoptosis [3, 35‐38]. This leads to T-cell exhaustion and adaptive tumor immunosuppression [39].

Cytokines also have a role. Cytokine scores were recently found to be associated with overall survival in CheckMate 017 and 057 (both for nivolumab and docetaxel treated patients) [40]. In this present study, 9 chemokines were selected that chemoattractant dendritic cells [41‐43]. In other studies, dendritic cells were present in NSCLC tumors [42, 44] and dendritic cell infiltration was reported to be an independent prognostic factor for NSCLC [44]. The simulation models here captured the role of dendritic cells by predicting chemokine expression in the form of a functional dendritic cell index (Additional file 6: Table S4).

Fourteen molecules were included that have pleiotropic functions including immunosuppressive properties (Additional file 1: Table S1). IL6, can prevent dendritic cell maturation, prime tumor-specific T-cells via signal transducer and activator of transcription 3 (STAT3) signaling, inhibit NF-κB binding activity, and inhibit C-C chemokine receptor type 7 (CCR7) expression [45‐47]. IL10 can impair dendritic cell function and protect tumor cells from cytotoxic T-cell-mediated cytotoxicity by downregulating transporter-associated with antigen processing (TAP)1 and TAP2 [48, 49]. TGFβ can alter immune surveillance of regulatory T-cells. It represses CTL-mediated tumor cytotoxicity by altering the expression of perforin, granzyme A, granzyme B, Fas ligand (FASL), and IFNγ [49‐51]. VEGF is a marker of tumor invasion and metastasis and can inhibit maturation of dendritic cells [44, 52]. IDO, a tryptophan-metabolizing enzyme that limits tryptophan, inhibits the proliferation of lymphocytes, and contributes to peripheral immunologic tolerance [53‐56]. Increased IDO production by cancer cells down regulates natural killer (NK) receptors and induces NK cell apoptosis. It induces cell cycle arrest, decreases activation, and increases apoptosis in cytotoxic T-cells. Tryptophan 2, 3-dioxygenase 2 (TDO2) inhibits tryptophan 2,3-dioxygenase [57]. Prostaglandin E2 (PGE2) suppresses NK cell function through the E₂ prostaglandin receptor 4 (EP4) [58]. Lectin, galactoside-binding, soluble, 9 (LGALS9) mediates T-cell dysfunction and T-cell senescence [59]. Cluster of Differentiation 47 (CD47) is a negative regulator of dendritic cells binding to signal regulatory protein (SIRP) on dendritic cells and directly repressing dendritic cell phagocytosis, maturation, and production of IFNγ [47]. CTLA4 restrains the adaptive immune response of T-cells towards tumor-associated antigens [60‐62]. Gangliosides GM3 and GD2 induce monocyte apoptosis and impair differentiation to dendritic cells [63].

Using a 24 molecule expression profile allowed computational models to achieve a greater than 85.0% predictive correlation. However, this list was not exclusive and incorporating additional molecules into patient NSCLC-specific expression profiles may have merit and improve computational model accuracy. These included Lymphocyte activation gene-3 (LAG-3), T cell immunoglobulin-3 (TIM-3), and T cell immunoglobulin and ITIM domain (TIGIT) co-inhibitory receptors [64]. LAG-3 is a co-inhibitory receptor upregulated on activated CD4+ T cells, CD8+ T cells, and subsets of natural killer (NK) cells [64‐66]. It impairs T cell proliferation and cytokine production and alters NK cell cytotoxicity and cytokine production. TIM-3 is a cell surface molecule expressed on IFNγ-producing CD4+ T helper 1 cells, CD8+ T cytotoxic 1 T cells, NK cells, monocytes, and dendritic cells [64]. TIM-3 dampens the development of protective immunity and TIM-3 blockade improves cell function. In patients with NSCLC, co-blockade of the TIM-3 and PD-1 pathways suppresses tumor growth. TIGIT is another co-inhibitory receptor expressed on NK cells, T cells, and Treg cells [64]. CD155, CD112, and TIGIT ligands suppress immune responses through CD155 on dendritic cells. TIGIT is thought to work with PD-1 and TIM-3 to attenuate T cell responses and promote T cell dysfunction.

After the NSCLC patient-specific predictive computational simulation models were created and the profiles of 24 chemokines and immunosuppressive molecules were predicted, we created a decision tree to identify patients who would or would not respond to PD-1 immunotherapy. Decision cutoffs were established at 29.0% PD-L1 expression (Step 1), < 20.0% dendritic cell infiltration (Step 2a), > 60.0% dendritic cell infiltration (Step 2b), and immunosuppressive molecule expression as < PD-L1 with a margin of greater than 5.0% (Step 3) (Fig. 2). The decision tree was robust and had built-in redundancy. Basing the PD-L1 drug responder status on 3 separate predicted criteria allowed a responder/non-responder not identified at one step to be identified at a later step. Also the thresholds were specific. At 29.0% PD-L1 expression (Step 1), 9 non-responder patients were identified. Decreasing the PD-L1 expression cutoff from 29.0% to 25.0% identified only 6 non-responders. Increasing the PD-L1 expression cutoff from 29.0% to 35.0% identified a number of false negatives and setting the PD-L1 expression cutoff at 35.0% identified up to 13 non-responder patients: the three additional patient responders L8MTGU, P90A0O, and 26YMUF would be falsely identified as non-responders.

A diversity of signaling pathways are reported to be involved in the expression and regulation of PD-L1 [67‐69] and these pathways were observed in expression and regulation of PD-L1 in this study (Fig. 5). Responder patients had mutations around the rapidly accelerated fibrosarcoma (RAF)-rat sarcoma (RAS)-ERK pathway including KRAS/BRAF and MEK-related mutations that predicted the profiles to have stronger expression of PD-L1. However, the presence of KRAS cannot be the only criteria for predicting strong expression of PD-L1 and thus a likely PD-1 drug responder, since there were non-responder profiles that also had KRAS mutations. We observed that matched predicted and clinical non-responder patients 195P5D and J0T9TJ and mismatched predicted responder and clinical non-responder patient 6QFSVV all had KRAS mutations (Additional file 2: Table S2, Fig. 5).

Recent studies support the concept that NSCLC is not a homogeneous disease and at least 3 subtypes of KRAS mutations involving LKB1 or TP53 can be identified. The tumors with these mutations have different PD-L1 expression patterns (higher in KRAS mutations and TP53 mutations) and different sensitivities to immune checkpoint blockade. Thus the effects of KRAS mutations and KRAS co-mutations on PD-L1 expression was further assessed using 2 additional datasets [33, 34] beyond the Supplement Table 3 of the Rizvi et al study [15].

Dong, et al. [33] reported that TP53 and KRAS mutations may predict which patients would or would not respond to PD-1 immunotherapy. Modeling the dataset in their study, we predicted that KRAS+TP53 co-mutation (KP Subgroup) would lead to increased PD-L1 expression. Skoulidis, et al. [34] reported that KRAS mutations in lung adenocarcinoma were associated with co-mutations in STK11/LKB1 (the KL subgroup) [34]. KL tumors had high rates of KEAP1 mutations with lower PD-L1 expression. Modeling the dataset in their study, we predicted that KRAS+STK11+KEAP1 co-mutations (KL Subgroup) also would lead to reduced PD-L1 expression. We predicted that KRAS+CDKN2A/B co-mutation (KC Subgroup) would lead to reduced PD-L1 expression. There was a reduction in positive regulation due to reduction in AMPK, mTOR pathway and also due to KEAP1 loss of function. There was an increase in the WT TP53 mediated inhibitory regulation of PD-L1 expression. This was a novel finding based on network analysis.

Furthermore, PTEN deletion [70], PI3K mutations [71] and MYC overexpression [72] have also been recently characterized as oncogenic mechanisms leading to PD-L1 expression.

The techniques described in this retrospective study have application. Although the techniques were complicated and need more extensive validation with larger datasets, their utility in clinical practice is possible. Profiling of tumors is becoming more main stream for precision personalized medicine. The approach may not necessarily be expensive, but in fact provides more utility to the generated profiling data for most tumor samples.

Patient tumor cell genomics were found to influence cell signaling with downstream effects on the expression of 24 chemokines and immunosuppressive molecules. This allowed us to establish patient-specific profiles of these molecules that could be used to predict patient clinical responses with greater than 85.0% correlation among predicted and reported patient clinical responses. Developing a workflow incorporating immunosuppressive molecules could a) be used as a potential complementary assay to affirm IHC results or used as an alternate assay where IHC in unfeasible, b) affirm patient PD-1 and PD-L1 drug responder status, c) as a method to determine influencing factors on PD-L1 expression, and d) as a potential clinical decision support system facilitating selection of therapies based on individual patient mutational profiles. The latter application used shortly after cancer diagnosis and just before cancer treatment could generate important patient-specific treatment options that could assist clinicians in selecting appropriate mono-therapies or combination therapies.