Identification of Deleterious NOTCH Mutation as Novel Predictor to Efficacious Immunotherapy in NSCLC

Immune checkpoint inhibitors (ICI) have renovated the standard treatment for patients with non–small cell lung cancer (NSCLC), by virtue of the unprecedented prolongation of life (1–3). Despite this, durable response of ICIs merely occurs in a tiny minority, which necessitates further investigation into the biomarkers predicting the clinical benefit.

To date, two critical biomarkers, programmed death ligand 1 (PD-L1) expression and tumor mutational burden (TMB), have been validated prospectively in randomized controlled trials (RCT) concerning NSCLC (4–6). Meanwhile, retrospective analyses of genomic profiles identified potential biomarkers that are associated with better outcome such as DNA damage response (DDR) pathway comutations and TP53/KRAS comutations (7, 8). The genomic landscape that is associated with better clinical outcome of ICIs has not been fully explored.

The NOTCH pathway, a highly conserved signaling system, is regulated by short-range cell–cell interaction between NOTCH receptor (NOTCH1–4) and “canonical” ligand [Jagged1, Jagged2, delta like canonical NOTCH ligand 1 (DLL1), DLL3, or DLL4; ref. 9], or noncanonically through activation of other pathways such as NF-κB, WNT, TGFβ, and STAT3 (10).

Of great importance is the irreplaceable action of NOTCH in the development of multiple organs at early stage (11), including the longitudinal regulation of lung growth from trachea/bronchi differentiation during embryogenesis (proximal) to mature alveoli formation in postnatal period (distal; ref. 12). In developed lung, NOTCH modulates the homeostasis between secretory cells (Clara and goblet cells) and ciliated cells, extracellular matrix production and subepithelial fibrosis, myofibroblast differentiation, epithelial–mesenchymal transition (EMT), and pulmonary vascular remodeling, contributing to the initiation and progression of multiple pulmonary diseases, such as asthma and chronic obstructive pulmonary disease (11, 12). Of the NOTCH-related pathogenesis mentioned above, EMT, in which epithelial cells lose polarity together with cell–cell adhesion and acquire properties of mesenchymal cells (13), frequently associates with stemness property (as cancer stem cell; refs. 14, 15), and resistance to chemotherapy (16, 17), radiotherapy (17–19), and even targeted therapy in NSCLC (17, 20, 21).

As for immunotherapy, plenty of molecular research concerning tumor initiation, immunogenicity, and immune microenvironment strongly supports the possible association between NOTCH and immunotherapeutic efficacy. Firstly, tumor initiation via suppressing TP53 requires the regulation of NOTCH1 on MDM2 (22), a robust biomarker related to hyperprogression in immunotherapy (23). Secondly, tumor immunogenicity regulated by DDR pathways might be enhanced by NOTCH1 inhibition (22), possibly via the direct binding between NOTCH1 and ataxia-telangiectasia mutated (24, 25). Thirdly, in immune microenvironment, the NOTCH ligands on myeloid-derived suppressor cells (MDSC) interact with the NOTCH receptor on tumor cells and thereby improve the cancer stem cell (CSC) capacity (26), which in turn increases the expression of NOTCH ligands on MDSCs (27), constituting a positive feedback eventually resulting in immune tolerance (28). Based on these observations, we hypothesized that NOTCH mutation in NSCLC might predict the clinical benefit from immunotherapy.

Patients with NSCLC treated with PD-1/PD-L1 inhibitors in the Wuhan Union Hospital who had genomic profiling of circulating tumor DNA (ctDNA; 150-gene panel, Supplementary Table S1) before treatment were included in our 3DMed cohort. The patient characteristics of our cohort is shown in Supplementary Table S2. Another four independent public cohorts were also used in the present study, including Rizvi, POPLAR/OAK, Van Allen, and MSKCC cohorts.

The data for the four independent cohorts were retrieved from published articles (detailed features are displayed in Supplementary Table S3). (1) The Rizvi cohort contains 240 patients with advanced NSCLC treated with anti–PD-1/PD-L1 monotherapy or combination therapy with anti–CTLA-4, and their tumor tissues were profiled with MSK-IMPACT panel (341-gene panel, 0.98 Mb, 56 patients; 410-gene panel, 1.06 Mb, 164 patients; 468-gene panel, 1.22 Mb, 20 patients; ref. 29). (2) The POPLAR/OAK cohort of 853 patients with advanced NSCLC who had received 1 or 2 cytotoxic chemotherapies from a phase II trial, POPLAR (NCT01903993; ref. 30), and a phase III trial, OAK (NCT02008227; ref. 31), includes two different regimens, atezolizumab and docetaxel. All patients in the POPLAR/OAK trials implemented a genomic profiling of ctDNA with Foundation One panel (315-gene panel, 1.1 Mb; ref. 32). (3) The Van Allen cohort was defined as the NSCLC subpopulation of the pan-cancer research on microsatellite-stable (MSS) patients with clinically annotated outcomes to immune checkpoint therapy (33). All 56 MSS NSCLC samples and matched normal tissues were profiled by whole-exome sequencing (33). (4) The MSKCC cohort was identified as the NSCLC subpopulation of the pan-cancer research on the relation between TMB and immunotherapeutic overall survival (OS). Tumor tissues from 350 patients with NSCLC were profiled with MSK-IMPACT panel (341-gene panel, 0.98 Mb, 56 patients; 410-gene panel, 1.06 Mb, 239 patients; 468-gene panel, 1.22 Mb, 55 patients; ref. 34). Of note, overlapped patients identified between the Rizvi and MSKCC cohorts do not affect the independence of their clinical outcomes, for the reason that the Rizvi study displayed objective response rate (ORR) and progression-free survival (PFS) data and the MSKCC study published OS data (29, 34).

All human sample collection and usage were in accordance with the principles of the Declaration of Helsinki and approved by the Institution Review Board of Huazhong University of Science and Technology. The written consents were received from all the participated patients.

In the 3DMed cohort, the ORR was defined as the percentage of patients with confirmed complete response or partial response by RECIST v1.1. PFS was defined as the time from the start of anti–PD-1/PD-L1 treatment until disease progression (assessed by an investigator using RECIST v1.1) or death from any cause. In the other four independent cohorts, tumor response was assessed by thoracic radiologists (Rizvi, Van Allen) or investigators (POPLAR/OAK) using RECIST v1.1.

PD-L1 expression on tumor cells was assessed by VENTANA PD-L1 (SP263) assay in the 3DMed cohort, and the PD-L1 status was characterized as negative or positive. In the OAK trial, PD-L1 expression on tumor cells or immune cells was evaluated simultaneously by VENTANA PD-L1 (SP142) assay. Detailed definition of PD-L1 status in tumor cells (TC1/2/3) and immune cells (IC1/2/3) was described in the original article (31).

The testing of ctDNA in the 3DMed cohort followed the method that had been previously published (35).

PolyPhen-2 (Polymorphism Phenotyping v2) is a software tool from Harvard University which predicts possible impact of amino acid substitutions on the structure and function of human proteins using straightforward physical and evolutionary comparative considerations (36). PolyPhen-2 predictions were calculated for all resulting amino acid residue substitutions in human UniProtKB proteins with the maximum coding sequence (CDS) overlap and identity. PolyPhen-2 uses eight sequence-based and three structure-based predictive features which were selected automatically by an iterative greedy algorithm.

PolyPhen-2 calculates the naive Bayes posterior probability that a given mutation is damaging and reports estimates of false-positive (the chance that the mutation is classified as damaging when it is in fact nondamaging) and true-positive (the chance that the mutation is classified as damaging when it is indeed damaging) rates. Based on the final score assessing the harm of missense mutation, we set the cutoff value as 0.800 to categorized NOTCH missense mutation as higher impact group or lower impact group, and thereafter explored better predictive efficacy of deleterious NOTCH mutation (del-NOTCH^mut), compared with non-deleterious NOTCH mutation (non-del-NOTCH^mut).

For gene set enrichment analysis (GSEA; ref. 37), the javaGSEA Desktop Application (GSEA 4.0.1) was downloaded from http://software.broadinstitute.org/gsea/index.jsp. GSEA was used to associate the gene signature with del-NOTCH^mut and non-del-NOTCH^mut. The signatures tested in the present study are shown in Supplementary Table S4. The genes identified to be on the leading edge of the enrichment profile were subject to pathway analysis. Fold-change values were exported for all genes and analyzed with version 4.0.1 of GSEA, using the GSEA preranked module. The normalized enrichment score (NES) is the primary statistic for examining gene set enrichment results. The nominal P value estimates the statistical significance of the enrichment score. A gene set with nominal P < 0.05 was determined to be significantly enriched in genes.

The leading edge analysis allows for the GSEA to determine which subsets (referred to as the leading edge subset) of genes contributed the most to the enrichment signal of a given gene set's leading edge or core enrichment. In the present study, leading edge analyses were performed in GSEA 4.0.1 to discern whether a small number of DDR members contribute to multiple significance of enrichment of DDR signature in the del-NOTCH^mut group.

Cell-type identification by estimating relative subsets of RNA transcripts (CIBERSORT), an online method (https://cibersort.stanford.edu/index.php) for characterizing cell composition of complex tissues from their gene expression profiles (38), was applied to enumeration of hematopoietic subsets in mRNA mixtures from The Cancer Genome Atlas (TCGA) database. CIBERSORT outperformed other methods with respect to noise, unknown mixture content, and closely related cell types (38).

The association between del-NOTCH^mut and relevant immune-related genes was analyzed in TCGA database, where both DNA sequencing and RNA sequencing data are available. The list of DDR-related genes was determined by the leading edge analysis of GSEA in DDR pathways (Supplementary Table S5). The immune gene list was mainly based on a published article that summarized the genes related to activated T cells, immune cytolytic activity, and IFNγ release (8). Other immune genes were added according to two relevant clinical trials (39, 40). A list of 47 immune genes is provided in Supplementary Table S5. Statistical significance was determined by the DESeq2 method in RStudio 3.6.1, and the normal P value was given in the figure.

The significance with categorical variables (e.g., ORR, PD-L1 status) was evaluated by the Fisher exact test. The Kaplan–Meier method was performed to delineate the curve of PFS and OS, and the Log-rank method was used to assess their significance. Cox regression was implemented to calculate the HR on PFS and OS, in both univariable and multivariable analyses and interaction tests exploring the interaction effect between NOTCH mutation and treatment choice (atezolizumab vs. docetaxel). The variables with P value below 0.10 in the univariable analyses were included in the following multivariable analyses. The method of random-effect inverse-variance weighted was used to pool outcomes, which is calculated by HR and its 95% confidence interval (CI) to estimate the size of influence on the clinical benefits including ORR, PFS, and OS. Heterogeneity assessment between different studies was applied using the I² statistics. A result of P > 0.1 and I² < 50% indicates that no significant between-study heterogeneity was present. The significance with continuous variables (e.g., TMB/count and fraction of copy-number alteration) was assessed by one-way ANOVA or two-way ANOVA, with Tukey post test, with the requirement of homoscedasticity between different groups. If the data failed to meet the criteria for parametric test, nonparametric analyses would be implemented, that is, χ² test and rank-sum tests. All statistical analyses mentioned above were performed using IBM SPSS Statistics 22 or Stata/SE 15.1, and the graphs in the present study were drawn by GraphPad Prism 8. We set the nominal level of significance as 10% for heterogeneity test and 5% for the rest of the statistical analyses, and all 95% CIs were two-sided.

In the present study, 58 patients with NSCLC with ctDNA sequencing before anti–PD-1/PD-L1 treatment were included to investigate the association between ICI efficacy and genomic alterations. The baseline characteristics are described in Supplementary Table S2. In brief, this cohort was representative of the general population of patients with advanced NSCLC with median age of 59 (range, 36–72), major proportion of male (74.1%), and high percentage of ever-smokers (56.9%). In this cohort, 86.2% of the patients received anti–PD-1, and the rest underwent anti–PD-L1 treatment. The lines of anti–PD-1/PD-L1 immunotherapy varied from first to seventh (first-line, 20.7%; second-line, 39.7%; third-line or later, 39.7%), and the overall response rate reached 20.7% in total. The median PFS and OS were comparable with previously published trials, as 3.1 months and 16.0 months, respectively.

The genomic mutational landscape of 58 patients with NSCLC categorized according to response in the 3DMed cohort is displayed in Fig. 1A. Consistent with previous studies, higher mutational rates of TP53 and KRAS were shown in responders, relative to nonresponders (8). Besides these, NOTCH mutation was discovered to be enriched in responders as well (Fig. 1B).

Due to the limited efficacy of ICIs in patients with NSCLC with EGFR or ALK driver mutation, we excluded these individuals in the following exploration. The mutation of NOTCH1/2/3 was significantly associated with higher ORR (60.0% vs. 11.9%, P = 0.003), longer PFS (HR, 0.30; 94% CI, 0.12–0.78; log-rank P = 0.009, Fig. 1C), and OS (HR, 0.21; 95% CI, 0.05–0.91; log-rank P = 0.020, Fig. 1D) in patients with EGFR/ALK^WT NSCLC. These results suggest that NOTCH mutation might be associated with the clinical benefit of immunotherapy.

To further evaluate the predictive value of NOTCH family, another four public cohorts of ICIs with adequate information of the genomic alterations in tumor tissue or ctDNA and survival were analyzed, including the cohorts from Rizvi, POPLAR/OAK, Van Allen, and MSKCC (characteristics displayed in Supplementary Table S3).

NOTCH family consists of four members, with potentially distinct mechanisms underlying the development of NSCLC. Among the samples from the OAK/POPLAR and Rizvi/MSKCC cohorts, the mutational rates of NOTCH were 6.11% (NOTCH1), 5.19% (NOTCH2), 4.12% (NOTCH3), and 6.72% (NOTCH4), which exhibited similar distribution in lung squamous cell carcinoma and lung adenocarcinoma (Supplementary Fig. S1). Univariable and multivariable analyses of the impact from nonsynonymous mutation in each NOTCH gene on PFS and OS benefit from ICI were performed in the POPLAR/OAK, Rizvi, and MSKCC cohorts. In each cohort, the multivariable HR value of NOTCH4 mutation exceeded 1, and on the contrary, the HR values of NOTCH1, NOTCH2, and NOTCH3 mutations were inferior to 1 (Supplementary Fig. S2), which indicates the contradictory predictive value of NOTCH4 mutation to NOTCH1/2/3 mutation. Thus, we further focused on the potential of predictive function of NOTCH1/2/3 mutation in these cohorts.

Among the EGFR/ALK^WT population, the beneficial trend of NOTCH1/2/3 mutation in immunotherapeutic ORR (Fig. 2A–C), PFS (Fig. 2A–C), and OS (Fig. 2E–G) was observed in all cohorts. Pooled estimates demonstrated that compared with the NOTCH^WT group, the NOTCH1/2/3^mut group exhibited better ORR (RR, 2.2; 95% CI, 1.39–3.51; P = 0.001, Fig. 2D), longer PFS (HR, 0.61; 95% CI, 0.46–0.81; P = 0.001, Fig. 2D), and OS (HR, 0.56; 95% CI, 0.32–0.96; P = 0.035, Fig. 2H). Statistical analyses for heterogeneity were insignificant in all pooled estimates (P > 0.10), indicating the consistency of the association between NOTCH1/2/3 mutation and favorable benefit to ICIs across these cohorts.

As illustrated in Fig. 3A, missense mutations were predominant among all kinds of NOTCH mutations in the POPLAR/OAK and MSKCC cohorts. Unlike truncating mutations that strikingly affect tumor cells by the loss of gene expression, missense mutations might be deleterious or benign, on account of their effects on protein structure. To rule out the benign mutations that generally do not affect the protein function, PolyPhen-2 analysis was implemented to distinguish between the missense mutations of NOTCH with higher or lower probability of impact on protein structure (abbreviated as NOTCH^mis-high and NOTCH^mis-low, respectively). Because both truncating and mis-high mutations explicitly affect the biological function of the involved gene, we hereby combined these two kinds of alterations and defined them as del-NOTCH^mut. Relatively, mis-low mutations of NOTCH1/2/3 and all NOTCH4 mutations were identified as non-del-NOTCH^mut, and the patients without any NOTCH mutation were categorized as the control group (NOTCH^WT). The work flow is illustrated in Fig. 3B.

To investigate the possible mechanism underlying the predictive role of NOTCH mutation, we first aimed to ascertain whether co-occurrence takes place between del-NOTCH^mut with robust predictors in EGFR/ALK^WT NSCLC, including PD-L1 expression and higher TMB. We identified a cohort from the OAK trial where the data of both ctDNA and PD-L1 testing are available (n = 637). In this cohort, participants were divided into nine subgroups by the two variables, bTMB (bTMB-L: bTMB<8; bTMB-I: 8≤bTMB<16; bTMB-H: bTMB≥16) and PD-L1 expression (negative: TC0 and IC0; intermediate: TC1/2 and/or IC1/2; strong expression: TC3 or IC3). As illustrated in Fig. 3C, the digits in these nine squares representing nine subgroups are the incidence rates of del-NOTCH^mut, which was enriched in bTMB-H subgroups, irrespective of PD-L1 expression. Furthermore, the distribution of PD-L1 status (SP142 antibody) was parallel among three groups (Fig. 3D), but the level of bTMB was significantly higher in the del-NOTCH^mut and non-del-NOTCH^mut groups, compared with NOTCH^WT individuals (Fig. 3E). In addition, a similar trend of TMB was observed in the patients with EGFR/ALK^WT NSCLC of TCGA database.

Despite the similar levels of mutational burden discovered in the del-NOTCH^mut and non-del-NOTCH^mut groups, the underlying mechanisms to repair the miscoding DNA, i.e., DDR pathways, might be dissimilar between these two groups. GSEA revealed prominent enrichment of signatures related to DNA repair in the del-NOTCH^mut group, compared with the non-del-NOTCH^mut (P < 0.001, Fig. 3G) and NOTCH^WT (P < 0.001, Fig. 3G) groups. In detail, DDR could be defined as eight different pathways according to their diverse functions, including homologous recombination repair, mismatch repair, base excision repair, nucleotide excision repair, Fanconi anemia pathway, translesion DNA synthesis, nonhomologous end-joining (NHEJ), and checkpoint factors. Here, we further explored that the signatures of all pathways except NHEJ displayed enrichment in the del-NOTCH^mut group, relative to the NOTCH^WT (Fig. 3H) and non-del-NOTCH^mut (Fig. 3I) groups, whereas no difference was found between the non-del-NOTCH^mut group and the NOTCH^WT group (Fig. 3J). There are members shared among different DDR pathways, and to exclude the possibility that identical genes contribute to multiple significances in several DDR pathways, leading edge analyses were performed and merely limited DDR genes were found to affect various enrichments (Supplementary Fig. S3). In addition, the expression of all genes in the leading edge analysis (113 genes in total) was further analyzed in comparison between the del-NOTCH^mut group and the NOTCH^WT group. The heatmap and box plots of the most significant 30 DDR genes were displayed in Supplementary Fig. S4.

Taken together, these results demonstrate the potentially intensified activation of DDR pathways in the EGFR/ALK^WT NSCLC with del-NOTCH^mut, in comparison with the one with non-del-NOTCH^mut, despite the comparable level of TMB in these two clusters. TMB is the consequence of the confrontation between mutagenesis and DNA repair. The comparable TMB but higher activation of DDR system in the del-NOTCH^mut NSCLC might suggest the possibility of more drastic mutagenesis in the NSCLC harboring del-NOTCH^mut.

Using CIBERSORT, we estimated the degree of infiltrated immune cells, discovering an increase of M1 macrophage in the EGFR/ALK^WT NSCLC with del-NOTCH^mut, relative to those without NOTCH mutation (Supplementary Fig. S5). Despite no distinctions that were uncovered in other subsets of immune cells, the immune reaction to tumoral neoantigens might be higher in patients with del-NOTCH^mut, which was explored by GSEA in the following section.

Displayed in Fig. 4A, multiple significant enrichments of immune activation were revealed in the del-NOTCH^mut group, compared with the NOTCH^WT group, including antigen processing and presentation, B-cell receptor (BCR)/T-cell receptor (TCR) downstream signaling, activation of CD4/CD8 T cell and natural killer T (NKT) cell, inhibition of regulatory T cell (Treg), programmed cell death, and metabolism of steroid hormones. In contrast, the non-del-NOTCH^mut group has much less enrichment compared with the NOTCH^WT group. Detailed curves of GSEA are shown in Supplementary Figs. S6 to S9.

In addition, the IL pathways were also taken into consideration. GSEA of IL signatures unmasked the activations of IL1 and IL12 pathways in the del-NOTCH^mut group, and the activations of IL6, IL10, and IL12 families in the non-del-NOTCH^mut group, relative to the NOTCH^WT group. Of great importance, the IL10 signaling enriched in the non-del-NOTCH^mut NSCLC associated with an anti-inflammatory and immunosuppressive microenvironment, in accordance with the previous result of a Treg activation in the non-del-NOTCH^mut NSCLC.

To sum up, del-NOTCH^mut was positively associated with infiltration of M1 macrophage, antigen processing via degradation in proteasome, cross-presentation of antigen, BCR/TCR downstream signaling, activation of CD4 T cell/CD8 T cell/NKT cell, deactivation of Treg, programmed cell death of tumor cell, and metabolism of steroid hormones. These results delineate the hyperactive immune microenvironment and the robust immune reaction in the EGFR/ALK^WT NSCLC with del-NOTCH^mut, relative to the ones with non-del-NOTCH^mut or without any NOTCH mutation, which might be linked with the larger benefit from immunotherapy in the del-NOTCH^mut patients with EGFR/ALK^WT NSCLC.

As shown above, del-NOTCH^mut, compared with non-del-NOTCH^mut, exhibited greater association with potentially higher transcription of DDR genes and activated immune microenvironment, which is plausibly linked with better efficacy of immunotherapy. We next sought to validate this hypothesis in the POPLAR/OAK cohort, where the PFS/OS data of both ICI (atezolizumab) and chemotherapeutic agent (docetaxel) are available and of high credibility due to the RCT setting.

Previously in Supplementary Fig. S2, we ruled out the probability of better ICI outcome associated with NOTCH4 mutation. Here, we further distinguished NOTCH4 mutation by PolyPhen-2 to explore whether deleterious NOTCH4 mutation could be beneficial. The patients harboring deleterious NOTCH1, NOTCH2, or NOTCH3 mutations exhibited a trend of better ORR and PFS than the patients with wild-type (WT) NOTCH genes (Supplementary Fig. S10A–S10C). However, as illustrated in Supplementary Fig. S10D, the lines representing deleterious NOTCH4 mutation and wild-type NOTCH are mingled, and none of the patients with NOTCH4 mutation acquired objective response, regardless of the deleterious status of the mutation. Taken together, this analysis further supports the results in Supplementary Fig. S2 that unlike the mutations in NOTCH1/2/3, NOTCH4 mutation might not be associated with immunotherapeutic outcomes in EGFR/ALK^WT NSCLC.

To directly evaluate the utility of del-NOTCH^mut in clinical decision on immunotherapy or chemotherapy in advanced EGFR/ALK^WT NSCLC, the PFS and OS benefits from atezolizumab relative to docetaxel in the del-NOTCH^mut, non-del-NOTCH^mut, and NOTCH^WT groups were separately calculated and compared. As shown in Fig. 5A, in the intention-to-treat (ITT) population, limited ORR and PFS benefits were observed (ORR: 15.6% vs. 12.3%, P = 0.204; PFS: HR, 0.84; 95% CI, 0.72–0.98, P = 0.025), and this weak improvement became even milder and insignificant in the NOTCH^WT subpopulation (ORR: 14.5% vs. 13.3%, P = 0.726; PFS: HR, 0.91; 95% CI, 0.77–1.08, P = 0.279). In contrast, moderate benefit was observed in the non-del-NOTCH^mut group (ORR: 10.7% vs. 10.3%, P > 0.999; PFS: HR, 0.65; 95% CI, 0.37–1.13, P = 0.126), and remarkably, considerable benefit was achieved in the del-NOTCH^mut group (ORR: 29.4% vs. 9.3%, P = 0.008; HR, 0.45; 95% CI, 0.27–0.77, P = 0.004). As a result, the interaction effect on PFS between NOTCH mutation (del-NOTCH^mut vs. non-del-NOTCH^mut vs. WT) and treatment choice (atezolizumab vs. docetaxel) was significant (HR, 0.73; 95% CI, 0.57–0.93, P_interaction = 0.012), recognizing del-NOTCH^mut mutation as a predictive biomarker of PFS benefit from immunotherapy over chemotherapy.

Identical analyses were performed on OS benefit (Fig. 5B). Compared with docetaxel, atezolizumab monotherapy decreased the hazard of death by 33% in the NOTCH^WT group (HR, 0.67; 95% CI, 0.55–0.80, P < 0.001), by 49% in the non-del-NOTCH^mut group (HR, 0.51; 95% CI, 0.27–0.96, P = 0.038), and by higher proportion as 52% in the del-NOTCH^mut group (HR, 0.48; 95% CI, 0.28–0.85, P = 0.011).

Furthermore, multivariable analyses were performed to explore whether del-NOTCH^mut is independently associated with immunotherapeutic efficacy. In the POPLAR/OAK cohort where mutations were detected in ctDNA, del-NOTCH^mut independently associated with better immunotherapeutic PFS (HR, 0.57; 95% CI, 0.38–0.86, P = 0.006, Table 1) and OS (HR, 0.63; 95% CI, 0.40–0.99, P = 0.045, Table 1), but not chemotherapeutic benefit (PFS: HR, 1.04; 95% CI, 0.73–1.49, P = 0.840; OS: HR, 0.86; 95% CI, 0.58–1.29, P = 0.462, Supplementary Table S6). These results demonstrate that del-NOTCH^mut is a predictive, not prognostic, biomarker in the POPLAR/OAK cohort.

In the other two immunotherapeutic cohorts (Rizvi and MSKCC) where mutations were detected in tissue samples, TMB was significantly associated with immunotherapeutic outcomes in the univariable analyses and therefore was involved in the further multivariable analyses. The association between del-NOTCH^mut and immunotherapeutic benefits remained significant after the adjustment for TMB (Rizvi-PFS: HR, 0.56; 95% CI, 0.32–0.99, P = 0.047, Supplementary Table S7; MSKCC-OS: HR, 0.54; 95% CI, 0.29–1.00, P = 0.049, Supplementary Table S8), suggesting that the predictive utility of del-NOTCH^mut is independent of TMB in EGFR/ALK^WT NSCLC. Taking TCGA cohort as a comparison, which includes the patients with NSCLC with predominantly early stage and surgically resected tumors without receiving immunotherapy, we discovered no significant association between del-NOTCH^mut and both the PFS and OS in EGFR/ALK^WT NSCLC (Supplementary Fig. S11).

Taken together, del-NOTCH^mut detected by either ctDNA or tumor tissue was independently associated with significant improvement of immunotherapy and was not linked with prognosis, delineating the predictive, but not prognostic, value of del-NOTCH^mut in immunotherapeutic treatment against NSCLC.

In this study, NOTCH mutation, especially del-NOTCH^mut, was identified as tumor cell–intrinsic determinant of better response to ICI treatment in five cohorts, involving 1,557 patients with advanced NSCLC in total. Of great attention are their noticeable correlations with better clinical outcome of ICI, instead of chemotherapeutic agents, elucidating the predictive, but not prognostic, role in immunotherapy. Of all kinds of NOTCH mutations, deleterious, relative to non-deleterious alteration, might possess more predictive power. Furthermore, we distinguished greater transcription of genes related to DDR and immune activation as potential mechanisms underlying the predictive value of del-NOTCH^mut in NSCLC population (diagram shown in Supplementary Fig. S12).

This study represents one of the first reports to examine the association between ICI response and NOTCH family. As reported by a previous study mainly involving melanoma and NSCLC, NOTCH1 mutation seldom occurred in patients who hyperprogressed while on immunotherapy (23). Based on the similarity among different NOTCH members, we took all four members of NOTCH family into consideration and discovered that the mutation in NOTCH1/2/3, instead of NOTCH4, was associated with higher ORR and longer PFS/OS with ICI treatment in EGFR/ALK^WT NSCLC.

Of the NOTCH mutations in these cohorts, frameshift and nonsense mutations are overtly loss of function, whereas missense mutations might be damageable or merely benign as nonpathogenic passenger events. Earlier NSCLC studies in NOTCH1 implemented PolyPhen-2 to distinguish detrimental mutations, and confirmed their downregulating function on transcriptional activity of NOTCH1 by both reporter gene assay and electrophoretic mobility shift assay (41), suggesting the missense mutations of NOTCH might mainly be inhibitory in NSCLC, similar to truncating mutations. In addition, in light of the data in the Catalogue of Somatic Mutations in Cancer (COSMIC) database, the pattern of mutational loci in NSCLC is sporadic, unlike the enriched distribution in hematopoietic and lymphoid where NOTCH mutations are recognized to be oncogenic and gain of function, which further suggests that the NOTCH mutation in NSCLC may be dominated by inhibitory function. Inhibition of NOTCH1 was discovered to stabilize TP53 via downregulating the phosphorylation and protein level of MDM2 (42), a prominent biomarker linked with ICI-induced hyperprogression across multiple tumor types (23). These results indicate the possibility that deactivation of MDM2 might account for the predictive function of inhibitory NOTCH mutation in immunotherapy.

Further analyses in the POPLAR/OAK cohort comparing atezolizumab and docetaxel revealed that relative to the limited benefit in the NOTCH^WT group, moderate ICI benefit was shown in the non-del-NOTCH^mut group, and remarkable immunotherapeutic outcome was observed in the del-NOTCH^mut individuals (detailed comparisons shown in Supplementary Fig. S12). Many missense mutations may have little or no effect on protein function; this is likely the basis for the weaker correlation between clinical outcomes and identified non-del-NOTCH^mut.

Consistent with clinical benefit, similar PD-L1 staining but higher TMB was uncovered in both del-NOTCH^mut and non-del-NOTCH^mut groups. The PD-L1 antibody used in the OAK trial is SP142, which was indicated to be poorly consistent to other antibodies including FDA-approved standardized assays, 22C3 and SP263, especially in the high-staining samples (43, 44), which might reduce the credibility of the PD-L1 analysis in the present study. As for TMB, it is the consequence of the confrontation between mutagenesis and DNA repair. In the present study, similar level of TMB was revealed in NSCLC with del-NOTCH^mut or non-del-NOTCH^mut, whereas the DDR is much stronger in the NSCLC with del-NOTCH^mut, relative to non-del-NOTCH^mut, suggesting a possibility of fiercer mutagenesis in NSCLC with del-NOTCH^mut. Multivariable analyses in the Rizvi and the MSKCC cohorts, where tissue TMB was significantly associated with immunotherapeutic outcome, indicate the predictive utility of del-NOTCH^mut is independent to the level of TMB.

After the translation of mutated protein, immune procedures including processing and presentation of neoantigen, BCR and TCR downstream signaling, activation of CD4 T cell, CD8 T cell, and NKT cell, and inhibition of Treg contributed to the superior degree of programmed cell death and the success of immunity-induced tumor rejection. Here in the present study, we observed the association between del-NOTCH^mut and these procedures, which might be part of the mechanism of del-NOTCH^mut in predicting better immunotherapeutic outcome.

As for limitations, the retrospective setting and pooled-estimate methodology of this study might introduce multiple biases. The limitation from retrospective setting could be greatly minimized by the large sample size (5 cohorts involving 1,557 patients), by which the experimental features might be balanced, such as race, ICI regimen, treatment line, the platform/panel/used sample of NGS testing, etc. However, the pooled-estimate methodology probably weakens the credibility of the conclusion to some extent, where the included studies with larger sample size tend to possess more weight in the final pooled estimate. In the present study, the predictive utility of NOTCH mutation was mainly driven by small datasets, which introduce minor-to-moderate heterogeneity to the pooled estimates.

In our 3DMed cohort, merely NOTCH1–3, instead of NOTCH4, were involved in the gene panel, and the comparison among the 4 NOTCH genes was comprehensively analyzed in the public cohorts. Fortunately, the further analyses indicate that NOTCH4 mutation was not associated with better immunotherapeutic outcome, making it logical to put all the cohorts together for a meta-cohort analysis. In addition, our attempt to recruit functional NOTCH mutations was handicapped by the limited information available regarding the functions of different mutations and the lack of hotspots in NOTCH gene mutations, as illustrated in TCGA and COSMIC databases. To alleviate this limitation, we used PolyPhen-2 system which provides powerful evaluation of the mutational functions on NOTCH protein, helping us distinguish the del-NOTCH^mut out of all NOTCH missense mutations. However, this system fails to decipher the precise function of missense mutation as activation or deactivation. Despite the similar trend of increased ICI benefit in patients with missense or nonsense mutations, we cannot simply assume that missense mutations at different loci are all inactivating, due to the lack of direct functional data in the present study. Activating and inactivating mutations might lead to distinct influences on immunotherapeutic efficacy, which might be addressed by further molecular studies in cell line and xenograft model.

Despite the identification of del-NOTCH^mut, the predictive efficacy in cohort (POPLAR/OAK) using the similar detection methods with our cohort is not ideal, which might decrease the robustness of this biomarker. This poorer result might be explained from two angles. Firstly, the confounding variables significantly associated with the PFS and OS in the atezolizumab arm (including sex, race, histology, Eastern Cooperative Oncology Group, number of metastatic sites, and mutations of STK11, KEAP1, and TP53) might influence the predictive efficacy of NOTCH mutation in univariable analysis. After adjusting these factors, the association between del-NOTCH^mut and OS became significant. Secondly, unlike routine administration of ICI in our cohort, the POPLAR/OAK trial allowed the continuation of atezolizumab despite radiological progression, if the investigator deemed the patient to be receiving clinical benefit, which might cover part of the predictive utility of NOTCH mutation in the OS benefit. Taken together, the poorer result in the POPLAR/OAK cohort might be partially attributed to its special procedure and the confounding factors. The significant multivariable P values among multiple cohorts demonstrate the robustness of del-NOTCH^mut in predicting favorable ICI benefit.

In summary, our data identified NOTCH mutation, especially del-NOTCH^mut, as a novel, frequent determinant of sensitivity to immune checkpoint blockade in EGFR/ALK^WT NSCLC. More importantly, our results unravel a possibility of personalized combination immunotherapy as adding NOTCH inhibitor to ICI regimen in NSCLC, for the optimization of ICI treatment in clinical practice.

No potential conflicts of interest were disclosed.

Conception and design: Z. Song, Y. Xu, G. Wang, Z. Lou, Y. Zhang, L. Liu

Development of methodology: Y. Xu, G. Wang

Acquisition of data (provided animals, acquired and managed patients, provided facilities, etc.): K. Zhang, X. Hong, Y. Xu, Y. Zhang, X. Zhao, Z. Zhao, J. Zhao, M. Huang, D. Huang, C. Qi, C. Gao, S. Cai, F. Gu, Y. Hu, C. Xu

Analysis and interpretation of data (e.g., statistical analysis, biostatistics, computational analysis): K. Zhang, Y. Xu, C. Li, G. Wang, W. Wang

Writing, review, and/or revision of the manuscript: K. Zhang, Y. Xu, C. Li, G. Wang, Y. Zhang, X. Zhao, Z. Zhao, J. Zhao, M. Huang, D. Huang, C. Qi, C. Gao, S. Cai, Y. Hu, C. Xu, L. Liu

Administrative, technical, or material support (i.e., reporting or organizing data, constructing databases): Z. Song, S. Cai, L. Liu

Study supervision: W. Wang, L. Liu

This work was supported by the National Key Research and Development Program of China (2016YFC1303800) and National Natural Science Foundation of China (81773056).

The costs of publication of this article were defrayed in part by the payment of page charges. This article must therefore be hereby marked advertisement in accordance with 18 U.S.C. Section 1734 solely to indicate this fact.