In this observational study, prospective clinical sequencing was carried out in tissue samples from 55 patients with metastatic NSCLC treated at two oncological centers, of whom 35 had received an ICI in the routine clinical setting from January 2018 to November 2019. Patients were considered eligible if complete data on clinical features and treatment outcomes were available. Targeted DNA sequencing was performed with the FoundationOne CDx assay. The FoundationOne CDx assay detects mutations and copy number variations (CNVs) in 324 cancer-related genes, and selected rearrangements. The assay also provides information on MSI and TMB. Regarding TMB, it is reported with the following cut-offs: low (1–5 mutations/mb), intermediate (6–19 mutations/mb) and high (≥20 mutations/mb). Detailed information on the assay and the variant calling pipeline are available at https://www.rochefoundationmedicine.com/f1cdxtech. The immunohistochemical assessment of PD-L1 was performed in formalin-fixed paraffin-embedded tissues with the monoclonal mouse anti-PD-L1 antibody clone 22C3 (Dako). PD-L1 expression was evaluated in tumor cells and considered positive when expressed in ≥1% of neoplastic cells.

For external validation, we used three different and publicly available datasets. The Memorial Sloan Kettering Cancer Center (MSKCC) clinical sequencing cohort (tissue-based next-generation sequencing (tNGS)) was downloaded from cBioPortal (last accessed on the February 12, 2020).6 11 Data from blood-based NGS (bNGS) carried out in patients enrolled in the randomized phase II POPLAR trial and the randomized phase III OAK trial, both comparing atezolizumab versus docetaxel as second-line treatment in patients with NSCLC, were obtained as an excel file format, being available as supplementary information.12–14 The three datasets were merged into a metadaset containing 779 ICI-treated patients evaluable for overall survival (OS) and 669 patients evaluable for progression-free survival (PFS). We also analyzed the MSKCC pan-tumor cohort on exclusion of patients with NSCLC (n=1.311) to evaluate whether results from NSCLC can be extended to other tumor types. We did not use any variant interpreter to generate the model tested in survival analyses. However, the functional consequence of the detected mutations was evaluated in all the three cohorts using Fathmm (fathmm.biocompute.org.uk/), PROVEAN (provean.jcvi.org), SIFT (sift.bii.a-star.edu.sg) and PolyPhen (genetics.bwh.harvard.edu) (online supplementary files 1 and 2). In order to ensure the reproducibility of the results (and the generation of the model) in the various datasets, we did not consider CNVs.

Survival curves were estimated with the Kaplan-Meier product-limit method and compared by log-rank test. The nature of clinical variables included in multivariate Cox regression models for PFS and OS was determined on the basis of the possible association with the outcome of interest, and taking into account some differences in the nature of variables reported in the various databases. On this ground, we considered the following variables: smoking status, gender, TMB and histology. The related estimates were reported as HR and 95% CI. The relationship between categorical variables was assessed with the Pearson’s χ² test of independence. Level of significance was defined as p<0.05. Statistical analyses were carried out using SPSS V.21.0 (SPSS, Chicago, Illinois, USA).

Among the 55 patients whose tissue samples underwent targeted sequencing in our observational study (Rome cohort), 20 were excluded as they did not receive immunotherapy (figure 1, panel A). Basal characteristics of the 35 ICIs-treated patients, along with the treatment received, disease evolution pattern and mutational landscape are illustrated in figure 1 (panel B) and summarized in table 1. Focusing on the mutations enriched in tumors from patients who obtained a DCB (defined as stable disease or tumor response lasting ≥6 months), we noticed four NOTCH gene mutations (patients 1–4). However, we also observed that a NOTCH3 mutation was present in a patient with rapidly progressing disease (patient 5). On this ground, we turned our attention to the pattern of co-occurrence/mutual exclusivity with other altered molecular networks. With this approach, we observed that while NOTCH-mutant tumors associated with DCB also harbored mutations in one or more homologous repair (HR) genes (ATR, ATM, BRCA1, BRCA2 and MUTYH), the only NOTCH-mutant tumor characterized by unresponsiveness did not carry any alterations in the HR machinery. Next, we noticed a trend towards higher TMB in HR-mutant tumors as compared with wild-type cases (online supplementary figure 1). Hypothesizing that co-occurring NOTCH and HR mutations may delineate an uncommon genomic background associated with sensitivity to ICIs, we gathered data from three different studies exploiting either tNGS and bNGS. These datasets contained largely overlapping HR genes (ATR, ATM, BRCA1, BRCA2, CHEK1, CHEK2, FANCA, FANCC, FANCG, MUTYH, PALB2, POLE, POLD, RAD50, RAD51 and its paralogs), which enabled us to recapitulate our model, where NOTCH^mut/HR^mut tumors were defined by a mutation in NOTCH1, NOTCH2 or NOTCH3 co-occurring with at least one mutation in the aforementioned HR genes.

In the metadaset, patients whose tumors carried the NOTCH^mut/HR^mut signature had significantly longer PFS as compared with their negative counterparts (log-rank p=0.002, figure 2, panel A and B). The performances of the model were strictly dependent on the presence of alterations in both pathways, as documented by the drop of survival curves in patients with alterations in either NOTCH or the HR pathway (panel A). Importantly, the NOTCH^mut/HR^mut signature was associated with a higher metastatic burden (figure 2, panel C). On this basis, a ‘Janus-faced’ significance of the model may be envisioned, namely, an increased immune responsiveness coupled with more aggressive biological traits. In the multivariate Cox regression model, the genomic signature was associated with a decreased risk of disease progression (HR 0.51, 95% CI 0.34 to 0.76, p=0.001) (figure 2, panel D, blue box). The model retained its significance when separately evaluated (univariate analyses) in the MSKCC cohort (tNGS) and the POPLAR/OAK cohort (bNGS) (figure 2, panel D, red box). Patients with NOTCH^mut/HR^mut tumors also had longer survival (figure 3, panel A and B), as further confirmed in the multivariate Cox regression model (HR 0.53, 95% CI 0.34 to 0.85, p=0.008) (figure 3, panel C). Survival outcomes observed in the OAK/POPLAR (atezolizumab) and MSKCC cohorts (anti-PD-(L)1 monotherapy or in combination with anticytotoxic T-lymphocyte antigen-4 antibody) were comparable (online supplementary figure 2).

Interestingly, in chemotherapy-treated patients (docetaxel arm of the OAK/POPLAR trials, crossover not allowed, n=424) co-occurring NOTCH and HR mutations had an opposite clinical significance (online supplementary figure 3, panel A and B). Likewise, in the Rome cohort the two patients with tumors harboring co-existing mutations who experienced a DCB with immunotherapy in the second-line setting had a fairly short PFS with first-line chemotherapy (online supplementary figure 3, panel C). This observation was consistent with the clinical course of the disease recorded in the three patients with co-existing alterations who exclusively received chemotherapy (online supplementary figure 3, panel C). Thus, a treatment-dependent prognostic significance of the NOTCH^mut/HR^mut model may be envisioned.

Finally, we verified whether the NOTCH^mut/HR^mut model retained its predictive capability in ICI-treated patients with other tumor types. Interestingly, the model was associated with longer OS also in the MSKCC pan-cancer dataset (ICI-treated patients) on exclusion of patients with NSCLC (n=1311, log-rank p=0.002, figure 3, panel D). In the pan-cancer cohort, cutaneous melanoma, bladder urothelial carcinoma and renal clear cell carcinoma are the most represented tumors (beyond lung adenocarcinoma). The frequency and type of NOTCH and HR mutations at the pan-cancer level is provided as online supplementary figure 4. For completeness, the mutational frequency of individual genes in the OAK/POPLAR and MSKCC studies is reported as online supplementary table 1.