Tumour Mutational Burden and Immune Checkpoint Inhibitor Response in Non-small Cell Lung Cancer: A Continuous Modelling Approach

We utilised data from both real-world use and randomised controlled trial (RCT) evaluation of ICIs, and evaluated both tissue-based TMB (tTMB) and blood-based TMB (bTMB; utilising circulating tumor DNA). The real-world cohorts included the Memorial Sloan Kettering-Integrated Mutation Profiling of Actionable Cancer Targets (MSK-IMPACT) cohort [23] and the Strata Clinical Molecular Database (SCMD) cohort [22]. The MSK-IMPACT cohort included patients diagnosed with advanced NSCLC (2015–2018) treated with an ICI at the Memorial Sloan Kettering Cancer Center [23] who were evaluated for primarily pre-treatment tTMB (5% of patients had post-treatment sequencing of tTMB) via genomic profiling using the MSK-IMPACT next-generation sequencing (NGS) platform [6]. If patients in the MSK-IMPACT cohort were treated with multiple ICIs, only data from their first ICI were utilised [23]. The SCMD cohort included patients in the Strata Clinical Molecular Database who were treated for advanced NSCLC with an ICI and evaluated for pre-ICI treatment tTMB using the StrataNGS test (Strata Oncology, Ann Arbor, MI, USA) [22]. Randomised controlled trial cohorts included the MYSTIC (NCT02453282) [17, 21], OAK (NCT02008227) and POPLAR (NCT01903993) trials [7]. The MYSTIC trial evaluated durvalumab ± tremelimuab against platinum-based doublet chemotherapy as first-line therapy in advanced NSCLC and bTMB was evaluated using the GuardantOMNI NGS assay [17, 21]. OAK and POPLAR evaluated atezolizumab against docetaxel in previously treated patients with advanced NSCLC and bTMB was evaluated using a custom assay previously described [7, 13]. The Guardant OMNI NGS bTMB assay and the custom OAK/POLAR bTMB assays have all been positively correlated with the Food and Drug Administration-approved FoundationOne CDx NGS tTMB assay [7, 11, 21], with a Guardant OMNI bTMB of 14.5 mutations/megabase equivalent to a Foundation One CDx tTMB of 10 mutations/megabase in the MYSTIC cohort [21].

Overall survival (OS) was defined as the time between the commencement of treatment (or the date of randomisation for RCT cohorts) and death due to any cause. Patients alive at the time of the last follow-up were censored at the time last known to be alive. Progression-free survival (PFS) was defined as the time between the date of commencement of treatment (or date of randomisation for RCT cohorts) and the date of the first documented disease progression, as assessed by the investigator using RECIST v1.1, or date of death due to any cause, whichever occurs first. Objective response rate was defined according to RECIST v1.1 criteria and included both complete and partial best overall response. Data on OS, PFS and objective response outcomes were available for the MSK-IMPACT cohort and the OAK and POPLAR RCTs. For the SCMD cohort, only data on OS and PFS (defined in terms of time to next therapy) outcomes were available [22], and for the MYSTIC RCT only the OS outcome was available [17] (Fig. 1 of the Electronic Supplementary Material [ESM]). Notably, programmed death-ligand 1 status was only available for OAK and POPLAR trials.

Restricted cubic splines (R rms package) were utilised to model the potentially non-linear relationship between TMB and objective response/survival [8, 16]. Restricted cubic splines enable flexible modelling of non-linear relationships via piecewise cubic polynomials. Splines were selected based on their ability to recover complex non-linear relationships [3], and 4 knots is a recommended compromise between flexibility and the risk of overfitting [8]. The relationship between the objective response rate (ORR) and TMB was evaluated using logistic regression. The relationship between TMB and survival outcomes (OS and PFS) was evaluated using Cox proportional hazards regression stratified by treatment. The prognostic association between TMB and survival outcomes was reported in terms of estimated survival at 24 months for OS and 18 months for PFS. The survival point estimate and the associated 95% confidence interval were plotted over the range of TMB values to highlight how estimated survival varied with TMB. The likelihood ratio test was used to evaluate the statistical significance of the overall relationship between TMB and response/survival for each treatment. Because of the differences in TMB assays utilised for the four cohorts, each cohort was analysed and reported separately. Sensitivity analyses were conducted utilising a generalised additive model with penalised thin plate regression splines (R mgcv package)—a more complex smoothing approach in which the smoothness was optimised using generalised cross-validation [24]. Because of a small proportion of extreme TMB values, the TMB value was winsorised at the 97.5% percentile. Tumour mutational burden values were modelled in units of mutations per megabase.

As an extension of the primary analyses focusing on prognosis, the RCT treatment effect of ICI (vs chemotherapy) across the range of TMB values was reported as a treatment hazard ratio. A Cox proportional hazards model was utilised with restricted cubic splines (4 knots) for TMB and a treatment-TMB interaction term was included. A sensitivity analysis utilised the subpopulation treatment effect pattern plot (R stepp package)—a non-parametric method involving tail-oriented subsets of TMB values and the Kaplan–Meier survival estimator [4].

All clinical trials used in the analyses were performed in accordance with the Declaration of Helsinki, with patient consent affirmed in each of the aforementioned trials. The secondary analysis of anonymised patient data was deemed as minimal risk research by the Southern Adelaide Local Health Network, Officer for Research and Ethics, and was exempted from review. All analyses were conducted within the R software environment (version 4.1).

The MSK-IMPACT cohort included 666 patients with advanced NSCLC evaluable for tTMB. Median [interquartile range] tTMB was 7 [4–12], and median follow-up was 21 months (Figs. 2 and 3 of the ESM). The SCMD cohort included 302 patients with advanced NSCLC evaluable for tTMB. Median [interquartile range] tTMB was 7 [3–13], and median follow-up was 17 months (Fig. 4 of the ESM). The MYSTIC RCT had 809 patients evaluable for bTMB. Median bTMB was 13 [8–20], and median follow-up was 30 months (Fig. 5 of the ESM). The OAK and POPLAR RCTs collectively had 779 EGFR/EML4-ALK wild-type patients evaluable for bTMB. Median bTMB was 7 [4–15], and median follow-up was 21 months (Fig. 6 of the ESM).

For patients treated with ICIs, progressively higher TMB was associated with a progressively higher ORR (Fig. 1) and 18-month PFS (Fig. 2) in both the real-world and RCT cohorts. Specifically, for the MSK-IMPACT cohort, there was a significant association between tTMB and ORR (p < 0.001), and tTMB and PFS (p < 0.001). For the SCMD cohort, there was a significant association between tTMB and PFS (p = 0.023). For the OAK and POPLAR RCT cohort treated with atezolizumab, there was a significant association between bTMB and ORR (p = 0.017), and bTMB and PFS (p < 0.001).

In contrast, for patients treated with docetaxel in the OAK and POPLAR RCTs, there was no evidence of improved ORR at progressively higher bTMB values (p = 0.231, Fig. 1B). For patients treated with docetaxel, there was a statistical association between bTMB and PFS (p < 0.001); however, in contrast to patients treated with ICIs, this significant association was driven by favourable PFS at lower values of TMB rather than improved PFS at higher TMB (Fig. 2C).

The relationship between TMB and OS was more complex and there was heterogeneity across the patient cohorts evaluated. At lower values of TMB, the relationship between TMB and OS differed for the four ICI-treated cohorts (Fig. 3). However, in the upper range of TMB values, there was a consistent and progressive improvement in OS with higher TMB levels in all four cohorts treated with ICIs. For MSK-IMPACT, there was a significant prognostic OS association (p < 0.001) with ICI treatment and TMB, with SCMD cohorts (p < 0.001) sharing this association. In contrast, for RCT arms treated with chemotherapy alone, there was no indication of improvement in OS with higher TMB.

Sensitivity analyses using generalised additive modelling with penalised thin plate regression splines yielded similar results to the above-described Cox proportional hazards models with restricted cubic splines (4 knots). Specifically, the general nature of the continuous relationship between TMB and clinical endpoints across was consistent with the primary analyses for all cohorts (Figs. 7–9 of the ESM).

An interaction analysis was performed for the RCT cohorts using restricted cubic splines and the subpopulation treatment effect pattern analysis. This indicated that comparative to the respective chemotherapy treatment arms, the ICI treatment effect was progressively greater with progressively higher TMB values (Figs. 10 and 11 of the ESM).