Therapeutic guidance of tumor mutation burden on immune checkpoint inhibitors in advanced non-small cell lung cancer: a systematic review and comprehensive meta-analysis

Abstract

Background

Tumor mutation burden (TMB) remains a promising but ambiguous predictive biomarker for the efficacy of immune checkpoint inhibitors (ICIs). We investigated the predictive value of TMB in patients with advanced non-small cell lung cancer (NSCLC) treated by ICI-containing therapies under strictly matched clinical settings.

Methods

PubMed, Embase, Cochrane Central, ClinicalTrials.gov, and bioRxiv databases were searched till October 16, 2021. All randomized controlled trials (RCTs) that compared patients with high TMB (TMB-H) and low TMB (TMB-L) and provided hazard ratio (HR) and corresponding 95% confidence interval (CI) in advanced NSCLC patients receiving ICIs were included, and mirror-based meta-analysis was performed (Part1). Bayesian network meta-analysis was conducted to investigate the efficacy of distinct first-line regimens in TMB-H and TMB-L groups (Part2). Public cohorts were used for validation and further exploration (Part3).

Results

Twelve RCTs (n=5527) and 5 public cohorts (n=573) were included. In Part1, TMB-H patients generally exhibited a more significant progression-free survival (PFS) benefit from ICI-containing therapies compared to TMB-L patients (HR=0.58, 95% CI: 0.49-0.67, P < 0.0001). In Part2, anti-PD-1 plus chemotherapy ranked best for PFS in both TMB-H and TMB-L groups. Anti-PD-L1 plus anti-CTLA-4 therapies indicated better PFS and overall survival (OS) benefit than single ICI and chemotherapy in the TMB-H group, but ranked worst in the TMB-L group. Finally, TMB was validated to be an independent predictive biomarker from programmed cell death-ligand 1 (PD-L1) expression in Part3, which could further distinguish beneficiaries of ICI-containing therapies with PD-L1 < 50%.

Conclusion

TMB-H could be a predictive biomarker independent of PD-L1 expression to identify beneficiaries of ICI-containing therapy in advanced NSCLC patients.

1. Introduction

Immunotherapy provided by immune checkpoint inhibitors (ICIs) has revolutionized the treatment landscape of advanced non-small cell lung cancer (NSCLC); however, a considerable fraction of patients cannot derive benefit from ICIs.¹ Hence, the identification of the responders to ICIs remains an urgent need.² Tumor mutation burden (TMB), a promising biomarker for ICIs,^3,4 was investigated and approved by US Food and Drug Administration (FDA) as a tumor-agnostic companion diagnostic. Nevertheless, the utility of TMB varied across different tumor types⁵ and the predictive value of TMB for ICIs in advanced NSCLC patients has been deemed controversial in several retrospective studies^6,7 and clinical trials. In CheckMate-026 study, patients with high TMB showed a favorable progression-free survival (PFS) with nivolumab than those who received platinum-based doublet chemotherapy, while no significant difference existed among patients with low or medium TMB.⁸ Also, in Checkmate 227 study, by comparing first-line nivolumab plus ipilimumab with platinum doublet chemotherapy, a prolonged PFS was noted in NSCLC patients with high TMB regardless of programmed cell death-ligand 1 (PD-L1) expression.⁹ However, in KEYNOTE-189 and KEYNOTE-407 trials, first-line pembrolizumab plus chemotherapy showed better PFS than chemotherapy irrespective of TMB levels.

To further clarify the predictive role of TMB in NSCLC, some systematic reviews were previously conducted on its association with the efficacy of immunotherapy. Zhu et al.¹⁰ included eight studies that compared the efficacy of anti-PD-(L)1 versus chemotherapy according to TMB levels, and the efficacy of anti-PD-(L)1 in patients with high TMB (TMB-H) versus low TMB (TMB-L). The study demonstrated a significant PFS instead of overall survival (OS) benefit from ICIs as compared to chemotherapy in the TMB-H subgroup, and a significant better PFS benefit from anti-PD-(L)1 in the TMB-H subgroup than in the TMB-L subgroup. However, another study¹¹ showed superior OS from ICIs than chemotherapy in the TMB-H subgroup. Wu et al.¹² included 29 studies covering diverse tumor types and revealed that advanced NSCLC patients with high TMB showed longer PFS instead of OS after receiving immunotherapy than those with low TMB. Osipov et al.¹³ reported significant association between TMB and objective response rate (ORR) for ICIs by meta-regression analysis of 117 clinical trials. The above-mentioned conclusions should be drawn with caution due to the potential bias caused by the confounding factors related to tumor types, control arms, treatment lines, and the compared regimens. As such, comparisons with strictly matched clinical settings could provide a possible approach to better reflect the difference between two comparators.¹⁴

In this study, we aimed to measure the predictive value of TMB levels in various ICI-contained therapies with strictly matched clinical settings. In addition, we clarified the potential of TMB in guiding therapeutic selection based on survival when making first-line decisions. Finally, the study results above have been validated and the clinical availability of the combined use of TMB and PD-L1 expression will be explored further.

2. Materials and methods

This study included three parts (Fig. 2; Supplementary Fig. 1). Part 1 focused on the comparisons on the impacts of TMB levels (TMB-H vs TMB-L) on the efficacy of ICI-containing therapy versus chemotherapy. The same method was used to estimate the efficacy of anti-PD-L1 versus anti-PD-1. Part 2 dealt with the network meta-analysis of first-line therapies in both TMB-H and TMB-L groups. Part 3 was about the multi-cohort validation of the effect of TMB on ICI-contained therapy and exploration for the correlation between TMB and PD-L1 expression. This study followed the PRISMA reporting guideline¹⁵ and the PRISMA statement for network meta-analysis.¹⁶ A prospective protocol was registered on the PROSPERO online platform (CRD42020194499).

Fig. 2.

Flow diagram of the study. (A) Part 1a, Mirror comparisons of the impact of TMB-High vs TMB-Low on the ICI-containing therapy compared to chemotherapy. (B) Part 1b, Mirror comparisons of progression-free survival of anti-PD-L1 vs anti-PD-1 inhibitors. (C) Part 2, Network diagram of comparisons on progression-free survival of all first-line treatments. The circle size is proportional to the number of patients with TMB-Low receiving the regimen (in red brackets), and the number of patients with TMB-High receiving the regimen is given in yellow brackets. The number of patients receiving ICI and chemotherapy in Impower110 TMB-Low group is unavailable. (D) Part 3, Multi-cohort validation of the effect of TMB on ICI-contained therapy. Abbreviations: 1L, first-line treatment; ≥2L, second or more line treatment; bTBM, blood-based tumor mutation burden; CGP, cancer gene panel; Chemo, chemotherapy; CTLA-4, cytotoxic T-lymphocyte-associated protein 4; Doce, docetaxel; HR, hazard ratio; ICIs, immune checkpoint inhibitors; P-Chemo, platinum-based doublet chemotherapy; PD-1, programmed cell death 1; PD-L1, programmed cell death-ligand 1; PFS, progression-free survival; tTMB, tissue-based tumor mutation burden; WES, whole exosome sequencing.

2.1. Literature search

Articles were systematically searched from PubMed, Embase, Cochrane Central databases, ClinicalTrials.gov, and bioRxiv before October 16, 2021 by using the following key words: “anti-PD-1 or anti-PD-L1 or anti-CTLA-4 or nivolumab or pembrolizumab or atezolizumab or durvalumab or avelumab or cemiplimab or tremelimumab or ipilimumab or toripalimab or tislelizumab or camrelizumab or sintilimab” and “tumor mutation burden” and “non-small cell lung cancer” and “randomized controlled trial”. The detailed search strategy is provided in Supplementary Table 1. Abstracts from all major conference proceedings, including American Society of Clinical Oncology (ASCO), American Association for Cancer Research (AACR), European Society for Medical Oncology (ESMO) et al. and references from related articles were manually searched.

2.2. Study selection

Inclusion criteria: 1. Study type: phase 2/3 randomized controlled trials (RCTs) or exploratory analysis of RCTs. 2. Population: histologically confirmed stage IIIB-IV advanced/metastatic NSCLC (squamous or non-squamous or both) patients. 3. Intervention: ICI-containing therapies or chemotherapy. 4. End-point: PFS or OS. 5. Data provided: hazard ratio (HR) and corresponding 95% confidence interval (95% CI) for PFS/OS of ICI-containing therapies versus chemotherapy in the TMB-H and TMB-L subgroups.

Exclusion criteria: 1. Study type: non-RCTs, observational studies, reviews, meta-analysis, case reports. 2. Population: stage I-IIIA patients receiving neo-adjuvant or adjuvant therapy. 3. End-point: ORR or patient-reported outcomes. 4. Survival data according to TMB status were unavailable. 5. For duplicated reports of the same study, only the most recent and complete publication was included.

2.3. Data extraction, risk of bias assessment, and study outcomes

The study name, first author, publication year, National Clinical Trail (NCT) number, phase of the trial, number of analyzed patients, follow-up time, histology, line of therapy, treatment and control, PD-L1 status, test method, definition and cut-off point of TMB were independently extracted by two authors (Zhao J and Wang GQ). HRs with 95% CIs of ICI-containing treatments compared with controls in the TMB-H and TMB-L subgroups were extracted. The Kaplan-Meier curves for PFS of patients with TMB ≥ 10 mutations (mut) /Mb and PD-L1 expression ≥ 1% in the CheckMate-227 trial were digitized using an Engauge Digitizer (www.digitizer.sourceforge.net) to estimate the unavailable HR by the method of Tierney et al.¹⁷ Risk of bias of each included study was assessed using the Cochrane Risk of Bias Tool 2.0 for RCTs,¹⁸ involving five domains of sources of bias: the randomization process, deviations from the intended interventions, missing outcome data, measurement of the outcome, selection of the reported result. Each aspect was assessed based on the detailed items and scored as low, moderate, or high risk of bias. The primary outcome was PFS, and the secondary outcome was OS.

2.4. Data synthesis and statistical analysis

Part 1: Relative HR with 95% CI of the TMB-H to TMB-L subgroup was calculated for each mirror group in Part 1a. HR with 95% credible interval (95% CrI) of indirect comparison of anti-PD-L1 versus anti-PD-1 was calculated with a random-effect model by Markov chain Monte Carlo (MCMC) methods, and the same detail is presented in Part 2 using the Bayesian framework approach in Part 1b. HRs with 95% CIs of multiple studies for specific intervention within the mirror group were pooled first with a random effects model by generic inverse variance method, using R package meta (version 4.14-0) and HRs with 95% CrIs derived from each mirror group were finally pooled with the same approach. We also conducted preplanned subgroup analyses by line of therapy, regimen, test method of TMB, and PD-L1 expression in Part 1a. The funnel plot, Begg's test, and Egger test were used to assess publication bias, and P value < 0.10 indicated significant asymmetry and publication bias. Heterogeneity across the included studies was assessed by I² statistic and Q test, and I²>50% or P value < 0.10 indicated significant heterogeneity. Sensitivity analysis was conducted by excluding one study each time and repeating the analyses to assess the stability of pooled results.

Part 2: Comparison of all first-line treatments was performed with Bayesian network meta-analysis by MCMC methods in a random-effect model. Four independent Markov chains were generated and run for 5000 adaptation and 50,000 inference iterations per chain to obtain the posterior distribution. Relative effects of all treatments were reported as HRs with 95% CrIs. The rank profile of each treatment was estimated, and the surface under the cumulative ranking curve (SUCRA) based on it was adopted to measure overall ranks of all treatments. Heterogeneity was assessed with I² statistic. The fit of consistency and inconsistency models, measured by deviance information criteria (DIC) were compared to investigate global inconsistency. Trace plots, Posterior Probability Density plots, Brooks-Gelman-Rubin diagnosis plots were presented to evaluate the convergence of iterations. These analyses were conducted using R package gemtc version 0.8-7 and JAGS software version 4.3.0.

Part 3: Kaplan-Meier curves for PFS were compared by the log-rank test, and HRs with 95% CIs were calculated through a Cox proportional hazards regression model using R package survival (version 3.1-12). Odds ratios (ORs) with 95% CIs for ORR were calculated with a logistic regression model using R package stats (version 4.0.2). The χ2 test or Fisher's exact test was used to test the difference of ORR between two groups.

All statistical tests were two-sided and P value < 0.05 was considered to be statistically significant. All analyses were performed with R 4.0.2.

3. Results

3.1. Study selection and characteristics

A total of 938 publications were retrieved through the literature search, and 114 potentially eligible studies were identified by reviewing the titles and abstracts after excluding those of duplications, reviews, case reports, meta-analysis, non-randomized controlled trials, with irrelevant topics or no usable data. After a full-text review, 13 publications^8,9,19, 20, 21, 22, 23, 24, 25, 26, 27, 28, 29 from 12 clinical trials involving 5527 patients were included. The flowchart of study selection is shown in Fig. 1.

Fig. 1.

PRISMA Flowchart.

The characteristics of the included studies are summarized in Supplementary Table 2. All the studies were exploratory analysis of prospective RCTs and most of them mentioned similar baseline characteristics of patients in the TMB-H and TMB-L subgroups. Generally, there were 9 clinical trials for tissue-based TMB (tTMB) and 7 clinical trials for blood-based TMB (bTMB), including 4 clinical trials investigating both tTMB and bTMB: MYSTIC, KEYNOTE-189, CheckMate-9LA, and RATIONALE307. TMB was tested by whole exosome sequencing (WES) in 5 clinical trials and through the cancer gene panel (CGP) in 7 clinical trials. The risk of bias of the studies were assessed and presented in Supplementary Table 3. The Begg test and Egger test provided P values of 0.91 and 0.80 for studies on PFS and of 0.82 and 0.84 for studies on OS, respectively, indicating no significant publication bias for the included studies, with the funnel plots visualized in Supplementary Fig. 2.

3.2. Mirror comparisons of the effect of TMB-H versus TMB-L on ICI-containing therapy

To demonstrate the effect of TMB on ICI-containing therapy, we constructed 10 mirror comparisons between the TMB-H and TMB-L subgroups (Part 1a. Fig. 2A). More significant PFS benefit was observed in patients with high TMB than in those with low TMB when treated with ICIs compared with chemotherapy, which was consistently observed with tTMB and bTMB (tTMB: HR = 0.57; 95% CI, 0.47-0.68; bTMB: HR = 0.59; 95% CI, 0.48-0.71). No significant heterogeneity existed across studies with either tTMB or bTMB (I² = 0%, P = 0.86) (Fig. 3A). In such, the tTMB and bTMB data were pooled together for further analysis and only tTMB-related data were included if both tTMB and bTMB data were provided in one study, thus providing a consistently superior PFS (HR = 0.58; 95% CI, 0.49-0.67; P < 0.0001, Fig. 3B).

Fig. 3.

Progression-free survival of ICI-containing therapy compared with chemotherapy for patients with TMB-High vs TMB-Low. (A) Pool analysis based on tTMB and bTMB. (B) Subgroup analysis by regimen Squares indicate mirror-specific relative HRs of TMB-High vs TMB-Low subgroup when treated with ICI-containing therapy compared to chemotherapy. Horizontal lines indicate 95% CIs. Diamonds represent the meta-analytic pooled HRs (95% CIs) of subgroups and all patients. MYSTIC¹ refers to the comparison of durvalumab to platinum-based doublet chemotherapy, and MYSTIC² refers to the comparison of durvalumab plus tremelimumab to platinum-based doublet chemotherapy. Abbreviations: Ate: Atezolizumab; bTBM, blood-based tumor mutation burden; Chemo, chemotherapy; CI, confidence interval; Doce, docetaxel; Durv, Durvalumab; HR, hazard ratio; ICI, immune checkpoint inhibitor; Ipi, ipilimumab; Nivo, nivolumab; P-Chemo, platinum-based doublet chemotherapy; Pem, pembrolizumab; Tisle, tislelizumab; TMB-H, TMB-High; TMB-L, TMB-Low; Tre, tremelimumab; t-TMB, tissue-based tumor mutation burden.

In the preplanned subgroup analysis of PFS, patients with high TMB exhibited better PFS benefit than those with low TMB when ICIs were applied either as first-line treatment (HR = 0.56; 95% CI, 0.47-0.67) or second- or post-line treatment (HR = 0.61; 95% CI, 0.45-0.81) (Supplementary Fig. 3A), and either as ICI-mono therapy (HR = 0.57; 95% CI, 0.46-0.69) or ICI-combo therapy (HR = 0.59; 95% CI, 0.47-0.74) (Fig. 3B). The disparity was consistently observed under the stratifications of the TMB test method (Supplementary Fig. 3B) or PD-L1 expression (Supplementary Fig. 3C).

In addition, the TMB-high subgroup presented greater OS benefit from ICIs over chemotherapy than the TMB-low subgroup with statistical significance for tTMB (HR = 0.76; 95% CI, 0.63-0.93) and bTMB (HR = 0.74; 95% CI, 0.58-0.94) (Supplementary Fig. 4A). Overall, TMB-high patients indicated significant OS benefit than TMB-low patients (HR = 0.79; 95% CI, 0.67-0.92; P = 0.0029, Supplementary Fig. 4B). The series of preplanned subgroup analyses exhibited superior and some significant OS benefit in TMB-high versus TMB-Low patients (Supplementary Fig. 4B). Sensitivity analysis suggested the high stability of the pooled effect size across studies (Supplementary Fig. 5).

To identify the potential difference of the impact of TMB on the efficacy of anti-PD-L1 versus anti-PD-1 inhibitors, a series of mirror comparisons were constructed with TMB-high or TMB-low patients (Part 1b, Fig. 2B). No statistical discrepancy was observed in PFS between anti-PD-L1 and anti-PD-1 in either TMB-high (HR = 1.09; 95% CrI, 0.64-1.84) or TMB-low (HR = 0.98; 95% CrI, 0.68-1.43) subgroup (Fig. 4). In addition, OS conferred no significant difference between PD-1 and PD-L1 inhibitors under both TMB-high and TMB-low settings (Supplementary Fig. 6).

Fig. 4.

Progression-free survival of anti-PD-L1 vs anti-PD-1 in TMB-High and TMB-Low subgroups. Squares indicate mirror-specific HRs of indirect comparison of anti-PD-L1 vs anti-PD-1. Horizontal lines indicate 95% CrIs. Diamonds represent the meta-analytic pooled HRs (95% CrIs) in the TMB-High and TMB-Low groups. MYSTIC¹ refers to the comparison of durvalumab to platinum-based doublet chemotherapy and MYSTIC² refers to the comparison of durvalumab plus tremelimumab to platinum-based doublet chemotherapy. CheckMate-227³ refers to the comparison of nivolumab to platinum-based doublet chemotherapy and CheckMate-227⁴ refers to the comparison of nivolumab plus Ipilimumab to platinum-based doublet chemotherapy. Abbreviations: Crls, credible interval; HR, hazard ratio; PD-1, programmed cell death-1; PD-L1, programmed cell death-ligand 1; PFS, progression-free survival; TMB, tumor mutation burden.

3.3. Network meta-analysis of various first-line therapies in the TMB-high and TMB-low subgroups

To better clarify the potential of TMB in guiding first-line therapies, we performed network meta-analysis assessing the survival benefit of all kinds of regularly available first-line regimens stratified by TMB levels (Part 2, Fig. 2C and Supplementary Fig. 1B). In the TMB-High subgroup, both PD-1 inhibitor plus chemotherapy and anti-PD-L1 plus anti-cytotoxic T-lymphocyte associated antigen-4 (CTLA-4) demonstrated significantly prolonged PFS than platinum-based chemotherapy (HR = 0.44; 95% CrI, 0.28-0.67; and HR = 0.49; 95% CrI, 0.24-0.99, respectively). In the TMB-Low subgroup, PD-1 inhibitor plus chemotherapy showed superior PFS than platinum-based chemotherapy (HR = 0.59; 95% CrI, 0.41-0.82), anti-PD-1 (HR = 0.41; 95% CrI, 0.24-0.65), anti-PD-L1 (HR = 0.54; 95% CrI 0.32-0.89), anti-PD-L1+anti-CTLA-4 (HR = 0.40; 95% CrI, 0.22-0.71), and platinum-based chemotherapy exhibited prolonged PFS than anti-PD-1 (HR = 0.69; 95% CrI, 0.46-0.98) (Fig. 5A). Trends of OS differences without significance were observed between any two first-line regimens according to the TMB level (Supplementary Fig. 7A).

Fig. 5.

Pooled estimates and plots of SUCRA for progression-free survival for all of the first-line treatments. (A) Pooled estimates for PFS in the TMB-High subgroup (upper triangle) and the TMB-Low subgroup (lower triangle). Data in each cell are HR (95% CrI) of the comparison of column-defining treatment vs row-defining treatment. HRs less than 1 favor column-defining treatment and significant outcomes are given in bold. (B) Plots of SUCRA for PFS in the TMB-High subgroup. (C) Plots of SUCRA for PFS in the TMB-Low subgroup. SUCRAs of each treatment are calculated based on the Bayesian ranking profiles in eFigure8 in Supplement Material and summarized in eTable4 in Supplementary Material. It suggests better efficacy of treatment with larger value of SUCRA, ranging between 0 (0%) and 1 (100%). Abbreviations: Chemo, chemotherapy; Crls, credible interval; CTLA-4, cytotoxic T-lymphocyte-associated antigen 4; HR, hazard ratio; PD-1, programmed cell death-1; PD-L1, programmed cell death-ligand 1; PFS, progression-free survival; SUCRA: surface under the cumulative ranking curve; TMB, tumor mutation burden.

Based on the Bayesian ranking profiles (Supplementary Fig. 8), SUCRAs were subsequently plotted to rank the survival benefit of all regimens, with SUCRAs and the probability of ranking best (PbBT) summarized in Supplementary Table 4. PD-1 inhibitor plus chemotherapy ranked best for PFS and third for OS in the TMB-High group and ranked best for both PFS and OS in the TMB-Low group. Anti-PD-1 plus chemotherapy ranked best for OS and second for PFS in the TMB-High group, while worst for both PFS and OS in the TMB-Low group. Platinum-based chemotherapy ranked worst for both PFS and OS in the TMB-High group (Fig. 5B and C; Supplementary Fig. 7B and C; Supplementary Table 4).

I² statistic of each network meta-analysis implied relatively low heterogeneity, and approximate DIC of consistency and inconsistency models indicated good consistency (Supplementary Table 5). Visual inspection of trace plots and Brooks-Gelman-Rubin diagnosis plots proved established convergence and stability of inference iterations (Supplementary Fig. 9 and 10).

3.4. Multi-cohort validation of the effect of TMB on ICIs therapy

To validate the impact of TMB on ICI therapy in real-world, five public cohorts30, 31, 32, 33, 34 involving 573 NSCLC patients receiving ICIs with available TMB data were included (Part 3; Fig. 2D), and baseline characteristics of all the patients are summarized in Supplementary Table 6.

Generally, superior PFS after receiving ICIs was observed in the TMB-high subgroup than in the TMB-low subgroup with a series of cut-off points adopted, with 180 selected as the optimal cut-off point (Supplementary Fig. 11) for TMB tested by WES (HR = 0.62; 95% CI, 0.48-0.81; P < 0.0001) and the median value (7.4 mut/Mb) of TMB in Rizvi cohort was used as the cut-off point for TMB tested by CGP. Five cohorts with TMB tested by WES or CGP were pooled for further analysis. The multivariable Cox proportional hazards regression model demonstrated that PD-L1 ≥ 50%, TMB-H, and intervention type remained significantly associated with PFS (PD-L1 ≥ 50%: HR = 0.59; 95% CI, 0.45-0.79; P < 0.001; TMB-H: HR = 0.69; 95% CI, 0.53-0.90; P < 0.01) and ORR (PD-L1≥50%: OR = 2.96; 95% CI, 1.76-5.00; P < 0.0001; TMB-H: OR = 2.76; 95% CI, 1.61-4.82; P < 0.001) benefit (Supplementary Table 7).

To better explore the clinical availability of TMB used in combination with PD-L1 expression, 327 patients with both TMB and PD-L1 expression data were extracted for further analysis. PD-L1 expression was associated with PFS after receiving ICIs in the overall population and the TMB-low subgroup (log-rank P < 0.01) but not in the TMB-high subgroup (Fig. 6A-C). TMB-H could distinguish beneficiaries of ICIs with unselected PD-L1 expression (HR = 0.68; 95% CI, 0.56-0.82; P < 0.0001) and those with PD-L1 < 50% (HR = 0.48; 95% CI, 0.35-0.66; P < 0.0001), but not those with PD-L1 ≥ 50% (Fig. 6D-F).

Fig. 6.

Multi-cohort validation of the impact of TMB on progression-free survival after receiving ICIs. (A-C) PFS after receiving ICIs by PD-L1 expression status in all patients (A), TMB-High subgroup (B) and TMB-Low subgroup (C); HR refers to the comparison of PD-L1 ≥50% vs PD-L1<50%. (D-F) PFS after receiving ICIs by TMB level in all patients (D), PD-L1<50% subgroup (E) and PD-L1≥50% subgroup (F). HR refers to the comparison of TMB-High vs TMB-Low. (G-I) PFS after receiving combination therapy (anti-PD(L)-1 plus anti-CTLA-4) compared with monotherapy (anti-PD(L)-1) in all patients (G), TMB-High subgroup (H) and TMB-Low subgroup (I). HR refers to the comparison of combination therapy vs monotherapy. Abbreviations: CTLA-4, cytotoxic T-lymphocyte-associated antigen 4; HR, hazard ratio; ICIs, immune checkpoint inhibitors; PD-1, programmed cell death-1; PD-L1, programmed cell death-ligand 1; PFS, progression-free survival; TMB, tumor mutation burden.

Similar results of TMB level combined with PD-L1 status were observed for ORR (Supplementary Fig. 12A and B). Further analysis demonstrated that patients treated with combination therapy had better clinical outcomes than those receiving monotherapy in the overall population (PFS: HR = 0.61; 95% CI, 0.47-0.78, P < 0.0001; ORR: 33% vs. 26%, P = 0.11) and in the TMB-high subgroup (PFS: HR = 0.43; 95% CI, 0.28-0.66; P < 0.0001; ORR: 59% vs. 33%, P = 0.001) but not in the TMB-low subgroup (Fig. 6G-I, Supplementary Fig. 12C).

4. Discussion

The clinical value and application strategy of TMB remain ambiguous. Here, we provided evidence indicating that TMB could be used to identify beneficiaries of ICI-containing therapies in patients with NSCLC, based on a comprehensive meta-analysis and validation by public datasets.

In this study, patients with high TMB obtained consistently superior survival benefit than those with low TMB under ICI-containing therapy versus chemotherapy. These results strongly indicated that TMB could stratify beneficiaries of ICI-containing therapies in patients with NSCLC. Intriguingly, the consistent survival benefits of ICIs over chemotherapy were observed for both tTMB and bTMB, similar to the observations based on tissue and blood paired testing samples in previous studies,35, 36 which further indicated that bTMB could be a promising and clinically accessible biomarker for predicting benefit from ICIs.21, 37,³⁷ Moreover, TMB estimated by CGP or WES delivered similar predictive performance in our analysis, which supported CGP as a routine technology to test TMB for clinical usage.33, 38, 39

In our previous study,¹⁴ anti-PD-1 exhibited superior efficacy versus anti-PD-L1 based on mirror comparisons in patients with NSCLC; however, the impact of TMB levels on the efficacy of anti-PD-1 versus anti-PD-L1 remained unclear. In the present study, we found that anti-PD-L1 versus anti-PD-1 showed no statistically different survival outcomes in both TMB-high and TMB-low subgroups, implying that TMB might not be associated with the efficacy of anti-PD-L1 versus anti-PD-1 antibodies. However, the results should be interpreted carefully because of the limited number of mirrors and probably ambiguous match on PD-L1 status and TMB test.

With various first-line regimens available, how to determine the optimal treatment strategy remains a major concern.²⁴ In our network meta-analyses, anti-PD-1 combined with chemotherapy ranked first on PFS among all first-line therapies regardless of TMB. However, in the clinical practice, ICI-combined with chemotherapy may produce severe and even life-threatening adverse events and shorter survival as compared to ICI-only treatment. Thus, our results may provide some guidance for first-line treatment selection when ICI plus chemotherapy is unavailable or unsuitable for patients. In the TMB-High group, anti-PD-L1 plus anti-CTLA-4 indicated superior PFS and OS benefit than single ICI treatment and platinum-based chemotherapy, while anti-PD-L1 plus anti-CTLA-4 was the worst choice of therapy for TMB-Low patients. Anti-PD-1 plus anti-CTLA-4 showed a moderate survival benefit for both TMB-High and TMB-Low patients. Coincidently, CA209-538 pan-cancer population treated with nivolumab/ipilimumab combination regimen did not report a predictive value of TMB.⁴⁰ To our knowledge, this is the first study to explore and describe the sequence of the benefit derived from all kinds of first-line regimens stratified by TMB. With accelerating accumulated evidence, the optimal first-line regimens should be further discussed to guide clinical practice.

As the most evidenced predictive biomarkers for immunotherapy, TMB and PD-L1 expression have been identified to be independently representing mostly distinct populations. In our analyses in five public cohorts, PD-L1 ≥ 50% or TMB-high could serve as mutually complementary biomarkers, which highlighted the importance of the combined application as well as the sequential selection of these two biomarkers in clinical practice.

While previous meta-analysis on the predictive value of TMB included retrospective studies,10, 11, 12^,41 the data in our study were derived from exploratory analysis of RCTs, which minimized the heterogeneity and ensured the quality of the studies. In addition, the mirror principle-based comparisons used in our study not only further reduced the confounding factors and enhanced the comparability,¹⁴ but also delivered quantitative effect over TMB-high and TMB-low subgroups and between anti-PD-L1 versus anti-PD-1 therapies.

5. Limitations

Our study has several limitations. First, the limited number of the included studies may introduce potential bias. In addition, some of the studies were derived from meeting abstracts, whose baseline characteristics of patients and design were incomprehensive, and the results should be interpreted with caution. Secondly, with regard to TMB as a biomarker, the refinement of the calculation algorithm across multiple next generation sequencing (NGS) platforms and the establishment of robust cancer-specific cutoff value⁵ are required to standardize the use of TMB in clinical practice.^42,43 Third, TMB was found to exhibit a predictive value only in cancer types showing positive correlation between CD8 T cell levels and neoantigen load;⁴⁴ thus, the complex interaction within tumor microenvironment may require the combined use of multiple biomarkers45, 46 or the construction of prediction models47, 48 to achieve the optimal predictive performance. Lastly, the multi-cohort validation provided clues for the future combined utilization of PD-L1 and TMB as predictive indicators. Future accumulations of real-world evidence are needed to better reform the practice of united biomarkers.⁴⁹

6. Conclusions

Our comprehensive analyses demonstrated TMB as a favorable predictor to identify NSCLC patients who can benefit from ICI-containing therapies versus platinum-based chemotherapy. Large-scale clinical trials in prospective settings to investigate the prediction value of TMB are warranted to establish guidelines in the future.

Declaration of Competing Interest

All data generated or analyzed during this study are included in this published article. Dr. G. Wang is an employee of Burning Rock Biotech. There was no financial compensation outside of salary. No other conflicts were reported.