Tumor Mutational Burden as a Predictive Biomarker in Solid Tumors

A major advance in cancer treatment is the development of immune checkpoint inhibitors (ICI) that have produced durable responses and improved survival in multiple solid malignancies (1–8). However, a majority of patients treated with ICIs do not derive benefit, and therefore identification of predictive biomarkers of ICI response are needed to enable more selective use of ICIs as well as to elucidate and overcome mechanisms of treatment resistance. Tumor mutational burden (TMB) is broadly defined as the number of somatic mutations per megabase of interrogated genomic sequence. TMB is believed to be a key driver in the generation of immunogenic neopeptides displayed on major histocompatibility complexes (MHC) on the tumor cell surface that influence patient response to ICIs. Tumor-specific neoantigens arise from somatic mutations (9, 10) and can play a pivotal role in tumor-specific T cell–mediated, antitumor immunity after inhibition of checkpoint signals (11–14). In addition to neoantigen quantity, evidence suggests that quaIity may also be important in that high-quality neoantigens might include expressed clonal neoantigens in essential genes, which bind to multiple HLA alleles and cannot be repressed or deleted by virtue of their genomic position (15). Accumulating evidence suggests that TMB may be a predictive biomarker of tumor response to ICIs in several cancer types (10, 16, 17). The most robust initial responses to ICIs were observed in melanoma and non–small cell lung cancer (NSCLC), which typically have high mutation burden owing to the mutagenic effects of ultraviolet light and tobacco smoke, respectively (18). Subsequently, significant associations between high TMB and response to ICIs were reported in other solid tumor types (19–21). However, outliers have been observed that include renal cell carcinoma and Merkel cell carcinoma that responded better than expected on the basis of TMB alone, suggesting the importance of other as yet undefined factors (16, 22, 23). In a study in patients with previously treated, unresectable, or metastatic solid tumors (KEYNOTE-158), TMB-high status (≥10 mut/Mb) was associated with a clinically meaningful improvement in efficacy of the anti–PD-1 antibody pembrolizumab (24). Responses were observed across tumor types, and microsatellite instability–high (MSI-H) status did not account for all of the increased clinical benefit in the TMB-high subgroup (24). On the basis of these data, the FDA approved pembrolizumab monotherapy for the subgroup of solid-tumor patients with TMB ≥10 mut/Mb who are treatment-refractory and lack satisfactory alternative treatment options. Although data from KEYNOTE-158 demonstrated a role for TMB in selection of patients for cancer treatment, important issues remain, including the selection and implementation of a fixed TMB cutoff based on pan-cancer data. In this regard, observed cancer type–related differences in TMB distributions are relevant to determining optimal TMB cutoffs to enable its use as a predictive biomarker for immunotherapy.

TMB is accurately measured using whole-exome sequencing (WES), but this approach is impractical for use in the clinic. Although panel-based sequencing of tumor tissue is commonplace in clinical practice, differences in panel size, mutation types, and bioinformatic platforms exist. TMB is associated with certain other biomarkers, including MSI-H that is detected in a subset of human cancers and is due to deficient DNA mismatch repair (dMMR). Tumors with MSI-H/dMMR typically display high TMB (25–27), and MSI-H/dMMR is an established predictive biomarker for the efficacy of ICIs. Expression of PD-L1 on tumor cells and immune cells has become a widely used predictive biomarker for responsiveness to ICIs in several cancer types (6, 28–31). However, TMB levels and response to immunotherapy in many cancers is independent of the level of PD-L1 expression (4, 32, 33), suggesting a potential role of TMB to identify additional subgroups of patients who may benefit from ICIs (34). In this review, we will discuss the need for a consensus definition of TMB as well as the need to standardize TMB measurement among gene panels. Furthermore, we will review TMB variability across solid tumors with implications for cutoff selection, its association with benefit from ICIs, and strategies to optimize TMB as a predictive biomarker for ICIs in clinical practice. We also discuss the current conundrum of supportive retrospective evidence, but the relative paucity of prospective data confirming the clinical utility of TMB.

TMB can be assessed using a number of next-generation sequencing (NGS) platforms, including whole-genome sequencing (WGS), WES, or targeted panel sequencing. WES is the “gold standard” for measuring TMB, allowing for the detection of (nonsynonymous) somatic coding mutations present within the entire exome. WES targets approximately 30 Mb of coding regions, covering all approximately 22,000 genes and making up approximately 1% of the genome. The definition of TMB varies by the measurement method utilized. Although TMB has been accurately measured by WES in several studies, this is currently not feasible in clinical practice due to its high cost, relatively long turnaround time, and the need for sufficient tissue samples. Multiple commercially available gene panels designed for TMB estimation cover between approximately 0.80 and 2.40 Mb, representing <5% of the total coding sequence (35–44). The number of genes in each of the gene panels ranges between 324 and 595 genes. Importantly, panels are not necessarily limited to the coding regions because many panels include intronic regions needed for gene fusion detection. Shown in Fig. 1 are confidence intervals relative to the level of TMB of four theoretical gene panels (size from 0.5 Mb to 4 Mb) that vary markedly according to panel size. The coefficient of variation (CV) of TMB derived from panel sequencing decreases in a manner that is inversely proportional to both the square root of the panel size and the square root of the TMB level; for example, halving the CV requires a 4-fold increase in panel size (45). Multiple NGS panels that are commercially available include Memorial Sloan Kettering-Integrated Mutation Profiling of Actionable Cancer Targets (MSK-IMPACT) and FoundationOne CDx (Foundation Medicine, Inc.), both of which are approved by the FDA, that cover approximately 1.14 Mb over 468 genes or approximately 0.8 Mb over 324 genes, respectively (Table 1). Each commercial laboratory uses their own bioinformatic algorithms and workflow that were optimized using sequencing methods, mutation types, and filters that best suit their own panel specifications.

TMB outputs from gene panel assays are usually normalized to mutations per Mb because they differ in the number of genes and target region size. Both MSK-IMPACT and FoundationOne CDx panels detect (nonsynonymous) somatic coding mutations per megabase of tumor genome examined, inclusive of frameshift, point mutations, and small insertions and deletions (indels; see Supplementary Material). While synonymous mutations are detected by these panels (not reported by MSK-IMPACT), they are not involved in neoantigen production, although their inclusion may reduce sampling noise and improve the approximation of TMB across the whole genome if tumor–normal pairs are sequenced (35). MSK-IMPACT and FoundationOne CDx have been shown to be moderately concordant with WES in TMB assessment (35, 45, 46). Importantly, differences such as the location and size of the sequenced region, the number of sequenced genes, the mutation types detected, as well as differences in the definition of TMB among panels, create confusion in the interpretation of TMB and in comparing TMB values across test platforms. Genetic changes in cancer include nonsynonymous (including missense, nonsense, frameshift, and splice-site mutations) and synonymous point mutations, indels, and gene copy-number alterations (CNA) and gene fusions. Indel calling can vary depending on the bioinformatic pipeline used, and whether indels generate a higher number of immunogenic neoantigens needs to be determined. Calculation of TMB from panel-based sequencing data has important limitations (35). There is a need to harmonize the types of mutations analyzed, as missense mutations are included in all panels, but other types can vary. Panel sequencing–derived TMB measurement extrapolates the total number of mutations in the coding sequence by analysis of a limited panel of genes. For clinical purposes, evidence suggests that gene panels of at least 1 Mb are needed for TMB measurement (35, 47, 48). Even for large panels, the stochastic error related to panel size represents the largest of all contributions to total TMB variance (49). One can convert mutation number from WES data to mut/Mb that is reported in gene panel sequencing data (50); however, it is dependent on multiple sequencing-related parameters (e.g., sequencing methodology, WES enrichment kits, bioinformatics pipelines, etc.), such that a static conversion rate does not exist.

Other relevant factors include the bioinformatic protocols used to calculate TMB and methods of filtering germline mutations (43, 47, 48, 51, 52). Germline mutation filtering is an important step in panel-based TMB measurement because only somatic mutations in the tumor can be recognized by the immune system. Germline mutation filtering can be performed in silico using bioinformatic pipelines, or, alternatively, paired normal tissue or blood samples can be sequenced and used as a filtering tool. In panel-based sequencing, fewer total mutations were called when a patient-matched normal tissue was used in variant calling because germline single-nucleotide variants (SNV) were appropriately filtered out (53). TMB variability introduced by errors in somatic mutation detection is only moderate compared with stochastic error related to panel size and other confounders (49, 54). Panel size is the most important contributor to TMB variance because panels represent only a very small proportion of the exome, especially in tumors of small to medium TMB (up to 10 mut/Mb; ref. 54). With regard to gene panel composition and its contribution to TMB variability, an in silico study found that TMB variability increased by 8% when using 1 MB panels composed of oncogenes and tumor suppressor genes (TSG) compared with 1 MB panels composed of random genes (45). Thus, including oncogenes and TSGs only slightly increases TMB variability. Bioinformatic pipelines usually include negative filtering for cancer hotspot mutations, further mitigating the influence of the panel composition.

TMB is a continuous variable, and variability of TMB (ranging from 0.001/Mb to more than 1000/Mb) has been observed across and within cancer types (18, 35, 41). Studies indicate that some cancer types have less variability in TMB, such as lung and head and neck cancers, and some have greater variability, such as colon, bladder, and uterine cancers (55). Cancers related to chronic mutagenic exposures, such as lung cancer (tobacco) and melanoma (UV light), exhibit the highest TMB, whereas leukemia and certain childhood cancers have the lowest TMB (18). In an analysis of 24 cancer types using The Cancer Genome Atlas (TCGA) database, only three cancer types, adenocarcinoma of the colorectum, stomach, and uterus, harbor bimodal or multimodal TMB distribution (ref. 56; Fig. 2A). Also shown are the percentage of tumors with TMB above 10 mut/MB and the contribution of MSI-H tumors. In these cancers, TMB distribution is shaped by the occurrence of hypermutation in MMR-deficient and/or POLE/POLD1–mutated tumors and permits relatively clean dichotomization. For most other cancer types, including adenocarcinoma or squamous cell carcinoma of the lung or cutaneous melanoma, TMB is unimodally distributed with a dense point cloud of TMB scores scattered around the cutoff (ref. 56; Fig. 2A). Eleven cancer types include a high or at least moderate percentage of tumors above the cutoff of 10 mut/Mb: melanoma (71%), lung squamous cell carcinoma (50%), lung adenocarcinoma (44%), uterus adenocarcinoma (39%), transitional cell carcinoma (36%), stomach adenocarcinoma (29%), head and neck squamous cell carcinoma (19%), colorectal carcinoma (18%), cervix carcinoma (15%), esophagus adenocarcinoma (9%), and sarcoma (6%; Fig. 2A). The percentage of tumors above the cutoff for the remaining 13 cancer types was low (<5%). For uterus, stomach, and colorectal carcinomas, the majority of tumors above the cutoff were MSI-H (77%, 69% and 78%). Accordingly, 9% of adenocarcinomas of the uterus and stomach and 4% of colorectal carcinomas were above the cutoff and microsatellite stable (MSS)/MSI-L. Additional data for TMB distributions are provided in sections for individual tumor types.

Thousands of somatic mutations can now be identified in single cancer samples, offering the possibility of deciphering various mutational signatures even when they are caused by several mutational processes (18, 57) that can differ by cancer types and between individual tumors. Using the TCGA pan-cancer cohort, analysis of TMB and mutational signatures was performed in 24 cancer types (8,273 tumors), and a new type of heat map was generated to analyze the contribution of single base substitution–derived mutational signatures to TMB and, in particular, to hypermutation and ultra-hypermutation (Fig. 2B). To gain insight into the biological processes underlying high TMB in specific tumors, we took advantage of a method of cancer genome analysis recently developed by Alexandrov and Stratton (18). Ultra-hypermutation (>100 mut/MB) or hypermutation (>10 mut/MB) can result either from the activity of a single mutational process or by accumulation over multiple such processes. Mutational processes that can cause a very high TMB and hypermutation include: POLE/POLD1 mutation, mismatch repair deficiency, UV light, tobacco smoking, AID/APOBEC activation, and the two clock-like mutational processes (SBS1, SBS5; Fig. 2B). TMB is to a large extent independent of PD-L1 status in most cancers (34), as shown in an analysis of 24 cancer types included in the TCGA database (Fig. 2C).

A TMB cutoff is a function of the gene panel (genomic footprint and bioinformatics platform) that is used in a given study. Because data obtained from a given gene panel cannot be directly applied to another panel without a conversion algorithm, direct comparisons of results between panels can be very problematic. As indicated, there are cancer type–related biological TMB distributions, and accordingly these distributions may be relevant to evaluation of the clinical utility of TMB and for determining optimal TMB cutoffs. Because of diverse TMB distributions, optimal TMB cutoff values to discriminate potential responders versus nonresponders to ICIs may vary significantly among cancer types (35, 58). Cohort-specific TMB cutoffs have been defined differently across studies, tumor types, and testing platforms, and using variable bioinformatics methods. Importantly, cancer type–related biological TMB distributions are distinct from prognostic cutoffs whereby the former may not predict the latter.

The rationale for the association between TMB and benefit from immunotherapy is based on the hypothesis that tumor mutation–specific neoantigens can be displayed on MHCs on the tumor cell surface, and then recognized by tumor-infiltrating T cells. Accordingly, a higher TMB will generate more neoantigens that can then trigger intratumoral T cells whose ability to attack and destroy tumor cells is enabled by ICIs (10, 17, 59). The first evidence to support this hypothesis came from studies of melanoma and NSCLC treated with anti-CTLA4 and anti–PD-1 antibodies, respectively, whereby a higher nonsynonymous mutation burden was associated with improved objective response rate (ORR) and progression-free survival (PFS; refs. 10, 17). Using an optimized receiver operating characteristic (ROC) in patients with melanoma, TMB levels that were divided into high (>23.1 mut/Mb), intermediate (3.3–23.1 mut/Mb), and low (<3.3 mut/Mb) groups showed superior prediction of ICI efficacy compared with a binary classification (60). Evaluation of TMB in an NSCLC (CheckMate 568) study using the FoundationOne CDx assay utilized ROC curves to determine an optimal TMB cutoff in patients receiving first-line therapy with nivolumab plus ipilimumab. The ORR was increased in patients with higher TMB, and the benefit was observed to plateau with a TMB threshold of ≥10 mut/Mb. This TMB cutoff of ≥10 mut/Mb was subsequently evaluated in the first prospective phase III trial known as CheckMate 227 in patients with NSCLC where TMB served as the co–primary efficacy endpoint (61–63). Patients whose tumors had a prespecified TMB cutoff of ≥10 mut/Mb had significantly prolonged PFS, but not overall survival (OS; independent of PD-L1 expression), with a combination of nivolumab + ipilimumab versus chemotherapy (63). Despite the negative data for OS, PFS is considered to be more informative for predictive biomarker evaluation because OS is influenced by therapies given after progression on ICIs.

Multiple studies have shown an association between TMB level and the efficacy of immunotherapy in melanoma, lung cancer, and urothelial cancer, although TMB only weakly discriminated responders from nonresponders (AUC 0.6–0.7) among these tumor types (6, 10, 17, 59, 61, 64). In a pan-cancer analysis of 151 patients treated with anti–PD-1/PD-L1 monotherapy, tumor response rate and TMB level were linearly associated and there was also a significant association with the dichotomized TMB level (≥20 mut/Mb), which was consistent across tumor types (65). A nonrandomized and open-label phase II study of pembrolizumab monotherapy was conducted in patients with multiple advanced solid tumors who progressed on or were intolerant to one or more lines of prior therapy (KEYNOTE-158). In this study, 790 patients had evaluable TMB data, and of these, 102 (13%) were TMB-high (≥10 mut/Mb per FoundationOne CDx). The most common tumor types that showed TMB-high included small cell lung cancer (33%) and carcinomas of the uterine cervix (16%), endometrium (15%), and anus (14%). The most common non–TMB-high tumors were mesothelioma (12%) and carcinomas of neuroendocrine origin (12%), salivary gland (11.0%), and endometrium (10%). The ORR and PFS were superior in patients with TMB-high versus TMB-low tumors (ORR: 29% vs. 6%; PFS at 12 months: 26% vs. 13%). Among the TMB-high group, 88/102 (86%) tumors were MSS, indicating that MSI-H status did not account for the predictive utility of TMB-high (24). On the basis of these impressive data, the FDA approved pembrolizumab monotherapy for the treatment of patients with solid tumors showing TMB-high status (≥10 mut/Mb), as determined by an FDA-approved test, who had progressed following prior treatment and lacked alternative treatment options. Here, we analyzed 24 cancer types (8,273 tumors) from the TCGA database whereby TMB-high (≥10 mut/Mb) was found in more than 20% of cutaneous melanomas, NSCLC (squamous carcinoma and adenocarcinoma), transitional cell carcinoma, and adenocarcinomas of stomach and uterus (of which most are MSI-H; Fig. 2A). In addition, more than 10% of squamous carcinomas of the head/neck and cervix and colorectal adenocarcinomas (most MSI-H) were TMB-high. Using this cutoff for TMB, many solid tumors are eligible for treatment with ICIs with the potential for clinical benefit. Although the KEYNOTE-158 data and related FDA approval are practice-changing, a pan-cancer and prespecificed TMB-high cutoff may not be an optimal approach for individual tumor types.

We acknowledge that cutoff values for TMB are critical for use in clinical decision-making and propose three potential approaches to determine cutoff values of TMB as a predictive biomarker: (1) a single cutoff for all cancer types, that is, one-size fits all that is ideally determined in a pan-cancer trial, (2) cancer-specific cutoffs that increase complexity and cost, and (3) a variable cutoff that relates to some upper percentiles for each individual cancer type (58). A pan-cancer TMB cutoff, as approved by the FDA for use of pembrolizumab, is intended to enrich for responders to ICIs and thus enables patient selection for such therapy. However, a static cutoff for TMB may not be an optimal approach to identify tumor type–specific responders to ICIs, and thus fine-tuning of TMB cutoffs by tumor type is an area of future research. A further limitation of a pan-cancer TMB cutoff approach is illustrated using a dataset where genomic correlates of response to ICIs were examined in MSS solid tumors (66). Using this dataset of MSS solid tumors, we analyzed the impact of TMB cutoff selection on sensitivity and specificity of ICI response prediction. As shown in Fig. 3A–D, a cutoff of 10 mut/Mb corresponded to a sensitivity of approximately 75% for melanoma and lung and bladder cancers, although specificity varied widely and was much lower for melanoma (44%) than for lung and bladder cancers (89% and 70%). Using this cutoff, 20% to 25% of responders would be missed. Decreasing the cutoff to 5 mut/Mb would increase sensitivity, but at the expense of decreasing specificity to 29%, 68%, and 20% for the cancer types under consideration. This analysis raises concern for use of a tumor type–agnostic designation for high TMB to predict ICI response. If a cutoff is selected for each specific cancer type, then complexity is increased in that prospective clinical trials are needed to validate TMB cutoffs in individual tumor types. The third approach requires a comparator set/benchmark against which percentiles are chosen. A universal standard would be needed (how many cases, which tumor types, which spectrum of TMB). Regardless, all three approaches and their determined cutoffs are a function of the gene panels used such that comparability is ideally achieved only if the same panel is used. Other potential options in need of further research are probabilistic approaches of low-intermediate-high cutoffs. A current limitation for in-depth comparative study of the three approaches as well as for TMB cutoff optimization is the limited number of patients with high-quality outcome data and available TMB data. Although statistical methods such as Subpopulation Treatment Effect Pattern Plot (STEPP) and Cutoff Finder for cutoff optimization are available (67, 68), clinical studies are typically well powered for comparison of immunotherapy with a reference therapy, but underpowered for biomarker analysis with a fixed cutoff and even more underpowered for cutoff optimization.

Consistent with the known high TMB in cutaneous melanoma, metastatic melanoma also showed an increase in median nonsynonymous mutation burden assessed by WES (18, 64). Importantly, TMB differed significantly among melanomas harboring mutations in BRAF, NRAS, NF1, or triple wild-type (WT) tumors with median TMB values of 12.0, 17.6, 62.7, and 2.2 mut/Mb, respectively (P < 0.001; ref. 60). Melanomas with NF1 mutations are associated with chronic UV damage and thus have a high TMB. Of note, TMB differed among melanoma subtypes, with cutaneous and occult melanomas having higher TMB than did acral and mucosal subtypes, as well as higher response rates to anti–PD-1 blockade (69).

Among 32 of 110 (29%) patients with metastatic melanoma treated with anti–PD-1/PD-L1 antibodies, median TMB values were significantly higher in responders versus nonresponders (45.6 vs. 3.9 mut/Mb, P = 0.003). Furthermore, the ORR to ICI treatment was significantly increased in tumors with high TMB (>23.1 mut/Mb) versus intermediate (3.3–23.1 mut/Mb) versus low (<3.3 mut/Mb) groups (ORR: 82% vs. 36% vs. 10%, P = 0.003; ref. 60; Table 2). Results were confirmed in an independent validation cohort (n = 33; P = 0.002; ref. 60). Certain genomic alterations such as mutations in NF1 or BRCA2 were more common in responders to ICI treatment compared with triple WT tumors (60). In a pooled dataset of 300 patients with cutaneous metastatic melanoma, TMB was categorized as high (7.1 mut/Mb) in 21% of tumors and was associated with a higher ORR to anti–PD-1/PD-L1 treatment compared with lower TMB (33% vs. 21%; ref. 70; Table 2). Responders with mucosal or acral melanomas had a lower TMB than did progressors with cutaneous or occult melanoma, suggesting that melanoma subtype may confound the association between TMB and response to anti–PD-1 therapy (69). In this regard, TMB failed to predict benefit from ICIs in a multivariable model adjusting for melanoma subtype (cutaneous, occult, acral, and mucosal). Among aggregate data in patients with melanoma, only a modest difference in clinical benefit by TMB level has been observed. Further data in molecular subgroups remain of interest, with the caveat that large sample sizes are needed for such comparisons.

Lung cancers have the highest somatic mutation burden among solid tumors that is believed to be due to direct exposure to mutagens in tobacco smoking (18, 71–73). Lung cancer in smokers was associated with significantly higher median TMB (10.5 mut/Mb) compared with never-smokers with lung cancer (0.6 mut/Mb; ref. 74). Of note, smoking status is inversely related to prevalence of targetable oncogenic driver mutations in EGFR, ALK, and ROS1 genes in lung adenocarcinoma (75). Among lung cancers (squamous carcinoma and SCLC) associated with cigarette smoking, mutations in BRAF, KRAS, PTEN, and PIK3CA are most common (76). In a large lung cancer database, TMB was found to be relatively similar across lung cancer histologies, although squamous cell carcinomas had a slightly higher mean TMB (n = 1,324, 11.3 mut/Mb) compared with adenocarcinomas (n = 7,925, 9.1 mut/Mb) and SCLC (n = 640, 10.3 mut/Mb; ref. 75). Variability of TMB has been identified among molecular subgroups of lung cancer. In this regard, tumors with ALK/ROS1 (n = 489), EGFR (n = 1,775), and MET exon 14 (n = 286) had mean TMB levels (mut/Mb) of 3.1, 4.5, and 6.2, respectively (75).

The ability of TMB to predict efficacy of the anti–PD-1 antibody pembrolizumab was first shown in patients with advanced NSCLC from two independent cohorts (10), and led to multiple clinical trials. In patients with metastatic or recurrent NSCLC, the CheckMate 026 study compared nivolumab versus platinum-doublet chemotherapy as first-line treatment. Patients whose tumors had a higher TMB experienced an improved ORR (47% vs. 28%) and longer PFS [median (m) PFS: 9.7 vs. 5.8 months, HR 0.62; 95% CI, 0.38–1.00] in the nivolumab study arm (ref. 77; Table 2). In the subsequent CheckMate 568 study of nivolumab plus ipilimumab in treatment-naïve patients with metastatic NSCLC, TMB data were available in 98 patients whereby ORRs (44% vs. 12.0%) and mPFS (7.1 months vs. 2.6 months) were improved in those whose tumors had high versus low TMB (≥10 mut/Mb; ref. 61). In the phase III MYSTIC trial, 1,118 patients with treatment-naïve, metastatic NSCLC (without EGFR or ALK mutations) were randomized to durvalumab (PD-L1 inhibitor) alone or combined with tremelimumab (CTLA4 inhibitor), or platinum-based doublet chemotherapy (78). Among 488 patients with tumor cell PD-L1 expression ≥25%, the primary endpoint was not met in that neither study arm showed a survival benefit compared with the chemotherapy arm. Among 460 patients with TMB data, high TMB (≥10 mut/Mb) was associated with longer median OS that did not achieve statistical significance (78). Importantly, TMB levels from 352 matched tumor and blood samples tested were correlated (Spearman ρ = 0.6; Pearson r = 0.7), indicating that TMB testing in blood may be a useful adjunct when tissue testing is not feasible (78). In the first prospective trial (CheckMate 227) that examined TMB as a co–primary endpoint biomarker, nivolumab plus ipilimumab was compared with platinum-doublet chemotherapy in patients with advanced NSCLC in the first-line setting. Among patients with TMB data (n = 1,004), tumors with high versus low TMB (prespecified cutoff of ≥10 mut/Mb) demonstrated an ORR of 45.3% versus 26.9% and superior PFS of 7.2 versus 5.5 months (HR 0.58; 95% CI, 0.41–0.81, P < 0.001) for nivolumab plus ipilimumab versus chemotherapy, respectively, that was independent of PD-L1 expression (ref. 63; Table 2). No difference was observed for OS in this study, yet PFS is a preferred endpoint for predictive biomarker evaluation because OS includes outcome of all subsequent treatments. In the CheckMate 568 study (61), TMB ≥10 mut/Mb was a cutoff point for ORR and was validated as a predictive biomarker when prospectively applied to CheckMate 227 (63), where it showed improved PFS for nivolumab and ipilimumab compared with standard chemotherapy. In these studies, PD-L1 and TMB were independent predictive biomarkers.

Although a modest association of TMB with efficacy of dual checkpoint blockade was demonstrated in the CheckMate 568 and 227 studies, inconsistent data have been reported for the role of TMB as a predictor of response and survival to dual checkpoint blockade across clinical trials. Among 17 patients from a pan-cancer cohort treated with a combination of anti–PD-1 and anti-CTLA4 antibodies, TMB levels were not associated with treatment efficacy (65). In treatment-naïve patients with metastatic NSCLC (N = 955) randomized to tremelimumab plus durvalumab versus platinum-based chemotherapy (NEPTUNE trial), TMB level (TMB high ≥20 mut/Mb) determined in blood samples was not shown to predict patient survival (https://www.astrazeneca.com/media-centre/press-releases/2019/update-on-the-phase-iii-neptune-trial-of-imfinzi-plus-tremelimumab-in-stage-iv-non-small-cell-lung-cancer-21082019.html). The explanation for the more modest association of TMB with clinical outcome among patients treated with dual checkpoint inhibitors is unclear, but relevant factors include retrospective design; relatively small numbers of cases with TMB data, especially the TMB-high population that can limit statistical power; varied DNA sources, that is, blood versus tissue; and multiple platforms for TMB measurement as well as variable TMB cutoffs for correlation with outcomes.

The role of TMB as a predictive biomarker in extensive- stage SCLC was studied in patients who had failed at least one prior chemotherapy regimen. In patients treated with nivolumab alone or combined with ipilimumab (CheckMate032), the ORR by treatment arm increased stepwise in patients whose tumors showed high versus medium versus low TMB levels (ref. 59; Table 2). Importantly, dual ICI treatment was associated with an impressive ORR of 46.2% and an estimated 1-year OS rate of 62.4% in these previously treated patients with high TMB tumors (59).

Cigarette smoking is the most common risk factor and is estimated to be responsible for approximately 50% of all urothelial carcinomas (79). Urothelial carcinoma carries the third-highest mutation rate among solid tumors (80). In an analysis of 472 urothelial carcinomas using a panel of 237 cancer genes, a median TMB of 10.9 mut/Mb was found. TMB was increased in high-grade versus low-grade cancers with or without muscle invasion (81). Mutation in the apolipoprotein B editing enzyme (APOBEC) was common in all stages and locations of urothelial carcinoma, was strongly associated with TMB, and was more frequent in muscle-invasive bladder cancers (MIBC) and high-grade non–muscle-invasive bladder cancers (81).

In the CheckMate 275 study, patients with unresectable locally advanced or metastatic urothelial carcinoma who failed at least one platinum-based regimen received nivolumab monotherapy. Of 270 patients, 139 (51%) had evaluable TMB. High TMB (≥13 mut/Mb) was associated with higher ORR, longer PFS, and longer OS in patients treated with nivolumab (ref. 19; Table 2). In a phase II single-arm trial I of atezolizumab monotherapy in this same study population, median TMB (measured in 150 patients by a 315-gene panel) was significantly increased in responders versus nonresponders (12.4 mut/Mb vs. 6.4 mut/Mb, P < 0.0001; ref. 6; Table 2). The role of TMB as a predictor for immunotherapy outcome was also explored in the neoadjuvant setting. In the PURE-01 study of neoadjuvant pembrolizumab in patients with MIBC, median TMB was higher in patients who achieved a pathologic complete response (18.4 mut/Mb vs. 8.4 mut/Mb; ref. 82). These data suggest that TMB may enrich for responders to ICIs in patients with urothelial carcinomas.

Etiologic factors associated with squamous cell carcinoma of the head and neck (SCCHN) include tobacco and alcohol consumption and virus infection [human papillomavirus (HPV) and Epstein-Barr virus (EBV); ref. 16]. Of 126 patients with SCCHN who received anti–PD-1/PD-L1 treatment, 64% had evaluable TMB that ranged from 1.5 to 76.0 mut/Mb (median 7.6 mut/Mb), and 13% of tumors showed TMB values greater than 20 mut/Mb (20). The median TMB level was significantly higher among responders versus nonresponders (17.7 vs. 7.1, P < 0.01) to ICIs. Interestingly, patients with non–virus-mediated versus virus-mediated tumors had a higher median TMB (8.2 vs. 4.7, P < 0.01), which is consistent with prior studies (83, 84). TMB predicted anti–PD-1/PD-L1 response among HPV- and EBV-negative tumors, but not among HPV- or EBV-positive tumors. Moreover, TMB correlated with longer OS among virus-negative patients, but not among their virus-positive counterparts. Among responders to anti–PD-1/PD-L1 treatment, the most common mutations were in NOTCH1, TP53, KMT2D, and SMARCA4 genes (ref. 20; Table 2). These data support TMB as a promising biomarker for ICI efficacy in SCCHN.

Among various gastrointestinal (GI) cancer types (4,125 patients), mean level of TMB ranged from 5.1 to 13.0 mut/Mb (592 gene panel; Table 2; ref. 85). Adenocarcinomas of the right colon and small bowel exhibited the highest mean TMB (13 and 10.2 mut/Mb, respectively), whereas pancreatic adenocarcinoma, pancreatic neuroendocrine tumors, and gastrointestinal stromal tumors had the lowest TMB levels (6.1 mut/Mb, 5.8 mut/Mb, and 5.1 mut/Mb, respectively). Cancers with the highest known TMB levels are ultramutated and are caused by mutations in polymerase epsilon (POLE) that impair DNA proofreading (86). Such tumors are nearly exclusively MSS, have TMB values ranging from 122 mut/Mb to 303 mut/Mb, and comprise approximately 1% to 2% of all MSS colorectal cancers (26, 43, 87). Also hypermutated are GI cancers with dMMR resulting in MSI-H. MSI-H have a significantly higher median TMB level compared with MSS tumors (25–27), yet lower than those with POLE mutation. Analysis of MSI-H cancers by WES showed a mean of 1,782 somatic mutations per tumor compared with 73 mutations per tumor in patients with MSS cancers (P = 0.007; ref. 5).

To date, results for TMB and outcome from ICIs in GI cancers is limited, with the exception of the subset of tumors with MSI-H, especially colorectal cancer. Among MSI-H cancers, high ORRs and prolonged PFS have been observed and are discussed below under the heading “TMB and Tumor MSI Status.” To date, only limited data exist among MSS GI cancers treated with ICIs (5, 21, 26, 88, 89). In a study of 54 advanced gastric cancers, median TMB for MSS cases was 6.6 mut/Mb (range 0–30.0) and there was one MSI-H tumor. Evaluation of the predictive utility of TMB for the anti–PD-1 antibody toripalimab revealed that patients with TMB-high (≥12 mut/Mb) versus TMB-low tumors had better ORR (33.3% vs. 7.1%, P = 0.017), similar PFS (2.5 vs. 1.9 months, P = 0.055), but significantly improved OS (14.6 vs. 4.0 months, P = 0.038; ref. 21). Among 17 patients with MSS advanced hepatocellular carcinoma treated with an anti–PD-1 antibody, 1 patient (TMB 15 mut/Mb) had a sustained complete response to nivolumab lasting >2 years. However, TMB did not segregate patients by response criteria (89). Future studies are needed to evaluate ICI efficacy in patients with GI cancer with MSS and high TMB, which will also enable comparison of results with MSI-H cancers.

Evidence indicates that breast cancers typically have a lower TMB compared with NSCLC and melanoma, although TMB has been shown to vary both within and across breast cancer subtypes (90–92). Using WES and panel-based sequencing data from 3,969 patients from six cohorts, median TMB was 2.63 mut/Mb with 5% of patients showing high TMB (≥10 mut/Mb). Among breast cancer subtypes, median TMB was significantly higher in triple-negative breast cancer (TNBC; 1.8 mut/Mb) compared with hormone receptor–positive (1.1 mut/Mb, P < 0.001) or HER2-positive cancers (1.3 mut/Mb, P = 0.003; ref. 91). Although these differences in TMB were statistically significant, they are within the error range of large gene panels.

To date, only limited data are available for the evaluation of TMB as a predictive biomarker in patients with breast carcinoma treated with ICIs. In a single-center cohort of patients with metastatic TNBC (N = 62) treated with an anti–PD-1/PD-L1 antibody, patients whose tumors had high TMB (≥10 mut/Mb) had a 2-fold increase in likelihood of response compared with those with lower TMB (ref. 93; Table 2). In a prospective clinical trial known as TAPUR, patients with heavily pretreated metastatic breast cancer were treated with pembrolizumab, which was associated with a modest benefit in patients with high TMB (≥9 mut/Mb) tumors (ref. 94; Table 2). On the basis of available data, TMB appears to be less relevant to ICI treatment outcome in breast cancer given that only 5% of these tumors have high TMB (≥10 mut/Mb). This TMB cutoff was associated with pembrolizumab benefit in a recent pan-cancer study (KEYNOTE-158) that led to FDA approval of this antibody for this high-TMB subgroup.

Studies have shown that TMB is to a large extent independent of PD-L1 status in most cancers (34) and might therefore identify additional subgroups of patients who may benefit from ICIs (46, 59, 65, 77). Among 24 cancer types included in the TCGA database, TMB and PD-L1 mRNA expression were examined and significant correlations were observed in only 5 of 24 cancer types (Fig. 2C). These comprised adenocarcinoma of the colorectum, stomach, and uterus (R = 0.26, R = 0.22, and R = 0.19) as well as transitional cell carcinoma (R = 0.19) and breast cancer (R = 0.14). Across cancer types, we observed a weak but highly significant positive correlation (R = 0.13). In a retrospective analysis of 11,348 patients across 26 cancer types, 7.7% of tumors were found to be TMB-high (≥17 mut/Mb), and of these 44% were also PD-L1 immunopositive (≥1%; PD-L1 antibody: SP142; ref. 38). Among 4,125 GI cancers examined, squamous cell carcinomas of the esophagus and anus showed high PD-L1 expression yet low or negative TMB or infrequent MSI-H. Other tumor types such as right-sided colon cancers and small-bowel adenocarcinomas showed high TMB or MSI-H, yet low PD-L1 expression (85).

Multiple studies have analyzed PD-L1 expression as a potential predictive biomarker for response to ICIs targeting PD-1 or its ligand, PD-L1. Determination of PD-L1 expression is approved by the FDA as a companion diagnostic test for pembrolizumab treatment in patients with NSCLC, gastroesophageal junction and gastric cancer, cervical cancer, and urothelial cancer (28, 29, 95–98). However, this is not the case in other tumor types (99, 100), and therefore the predictive utility of PD-L1 remains limited (101). Whereas TMB and MSI-H describe features of the tumor, PD-L1 expression depends on the specific cell types examined and the score that is applied (tumor only, immune cell only, or their combinations), while quantification of lymphocyte densities or specific immune signatures, for example, IFNγ or T-cell clonality, highlight a specific state of the local immune environment and effector compartment. Both quantitative and qualitative measurements of the tumor compartment and the effector compartment are important for understanding outcomes of immunotherapy, and information on both compartments is likely needed for arriving at meaningful conclusions in a clinical setting.

Evidence suggests that complementary utilization of both TMB and PD-L1 may predict responsiveness to ICIs better than either alone (see Supplementary Data; refs. 46, 61). In patients with NSCLC, the CheckMate 568 study demonstrated superior ORR for nivolumab plus ipilimumab independent of PD-L1 status in patients with NSCLC with TMB-high (≥10 mut/Mb) versus TMB-low tumors (ORR: PD-L1 ≥ 1%, 48% vs. 18%; PD-L1 < 1%, 47% vs. 5%). Importantly, patients with tumors that had PD-L1 expression < 1% and TMB < 10 mut/Mb had an ORR of only 5%, suggesting that using the combination of PD-L1 and TMB testing can identify a subgroup of patients who have a low likelihood of benefit from the combination of PD-1 and CTLA4 checkpoint blockade (61).

MSI-H tumors show hypermutation including frameshift mutations that generate numerous neopeptides (71). Tumors with MSI-H due to dMMR typically have high TMB levels. WES revealed a mean of 1,782 somatic mutations per MSI-H tumor as compared with 73 in MSS tumors (P = 0.007), suggesting that a markedly increased number of mutation-associated neoantigens is responsible for enhanced anti–PD-1 response (5). Among 4,125 GI cancers of 14 different types, TMB-high (≥17 mut/Mb; 592 gene panel) was strongly correlated with MSI-H status in most cancer types, indicating that high TMB and MSI-H are inextricably linked. Exceptions among TMB-high tumors included squamous cell carcinomas of the anus and esophagus that were generally MSS (85). A majority of MSI-H/dMMR cancers are colorectal cancers or endometrial cancers (102–105).

MSI-H/dMMR has been shown to be a predictive biomarker for treatment with ICIs (106). In this regard, pembrolizumab was approved by the FDA for the treatment of MSI-H cancers agnostic of primary tumor type. Pembrolizumab is standard of care for treatment-refractory, MSI-H solid tumors, and more recently pembrolizumab was shown to be superior to and less toxic than chemotherapy in MSI-H colorectal cancer in the first-line setting (107). Among 124 patients with MSI-H/dMMR colorectal cancer treated with pembrolizumab (KEYNOTE-164), the ORR was 33% overall and median PFS was 2.3 months (≥2 prior lines of therapy) or 4.1 months (≥1 prior therapy; ref. 108). In another phase II study, nivolumab (anti–PD-1) provided a response rate of 31% and a 12-month OS rate of 73% in heavily pretreated patients with metastatic MSI-H/dMMR colorectal cancer, while its combination with ipilimumab demonstrated a response rate of 55% and 12-month OS rate of 85% (8). In a separate study, 22 patients with MSI-H metastatic colorectal cancers were treated with an anti–PD-1 or PD-L1 antibody. Median TMB level in responders was significantly greater than in nonresponders (54 mut/Mb vs. 29 mut/Mb, P < 0.001). Of the 13 patients with TMB-high tumors, defined as ≥ 37–41 mut/Mb, an objective response was observed in all, whereas only 3 of 9 (33%) patients with low-TMB tumors had disease control (ref. 25; Table 2). These data suggest that high TMB in MSI-H/dMMR tumors is associated with increased and durable responses to ICIs (5, 109, 110), and that TMB may further identify responders to ICIs within MSI-H cancers (25). However, TMB can occur due to multiple and distinct mutational processes, and the relative contribution of mutation load versus mutational process has yet to be clarified. In addition to variability in TMB observed in MSI-H cancers (25), heterogeneity in densities of tumor-infiltrating lymphocytes (TIL) are also found in this tumor subgroup that can prognostically stratify these tumors (111, 112). Further study is needed to examine the relationships of TMB with neoantigen load and TIL density.

Evidence indicates that a subset of MSS cancers have high TMB (85, 88, 113). In the study of 4,125 patients with GI cancers, anal squamous cell carcinomas exhibited the highest prevalence (8.3%) of MSS/TMB-high (≥17 mut/Mb) followed by esophageal squamous cancer (3.5%; ref. 85). In a study of MSS cancers of 14 different histologies, treatment with an ICI produced longer median PFS (26.8 vs. 4.3 mo., P = 0.0173) in TMB-high (≥20 mut/Mb) versus TMB-low tumors (113). In other studies, the prevalence of TMB-high (≥11.7 mut/Mb) in MSS colorectal cancers was 2.9% (164/5,702), (26) and in a study of multiple cancer types, TMB-high (≥20 mut/Mb) was identified in 5.4% (7,972/148,803) of cancers (113). Importantly, the prevalence of MSI-H was only 1.5% (2,179/148,803) in this study, so it was significantly exceeded by the TMB-high subgroup (113). These data suggest that mechanisms besides DNA repair defects, such as DNA replication mutations (POLD1 and POLE) or TP53 mutations (35, 114, 115), may underlie their increased TMB. Furthermore, evidence suggests that hypermutation with high mutation-specific neoantigenic load is a critical factor responsible for antitumor efficacy of ICIs because both MSI-H and MSS POLE tumors show relatively high response rates to ICIs. Of clinical relevance is that patient selection for ICIs based on TMB status may potentially expand the candidate pool for cancer immunotherapy. In this regard, data from the pan-cancer cohort of predominantly MSS solid tumors (KEYNOTE-158) found an ORR of 29% for patients with MSS tumors and TMB ≥ 10 mut/Mb (24).

Alterations in DNA damage response and repair (DDR) genes are associated with genomic instability and increased somatic tumor mutational burden, which may enhance immunogenicity through increased tumor-specific neoantigen load (116, 117). The relationship between TMB and DDR genes has been explored in various cancer types including NSCLC, urothelial cancer, and GI cancers (81, 118–120). Recent evidence revealed deleterious somatic DDR mutations in approximately 50% of patients with NSCLC or urothelial carcinomas (81, 120). Patients with DDR mutations had significantly increased tumor TMB levels (81, 120) and longer PFS and OS independent of covariates (120). The prevalence of DDR alterations was 17% among 17,486 GI carcinomas (119), of which ARID1A (9.2%) and ATM (4.7%) were most common followed by BRCA2 (2.3%), BRCA1 (1.1%), and CHEK2 (1.0%). DDR mutations were associated with increased TMB, and of DDR-altered/TMB-high cases, 87% were also MSI-H. Of note, MSI-H and high TMB (≥20 mut/Mb) were found in 19% and 21% of DDR-mutated cases, respectively. Even among MSS tumors, TMB-high was significantly more frequent in DDR-mutated versus nonmutated cases (119). An important caveat is that tumors with high TMB and/or MSI-H are more likely to harbor DDR mutations, which suggest the potential for confounding.

Observed cancer type–related biological TMB distributions offer the potential to determine tumor-specific and optimal TMB cutoffs. Various strategies to optimize TMB as a predictive biomarker for ICIs are being explored (Table 3). A novel three-tier (high, intermediate, and low) TMB classification scheme was introduced to reduce the possibility of misclassification by the current two-tier (high, low) TMB classification scheme. A so-called “gray zone” for TMB was identified to potentially aid future panel designs, trial design, and clinical decision-making (45). Transforming unadjusted TMB values into standardized z-scores (converts the right-skewed TMB distributions to normal distributions) has been proposed to standardize and compare TMB across panels from different platforms (121). However, this would require similar TMB mean and standard deviation from datasets of comparable cohorts. Other strategies include screening for actionable mutations or biomarkers, refining immunotherapy response prediction (such as negative predictors of response and variants predisposing to toxic effects), aligning panel-based TMB values to a WES-based TMB reference to ensure consistency across assays, standardizing bioinformatic algorithms used for mutation calling and filtering, using variant allele frequency as a proxy for clonality to further refine TMB quantification, and allowing calibration of results from different studies. Friends of Cancer Research and Quality Assurance Initiative Pathology are two international organizations that have proposed approaches to standardize and harmonize TMB assessment across assays and centers globally (54, 55, 122, 123). The Friends of Cancer Research TMB Harmonization Consortium made the following recommendations: (i) TMB should be reported in mutations/megabase (mut/Mb) to keep these values consistent and comparable among different studies; (ii) validation studies for TMB estimation should be standardized to include assessment of analytic accuracy, precision, and sensitivity; and (iii) consistency across panels should be ensured through alignment of panel TMB values to WES-derived universal reference standards (55).

Determination of TMB in commercially available gene panels relies mainly on missense mutations. Although gene-panel sequencing platforms generally detect and report indels (Table 1), indel calling can vary based on the bioinformatics pipeline (124). Data in colorectal cancer and melanoma suggest that frameshift indels generate a higher number of immunogenic neoantigens than do nonsynonymous SNV mutations (47, 125, 126), a suggestion that awaits confirmation. Indel load was strongly associated with ICI response, although controversy exists (126, 127). Furthermore, the proportion of indels in conjunction with TMB values can identify different tumor types and genetic subgroups, including MSI-H cases (ref. 127; Table 3). Evidence suggests that underlying biology is likely driving and shaping TMB with prime examples of hypermutation being due to MSI-H (due to defective DNA mismatch repair) or MSS POLE (exonuclease domain mutations). Both MSI-H and POLE tumors show relatively high response rates to ICIs, suggesting that TMB is a proxy or parameter governed by tumor biology. The relative contribution of mutation load versus mutational process to TMB has yet to be clarified. Considering mutational signatures in the context of TMB might help to decipher the biology that creates TMB. Furthermore, high TMB in MSI-H tumors can be identified and was associated with increased and durable responses to ICIs (5, 109, 110), suggesting that stratification of MSI-H tumors using TMB may distinguish responders from nonresponders to ICIs (25). Data suggests that complementary utilization of both TMB and PD-L1 may predict responsiveness to ICIs better than either alone (46, 61), and further evaluation of this approach in selected tumor types is warranted. Other studies indicate that tumors that have high levels of TMB and inflammatory markers [T cell–inflamed gene-expression profile (GEP) or PD-L1] represent a population with the highest likelihood of response to ICIs (128, 129). In this regard, TMB and GEP exhibited joint predictive utility in identifying responders and nonresponders to pembrolizumab. TMB and GEP were independently predictive of response and demonstrated a low correlation, suggesting that they capture distinct features of neoantigenicity and T-cell activation (128, 129). Pretreatment tumor burden may also influence the efficacy of ICIs. A hypothesis has been proposed that evaluates the ratio of TMB to tumor burden that could provide a more effective prediction of ICI efficacy, and warrants testing in clinical trials (130).

Evidence indicates that relative TMB can be determined through sequencing analysis of cell-free DNA (cfDNA). Measurement of TMB in blood plasma has been developed and, in limited studies, has been shown to provide predictive information for immunotherapy response in patients with NSCLC (131, 132). The number of mutations detected in cfDNA was positively correlated with ICI efficacy and OS across various cancer types (n = 69; ref. 133). Furthermore, retrospective analysis of cfDNA from two randomized trials (n = 211 and 583) demonstrated that cfDNA-derived TMB is associated with improved survival in patients with NSCLC treated with an anti–PD-L1 antibody (132). Although a blood-based assay has clear advantages for clinical application, further development and evaluation of this assay technology, in addition to its comparison with tissue-based approaches, is awaited.

Accumulating evidence suggests that TMB may serve as a predictive biomarker for immunotherapy in multiple solid tumors, although further prospective validation is needed. Use of TMB is now a component of routine oncologic practice based on recent FDA approval of pembrolizumab for TMB-high solid tumors. The TMB-high subgroup was identified using a prespecified cutoff of at least 10 mut/MB, and issues remain regarding optimal cutoffs per tumor type based, in part, on baseline TMB distributions. Furthermore, prospective studies are needed for further evaluation and validation of the predictive utility of TMB in clinical practice. TMB has been shown to be a predictive biomarker for ICIs in both MSI-H and MSS cancers, and patient selection based on TMB levels may better select patients for ICIs or potentially expand the candidate pool for immunotherapy. Although current evidence indicates that TMB is associated with ICI efficacy, the mechanisms underlying the association between TMB and benefit from immunotherapy are incompletely understood. TMB is independent of PD-L1 status in most cancer types, and the complementary utilization of TMB, PD-L1, and MSI-H has the potential to predict ICI responsiveness better than each alone. As TMB is a novel biomarker, there is an urgent need to harmonize and standardize TMB measurement, testing platforms, and reporting of TMB. The Friends of Cancer Research TMB Harmonization Consortium is working to establish guidelines to harmonize TMB across diagnostic platforms, and results of this effort are eagerly awaited. Larger datasets for TMB and clinical outcome of ICI-treated patients will facilitate the optimization of TMB cutoffs within specific cancer types and potentially extend the approval of immune therapies to larger patient populations. Furthermore, prospective and randomized studies are needed to validate a TMB-high cutoff and to explore optimal TMB cutoffs in specific tumor types. TMB combined with other potential biomarkers and computational assistance is paving the way toward a precision immunotherapy approach.

J. Budzcies reports grants from German Cancer Aid outside the submitted work. A. Stenzinger reports personal fees from AstraZeneca (advisory board), Eli Lilly (advisory board), Illumina (advisory board), Janssen (advisory board), MSD, Novartis (advisory board), Pfizer, Roche, Seattle Genetics, Thermo Fisher Scientific (advisory board); grants and personal fees from Bayer (advisory board), BMS (advisory board); and grants from Chugai during the conduct of the study. No potential conflicts of interest were disclosed by the other authors.

The study was supported, in part, by an NCI grant (grant number R01 CA210509-01A1; to F.A. Sinicrope). D. Sha was supported by the China Scholarship Council and Shandong Provincial Hospital Affiliated to Shandong First Medical University (Jinan, P.R. China).