Skip to main content

04-02-2019 | Non-small cell lung cancer | Article

Comparing and contrasting predictive biomarkers for immunotherapy and targeted therapy of NSCLC

Nature Reviews Clinical Oncology

Authors: D. Ross Camidge, Robert C. Doebele, Keith M. Kerr

Publisher: Nature Publishing Group UK


The era of personalized medicine for advanced-stage non-small-cell lung cancer (NSCLC) began when biomarker-based evidence of molecular pathway and/or oncogene addiction of the tumour became mandatory for the allocation of specific targeted therapies. More recently, the immunotherapy revolution, specifically, the development of immune-checkpoint inhibitors (ICIs), has dramatically altered the NSCLC treatment landscape. Herein, we compare and contrast the clinical development of immunotherapy and oncogene-directed therapy for NSCLC, focusing on the role of predictive biomarkers. Immunotherapy biomarkers are fundamentally different from oncogene biomarkers in that they are continuous rather than categorical (binary), spatially and temporally variable and reliant on multiple complex interactions rather than a single, dominant determinant. The performance of predictive biomarkers for ICIs might be improved by combining different markers to reduce the assumptive risks associated with each one. Novel combinations with chemotherapy and ICIs complicate biomarker discovery but do not decrease the value of the markers identified. Perfectly predictive biomarkers of benefit from immunotherapy are unlikely to be identified, although exclusionary biomarkers of minimal benefit or an unacceptable risk of toxicity might be feasible. The clinical adoption and applicability of such biomarkers might vary depending on line of treatment, the available therapeutic alternatives and health economic considerations.


The personalized medicine revolution in the treatment of advanced-stage non-small-cell lung cancer (NSCLC) happened when clinical decision-making regarding the use of specific targeted therapies, predominantly tyrosine kinase inhibitors (TKIs), became dependent upon evidencing addiction of the tumour to a given molecular pathway and/or oncogene using predictive biomarkers17. Although currently applicable to only a minority of patients with NSCLC, this paradigm has resulted in prominent successes that have led to approved molecularly specific, biomarker-defined indications for targeted therapies. These indications include the use of selective TKIs for the treatment of EGFR-mutant, ALK-rearranged (ALK+), ROS1-rearranged or BRAFV600E-mutant advanced-stage NSCLC17. More recently, the immunotherapy revolution, specifically, the development of anti-programmed cell death 1 (PD-1) and anti-programmed cell death 1 ligand 1 (PD-L1) antibodies as immune-checkpoint inhibitors (ICIs), has also dramatically altered the treatment landscape of NSCLC814. Herein, we compare and contrast the past, current and possible future clinical development of immunotherapy and oncogene-directed targeted therapy for NSCLC, focusing on the role of clinically applicable predictive biomarkers.

Predictive biomarkers for TKI therapy

Erlotinib was the first EGFR TKI to gain full approval from the FDA, initially for unselected patients with advanced-stage NSCLC after progression on chemotherapy15. This approval was based on a significant overall survival (OS) improvement observed in the BR.21 trial16, involving 731 patients who had received one or two lines of chemotherapy and were ineligible for further chemotherapy (median 6.7 months with erlotinib versus 4.7 months with placebo; HR 0.70; P < 0.001). The objective response rate (ORR) and progression-free survival (PFS) were also improved with erlotinib16.
Eventually, findings from multiple phase III trials of erlotinib and other EGFR TKIs demonstrated superior ORRs and PFS compared with those observed with standard first-line platinum-based chemotherapy but only among patients with NSCLC harbouring an activating somatic mutation in EGFR13 (Table 1). However, several years of study were required after the first approval of erlotinib in 2004 for the treatment of otherwise unselected patients with NSCLC to unequivocally attribute the clinical benefit solely to this EGFR-mutant subgroup. Clinical features associated with therapeutic benefit were initially identified — notably, female sex, adenocarcinoma histology, Asian ethnicity and a limited or no history of smoking1621. Different putative predictive molecular markers were also explored with both inclusionary (EGFR expression by immunohistochemistry (IHC) and increased EGFR copy number by fluorescence in situ hybridization (FISH)) and exclusionary approaches (absence of a KRAS mutation)1721. In retrospect, the promise associated with each of these proposed biomarkers can be explained by their underlying ability to variably enrich for EGFR mutations in the selected population (Fig. 1).
Table 1
Efficacy outcomes of key trials of molecularly targeted therapy or PD-1 and/or PD-L1 inhibitor monotherapy in NSCLC
NSCLC population
Study interventions
ORR (%)
Median DoR
Median PFS (months) (HR, 95% CI)
Median OS (months) (HR, 95% CI)
ALK-targeted therapy
Second-line crizotinib versus docetaxel or pemetrexed chemotherapy
65 versus 20
32.1 weeks versus 24.4 weeks
7.7 versus 3.0 (0.49, 0.37–0.64; P <  0.001)
20.3 versus 22.8 (1.02, 0.68–1.54; P = 0.54)
First-line crizotinib versus platinum-based and pemetrexed chemotherapy
74 versus 45
11.3 months versus 5.3 months
10.9 versus 7.0 (0.45, 0.35–0.60; P <  0.001)
Not reached (0.82, 0.54–1.26; P = 0.36)
EGFR-targeted therapy
LUX-Lung 3
First-line afatinib versus platinum-based and pemetrexed chemotherapy
56 versus 23
11.1 months versus 5.5 months
11.1 versus 6.9 (0.58, 0.43–0.78; P = 0.001)
Not reached (1.12, 0.73–1.73; P = 0.60)
First-line erlotinib versus platinum-based and gemcitabine chemotherapy
62.7 versus 33.6
11.0 versus 5.5 (0.34, 0.22–0.51; P <  0.0001)
26.3 versus 25.5 (0.91, 0.63–1.31; P = 0.607)
First-line gefitinib versus platinum-based and docetaxel chemotherapy
62.1 versus 32.2
9.2 versus 6.3 (0.489, 0.336–0.710; P <  0.0001)
30.1 versus not reached (1.64, 0.75–3.58; P = 0.211)
PD-1 or PD-L1 inhibitor monotherapy
PD-L1+ or PD-L1
Pembrolizumab 2 or 10 mg/kg (in multiple lines of treatment)
12.5 monthsa
3.7 (NA)a
12 (NA)a
PD-L1+ (≥1% expression)
Second-line pembrolizumab (2 mg/kg and 10 mg/kg groups) versus docetaxel
18 and 18 versus 9
Not reached and not reached versus 6 months
3.9 (0.88, 0.74–1.05; P = 0.07) and 4.0 (0.79, 0.66–0.94; P = 0.004) versus 4.0
10.4 (0.71, 0.58–0.88; P = 0.0008) and 12.7 (0.61, 0.49–0.75; P <  0.0001) versus 8.5
PD-L1+ (≥50% expression); EGFR-wild-type and ALK-wild-type
First-line pembrolizumab (200 mg) versus platinum-doublet chemotherapy, plus pemetrexed maintenance as appropriate
44.8 versus 27.8
Not reached versus 6.3 months
10.3 versus 6.0 (0.50, 0.37–0.68; P <  0.001)
Not reached (0.60, 0.41–0.89; P = 0.005)
PD-L1+ (≥1% expression); EGFR-wild-type and ALK-wild-type
First-line pembrolizumab (200 mg) versus platinum-doublet chemotherapy, plus pemetrexed maintenance as appropriate
27.3 versus 26.5
20.2 months versus 8.3 months
5.4 versus 6.5 (1.07, 0.94–1.21)
16.7 versus 12.1 (0.81, 0.71–0.93; P = 0.0018)
Previously treated NSCLC
Nivolumab 1, 3 or 10 mg/kg (in multiple lines of treatment)
17 monthsa
2.3 (NA)a
9.9 (NA)a
CheckMate 017
Squamous NSCLC
Second-line nivolumab (3 mg/kg) versus docetaxel
20 versus 9
Not reached versus 8.4 months
3.5 versus 2.8 (0.62, 0.47–0.81; P <  0.001)
9.2 versus 6.0 (0.59, 0.44–0.79; P <  0.001)
CheckMate 057
Non-squamous NSCLC
Second-line nivolumab (3 mg/kg) versus docetaxel
19 versus 12
17.2 months versus 5.6 months
2.3 versus 4.2 (0.92, 0.77–1.11; P = 0.39)
12.2 versus 9.4 (0.73, 0.59–0.89; P = 0.002)
CheckMate 026
PD-L1+ (≥5% expressionb); EGFR-wild-type and ALK-wild-type
First-line nivolumab (3 mg/kg) versus platinum-doublet chemotherapy, plus pemetrexed maintenance as appropriate
26 versus 33
12.1 months versus 5.7 months
4.2 versus 5.9 (1.15, 0.91–1.45; P = 0.25)
14.4 versus 13.2 (1.02, 0.80–1.30)
PD-L1+ or PD-L1 (on tumour cells and/or tumour-infiltrating immune cells)
Second-line or third-line atezolizumab versus docetaxel
15 versus 15
14.3 months versus 7.2 months
2.7 versus 3.0 (0.94, 0.72–1.23)
12.6 versus 9.7 (0.73, 0.53–0.99; P = 0.04)
PD-L1+ or PD-L1 (on tumour cells and/or tumour-infiltrating immune cells)
Second-line or third-line atezolizumab versus docetaxel
14 versus 13
16.3 months versus 6.2 months
2.8 versus 4.0 (0.95, 0.82–1.10)
13.8 versus 9.6 (0.73, 0.62–0.87; P = 0.0003)
JAVELIN Lung 200
PD-L1+, EGFR-wild-type and ALK-wild-type, non-squamous diseasec
Avelumab versus docetaxel (after prior treatment with platinum-doublet chemotherapy)
19 versus 12
Not reached versus 6.9 months
3.4 versus 4.1 (1.01, 96% CI 0.80–1.28; P = 0.53)
11.4 versus 10.3 (0.90, 96% CI 0.72–1.12; P = 0.16)
DoR, duration of response; NA, not available; NR, not reported; NSCLC, non-small-cell lung cancer; ORR, overall response rate; OS, overall survival; PD-1, programmed cell death 1; PD-L1, programmed cell death 1 ligand 1; PFS, progression-free survival. aOutcome reported for entire patient population, across all doses, PD-L1 expression levels and lines of treatment. bPatients with PD-L1 expression ≥1% were enrolled, but the primary PFS end point and other CheckMate 026 outcomes reported in this table were focused on the patients with PD-L1 expression ≥5%. cThe outcomes of JAVELIN Lung 200 reported in this table are from the primary analyses performed in the PD-L1+ population, but patients with PD-L1 NSCLC were also included in this study.
In 2013, both erlotinib and afatinib were approved by the FDA for the frontline treatment of advanced-stage NSCLC specifically harbouring EGFR exon 19 deletions or exon 21 (L858R) mutations. In addition, consistent with the wholesale adoption of a personalized medicine approach, the indications for EGFR TKI use in those without a known EGFR mutation began being removed from the National Comprehensive Cancer Network (NCCN) guidelines in 2015 (ref.22), and from October 2016 the FDA-approved indications for erlotinib were restricted to those with EGFR exon 19 deletions or L858R mutation23. By contrast, subsequent oncogene-directed targeted therapies have been developed in molecularly defined subgroups prospectively, before their entry into routine clinical practice. Preclinical data have been used to define specific subgroups to explore from the start of the clinical drug development process, leading to several rapid, novel, biomarker-based drug approvals for patients with NSCLC (including those with ALK+, ROS1-rearranged or BRAFV600E-mutant disease)47. Moreover, the detection of the EGFRT790M resistance mutation, which is selected for in EGFR-mutant NSCLC by first-generation and second-generation EGFR TKIs, forms the basis for the initial licensing of osimertinib, a third-generation EGFR TKI with activity against EGFR harbouring an original activating mutation and/or the T790M mutation24. The NCCN also lists many other desirable molecular changes (specific mutations, gene rearrangements and gene amplifications) to test for in NSCLC specimens, including in MET, RET, NTRK and HER2, on the basis of preliminary evidence of clinical efficacy for different targeted agents in molecularly defined patient subgroups, most of which predominantly occur in lung adenocarcinomas22.

Predictive biomarkers for immunotherapy

FDA licences for nivolumab and pembrolizumab (anti-PD-1 antibodies) and for atezolizumab (an anti-PD-L1 antibody) in the second-line treatment of advanced-stage NSCLC have been granted on the basis of improvements in OS versus that observed with docetaxel812 (Table 1). PD-L1 expression on tumour cells and/or infiltrating immune cells has been quantified retrospectively using a range of IHC-based assays in most of the pivotal studies of these agents. Only the trial of pembrolizumab10 required a PD-L1 tumour proportional score (TPS) ≥1%, defined using the 22C3 PD-L1 IHC assay, for eligibility, and thus this requirement is part of the FDA approval. The trials of the other agents demonstrated superior outcomes in unselected patient populations; therefore, specific PD-L1 levels are not mandated in their second-line licences.
Avelumab (another anti-PD-L1 antibody) has also been compared with docetaxel in the JAVELIN Lung 200 trial, but the primary OS end point was not met (HR 0.90, 96% CI 0.72–1.12; P = 0.16)25. In analyses of PD-L1 subgroups defined using the 73–10 PD-L1 IHC assay, however, improvements in OS were reported for the 40% of patients with PD-L1 levels ≥50% (HR 0.67, 95% CI 0.51–0.89; P = 0.0052) and the 30% of patients with ≥80% PD-L1 expression (HR 0.59, 95% CI 0.42–0.83; P = 0.0022)25.
In the first-line setting, nivolumab and pembrolizumab have been compared with first-line platinum-based doublet chemotherapy in EGFR-wild-type and ALK-wild-type and variably PD-L1-enriched patient populations (Table 1). In the KEYNOTE-024 trial13 involving patients with a PD-L1 TPS ≥50% (~30% of population evaluable for PD-L1 expression), pembrolizumab significantly improved PFS (the primary end point), as well as the ORR and OS. By contrast, nivolumab did not improve PFS in patients with PD-L1 levels ≥5% (as defined using the 28–8 PD-L1 IHC assay) in CheckMate 026 (ref.14).
In the PACIFIC trial26 involving otherwise unselected patients with stage III NSCLC that had not progressed within 1–42 days after completion of chemoradiotherapy, treatment with the anti-PD-L1 antibody durvalumab for up to 12 months significantly prolonged PFS (median 16.8 months versus 5.6 months with placebo; HR 0.52; P <  0.0001) and time to distant metastasis or death (HR 0.52; P <  0.0001). These data led to FDA approval of durvalumab in this setting, irrespective of PD-L1 expression, in February 2018.

Comparing and contrasting biomarkers

Two main differences exist between most of the predictive biomarkers used in selecting patients with oncogene-addicted subtypes of NSCLC for TKI treatments and the markers used to enrich for patients who derive clinical benefit from PD-1 or PD-L1 inhibitors: first, the nature of the markers themselves — categorical (binary) versus continuous and constitutive versus induced (Fig. 2); and, second, the number and riskiness of assumptions made in predicting subsequent drug sensitivity in the biomarker-selected population (Fig. 3).

Driver oncogenes versus PD-L1 expression

For all the oncogene-addicted NSCLC subtypes with FDA-approved targeted therapies, the underlying abnormality is a genetic alteration. Each mutation or gene rearrangement test — or in the case of ALK rearrangements, an alternative ALK IHC assay — requires some degree of thresholding to accurately distinguish signal from noise17,27. However, beyond this threshold, the results are categorical and binary: the abnormality is either present or absent (Fig. 1). In addition, because the alteration being assessed is considered a driver abnormality, all cells within the tumour are assumed to harbour the genetic aberration regardless of their location — that is, the aberration is ‘truncal’28. Many relevant driver oncogenes can also now be detected in circulating cell-free tumour DNA29, including an FDA-approved plasma assay for EGFR mutations30.
Active targeted therapy can exert selective pressure that promotes the outgrowth of subclones harbouring molecular variants mediating drug resistance31. By contrast, treatments that are not focused on the specific oncogenic pathway, such as chemotherapy and radiotherapy, are not expected to select for changes directly associated with alterations in, or bypass tracks around, these driver abnormalities. Consistent with this hypothesis, the efficacy of oncogene-directed therapies targeting EGFR or ALK is highly comparable regardless of prior chemotherapy2,4,32.
When a relevant genetic alteration is present, clinical benefit from the corresponding targeted therapy is highly likely because the assumption that the detected genetic change will have led to an oncogene-addicted state in the tumour is very likely to be true (Fig. 3a). Some co-mutations, such as TP53 mutations, have been associated with decreased survival of patients with NSCLC harbouring targetable ALK, ROS1 or EGFR abnormalities, although the benefits of targeted therapy, including objective responses and prolonged PFS, even among patients with such co-mutation, are still notable33.
Rare exceptions to this categorical picture of oncogene-addicted NSCLC biomarkers do exist. For example, challenges remain in accurately defining a form of MET gene amplification (which is a continuous variable, measurable in multiple different ways) that reflects true dependency on MET and therefore susceptibly to MET TKI therapy3436.
By contrast, many high-risk assumptions are inherent in any potential biomarker predictive of benefit from anti-PD-1 or anti-PD-L1 therapy in patients with NSCLC (Fig. 3b). For PD-1–PD-L1 inhibition to be effective, genetic changes resulting in altered amino acid sequences (non-synonymous mutations) are required. Additionally, the mutated peptides must be presentable to the immune system as neoantigens in the context of the MHC proteins of a cell. MHC class I (MHCI) complexes are most commonly implicated in anticancer immunity3740; however, MHCII complexes can also fulfil this role in some cancers. For example, patients with classical Hodgkin lymphoma, in which β2-microglobulin is often inactivated, leading to absent cell surface expression of MHCI, have high response rates to PD-1 inhibition, and MHCII expression is associated with clinical benefit41. For a response to PD-1–PD-L1 inhibition, a broadly pro-inflammatory, permissive immune environment must also be present in the tumour — for example, with all other relevant immune checkpoints appropriately stimulated or inhibited42. Finally, if all these assumptions are met, the tumour cells must then be exploiting the PD-1 axis as their dominant mechanism of immune evasion.
Expression of the drug target, assessed through IHC for PD-L1, has been the major initial molecular determinant of clinical benefit explored for immunotherapy; however, responses can occur even in tumours with no PD-L1 staining and, conversely, irresponsiveness and overt disease progression can be observed even in those with high tumoural PD-L1 levels43 (Table 2). PD-L1 expression apparently not being necessary for a response to PD-1–PD-L1 inhibition could reflect three main issues: the biology of PD-L1 expression, the sensitivity of the assays being used and the potential for an effect of the agent on another relevant therapeutic pathway (for example, the potential for PD-1 inhibition to also be active against PD-L2-mediated immunosuppressive signalling).
Table 2
Potential explanations why PD-L1 expression might not predict benefit from PD-1 or PD-L1 inhibition
Potential explanations
PD-L1 expression apparently not necessary
PD-L1 absent by IHC but clinical benefit seen from inhibition of PD-1 or PD-L1
• Spatial and/or temporal variability in PD-L1 expression within tumour (sampling error)
• Incomplete sensitivity of IHC in the detection of PD-L1, with variation between assays (false-negative result)
• PD-L2, the alternative ligand for PD-1, could provide a bypass mechanism for immunosuppression, leading to responses of PD-L1 tumours to anti-PD-1 antibodies, although in theory, not to anti-PD-L1 antibodies
PD-L1 expression apparently not sufficient
PD-L1 present by IHC but no clinical benefit from inhibition of PD-1 or PD-L1
• Elevation in PD-L1 expression for reasons other than in response to a primed immune attack (for example, intrinsic induction in some oncogene-addicted NSCLCs)
• Engagement of other immune checkpoints in addition to the PD-1–PD-L1 axis and/or immune suppression or deficiencies with different causes
• The measured extent of PD-L1 positivity (a continuous variable) might be insufficient for a response to PD-1 or PD-L1 inhibition, reflecting substantial heterogeneity in the underlying tumour biology (including neoantigen profiles and mechanisms of immune escape)
IHC, immunohistochemistry; NSCLC, non-small-cell lung cancer; PD-1, programmed cell death 1; PD-L, programmed cell death 1 ligand.
With regard to the PD-L1 biology, expression of this protein on tumour cells is considered to be induced in response to IFNγ released by activated T cells and, consequently, is known to be both spatially and temporally variable44. As such, the absence of PD-L1 staining by IHC at a given time does not preclude expression of this protein either contemporaneously in another section of the same lesion or at a later time point, for example, following treatment with chemotherapy or radiotherapy4448.
Considering the issue of analytical sensitivity and specificity, concordance studies suggest that the different PD-L1 IHC assays (comprising both the anti-PD-L1 antibodies and their associated detection system) associated with nivolumab (28–8), pembrolizumab (22C3) and durvalumab (SP263) produce comparable results4448. By contrast, the system used for atezolizumab (SP142) consistently stains fewer tumour cells than these assays, whereas the IHC test used with avelumab (73–10) stains more tumour cells47. Consequently, all other variables being equal, for patients with the same percentage of PD-L1+ tumour cells, the 73–10 assay might be expected to be associated with a lesser degree of therapeutic benefit and the SP142 with a greater level of benefit (including in patients with tumours without detectable PD-L1 expression) than with, for example, the 22C3 assay. For this reason, whereas a TPS of ≥50% defines the 30% of NSCLCs with the highest level of PD-L1 expression detected using, for example, the pembrolizumab-associated 22C3 assay, the same percentage of PD-L1hi NSCLCs is defined by a TPS ≥80% when the avelumab-associated 73–10 assay is used13,25,43. However, the available clinical evidence of assay comparability across four assays (not including 73–10) indicates that while 22C3, 28–8 and SP263 all performed similarly in predicting responses in 40 patients treated with second-line nivolumab, the SP142 test actually performed less well, including when only the patients with a PD-L1 level <1% were examined48.
Finally, relating to PD-L1-independent mechanisms of action, PD-L2 is an alternative ligand for PD-1 and could provide a bypass mechanism for downregulation of anticancer immune responses42. This theoretical activity of PD-L2 could lead to treatment responses of PD-L1 tumours to anti-PD-1 agents, although not to anti-PD-L1 agents. However, limited or no evidence from clinical trials supports this hypothesis, and PD-L2 is somewhat infrequently expressed on lung tumour cells, being more restricted in its expression predominantly to professional antigen-presenting cells49.
PD-L1 expression by IHC not being sufficient to ensure a response to therapy could also have multiple explanations. First, PD-L1 levels might be elevated for reasons other than immune-related induction. So-called intrinsic induction, which can occur downstream of specific signalling pathways, might upregulate tumoural PD-L1 expression in the complete absence of infiltrating immune cells50. For example, although 47% of crizotinib-naive ALK+ NSCLC specimens had PD-L1 levels ≥5% and 26% had levels ≥50%, high levels of CD8+ tumour-infiltrating lymphocytes (TILs), defined as ≥5% of tumour cells with coincident CD8+ T cells identified by IHC, were not detected in any of these tumours51. Unsurprisingly, the response rate to anti-PD-1 or anti-PD-L1 monotherapy in patients with ALK+ NSCLC, even in those with high PD-L1 levels, in this study and all others reported to date was 0%5153. Exactly what the function of PD-L1 expression might be in these tumours — if any — is not clear. Nor is it clear why, if PD-L1 expression is simply a downstream consequence of oncogene activation, expression of this protein is elevated in some but not all tumours driven by the same oncogene.
Intrinsic induction of PD-L1 expression also occurs in EGFR-mutant NSCLC, although true immune induction might rarely be responsible for upregulation of PD-L1. In the aforementioned retrospective study51, in which 16% of TKI-naive EGFR-mutant NSCLC specimens had PD-L1 levels ≥5% and 11% had levels ≥50%, coincident high levels of CD8+ TILs were noted in 2% of patients. Consistent with this low frequency of immune induction, trials of second-line nivolumab, pembrolizumab or atezolizumab all revealed no significant difference in OS compared with that reported with docetaxel in the EGFR-mutant subgroup; trends in favour of docetaxel were observed in trials of nivolumab and atezolizumab912. Nevertheless, responses to immunotherapy in some patients with EGFR-mutant NSCLC have been reported52.
The second reason why PD-L1 expression might not be sufficient to ensure therapeutic benefit from PD-1–PD-L1 inhibition could reflect more than one immune checkpoint being engaged within the tumour42; existing IHC evidence indicates that multiple checkpoint proteins, including PD-L1, IDO1, LAG3, TIM3 (also known as HAVCR2) and OX40 (also known as TNFRSF4), can be co-expressed54. Moreover, host immune functionality might be impaired for reasons beyond the activation of immune checkpoints55. Even among the 33–47% of patients with NSCLC and PD-L1 levels ≥50% included in the CheckMate 026 trial, first-line nivolumab failed to improve either PFS or OS when compared with platinum-based doublet therapy14. By contrast, both PFS and OS improvements were achieved with pembrolizumab in a similar patient population in the KEYNOTE-024 trial13. The reasons why the efficacy of PD-1 inhibition was so different, even in this patient subgroup, between these first-line trials of two drugs with comparable second-line efficacy continue to be debated. Differences in the rate of crossover to anti-PD-1–PD-L1 therapy in the control arms of the trials could have influenced the OS results, although these differences could not have affected the PFS outcomes. Whether relevant imbalances existed between the two treatment arms of CheckMate 026 or whether the patients enrolled in this trial were somehow less able to manifest an immune response than those included in KEYNOTE-024 has to be considered56. For example, in CheckMate 026, a considerable proportion of patients (38%) had received prior radiotherapy, despite only 13% having central nervous system (CNS) disease14; although comparable data were not reported for the KEYNOTE-024 cohort, this figure seems remarkably high for a population of patients with NSCLC in the first-line setting. Despite radiation having some theoretical immune benefits, such as antigen induction and/or release26, the use of prior radiotherapy in a substantial proportion of the CheckMate 026 cohort might have had direct effects on immune functionality within this population owing to the potential of radiation to negatively affect the circulating immune cell pool57 or might have been an indicator of the poor fitness and thus potentially impaired immune functionality, for other reasons, of the patients included in the trial.
The final reason to explore when discussing why PD-L1 expression might not be sufficient to ensure a response to immunotherapy relates to the use of different cut-points for defining positivity when evaluating this continuous variable during the clinical development of PD-1–PD-L1 inhibitors58 (Fig. 2). Considerable data suggest that higher levels of PD-L1 expression are associated with greater therapeutic benefit from these agents, with no apparent ceiling on this effect. Within the single-arm trial of pembrolizumab in patients with NSCLC (KEYNOTE-001), the large study cohort was divided into subgroups defined by quartiles of PD-L1 expression levels, as well as a PD-L1 subgroup43. The proportions of patients within the different expression groups were not equal: 39% were PD-L1-negative, 31% had a PD-L1 TPS of 1–24%, 7% had a PD-L1 TPS of 25–49%, 9% had a PD-L1 TPS of 50–74% and 15% had a PD-L1 TPS of 75–100%43. Response rates were 8%, 13%, 19%, 30% and 45%, respectively, across these subgroups43, suggesting a strong association between the extent of PD-L1 expression and response to therapy. The lack of a ceiling on the predictive potential of the PD-L1 TPS is also illustrated by comparing the benefit of first-line PD-1–PD-L1 inhibition in patients with a TPS of 50–74% versus 75–100%: among 112 patients, those with a PD-L1 TPS of 75–100% had a significantly higher ORR (47.1% versus 13.6%; P <  0.01), a significantly longer median PFS duration (5.1 months (95% CI 3.8–7.4) versus 2.5 months (95% CI 1.8–4.5); P = 0.02) and higher estimated 12-month OS (76.4% versus 54.4%)59.
The absence of a clearly definable binary cut-point in PD-L1 expression that can be used to distinguish those with no benefit from those with unequivocal therapeutic benefit has resulted in levels of PD-L1 positivity being set on a pragmatic basis within clinical trials. PD-L1 thresholds can be chosen on the basis of a need to enrich for a given expected extent of therapeutic benefit to ensure superiority over a specific comparator (for example, higher thresholds for studies with platinum-based doublet chemotherapy as the control treatment and lower thresholds for studies with docetaxel or salvage regimens as the comparator) while also adhering to the potential agenda of not wanting to shrink the intended target population too much, thereby maximizing beneficence and/or market size.
In several trials, rather than presenting data for groups of patients defined by upper and lower PD-L1 expression boundaries, the outcome data are presented for groups defined using only a lower inclusion boundary (for example, ≥1%, ≥5% or ≥10% PD-L1 staining on IHC) in what might be termed a bottom-up approach. Inevitably, this approach minimizes the visibility of biomarker-associated therapeutic benefit if the benefit is predominantly driven by responses of the subgroup with the highest PD-L1 levels, which remain included in each biomarker-positive group (Fig. 2). For example, in the CheckMate 017 trial of nivolumab versus docetaxel for the second-line treatment of squamous NSCLC8, although a numerical trend towards greater differences in the hazard ratio was observed as the lower boundary of the PD-L1 cut-point was increased (from ≥1% to ≥10%), this effect did not reach statistical significance. In the CheckMate 057 trial of nivolumab versus docetaxel for the second-line treatment of non-squamous NSCLC9, however, this effect was statistically significant, with the hazard ratio for OS dropping from 0.59 (Pinteraction = 0.06) at a PD-L1 level ≥1% to 0.4 at a PD-L1 level ≥10% (Pinteraction <  0.001), and the hazard ratio for PFS similarly dropping from 0.70 (Pinteraction = 0.02) to 0.52 (Pinteraction <  0.001). Importantly, the PFS and OS curves in the remainder subgroups — that is, the <1%, <5% and <10% subgroups — shown in the supplemental data of the trial publication minimally changed as the cut-point was shifted upwards9, again suggesting that the high PD-L1 (≥10%) subgroup, which was excluded from all these remainder subgroups, accounted for most of the reported therapeutic benefits. The idea that the hazard ratio improvement would continue and that the nonsignificant trend in the trials involving patients with squamous NSCLC would have become statistically significant if even higher levels of PD-L1 expression were explored seems likely, but these questions remain unanswered to date.
The potential for this bottom-up approach to obfuscate more than elucidate is perhaps best illustrated by the results of the KEYNOTE-042 trial60. In this trial, first-line pembrolizumab was compared with either carboplatin plus paclitaxel or carboplatin plus pemetrexed chemotherapy in patients with EGFR-wild-type and ALK-wild-type NSCLC and a PD-L1 TPS ≥1%. The co-primary end points of the study were met: pembrolizumab improved OS among those with a PD-L1 TPS ≥50%, ≥20% and ≥1% (HRs 0.69 (95% CI 0.56–0.85), 0.77 (95% CI 0.64–0.92) and 0.81 (95% CI 0.71–0.93), respectively)60. However, when upper boundaries to TPS groupings were used to avoid including the patients with a high TPS (≥50%) in every analysis, OS was not significantly different between the pembrolizumab and chemotherapy groups among patients with a TPS of 1–49% (HR 0.92, 95% CI 0.77–1.11)60. In addition, the Kaplan–Meier OS curves for the TPS 1–49% subgroup crossed at a time point just before 12 months, with OS favouring chemotherapy up to this point60, suggesting that although later outcomes might compensate in this subgroup overall, the initial use of pembrolizumab is detrimental to survival in a considerable proportion of these patients.
Recognizing that predictive benefit of PD-L1 expression might reflect a true continuous variable, with ongoing (if disproportional) increases in the benefit from PD-1–PD-L1 inhibition being seen as the proportion of tumour cells expressing PD-L1 increases, one of the most interesting questions to ask is ‘Why?’ Immune responses seem to be largely directed against non-driver mutations3740. Consequently, if passenger and/or non-truncal (branch) mutations are the origins of the altered amino acid sequences that lead to immune recognition of cancers, these neoantigens are likely to differ considerably within and between foci of cancer cells within the same individual. A higher proportion of the tumour expressing PD-L1 might, therefore, reflect a higher proportion of the tumour cells manifesting the same or even different non-synonymous mutations that are able to persist through their common use of the PD-1–PD-L1 axis as a dominant escape mechanism40. Beyond the associations with ORR, PFS and OS, correlating the percentage of PD-L1 positivity with both the depth and the duration of response would be another interesting avenue to explore in the future.

Tumour mutational burden

Tumour mutational burden (TMB) is increasingly being explored as an additional or alternative predictor of clinical benefit in PD-1–PD-L1 inhibition. The underlying hypothesis is that the chances of a neoantigen existing, being presented and leading to an immune response that is blocked by engagement of the PD-1–PD-L1 axis will be increased if more non-synonymous mutations are present in the tumour.
TMB can be assessed by quantifying the number of non-synonymous mutations across the whole exome or in a defined gene panel through next-generation sequencing (NGS)61. Initial studies of TMB were predominantly focused on relative levels, although absolute definitions are being pursued by commercial providers (for example, ten mutations per Mb has been adopted as a definition of a high TMB by Foundation Medicine within their NGS panel)6163. Importantly, TMB, like PD-L1 expression, is a continuous variable, and different assays and different definitions of a high TMB will alter the population size and potentially the extent of the associated treatment benefit in the identified biomarker-positive group.
In a subgroup analysis of the negative CheckMate 026 trial14, in which all patients had PD-L1 levels ≥1%, whole-exome-based TMB assessments demonstrated a nonsignificant trend towards a greater PFS benefit from nivolumab versus platinum-based doublet chemotherapy for those in the highest but not the combined lowest and middle TMB tertiles (HR 0.62 (95% CI 0.38–1.00) and HR 1.82 (95% CI 1.30–2.55), respectively)62. However, no trend in relation to an OS benefit was apparent. TMB is known to be higher in smoking-associated NSCLC than in NSCLC arising in never-smokers, and ORRs, PFS and OS with immunotherapy tend to be less favourable in never-smokers than in current or former smokers813,43,64. Therefore, smoking status has been proposed as a simple, clinically applicable surrogate for TMB. However, although smoking status was not broken down by pack-years or time from quitting, nivolumab was not associated with even a nonsignificant trend towards improvements in PFS or OS relative to chemotherapy in either current or former smokers in CheckMate 026 (ref.14).
In the large-cohort CheckMate 227 trial63, multiple separate first-line treatment arms, subdivided by PD-L1 levels <1% or ≥1%, were compared. The treatments included nivolumab plus the anti-cytotoxic T lymphocyte antigen 4 (CTLA-4) antibody ipilimumab or four or less cycles of platinum-based doublet chemotherapy (with optional pemetrexed maintenance after a platinum-pemetrexed doublet), with additional nivolumab plus doublet chemotherapy and nivolumab monotherapy arms included for only the PD-L1 <1% and ≥1% groups, respectively. Only partial data sets from the trial have been presented to date63. Multiple primary end points were assessed, including a later amendment of adding PFS among those with a high TMB, defined using an absolute cut-off of ten or more mutations per Mb detected using the Foundation Medicine panel, as an additional co-primary end point63. No correlation between the presence of high TMB and PD-L1 levels was found63. In the unselected population, the combination of nivolumab and ipilimumab was associated with superior PFS compared with chemotherapy (HR 0.83, 95% CI 0.72–0.96); in the subpopulation with a high TMB, the PFS hazard ratio was 0.58 (95% CI 0.41–0.81), and in those with a low TMB it was 1.07 (95% CI 0.84–1.35)63. The significant PFS benefit in the high TMB subgroup was demonstrated in both PD-L1 <1% and PD-L1 ≥1% groups (HR 0.48, 95% CI 0.27–0.85 and HR 0.62, 95% CI 0.44–0.88, respectively)63. However, whether combined CTLA-4 and PD-1 inhibition offers real value in this setting or whether TMB is the most appropriate biomarker for this combination is debatable.
The initial correlations between elevated TMB and longer PFS in single-arm studies of PD-1 inhibition and the observation of a trend towards a PFS benefit with nivolumab versus chemotherapy in the high TMB population in the aforementioned retrospective subgroup analysis of CheckMate 026 suggest that TMB is not a specific predictive biomarker of benefit from CTLA-4 inhibition or even from combined use of CTLA-4 and PD-1–PD-L1 inhibition62,63. Instead, a high TMB might simply be a biomarker that enriches for a population sensitive to any immune intervention (Fig. 3b). Consistent with this theory, in CheckMate 227 (ref.63),1-year PFS in patients with a PD-L1 level ≥1% and a high TMB (defined, in this case, as ≥13 mutations per Mb on the basis of prior analyses of the CheckMate 026 data) was 24% with nivolumab monotherapy versus 17% with chemotherapy, although the PFS hazard ratio was not statistically significant (0.95, 97.5% CI 0.61–1.48). In the CheckMate 026 all-comers population, 1-year PFS in patients with a PD-L1 level ≥5% was 22.0% and 23.2% for the nivolumab and chemotherapy groups, respectively14. In addition, the PFS differences with nivolumab plus ipilimumab versus nivolumab in the high TMB, PD-L1 ≥1% subpopulation of the CheckMate 227 trial were not statistically significant (HR 0.75, 95% CI 0.53–1.07)63. The ORR with nivolumab and ipilimumab in the ICI-naive, high TMB group of CheckMate 227 was 45.3%63. However, the ORR of a different combination of CTLA-4 and PD-L1 inhibitors (tremelimumab and durvalumab) in an otherwise unselected but PD-1–PD-L1 inhibitor-experienced NSCLC population was only 5%65. These data on combination ICI treatment after PD-1–PD-L1 inhibition suggest the potential for the addition of CTLA-4 inhibition to expand the population of patients with immunotherapy-sensitive NSCLC might be limited. Consequently, the effect of the high-TMB-based enrichment approaches used in CheckMate 227 in reducing the number of patients who do not benefit from immunotherapy in the denominator of any efficacy percentage calculation might be to emphasize even minor additional benefit achieved by combining immunotherapies.
Given the potential of high TMB-based selection approaches to reduce the number of patients who have a very low chance of benefiting from immunotherapy that receive such treatment, this pre-selection approach might be adopted for many different novel immune combinations going forward. Of note, however, a high TMB might not be equally predictive of benefit from immunotherapy when detected at different time points. Specifically, evidence for translational studies indicates that when the TMB is increased as a result of cytotoxic therapy (as could occur in patients included in trials of second-line or later-line therapy), a high TMB could reflect an increased abundance of branch mutations and thus either a lower likelihood of a unifying immune response or a greater likelihood of an acquired resistance mechanism manifesting clinically40.
TMB has also been determined through targeted genomic sequencing of circulating cell-free DNA (encompassing a panel of 394 genes), with higher blood TMB (bTMB) levels demonstrated to enable enrichment for patients who derive a greater PFS benefit from atezolizumab versus docetaxel66. This approach offers the potential for both easier access and quicker turnaround times than are possible with tissue-based TMB quantification67. For atezolizumab, an absolute bTMB cut-point of ≥16 was selected (accounting for 27% of those tested), which was able to produce a nonsignificant trend towards a lower PFS hazard ratio than that of the overall biomarker-evaluable population in the phase III OAK trial (0.65, 95% CI 0.47–0.92 versus 0.87, 95% CI 0.73–1.04)66. As with PD-L1 expression, however, no evidence of a plateauing of enrichment for either PFS or OS benefit at higher levels was found: a bTMB of ≥22, encompassing 14% of the tested population, was associated with a PFS hazard ratio of 0.57 (95% CI 0.35–0.91), and a bTMB of ≥26, accounting for 9% of the tested population, was associated with a PFS hazard ratio of 0.51 (95% CI 0.33–0.99)66.

Improving predictive immune biomarkers

In order to reduce the risks associated with each assumption surrounding the current predictive biomarkers for ICIs, multiple improvements could be envisioned, mostly related to combining biomarkers (Fig. 3c). Existing evidence indicates that combining biomarkers from the top (TMB) and towards the bottom (PD-L1) of the presumptive immune cascade (Fig. 3b) can result in predictive additivity or synergy. For example, in the PD-L1 TPS 1–49% group of the CheckMate 026 trial, the ORRs to nivolumab were 2.0-fold greater among patients in the highest TMB tertile versus those in the combined low and medium tertiles (32% versus 16%), and in the PD-L1 TPS ≥50% group, the ORRs were 2.2-fold higher (75% versus 34%)62. The PFS difference between arms was also most exaggerated in the 10% of patients who were in the highest TMB tertile and had PD-L1hi tumours62.
At the other end of the biomarker continuum, the negative predictive potential of combining biomarkers for lack of benefit is also starting to be explored as a potential treatment exclusionary approach. For example, in CheckMate 227, for patients with both a low TMB (<10 mutations per Mb) and a PD-L1 TPS <1%, the combination of nivolumab and ipilimumab was no better than platinum-doublet chemotherapy in terms of PFS (HR 1.17, 95% CI 0.76–1.81)68.
In terms of lessening the risks associated with the assumption that a high TMB will result in presentable immunogenic neoantigens, the predictive value of TMB biomarkers could potentially be increased if putative neoantigens are weighted, given that not all non-synonymous mutations are equally immunogenic69,70. Similarly, the utility of bioinformatics platforms that filter or weight neoantigens by their predicted potential to be presented in complex with the MHC molecules of an individual patient could be assessed. Such approaches have already been used to explain how MHCI genotype can restrict the driver oncogene landscape that an individual cancer can manifest by negatively selecting MHC-presentable immunogenic alterations69,70. MHC molecule expression within the tumour might have to be assessed separately from that in non-malignant tissue, however, as MHC allele loss has been described as a feature of lung cancer evolution, consistent with a role for this alteration in immune evasion71.
With regard to reducing the risks associated with the assumption that PD-L1 upregulation reflects dependence of the tumour on the PD-1–PD-L1 axis to evade an initiated, primed anticancer immune response, the combination of assessments of PD-L1 expression levels with phenotypic features of the tumour immune microenvironment is being actively explored. Quantifying PD-L1 expression on intratumoural immune cells in addition to tumour cells did not improve the predictive value of PD-L1 IHC for OS within the OAK trial72. This finding might, however, simply reflect limitations of the assay because agreement between pathologists on immune cell PD-L1 scoring is poor4448.
Gene expression data generated using mRNA extracted from formalin-fixed paraffin-embedded tumour samples have been used to evaluate the permissiveness of the immune microenvironment. An effector T (Teff) cell gene signature based on combined PDL1, IFNG and CXCL9 mRNA levels has been applied to the OAK trial data set11: for the high versus low gene expression groups (based on median mRNA levels) the PFS hazard ratios were 0.73 (95% CI 0.58–0.91) and 1.30 (95% CI 1.05–1.61), respectively; the OS hazard ratios were 0.59 (95% CI 0.46–0.76) versus 0.87 (95% CI 0.68–1.11)73. For both PFS and OS, the Teff cell signature was associated with better hazard ratios than those associated with PD-L1 positivity (levels >0%)73; however, comparisons using higher PD-L1 expression cut-points have not been reported. Similarly, a T cell-inflamed 18-gene mRNA expression profile has been associated with favourable ORRs and PFS with pembrolizumab across a range of different tumour types74. Non-responders fell into two broad categories: those with no evidence of an antitumour immune response, suggesting an earlier failure of immune recognition and/or activation; and those with evidence of tumour immune infiltration, in whom the tumour cells are presumably exploiting non-PD-1-dependent immune escape mechanisms alone or in combination with engagement of the PD-1–PD-L1 axis74. An IFNγ gene expression signature relating to a panel of four genes (encoding IFNγ, CD274, LAG3 and CXCL9) has also been associated with higher ORRs and longer PFS and OS durations among patients with NSCLC treated using durvalumab, independent of PD-L1 status75. Other gene expression signatures have also been proposed76. These gene signatures are continuous variables and, as with assessments of most individual immune checkpoints, what finding is defined as positive and why will always have to be questioned. With the Teff cell signature, for example, if a higher cut-point comprising only the patients in the highest quartile of gene expression levels was used in the analysis of the OAK trial data73, the hazard ratios for PFS and OS were 0.66 (95% CI 0.48–0.91) and 0.60 (95% CI 0.42–0.87), respectively, trending towards a greater effect at higher levels — despite the fact that the aforementioned group with higher than median levels of Teff cell signature gene expression will have contained overlapping data points with the highest quartile group (thus creating the same problems previously described for the bottom-up approach).
Direct evidence of an immune system poised for activation can offer a partial alternative to gene expression panels. For example, through flow cytometric analyses of CD45+CD3+CD8+ TILs from freshly excised melanoma specimens, the presence of ≥20% of TILs with high expression of CTLA-4 and PD-1 was associated with a response and favourable PFS with subsequent anti-PD-1 therapy77. However, such an approach is not currently feasible outside of a clinical trial, and whether the observed predictive potential is transferable to other tumour types remains unclear. In NSCLC, the analysis revealing the rarity of coincident CD8+ TILs with tumoural PD-L1 expression provided mechanistic insight into the limited efficacy of PD-1–PD-L1 inhibition in patients with ALK+ or EGFR-mutant disease51; however, beyond the simple presence or absence or immune cells, considerable opportunity also exists to finesse basic pathological descriptions of the tumour immune environment. For example, exactly where the immune cells are located and/or what types of immune cells are present might be important factors influencing outcomes of ICI treatment. These features could be described using basic diagnostic histology with haematoxylin and eosin staining plus simple, readily available IHC assays for established markers (such as CD8, CD68 and others), or analyses could be extended to include multi-colour immunofluorescence or multi-colour bright-field IHC.
Peripheral blood evidence of immune activation is also being explored, but whether the levels of particular immune cells in the blood before ICI therapy will predict for subsequent clinical benefit, distinct from their demonstrated roles as an early readout of such benefit when analysed following treatment initiation78, remains to be seen. One must also be cognizant of the risk that such blood-based assays might not necessarily reflect an antitumour immune response but rather an alternative, co-incident inflammatory reaction occurring elsewhere in the body.
The hard-wired truncal nature of some oncogenic mutations might also be usefully exploited to refine the use of ICIs if defined patient subgroups with varying responses to these agents can be identified. The limited efficacy of PD-1–PD-L1 inhibition in patients with ALK+ or EGFR-mutant NSCLC has already been discussed, and exclusion of these molecular subtypes was used to enrich patient populations of the KEYNOTE-024, KEYNOTE-042 and CheckMate 026 trials with those patients more likely to benefit from such treatment13,14,5153,60. KRAS mutations are present in ~25% of lung adenocarcinomas and have been associated with modestly increased responsiveness to immunotherapy and a more favourable OS hazard ratio in randomized studies versus chemotherapy (for example, HR 0.52 in the KRAS-mutant population versus 0.75 in the overall population and HR 0.98 in the known KRAS-wild-type population in CheckMate 057)9. However, potentially clinically relevant co-mutation subgroups within the KRAS-mutant NSCLC population have started to be described79. Among 174 patients with KRAS-mutant NSCLC who had received anti-PD-1 or anti-PD-L1 antibodies, the presence of a coexisting STK11 mutation (31% of patients) was associated with a lower ORR and shorter median PFS duration than the presence of a coexisting TP53 mutation (32%) or neither co-mutation (37%)80. Nevertheless, the ORR was still 7.4% among those with co-incident KRAS and STK11 mutations80 — not zero.
Predictive biomarkers for benefit from ICIs derived from analyses of the host microbiome and the possibility that disturbances of the gut flora by antibiotics could affect the outcomes of PD-1–PD-L1 inhibition have also been described81. The underlying hypothesis is that certain bacteria can enhance or impair host immune functionality, although the mechanistic basis for these observations remains under investigation. Such data nevertheless illustrate the broad range of factors that could ultimately be relevant to consider in predicting benefit from immunotherapy in the future.

Immunotherapy–cytotoxic combinations

Whether any signatures of the efficacy — or lack thereof — of PD-1 or PD-L1 inhibitors will also relate to combinations of these agents with standard cytotoxic approaches (chemotherapy and/or radiotherapy) also has to be considered. The rationale for immunotherapy–cytotoxic therapy is, in part, to increase the immunogenicity of the cancer through a broad cytotoxic insult; therefore, baseline assessments of the potential of a tumour for a response to immunotherapy, as described above for immunotherapy–immunotherapy combination approaches, might be less relevant. In the KEYNOTE-189 trial82, the addition of pembrolizumab to standard first-line carboplatin and pemetrexed significantly improved the ORR (47.6% versus 18.9%), PFS (HR 0.52, 95% CI 0.43–0.64) and OS (HR 0.49, 95% CI 0.38–0.64) of patients with advanced-stage non-squamous NSCLC. The OS and most of the PFS benefits were significant across all PD-L1 TPS groups; however, the effect size was not uniform, and higher levels of PD-L1 positivity were still associated with greater benefit, even with the immunotherapy–chemotherapy combination. For example, the OS hazard ratio for those with a PD-L1 TPS ≥50% was 0.42 (95% CI 0.26–0.68), and for those with PD-L1 TPS <1% it was 0.59 (95% CI 0.38–0.92)82. In the same two groups, the PFS hazard ratios were 0.36 (95% CI 0.25–0.52) and 0.75 (95% CI 0.53–1.05), respectively82. Similarly, in CheckMate 227 (ref.68), the modest PFS benefit from the addition of nivolumab to standard first-line chemotherapy in the PD-L1 <1% group (HR 0.74, 95% CI 0.58–0.94) was increased when only the high TMB subgroup was considered (HR 0.56, 95% CI 0.35–0.91) and was reduced to a nonsignificant trend when only patients with a low TMB were assessed (HR 0.87, 95% CI 0.57–1.33). In the low PD-L1 (<1% positivity), high TMB group, the ORR was far higher with chemotherapy plus nivolumab than with ipilimumab plus nivolumab (60.5% versus 36.8%)68. Among responders, however, the rate of responses lasting ≥1 year was markedly higher in the immunotherapy–immunotherapy combination arm than in the chemoimmunotherapy arm (93% versus 33%)68, suggesting that not all chemoimmunotherapy responses are true durable immune responses.
Another way of looking for synergy from the immunotherapy–cytotoxic therapy approach is to explore the benefit in groups with traditionally low responsiveness to immunotherapy. Among patients with stage III NSCLC included in the PACIFIC trial26, as in ICI monotherapy trials in those with stage IV disease, the EGFR-mutant subgroup did not conclusively benefit from durvalumab even though this agent was given after radical chemoradiotherapy (PFS HR 0.76, 95% CI 0.35–1.64). By contrast, in IMpower150 (ref.83), the combined EGFR-mutant or ALK+ disease subgroup did derive a significant PFS benefit from the addition of atezolizumab to carboplatin plus paclitaxel and bevacizumab (HR 0.59, 95% CI 0.37–0.94). When the data were stratified according to ALK rearrangements, common EGFR mutations (exon 19 deletions and L858R mutations) and other EGFR mutations, however, the PFS benefit remained significant only in those with the common EGFR mutations (n = 59; HR 0.41, 95% CI 0.22–0.78)83. To what extent these data reflect the general benefit of chemoimmunotherapy in this group or the effect of either the additional anti-angiogenic agent or as yet unidentified imbalances between the arms in these small subgroups (such as the presence or absence of treated or untreated CNS disease) remains to be determined. In the latest IMpower150 update84, a significant OS benefit was demonstrated in the wild-type population (HR 0.78, 95% CI 0.64–0.96), whereas the trend towards an OS benefit in the EGFR-mutant and ALK+ subgroup was not significant (0.55, 95% CI 0.29–1.03).
Overall, these data suggest that while the concept of re-initiating immune recognition by combining cytotoxic therapy and PD-1–PD-L1 inhibition is intriguing and seems to be more efficacious than either treatment alone, benefit from such combinations is not universal. Some predictive biomarkers for benefit from immunotherapy alone are being carried over to these new immunotherapy–cytotoxic therapy combination approaches; however, we might also need to explore the different kinds of cell death induced by our standard treatment approaches, whether they are all equally immunogenic and whether we can develop pre-cytotoxic therapy biomarkers that are predictive of benefit from such combinations85 — similarly to how we have tried to develop those for PD-1 or PD-L1 inhibitor monotherapy.


For some oncogene-addicted cancers, such as ALK+ NSCLC, even the first generation of predictive biomarkers have been able to perform as effective ‘give or not give’ decision-making tools with regard to specific targeted therapies. For others, such as the EGFR TKIs, narrowing in on the drug-sensitive population was an iterative process initially based on clinical characteristics and then a series of imperfect, partially confounded molecular biomarkers (Fig. 1).
Are predictive biomarkers for immunotherapy following the same slow path as those of EGFR TKIs? Combining assessments at different points in the immune pathway in order to lower the risks associated with each assumption might improve the predictive potential of immunotherapy biomarkers (Fig. 3c). Such improvements would enable a patient, a physician or a health system to more accurately decide whether PD-1–PD-L1 inhibition, or potentially some other immune combination, should be used versus an alternative therapy in any particular line of therapy. However, will a true give or not give predictive biomarker for PD-1–PD-L1 inhibition or other checkpoint inhibitors ever be identified?
Given the complexity and multifactorial nature of the anticancer immune response and the mechanisms of tumour immune evasion, compared with the simpler nature of an addictive oncogenic driver, a marker guaranteeing benefit seems unlikely. Likely, it will be easier and faster to define a biomarker for lack of responsiveness to PD-1–PD-L1 inhibition (akin to the initial exclusionary role of KRAS mutations in allocating EGFR TKIs) than to identify a marker of guaranteed therapeutic benefit (Fig. 4). The health economics of cost versus benefit of immunotherapy alone could ultimately lead some health systems to define a ‘never give’ group with a very low chance of a response, for instance, according to some composite assessment that might include minimal PD-L1 expression, no evidence of immune infiltration or activation, low potential antigenicity, and so on68. Yet, if health costs are not a barrier, in the absence of better next-line alternatives, human optimism might win out. For example, to date, the known response rate to anti-PD-1 or anti-PD-L1 monotherapy in patients with ALK+ NSCLC remains 0% across multiple independent studies5153. Despite this fact, many patients with ALK+ NSCLC, at least in the USA, have received a PD-1 or PD-L1 inhibitor, often based on their elevated PD-L1 level, as a preferred post-TKI line of therapy. Perhaps biomarkers predictive of harm from immunotherapy (hypertoxic signatures), in terms of pre-existing autoimmunity, solid organ transplants8688 or some definable signature for the risk of hyperprogression (defined as substantially more rapid disease progression than would have been predicted to result from a simple lack of efficacy of an agent on the basis of the prior growth dynamics of the tumour)8991, might also be used in such decision-making in the future.
If we do define an optimal predictive biomarker for giving, or not giving, anti-PD-1 or anti-PD-L1 therapy, two other questions present themselves. First, would the biomarker apply at any time point or only at a given time-locked decision point? Second, would it apply to only PD-1 or PD-L1 inhibitor monotherapy? As we have started to see with CTLA-4 inhibitor combinations with PD-1–PD-L1 inhibition63,65,68, the same challenges in defining who is sensitive to PD-1 or PD-L1 inhibitor monotherapy will continue to apply when trying to define who will be sensitive to a given immunotherapy–immunotherapy combination. Some predictive biomarkers towards the top of the immune cascade (such as those determining the chances of a presentable, immunogenic neoantigen existing) might be informative with regard to deciding on any immunotherapy (Fig. 3b). The potential predictive role for biomarkers lower in the cascade, such as PD-L1 expression, might be much more specific to a particular drug class or individual immune checkpoint.
In the future, just as we have already done during the evolution in our understanding of predictive biomarkers for treatment benefit among the oncogene-addicted subtypes of NSCLC, we will have to continue to ask who is actually benefiting, and from what aspect of any combination therapy used, in any immunotherapy trial with positive results — that is, whether synergy or additivity is being achieved with the combination therapy and in whom.


The work of D.R.C. and R.C.D. is partially supported by the University of Colorado Lung Cancer SPORE (P50CA058187).

Reviewer information

Nature Reviews Clinical Oncology thanks R. Rosell, M. Reck and other anonymous reviewer(s) for their contribution to the peer review of this work.

Competing interests

D.R.C. has ad hoc advisory roles with ARIAD, Arrys/Kyn, AstraZeneca, Bio-Thera DSMB, Celgene, Clovis, Daiichi Sankyo (interstitial lung disease adjudication committee), G1 Therapeutics (DSMB), Genoptix, Hansoh SRC, Hengrui, Ignyta, Lycera, Mersana Therapeutics, Novartis, Orion, Regeneron, Revolution Med, Roche/Genentech and Takeda and has received research funding from Takeda. R.C.D. is an advisory board member for ARIAD, AstraZeneca, Ignyta, Spectrum and Takeda; has received research sponsorship from Ignyta and licensing fees from Abbott Molecular, Ignyta and Rain Therapeutics; and owns stock in Rain Therapeutics. K.M.K. has been a consultant for and has received speaker’s honoraria from AstraZeneca, Boehringer Ingelheim, Bristol-Myers Squibb, Eli Lilly, Merck Serono, Merck Sharp and Dohme, Novartis, Pfizer, Roche and Ventana and has been a consultant for AbbVie.
Publisher’s note
Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations.

Related topics