Biomarker development for PD-(L)1 axis inhibition: a consensus view from the SITC Biomarkers Committee

Abstract

Therapies targeting the programmed cell death protein-1/programmed death-ligand 1 (PD-L1) (abbreviated as PD-(L)1) axis are a significant advancement in the treatment of many tumor types. However, many patients receiving these agents fail to respond or have an initial response followed by cancer progression. For these patients, while subsequent immunotherapies that either target a different axis of immune biology or non-immune combination therapies are reasonable treatment options, the lack of predictive biomarkers to follow-on agents is impeding progress in the field. This review summarizes the current knowledge of mechanisms driving resistance to PD-(L)1 therapies, the state of biomarker development along this axis, and inherent challenges in future biomarker development for these immunotherapies. Innovation in the development and application of novel biomarkers and patient selection strategies for PD-(L)1 agents is required to accelerate the delivery of effective treatments to the patients most likely to respond.

Introduction

Defining PD-(L)1 inhibitor resistance

While many patients treated with programmed death-ligand 1 (PD-(L)1) inhibitors achieve long-term benefit, the majority either fail to respond or experience relapse weeks to months after initiation of therapy. Furthermore, the cadence and manifestations of PD-(L)1 inhibitor resistance are heterogeneous and vary by indication and between individual patients but have been broadly classified as primary (occurring while on initial treatment) or secondary (acquired after an initial period of response to treatment). The field continues to grapple with defining core mechanistic similarities and differences between primary and secondary resistance. This complexity complicates the study of response to PD-(L)1 inhibitors, which in turn hinders the development of new therapies and robust biomarkers of response.

The Society for Immunotherapy of Cancer (SITC) has developed consensus recommendations on defining resistance to PD-(L)1 pathway blockade based on observed responses.1,5 Clinical definitions of primary resistance, secondary resistance, and resistance that develops after the discontinuation of therapy were proposed for both single-agent and combination PD-(L)1 regimens. Primary resistance was defined as “the inability of immune cells to mount an antitumor response upon initial drug exposure”. The task force concluded that two cycles of therapy with initial and confirmatory scans were required to define a lack of response. Primary resistance was therefore defined after approximately 6 weeks of exposure with no clinical benefit, or a best response of progressive disease (PD), or stable disease (SD) lasting for less than 6 months. SITC’s efforts also noted multiple caveats related to defining primary resistance, such as patients who discontinue therapy, variability between fast and slow growing cancers, and heterogeneity of response of distinct tumor lesions within an individual patient.¹ Secondary or acquired resistance was defined as a “documented, confirmed objective response or prolonged SD (more than six months) and subsequent disease progression in the setting of ongoing treatment”. There are also noted important caveats about secondary resistance such as excluding patients after discontinuation for adverse events, the recognition of a “gray zone” between primary and secondary resistance for small numbers of patients, and the need to consider new definitions for different disease types.

In addition to heterogeneity of the timing of response to PD-(L)1 blockade, emerging evidence points to lesion-to-lesion heterogeneity of response within individual patients, and the ensuing complexity in the definitions of radiographic disease progression. It has recently been shown that lesion-to-lesion response to programmed cell death protein-1 (PD-1) blockade within individual subjects can be highly heterogeneous.6,8 In metastatic melanoma, for example, only 10.9% (56/511) of patients displayed progression of every target lesion and most of these patients only had a single target lesion. In total, in patients with a Response Evaluation Criteria in Solid Tumors (RECIST) V.1.1 categorization of PD, only 23.5% (495/2111) of individual target lesions progressed during the trial. The most common cause of categorization as PD was the appearance of new metastatic lesions followed by unequivocal growth of non-target lesions. A net increase in tumor burden (ie, target lesion progression) was the least common cause of progression. Together the data suggests that random biomarker samples from patients with PD are likely from lesions that did not progress, thus complicating the study of mechanisms that drive progression. Several studies have explored the impact of treatment beyond progression with immunotherapy in patients with solid tumors.9,13 One recent study of patients treated beyond progression with PD-1 blockade showed that most growing lesions stabilize with continued PD-1 blockade,⁷ where only 16% of growing target lesions and 33% of new metastatic lesions showed clinically meaningful growth in the post-progression period. Similar results were observed in other indications. Together this suggests that many of the patient outcomes that RECIST V.1.1 defines as treatment failure simply display transient progression in a subset of lesions.

Despite the limitations and variables in defining primary versus secondary (ie, acquired) resistance, distinguishing patient responses between these categories should facilitate the discovery of distinct mechanisms of resistance to PD-(L)1 inhibitors, and in turn novel biomarkers, that can lead to novel clinical trial design based on said mechanisms. In agreement, SITC has recently published a commentary emphasizing the importance of the field coming together towards advancing novel PD-(L)1 development in this exact manner.³

The biology of PD-(L)1 resistance

There is a large ongoing effort to fully catalog the biological underpinnings and hallmarks of the mechanisms of resistance to PD-(L)1 monotherapies and associated biomarkers, as these could guide subsequent therapy to overcome resistance.¹⁴ The mechanisms of PD-(L)1 resistance are likely to vary based on the stage or classification of the cancer, with apparent differences in the composition of tumor microenvironment (TME) between primary and metastatic lesions.¹⁵ Different tumor types originating from distinct tissues may also employ different resistance mechanisms.¹⁶ Mechanisms of resistance (MoR) can be linked to a few broad categories previously defined across the tumor-immunity cycle, including failures in T-cell activation (MoR 1), barriers to accessing the TME (MoR 2), inhibitory immune suppressive activity (MoR 3), and tumor intrinsic resistance to killing (MoR 4) (figure 1).¹⁷

Figure 1. Tumor mechanisms of resistance (MoR) to PD-(L)1 therapeutics. Tumor MoR can be divided into different categories based on where the tumor-immunity cycle is disrupted. MoR1 represents failures of T-cell priming, activation, and expansion. MoR2 represents failures of T cells to traffic to the tumor. MoR3 represents suppression from the tumor or tumor-associated immune cells. MoR4 represents tumor resistance to T-cell-mediated cytolysis. APC, antigen-presenting cell; ECM, extracellular matrix; IDO, indoleamine-pyrrole 2,3-dioxygenase; LAG3, lymphocyte activation gene 3; MDSC, myeloid-derived suppressor cells; PD-L1, programmed death-ligand 1; TAM, tumor-associated macrophage; TCR, T-cell receptor; TGF, transforming growth factor beta; TIM3, T-cell immunoglobulin and mucin domain 3; TIGIT, T-cell immunoreceptor with immunoglobulin and ITIM domain; TME, tumor microenvironment; Tregs, T regulatory cells.

Failures in T-cell activation can have many biological causes, including lack of antigenicity and lack of adjuvancy that can result in T-cell failure to be primed and/or expand. Lack of tumor antigenicity to stimulate T cells via the T-cell receptor (ie, signal 1) can be caused by low (neo)antigen load or defects in the antigen processing and presenting machinery, including genomic or epigenetic changes at major histocompatibility complex (MHC) loci or defects in interferon pathways for induction of MHC expression.18,20 In addition, human leukocyte antigen (HLA) homozygosity can also limit the diversity of antigen expression in some patients.²¹ Finally, improper timing of antigen stimulation can also result in failure of T-cell priming and resistance to PD-(L)1, as PD-(L)1 blockade prior to T-cell priming results in the development of PD-1⁺CD38⁺CD8⁺ T-cell phenotypes that are resistant to anti-PD-1 treatment.²²

Antigen presentation for the induction of antitumor T effector cells requires an appropriate second signal providing costimulation via CD28, which can be antagonized either directly by cytotoxic T-lymphocyte associated protein 4 (CTLA-4), or indirectly by other negative costimulatory molecules such as lymphocyte activation gene-3 (LAG-3), T-cell immunoglobulin and mucin domain-3 (TIM-3) and/or T-cell immunoreceptor with immunoglobulin and ITIM domain (TIGIT).^{23 24} These can be upregulated during chronic antigen stimulation²⁴ as well as during PD-(L)1 therapy in tumors25,27 and have all been linked to some form of resistance.2327,29 Lack of necessary costimulation can also result from the absence of pathogen recognition receptors triggering molecules which stimulate costimulatory molecule expression on antigen-presenting cells.³⁰

CTLA-4 and LAG-3 both have inhibitory antibodies approved for clinical use.31,33 Most recently, LAG-3 expression is hypothesized to represent a major mechanism of resistance to PD-(L)1 therapy as the two molecules seem to be coordinately expressed. In patients with advanced melanoma, a combination of anti-PD-1 nivolumab and anti-LAG-3 relatlimab provided a 25% reduction in progression or death compared with anti-PD-1 monotherapy.³² In baseline samples from 90 patients with advanced non-small cell lung cancer (NSCLC) treated with PD-(L)1 axis blockers, elevated LAG-3 expression was significantly associated with shorter progression-free survival.³⁴ More recently, patients with NSCLC with a non-durable clinical benefit to PD-1 therapies had higher proportions of “burned-out” effector CD8⁺ tumor inflitrating lymphocytes (TILs) (termed Ebo and classified by CD45RO, EOMES, Fas, CD27, CD28, KLRG1, PD-1, LAG-3, TIM-3) than patients with durable clinical benefit.³⁵ It will be interesting to see if combination therapies including either LAG-3 or TIM-3 targeting agents can prolong the clinical benefit in patients with NSCLC and other diseases as has been observed for patients with melanoma.

In addition to signals 1 and 2, sufficient cytotoxic CD8⁺ T-cell expansion requires inflammatory cytokines (ie, signal 3) provided by CD4⁺ T cells to achieve clinical responses to PD-(L)1 blockade.³⁶ The lack of robust Th1 response, as characterized in practice by interferon gamma (IFN-γ) and IFN-γ-inducible chemokines such as CXCL9 and CXCL10, can result in a lack of antitumor effector CD8⁺ T cell, CD4⁺ T cells, and natural killer (NK) cells within the tumor for anti-PD-(L)1 to act on.³⁷ In addition to coordinating a robust antitumor TME, and induction of MHC molecules, IFN-γ signaling into tumor cells can induce expression of PD-L1 which may increase sensitivity to PD-(L)1 blockade.³⁸ Thus, either a lack of IFN-γ or defective IFN-γ signaling into tumor cells can lead to resistance to PD-(L)1 therapy.^{39 40}

If quality effector T cells are generated outside of the TME, then barriers to TME access can also contribute to the lack of efficacy of PD-(L)1 therapies as the second broad category of resistance. Indeed, multiple studies have interrogated and classified T-cell location and response. “Immune-excluded” tumors (immune cells aggregating at the tumor boundaries) and “cold” or “immune desert” tumors (no immune infiltration) have been associated with worse outcomes when compared with “hot” or “inflamed” tumors.^{41 42} The lack of inflammation may be due to the lack of trafficking signals (eg, chemokines or vascular cell adhesion molecules) and/or active mechanisms that prevent T-cell entry or motility (eg, T-cell exclusion by tumor-associated macrophages (TAMs) and cancer-associated fibroblasts).43,45 In addition, aberrant vessels with reduced permeability formed during erratic tumor angiogenesis can prevent both T cell and therapeutic antibody entry into tumors.^{46 47} Moreover, the resultant areas of hypoxia can induce the expression of LAG-3 and other immune checkpoints further promoting resistance.^{48 49} Interestingly, alteration in several tumor suppressor or oncogenic pathways such as WNT/β-catenin have been associated with a lack of infiltration by immune cells and decreased response to PD-(L)1 blockade.50,54 Other pathways associated with immune cell exclusion or depletion include mTOR,^{55 56} MAPK,57,59 and PTEN.^{60 61}

The actions of inhibitory cells and other suppressive mechanisms within the tumor represent the third overarching resistance category. These immune suppressive cells, as well as soluble factors, can promote resistance to PD-L1 by directly antagonizing the antitumor effector cells modulated by PD-(L)1 blockade. TAMs, T regulatory cells, and myeloid-derived suppressor cells (MDSCs) can all inhibit the action of antitumor T cells and NK cells.⁶² Thus, a low ratio of cells reflecting a productive adaptive immune response to “pro-tumor” phagocytic myeloid cells may contribute to resistance.^{63 64} In addition to functionally inhibiting effector cells via contact-independent mechanisms,⁵² these suppressive immune cells can also express PD-L1 themselves, with some reports suggesting that immune cell PD-L1 expression (eg, combined positive score; CPS) may be more predictive of response to PD-(L)1 therapy than that of tumor cell expressed PD-L1 (eg, tumor proportion score; TPS or tumor cell; TC expression).65,67 Suppressive cell types can also secrete inhibitory factors such as indoleamine-pyrrole 2,3-dioxygenase and adenosine. These alter the local metabolic environment and support TC consumption of nutrients at the expense of T cells, which prevents their proliferation and survival.⁶⁸ These and other soluble factors, including transforming growth factor beta, vascular endothelial growth factor (VEGF), and chemokine ligand 18, are generally anti-inflammatory and contribute to a suppressive TME that has been associated with resistance.⁶⁹ Other soluble factors can facilitate the recruitment of suppressive cells into the TME. For example, interleukin (IL)-8 recruits neutrophils into tumors and has been associated with non-response to PD-(L)1 therapy in multiple studies.70,72 While these represent plausible biomarkers that could be used to identify resistance to PD-(L)1 inhibitors, they also represent areas under active investigation for therapeutic targeting.

Finally, if effector cells of sufficient quality are generated, present within the tumor, and outcompete suppressive/inhibitory signaling, there remains the possibility that TCs have intrinsic insensitivity to killing as well as direct PD-1/PD-L1 pathway evasive mechanisms. TRAF2 loss lowers tumor necrosis factor cytotoxicity thresholds and increases T cell-mediated TC apoptosis.⁷³ In contrast, high TRAF2, anti-apoptotic proteins, intracellular scaffolding structure/stiffness⁷⁴ and TRAF2/CCND1 amplification have been associated with resistance to TC death.⁷⁵ Other mechanisms of resistance to effector cell killing include extracellular matrix stiffness,⁷⁶ serpin B9,^{77 78} autophagy,^{79 80} cellular CASP8 and FADD-like apoptosis regulator (c-FLIP),⁸¹ and caspase inhibition.⁸² All, except caspase inhibition, have been linked to PD-(L)1 resistance in clinical settings and preclinical combination therapy models. In addition to inhibiting T-cell effector activities, PD-L1 can act as an anti-apoptotic shield⁸³ and recruiter of MDSCs into the TME.⁸⁴

Other tumor intrinsic mechanisms of PD-(L)1 evasion include the secretion of exosomal vesicles with high levels of PD-L1 that can act as target decoys for therapeutic PD-1 antibodies. Patients with high circulating levels of exosomal PD-L1 are more resistant to PD-1 therapies.⁸⁵ Secreted PD-L1 decoy molecules have also been reported to mediate resistance to PD-(L)1 blockade therapy.^{86 87} Tumor intrinsic PI3K and PTEN signaling,⁸⁸ activating mutations in WNT/β-catenin,⁸⁹ Ras oncogene,⁹⁰ and hedgehog signaling⁹¹ have been associated with PD-L1 upregulation on TCs. While it may be counterintuitive that target expression could be associated with resistance to PD-(L)1 therapy, this type of tumor-driven expression of PD-L1 likely occurs in the absence of IFN-γ signaling, appropriate tumor antigen presentation (MoR 1, figure 1), or proximal TILs/antigen-presenting cells(APCs) (MoR 2, figure 1).⁹² Alternatively, the relevance of PD-L1 as a mechanism of immune evasion can be lost through reduced PD-L1 expression via mutations in IFN-γ receptor signaling pathway genes. Truncating mutations in JAK1/2 were associated with resistance to PD-1 blockade in patients with melanoma¹⁹ and loss-of-function mutations in these genes can occur in patients receiving immunotherapy.⁹³

While data are needed to support the clear association of the above mechanisms of resistance with primary versus secondary resistance outcomes, it is tempting to speculate which ones may be associated with each manifestation of resistance. Cold TMEs that lack effector T-cell infiltration (ie, no antitumor T cells in the tumor for PD-(L)1 agents to work on), tumors that lack immunogenic antigens or have defects in antigen presentation pathways (ie, lack of neoantigens or somatic mutations in B2M or JAKs), or tumors with mutations in IFN signaling pathways (ie, cannot upregulate PD-L1 or antigen-presenting molecules) are plausible candidates for mediating primary resistance. In addition, TCs that are intrinsically resistant to effector cell cytotoxicity are likely linked to primary resistance, as releasing checkpoint inhibition of T cells will not be sufficient. However, if such TCs represent a subclonal population within the tumor that is present during the initiation of treatment, they may be enriched after sensitive TCs are killed off by T cells and present as acquired resistance. Secondary resistance to PD-(L)1 blockade could be linked to adaptive upregulation of alternate checkpoints (ie, LAG-3, TIM-3, TIGIT) or expansion of countersuppressive immune cells during an immune attack. One could envision that these latter mechanisms could be engaged as the metabolism and tumor architecture within the TME changes during an antitumor immune response. Given the complexity of the immune response and the heterogeneity observed between patients and between lesions, it may ultimately be impossible to clearly segregate such mechanisms into drivers of primary versus secondary resistance.

Tissue collection considerations for PD-(L)1 biomarker development

There are several practical considerations to sampling tissue biopsies for the purposes of elucidating biomarkers for PD-(L)1 therapies (box 1). Typically, an incisional biopsy is taken prior to therapy, either at the time of diagnosis or at pretreatment from a single tumor lesion. The sampled lesion is generally not measured radiographically during the trial for fear of bias or change in tumor appearance. Thus, the treatment response of that individual lesion must be inferred from the response of other lesions, and as discussed previously, patients can have a heterogeneous response complicating the interpretation of biomarker data. Furthermore, any biopsy will, by definition, only be sampling a fraction of the tumor, and therefore cannot capture heterogeneity within a single tumor and/or between multiple metastatic tumor sites. In addition, the biopsy may completely miss the region of the tissue most important for mediating tumor-immune responses.

Box 1. Best practices for biopsy collection and processing for programmed death-ligand 1 biomarker discovery.

• Plan biomarker work before sample collection begins.

• Match preservation method to biomarker profiling strategy.

• Standardize the type of biopsy and timing of collection.

• Minimize variability in sample handling during collection and processing.

• Plan for how heterogenous biopsies will be sampled and analyzed.

• Develop a statistical analysis plan to ensure the study is appropriately powered.

• Consider the utility of liquid biopsy for serial sampling.

• If possible, biopsy on-treatment or at the time of progression to capture the unique biology of post-programmed cell death protein-1 tumor.

• Consider potential clinical utilization in practice and diagnostic development path early in study design.

Another challenge in developing biomarker assays is in the standardization of the biopsy collected. As the original materials collected from a patient affect all downstream analyses, and because the conditions under which biopsies are collected are often not uniform, technical artifacts can compromise future biomarker development work on these samples. The standard tissue preservation method is still formalin fixation and paraffin embedding (FFPE), but variations in the time between tissue collection and fixation, and the time in formalin fixation, can also lead to variability via tissue degradation, nucleic acid degradation, and/or protein epitope masking.⁹⁴ Furthermore, FFPE tissue preservation is not ideal for all downstream profiling methods, particularly RNA profiling assays like sequencing, where fresh frozen tissue tends to generate higher quality data but at the expense of providing high-resolution tissue architecture.⁹⁵ Guidelines from biobanking bodies like the International Society for Biological and Environmental Repositories and others, which are aimed at creating policy for correct tissue processing methods for sample preservation and harmonization, should ultimately be followed whenever possible to minimize sample variability across biomarker studies.⁹⁶

Even if specimens are perfectly preserved, another important factor leading to different results is the type of sample collected. Excisional biopsies or surgical resection, where a tumor lesion is removed in its entirety, generates a great deal of material for downstream profiling, but is more sensitive to post-collection processing artifacts due to the larger sample volume (PMID: 12466110). Also, as it is frequently not feasible or desirable to profile the entire tissue, the selection of a subset of the tissue for profiling may introduce another source of variability in downstream assay results. The standard profiling strategy (incisional biopsies where only a fraction of the tumor is collected) results in a small amount of tissue for profiling, necessitating a strict prioritization of the profiling assays that will be performed. Likewise, important regions of the tumor may be omitted from sampling because of tissue heterogeneity. While less invasive biopsy methods can provide earlier sampling for profiling response/non-response signals, they are not yet able to provide robust information on the TME. Less invasive core needle biopsy (CNB) is another type of tumor sample collection method that uses a hollow-core needle inserted into the tumor for sample withdrawal. Like excisional biopsies, CNB preserves TME architecture, but a very small fraction of the overall tumor tissue collected using this method can usually be expected to lead to a high likelihood of missing tumor margins and other important features of the tumor biology. This limits the amount of material available for downstream processing for response signal profiling. Discordance of numerous biomarkers has been observed in comparing CNB and surgery removed samples, as caused by tumor heterogeneity, air exposure, specimen fixation processes, and CNB procedures.⁹⁷ To address intra or extratumor heterogeneity from large or small specimens, regions of interest for other biomarkers can be selected in immune-enriched areas such as CD45 expression gating.98,100 Prospective studies using well-defined immune markers have demonstrated that single-region tumor samples using immunohistochemistry (IHC) and RNA sequencing had high concordance rates (83.3%) between tumor immune microenvironments within a tumor.¹⁰¹ Finally, an even less invasive type of tumor sample collection is non-terminal tumor sampling using fine needle aspiration (FNA), where a very fine needle is inserted into the tumor and cells aspirated.¹⁰² This can provide longitudinal TME sampling to identify biomarkers of response or resistance to different immunotherapies.¹⁰³ While minimally invasive, this produces a very small sample of often dissociated cells, where again the TME architecture is not preserved, making this approach frequently inadequate for meaningful analysis of the TME.¹⁰⁴ However, with pembrolizumab receiving full Food and Drug Administration (FDA) approval for treatment of select patients with a variety of mismatch repair deficient (dMMR)/microsatellite instability high (MSI-H) solid tumors, endoscopic ultrasound-guided fine-needle biopsy that can be used to determine dMMR and quantify PD-L1 expression may finally be able to gain some ground as a CDx technology.¹⁰⁵

The least invasive profiling method is liquid biopsy (LBx), for example, collecting blood from the patient, which can then be profiled for circulating TCs, circulating tumor DNA (ctDNA), other circulating factors, or changes in circulating immune cells that correlate with tumor status.¹⁰⁶ LBx offers three advantages over tissue biopsies: they produce a large amount of sample for testing, their collection is minimally invasive allowing repeat sampling as necessary, and they are a single sample that integrates signals from all lesions in the setting of metastatic disease, reducing the complexity of measuring a heterogeneous response.¹⁰⁷ While LBx has the potential to measure ctDNA, PD-L1 expression on circulating TCs, soluble plasma or serum PD-L1 and PD-1, PD-1 expression on circulating lymphocytes, and T-cell receptor (TCR) repertoire,¹⁰⁸ LBx contains no direct information about the TME, limiting its utility in elucidating some mechanisms of resistance (eg, MoR2, figure 1). With the advent of computational/artificial intelligence (AI)-driven methods that can perform deep profiling across several downstream platforms, the coupling of less invasive LBx and FNA samples may eventually pave the path forward to determining robust immune checkpoint inhibitor (ICI) response signals from patients.^{109 110} The least invasive LBx method is likely used to detect residual disease via ctDNA,¹¹¹ and could also provide the earliest signals of response/non-response due to its compatibility with frequent sampling in cohorts whose populations have differing response rates. Indeed, rising ctDNA levels from baseline to 6–7 weeks in solid tumor treated patients was shown to be 97.5% predictive of pembrolizumab resistance of solid tumors,¹¹² and in other cases was shown to either correlate with,¹¹³ or precede PD as assessed by radiologic imaging.^{114 115} However, LBx is not able to tease out MoR2 (figure 1), as access to the tumor or its location in the tumor (excluded vs infiltrated) could not be interrogated.

Current biomarkers of sensitivity to single-agent PD-(L)1 therapy

Despite intensive study into the mechanisms of resistance to PD-(L)1 blockade, the development of biomarkers able to predict response or resistance is complex. To date, there are multiple companion or complementary assays approved by the FDA to inform expectations of the likelihood of response to PD-(L)1 inhibitors. These fall into three categories: PD-L1 expression, MMR/MSI-H status, and tumor mutation burden (TMB) levels.¹¹⁶ However, the predictive power of all the approved assays is suboptimal and there remains an urgent need for more predictive biomarkers for PD-(L)1-based therapies, including monotherapy, combination therapies, and particularly biomarkers to predict immunotherapy response after PD-(L)1 failure.

PD-L1 IHC

As noted previously, PD-L1 may be expressed by TCs, multiple immune cell types, and other cell lineages within the TME and can be expressed constitutively or in an adaptive fashion in response to IFN-γ. It is the adaptive pattern of PD-L1 expression within the TME that is believed to be most closely associated with response to PD-(L)1 blockade,¹¹⁷ and the pattern of PD-L1 serves as a surrogate marker of an ongoing adaptive immune response to a tumor that may be potentiated by therapeutic antibody administration.

PD-L1 IHC is currently the most widely used biomarker for predicting response to PD-(L)1 blockade. Across multiple tumor types, approximately 45% of patients whose tumors are PD-L1 positive respond to therapy, while 15% of patients who are PD-L1 negative still respond, indicating that PD-L1 assay positivity may be used to help enrich for response, but is not essential.¹¹⁸ There are currently several commercially available PD-L1 IHC assays, which are commonly referred to by the antibody clone names that are used to detect PD-L1 protein in the assays. These include 22C3 (pharmDx, Dako North America), 28–8 (Agilent Dako North America), SP142 and SP263 (Ventana Medical Systems). Regarding staining performance, they are all generally equivalent in their ability to highlight TCs and immune cells (ICs), except for the SP142 assay, which shows a lower sensitivity for both.^{119 120} Additional challenges posed by PD-L1 IHC include limited standardized controls for broad use,¹²¹ the varied scoring systems for pathologist visual interpretation (scoring TC and/or IC expression with different thresholds of positivity for different tumor types and lines of therapy),122,125 and the fact that PD-L1 expression within the TME is spatially and temporally heterogeneous, among others.^{126 127} Furthermore, current PD-L1 IHC assays rely on absolute percentage cut-offs for measuring expression and do not account for the fact that the biomarker association with efficacy may not be uniform across the range of expression captured this way. Thus, the use of methods to capture the biomarker as a continuous variable may be necessary to optimize the predicative power.

Microsatellite instability/mismatch repair deficiency and tumor mutational burden

MSI/MMR and TMB are the other FDA-approved biomarkers for PD-L1 blockade therapy, with three approved indications for MSI/MMR status and one for TMB-high status.128,131 MSI is a molecular test measuring the frequency of small insertions or deletions (indels) within specific regions of the genome that are prone to copy errors when there are defects in the MMR pathway.¹³² dMMR is measured by loss of expression of one of the four genes in the MMR pathway. MMRd and MSI are highly correlated.¹³³ TMB status is a sequencing-based measurement of the mutations within the tumor and is not specific to defects in any one pathway.¹³⁴ Both MSI/MMR and TMB correlate with an elevation in variants (single nucleotide variants, insertions or deletions (indels)) present within a tumor, whether due to mutation in a DNA repair pathway, extensive ultraviolet radiation (UV) damage, or other causes. They are a surrogate for predicting the number of putative immunogenic neoantigens and thus likelihood of response to immunotherapy. In 2017, the FDA approved MSI/MMR status as a biomarker of response to pembrolizumab in solid tumors regardless of tissue of origin, representing the first tumor-type agnostic indication.^{128 130} This was followed by another tumor-type agnostic approval for pembrolizumab in 2020 for patients with high TMB, with the approval of the FoundationOne CDx test for patients with solid tumors harboring a TMB>10 mut/Mb.^{128 130}

MSI and TMB status may be determined by next-generation sequencing or PCR testing,¹³⁵ and IHC can be performed to detect the loss of proteins in the mismatch repair pathway¹³⁶ While complementary, utilization of TMB, PD-L1 and MSI-H in combination may have the potential to predict PD-(L)1 responsiveness better than each alone.¹³⁷ The limitation of MSI testing includes the simple incidence of this finding—there is a relatively low number of patients with this marker.^{138 139} TMB assessment has proven more challenging from a laboratory standpoint due to the lack of assay standardization and associated bioinformatic interpretation for this emerging biomarker.¹⁴⁰ Alternative approaches include the use of a limited set of genes to estimate overall TMB.¹⁴¹ Like PD-L1 IHC, TMB has been shown to have suboptimal sensitivity and selectivity as a biomarker for PD-(L)1 response,^{141 142} which is not surprizing given that antigens need to be properly presented in the context of MHC which is difficult to assess using existing methods.

Next-generation biomarkers

There are many additional approaches that are being explored for the prediction of immunotherapy response. These include assessments of the pretreatment tumor for evidence of host-tumor interactions, for example, gene expression studies or multiplex immunofluorescence (mIF)/IHC, assessments of the on-treatment tumor for the determination of tumor clearance, or assessments in the peripheral blood, the latter of which may potentially be performed at numerous points throughout the course of therapy, for example, with ctDNA.

Gene expression profiling has most commonly focused on an IFN-γ gene signature. Such a signature has been associated with response/resistance to anti-PD-(L)1 in multiple tumor types.³⁷ IFN-γ gene signature and TMB, as well as PD-L1 and TMB, have been shown to be independent biomarkers,^{143 144} and various combinations of these have demonstrated predictive performance for response to immunotherapy over the individual components alone.145,147 mIF/IHC is another form of combinatorial biomarker that allows for the simultaneous detection of multiple protein targets in tissue, allowing for quantitative assessments of spatial arrangements or complex cell phenotypes with single-cell resolution.^{67 148} A meta-analysis suggested that mIF/IHC assays have higher predictive value than PD-L1, TMB, or IFN-γ gene signatures alone or in combination for predicting response to anti-PD-1-based therapy.¹⁴⁷ Recently, a CD163+PD-L1− macrophage subset detected by multiplex IHC was shown to confer a tumor phenotype that was resistant to anti-PD-1-based therapy in patients with advanced melanoma.⁶⁷

On-treatment pathology specimens also provide a rich opportunity for determining which patients’ tumors are resistant to immunotherapy. In these instances, rather than using a patient’s diagnostic pretreatment biopsy to match them to a given therapy, a biopsy or surgery is performed after the patient has already received therapy and the pathology material is then used to assess whether the therapy is demonstrating any effect. This technique is used most frequently to characterize neoadjuvant treatment, where the definitive resection specimen from surgery has been exposed to therapy and the vast majority of the tumor is present for evaluation for evidence of treatment effect. Biomarker features in these specimens have been shown to predict 2–3 years event-free survival in multiple tumor types.^{127 149 150} This approach is also used to identify primary resistance in patients with advanced, metastatic disease.^{151 152}

Biomarkers of resistance to combination therapies with PD-(L)1 inhibitors

While the biology of resistance to single-agent PD-(L)1 therapies has been extensively studied and many potential mechanisms identified, the field’s understanding of resistance to combination therapies remains poor. MSI-H/TMB and PD-L1, rough proxies of tumor immunogenicity and an adaptive immune response, are clinically actionable biomarkers to single-agent anti-PD(L)1 therapies. However, in the setting of combination therapies, some patients may respond to one or both agents being used. In cases where the mechanism of each therapy is distinct or even orthogonal, biomarkers of response to single-agent PD-(L)1 may no longer be sufficient.

This concept has been observed in several clinical settings. In patients with NSCLC treated with pembrolizumab, low TMB was associated with resistance to single-agent treatment,153,155 whereas in patients treated with the combination of pembrolizumab and chemotherapy, TMB was no longer associated with response.^{156 157} Similarly, in patients with NSCLC treated with bevacizumab, carboplatin, and paclitaxel with or without atezolizumab, patients with PD-L1 low or negative and low Teff gene signature expression experienced an improved response to the atezolizumab-containing combination relative to control, although the response was enriched in patients with biomarker-high tumors.¹⁵⁸ These results suggest that the combination of chemotherapy with PD-(L)1 blockade conferred benefit to patients who would not normally be expected to respond to single agent PD-(L)1 blockade. Similarly, in patients with endometrial cancer treated with pembrolizumab and the tyrosine kinase inhibitor lenvatinib, patients responded irrespective of MSI status. Finally, while response to PD-(L)1 inhibitors, such as pembrolizumab and atezolizumab, is enriched in patients with urothelial carcinoma whose cancers express PD-L1, when pembrolizumab was used in conjunction with enfortumab vedotin, PD-L1 IHC no longer associated with response.¹⁵⁹

The role of PD-(L)1 biomarkers in combinations is still undergoing elucidation. Acknowledging the challenges of TMB as a biomarker as discussed above, in some contexts TMB was associated with the activity of nivolumab and ipilimumab combination therapy.¹⁶⁰ However, data from a Phase 3 study in first-line NSCLC were ambiguous on this point.¹⁶¹ TMB has not yet been reported in the definitive study of the anti-LAG-3 agent relatlimab with nivolumab, but efficacy benefit was enriched in the PD-L1-positive tumors relative to PD-L1 negative tumors in both the nivolumab monotherapy and nivolumab plus relatlimab arms.^{32 162}

The neoadjuvant setting provides an ideal paradigm for examining resistance to combination therapies, acknowledging that early-stage disease does not necessarily mirror the biology of advanced cancer. For example, a neoadjuvant study of nivolumab and ipilimumab in colorectal cancers demonstrated a pathological response in all dMMR tumors and 27% of mismatch repair-proficient tumors.¹⁶³ This is higher than the response rates seen with advanced colorectal cancer (69% in MMRd tumors¹⁶⁴). Similarly, the expansion of tumor-resident T-cell clones was more pronounced in the neoadjuvant than the adjuvant setting in stage III melanoma.¹⁶⁵ Acknowledging the likelihood that mechanisms of resistance may differ between the neoadjuvant and metastatic setting, neoadjuvant studies have begun to contribute to the field’s understanding of resistance to combination therapies. A study of neoadjuvant chemotherapy with nivolumab in NSCLC enabled in-depth characterization of IC subtypes in the peripheral blood associated with response and by extension resistance to therapy.¹⁶⁶ A study of neoadjuvant ipilimumab and nivolumab in melanoma found the combination of low IFN signature and low TMB was associated with a lower likelihood of response,¹⁴⁵ as did the absence of a CD4/IL-2 signature.¹⁶⁷ A separate study of nivolumab monotherapy versus ipilimumab plus nivolumab combination also identified that CD8 T-cell infiltration, immune gene signatures, and TCR clonality were associated with response.¹⁶⁸ Thus, advances in this area will likely include the application of composite biomarker approaches as we move beyond the narrow assessment of only single biomarkers.

As difficult as the study of resistance to single agent ICI has been, this challenge is amplified in the combination setting given the diversity of combination partners that have been and will be tested, and the distinct mechanisms of each. As of February 2022, over >5,600 clinical trials of checkpoint-based therapies were planned or actively ongoing.169,172 While the majority (>4,600) are in combination with chemotherapy, VEGF/R blockade, CTLA-4, or radiotherapy, over 2,500 studies were being conducted with other combination partners. Ideally, in addition to smaller academic studies, the sponsors of definitive clinical trials will choose to invest in the study of resistance mechanisms in larger studies. This may reduce the risk of false discovery from analyzing high-dimensional data in underpowered trials. In addition, the study of patients treated in the standard-of-care setting within larger health systems capable of tumor banking may enable such studies if not performed by industry sponsors. The fruitful examples of such studies outside of the immuno-oncology space may be instructive; for example, studies of resistance to first-generation epidermal growth factor receptor (EGFR) tyrosine kinase inhibitors,¹⁷³ or antibody-drug conjugates,¹⁷⁴ led to the successful development of improved agents. The collection and analysis of ctDNA may also facilitate the examination of DNA-based mechanisms of resistance in situations where collection of tumor tissue is prohibitive.175,177

Clinical actionability for post-PD-(L)1 therapy: challenges and potential solutions

Clinical actionability is the decision to prescribe a therapeutic intervention that can effectively prevent or delay clinical disease or burden to improve patient outcome.¹⁷⁸ In immuno-oncology, the complexities of the TME can influence both drug mechanisms of action and the accompanying molecular biomarkers.¹⁷⁹ As described above, the current biomarkers to predict response to PD-(L)1 are suboptimal, and to date no biomarkers offer guidance for treatment in the post-PD-(L)1 setting. On-treatment and post-progression biopsies offer hope because they capture the state of the tumor and TME after PD-1 therapy has been initiated and the TME may have experienced a concomitant change that can be associated with response or resistance. Therefore, these biopsies provide the greatest insight into the response or lack thereof. When collecting on-therapy biopsies, careful timing is needed to obtain a sample when the biomarker signal is likely to be maximized while ensuring that enough tumor tissue remains for profiling. This is further complicated when comparing cohorts predicted to have differing response rates. Additionally, there is the risk that the genetic makeup of the tumor and/or the cellular composition of the TME has changed in the time between when the original biopsy was collected and when PD-1 was determined to have failed. For example, biopsy samples obtained at different time points had an inverse correlation between PD-L1 expression and tumor volume.¹⁸⁰ Finally, biopsies from the time of progression can theoretically pose additional risk to the patient by augmenting the dissemination of metastatic TCs, seeding new lesions.¹⁸¹ Post-treatment biopsies are seldom reimbursed by health insurance,¹⁸² and consequently, post-treatment biopsy is not routine clinical practice, despite being a promising source of information about the immunological status of the tumor after PD-(L)1 treatment and failure.

Beyond challenges with sample collection, the overall process of developing biomarkers is complex. In the first place, there is often a misalignment of incentives between those involved throughout the process who believe some therapeutics could bring benefit to a broad population, and others who would like to witness the biomarkers used to direct or redirect therapies to those patient populations that are most likely to benefit. Because no diagnostic assay is perfectly predictive, the risk of false positives and false negatives must be weighed in decision-making for individual care. Consequently, diagnostic assays have found success in clinical indications and/or patient populations for which regulatory approval would otherwise not be possible. First, when the response rate in an unselected patient population is low but a given biomarker can identify a subset of patients with an increased likelihood of response, a diagnostic assay with a very low false negative rate can still be used to select the patients with any propensity to respond. The trade-off in such cases is that the low false negative rate is often accompanied by a high false positive rate (some patients who are biomarker positive may not respond to the treatment). Second, biomarkers can find success when the standard of care has moderate activity, and a follow-on therapy is seeking approval by showing improved performance in a biomarker selected subset of the population. In this case, a diagnostic can be developed to maximize the positive predictive value where the positive patients have the highest likelihood of benefit, with the trade-off being a higher negative predictive value (some patients who would respond are denied access to the therapy). This is the scenario under which cemiplimab was approved for metastatic NSCLC when patients’ tumors have a PD-L1 expression on ≥50% of TCs. In addition, regulatory considerations for diagnostics development foster a path to the simplest possible solution. For example, while many studies indicate that dual biomarkers are associated better with response to PD-(L)1 therapies, the prospective development of such selection methods does not have a regulatory precedent and compounds the complexities described above. However, this approach may be worthwhile to pursue as it would allow the assessment of both a biomarker related to efficacy from PD-(L)1 compounds, as well as another linked to the specific resistance mechanism the combination partner is intended to overcome. Composite biomarker approaches could also allow for simultaneous characterization of relevant ICs (PD-L1 CPS and other novel biomarkers associated with T cells) in addition to the existing tumor-focused biomarkers (eg, PD-L1 TPS, TMB, MMR, ctDNA).

In an ideal scenario, a single biomarker assay could be developed and provided to patients at the time of diagnosis (ie, a multidimensional assay or combination of assays) that would characterize the mechanisms of resistance at play in the tumor to inform about its propensity to respond to PD-(L)1 targeting monotherapy, combination therapies, or follow-on treatments at the time of PD-(L)1 failure. However, without a detailed understanding of these mechanisms of resistance or prospectively-tested multi-parameter diagnostic tests, this goal state remains elusive. Today, PD-(L)1-based combination therapies with chemotherapy, tyrosine kinase inhibitors, or CTLA-4 are the most commonly used immunotherapy treatments in solid tumors, with a myriad of additional combinations under development, which further complicates the field’s ability to define mechanisms of response and resistance. In a future state, the patient and physician may decide on therapeutic choices (including combination and/or stepwise approaches), based on indication-specific biomarker considerations.

Conclusion

As PD-(L)1 therapy continues to move into more disease settings (earlier lines of therapy, earlier stages of disease) there will be more patients treated with it, but it is accompanied by meaningful risks such as side effects (both acute and chronic) and opportunity cost to pursue other treatment options. In parallel, more promising IO agents are emerging and combination therapies are being pursued in the clinic. Thus, the need for effective biomarkers to predict response in the post-PD-(L)1 setting is greater than ever. The recognition that biomarkers play an important role in drug development is spurring more investment in biomarker discovery and assay development, but fundamental challenges to the field make the problem still difficult to overcome. Investment in greater access to samples and novel profiling strategies across the next generation of immune biomarkers studies will overcome these challenges and significantly advance the state of the art in patient care.