Abstract
PD-L1 expression levels derived from >16,000 samples guided the selection of tumor types likely to benefit from pembrolizuamb monotherapy in clinical trials. Although not fail-proof, FDA approvals for most of the prioritized indications speak to the power of RNA expression profiling and the value of large genomic datasets.
See related article by Ayers et al., p. 1564
In this issue of Clinical Cancer Research, Ayers and colleagues (1) showed how tumor types can be more effectively selected for clinical trial (KEYNOTE-012) based on gene expression signatures in large cohort of samples. Immune checkpoint inhibitors have drastically changed the treatment landscape of many cancer types. The PD-1 inhibitor pembrolizumab has been approved by the FDA for the treatment of eight different tumor types in a variety of settings, and for any tumors that are MSI-H or MMR deficient. The veritable plethora of currently ongoing clinical trials with immune checkpoint inhibitors may give the impression that little thought was put into the selection of cancer types included in these trials. In contrast, this article shows how gene expression signatures helped to prioritize tumor types after the initial successful phase I clinical trial with pembrolizumab.
In collaboration with the H. Lee Moffit Cancer Center, this investigative team from Merck assessed the expression of PD-L1 and coexpressed genes in about 16,000 cases representing more than 25 tumor types. They confirmed the robustness of the profiles using expression data from 10,000 cancers in The Cancer Genome Atlas (TCGA) cohort. A PD-L1 expression threshold of the 75th percentile or higher was selected on the basis of deviation from normal; association with response in melanoma, non–small cell lung cancer (NSCLC), and renal cell carcinoma; and separation on expression between MSI and MSS colorectal tumors. Different tumor types were then ranked in tiers (1 to 3) based on the proportion of cases with PD-L1 expression above this threshold. The highest proportion of cases with PD-L1 expression above the 75th percentile included both primary and metastatic head and neck cancers and primary squamous cell lung cancer, metastatic melanoma, and primary colorectal MSI. Tumors in the tier 2 group included metastatic melanoma and metastatic triple-negative breast cancer (TNBC) and primary bladder cancer. To enrich the analysis of tumors that are less prevalent in the United States, gastric cancer specimens from the Asian Cancer Research Group were analyzed and demonstrated that the proportion of cases with high PD-L1 expression was in the tier 2 group and differed on the basis of EBV and MSI status (Fig. 1). On the basis of these results, the KEYNOTE-012 clinical trial included patients with TNBC, head and neck cancer, bladder cancer, and gastric cancer.
Fraction of PD-L1 mRNA expression in large sample sets enabled prioritization of tumor types. Those in tier 1 were more likely to respond to pembrolizumab monotherapy. Better biomarkers are still needed to foretell which patient will actually do better or worse with the drug.
Fraction of PD-L1 mRNA expression in large sample sets enabled prioritization of tumor types. Those in tier 1 were more likely to respond to pembrolizumab monotherapy. Better biomarkers are still needed to foretell which patient will actually do better or worse with the drug.
Since the completion of KEYNOTE-012 and subsequent clinical trials, pembrolizumab has been approved by the FDA for three of the four selected indications that were prioritized for development. There is no approval yet for TNBC. Whereas the approval of three of these four indications is a testament to the strength of this genomics-driven approach, the lack of approval for TNBC speaks about the complexity of antitumor immunity and predicting which patients benefit from these immune checkpoint inhibitors based on PD-L1 expression alone.
Using a similar strategy, a clinical trial with the PD-L1 inhibitor atezolizumab identified an association with T-effector gene expression and overall survival in patients with NSCLC (2). Accordingly, this T-effector gene signature has been added as an endpoint in subsequent NSCLC clinical trials with atezolizumab (3). Others have explored tumor mutation burden as a surrogate for neoantigens. The survival benefit identified in patients with a high mutation burden treated with the PD-1 inhibitor nivolumab (4) has led to modifications of a clinical trial testing the combination of nivolumab and ipilimumab in NSCLC to include survival based on TMB categorizations (5).
At this point, many of the approved clinical indications for checkpoint inhibitor therapy are based on IHC positivity tumor cells. As such, pembrolizumab has been approved for the treatment of urothelial tumors with a combined positive score (CPS) of 10% or greater based on IHC, or for the treatment of gastric cancers with a CPS of 1% or greater. Pembrolizumab received accelerated approval by the FDA for the treatment of head and neck cancer regardless of PD-L1 protein status. Although these approvals suggest that pembrolizumab has efficacy against these tumors despite the absence of or low levels of detectable PD-L1 protein (or that testing might not be necessary in tumor types with high proportions of PD-L1–expressing tumors), the use of PD-L1 mRNA status as a biomarker was not presented. In light of the study by Ayers and colleagues, it may be valuable to determine whether a PD-L1 RNA–based assay can be more reliable at selecting for patients who will likely benefit from the treatment. Alternatively, pembrolizumab may be more effective than the comparative treatments regardless of tumor cell expression of PD-L1.
Ayers and colleagues note in their analysis that there was significant variability of PD-L1 mRNA expression and associated signatures between and within tumor types. Similarly, the assessment of primary and metastatic lesions demonstrated some discrepancies in the degree of PD-L1 RNA expression by tumor location. We have also shown that there can be significant heterogeneity of PD-L1 protein expression and tumor lymphocyte infiltration between lesions of multifocal lung cancer, and between paired primary lung cancers and brain metastases (6–8). The heterogeneity and dynamics of PD-L1 expression challenge its use as a predictive biomarker.
Ayers and colleagues also discuss how mutational load related to PD-L1 expression in their datasets, but they do not make clear whether this information informed their clinical trial design decisions. More recent work has suggested that the determination of tumor mutation burden based on nonsynonymous single nucleotide variants (SNV) may not adequately capture neoantigenic load because many of these passenger changes are not expressed and may not bind HLA molecules strongly. Furthermore, insertions and deletions (indels) that result in frameshifts in renal cell carcinoma may be much more immunogenic than nonsynonymous SNV mutations and may not be captured in current TMB estimations (9). Similarly, expression of chromosomal rearrangements may result in neoantigens that are not identified by standard sequencing approaches and are not included in current TMB calculations (10). A more inclusive determination of neoantigenic burden may be a better predictor of response based on the combined mutational signatures (expressed SNVs, indels, chromosomal rearrangements) of the tumors in question. Like PD-L1 mRNA and protein expression, neoantigenic load (as estimated by TMB) only represents one component of the complex mechanisms required for an effective antitumor response. In this regard, a more comprehensive cancer immunogram that assesses not only immune checkpoint expression and neoantigenic load, but also immune cell trafficking and tumor infiltration, presence of soluble inhibitors or inhibitory tumor metabolic factors and other factors (11).
Despite the unprecedented success of immune checkpoint inhibitors, many critical issues remain to be addressed. These treatments sometimes result in severely debilitating or even fatal side effects including colitis, pneumonitis, and myocarditis. In addition, there are patients who may experience hyperprogression of disease while on immune checkpoint inhibitors. Identification of risk factors and biomarkers that can predict which patients may experience severe autoimmune toxicities or hyperprogression of disease could result in general improvement of the therapeutic index for this class of therapy. To identify these biomarkers and risk factors, it may take equally large clinically annotated genomic datasets of patients treated with immune checkpoint inhibitors as those that were used to select tumor types to prioritize for clinical trials.
In summary, the work by Ayers and colleagues provides insight into how gene expression guided selection of tumor types for development with pembrolizumab in KEYNOTE-012. The resultant FDA approvals for three of the four selected indications accentuate the strength of this approach and highlight the value to generate and use large, comprehensive genomic datasets to help identify tumor types for inclusion in clinical trials. However, the complexity of the antitumor immune response challenges the use of a single biomarker to predict whom to treat with immune checkpoint inhibitors. Furthermore, studies are also urgently needed to preidentify patients who might not benefit from these agents so that they can pursue alternative therapeutic opportunities.
Disclosure of Potential Conflicts of Interest
A.S. Mansfield reports receiving commercial research grants from Novartis (paid to Dr. Mansfield's institution) and has served on advisory boards for AbbVie, Genentech, and Bristol-Myers Squibb (honoraria paid to Dr. Mansfield's institution). J. Jen is an employee of Celgene. No other potential conflicts of interest were disclosed.
Authors' Contributions
Conception and design: A.S. Mansfield, J. Jen
Analysis and interpretation of data (e.g., statistical analysis, biostatistics, computational analysis): A.S. Mansfield, J. Jen
Writing, review, and/or revision of the manuscript: A.S. Mansfield, J. Jen
Study supervision: J. Jen
Acknowledgments
A.S. Mansfield is supported by the NCI (grant no. P30 CA 15083). J. Jen is supported in part by the NCI (grant no. P30 CA 15083).