Gene expression profiling to predict responsiveness to immunotherapy

Response to immunotherapy in the context of the intratumoral genomic landscape

The immunologic and biological mechanisms underlying response or resistance to immune-checkpoint blockade remain poorly understood. Despite observations of clinical benefit from blockade in many cancer types, pretreatment correlates of response are incompletely characterized. Immune checkpoint blockade is hypothesized to function by attenuating immunosuppressive signals, thus releasing the brakes and activating antigen-specific immune-mediated destruction of tumor cells.(31, 32) This pro-inflammatory tumor microenvironment can be characterized through genomic and transcriptomic analyses and correlated with response to therapy.

Common characteristics of inflamed tumors responsive to anti-PD1 therapy include dense CD8+ T-cell infiltrates, a broad chemokine profile, PD-L1 expression on immune cells, a type 1 interferon (IFN) signature, and elevated expression of IFN-gamma induced genes.(33, 34) High cytolytic activity is also correlated with expression of IFN-stimulated chemokines (CXCL9, CLCL10 and CSCL11) capable of attracting T cells to the tumor site.(35) For anti-CTLA-4 therapy, early on-treatment proliferation of ICOS+ CD4+ T-cells can be used as a pharmacodynamic biomarker for monitoring of treatment efficacy.(36) In melanoma several cytokine and chemokine signaling genes, as well as the cytotoxic factors granzyme A and perforin 1, are differently expressed prior to ipilimumab therapy between responders and non-responders.(37, 38)

IHC-based studies support the notion that the density and location of CD8+, CD4+, PD-1+, and PD-L1+ in tumor cells and/or tumor-infiltrating immune cells in pretreatment biopsies can predict therapeutic response to immune checkpoint inhibitors.(39, 40) While pre-treatment PD-L1 expression in tumor biopsies does appear to predict response to anti-PD-1 therapies,(10, 34, 41) many tumors scored as PD-L1 positive do not respond, while some responses occur in PD-L1-negative tumors.(12, 42) Such conflicting results likely reflect the limitations related to IHC staining and scoring techniques, as well as to the complexity related to PD-L1 expression in a single biopsy across cells present in the tumor microenvironment, including infiltrating myeloid cells; therefore, better biomarkers are clearly needed to optimize therapeutic decisions. Expression of PD-L1/PD-L2 and CTLA-4 represent mechanisms of immune regulation and opportunities for gene expression-based targeting with immune checkpoint blockade.(34, 38) High pretreatment levels of CTLA-4 mRNA are predictive of response to both anti-CTLA-4 and anti-PD-L1 therapy. Further, amplification of the PD-L1/PD-L2 regions has been demonstrated to be associated with high cytolytic activity in cervical, gastric and colorectal cancer.(35) In Hodgkin’s lymphoma, amplification of PD-L1/PD-L2 is associated with an approximately 90% response rate to nivolumab.(43)

A gene expression profiling mechanism to characterize the tumor microenvironment has been created using a recently developed computational method to characterize tumor-infiltrating immune cell subsets (CIBERSORT) using RNA-sequencing (RNA-seq). This technique addresses many limitations of IHC, has enabled detailed transcript analysis of the tumor leukocytic infiltrate, and is capable of predicting prognosis from pretreatment biopsies.(44) Further, useful insight can be gained into the relationship between regulatory cell types (T-regulatory cells and myeloid derived suppressor cells) and checkpoint inhibition responses.(45)

Another significant benefit of transcriptomic analysis beyond quantitative IHC assessment is the extent of supplementary information assessable regarding inflammation associated gene signatures, oncogenic drivers, checkpoint protein expression and cell subtypes. Yet this approach lacks the spatial and structural fidelity (invading edge, density) afforded by detailed characterization of immune infiltrates which has established prognostic utility(46), therefore a complementary approach may be optimal.

Defects in antigen processing and presentations machinery (APM) are also frequently observed in the tumor microenvironment as an immune evasion mechanism.(35) However, in the report linking microsatellite instability with clinical outcomes following anti-PD1 therapy in colorectal cancer patients, despite 60% of mismatch-repair deficient tumors lacking MHC class I expression, 40% responded to anti-PD1 blockade.(47) Furthermore, MHC class II expression on tumor cells was also associated with response to anti-PD1 in two independent cohorts of patients with melanoma.(48) This suggests that transcriptomic and genomic factors may contribute to the response patterns of PD-1 therapeutics. Mutation of coding regions, promoter regions, epigenetic modification and alterations in IFN-gamma-signaling can all result in APM inactivation,(49) therefore both genomic and transcriptomic analysis are necessary to establish the integrity of the antigen presentation mechanism.

Response to immunotherapy in the context of DNA damage

The accumulation of large-scale cohort-based cancer genome sequence data in multiple cancer types has provided the opportunity to describe and classify mutational signatures, and to infer underlying mechanisms. Rather than being a random assortment of base changes spread across the genome, somatic tumor mutations reflect the sum of mutagenic exposures and mutational processes active during cancer evolution and progression. Following the pioneering work that defined five common mutational signatures active in a breast cancer cohort(50), a large Sanger Institute led collaborative analysed 4,942,984 mutations across 7042 tumors that incorporated 30 cancer types and identified greater than 20 distinct mutation signatures.(50–52) Many signatures were associated with the age of the patient at diagnosis, specific mutagenic exposures, or DNA damage response (DDR) defects, and approximately half were entirely novel signatures.

Both BRCA1 and BRCA2 proteins are integral to cross-linking DNA repair via homologous recombination,(53) therefore deficient cells accumulate DNA double-strand breaks and generate genomic instability with subsequent malignant transformation. Furthermore, they interact with Fanconi Anaemia pathways: FANCC, FANCG and PALB2.(54) Undoubtedly, mutational mapping of cancers, such as in this example, will be correlated with response to immunotherapies. The best example of this concept is in MSI instable colon cancer, where specific DNA mismatch repair machinery mutations have correlated with response to PD-1 inhibition.(47) In addition, the presence of extensive DNA damage is likely to increase the probability of neoantigen formation such that the most mutationally unstable tumors are likely to be the most strongly immunogenic.(52)

Response to immunotherapy in the context of mutational load

Through the development of next generation sequencing (NGS) technologies including whole genome sequencing (WGS) and RNAseq, a high mutational burden of greater than 100 non-synonymous single-nucleotide variants (nsSNV) per exome could be defined and found to be associated with improved clinical outcome in melanoma patients receiving anti-CTLA-4 treatment.(38, 55) This observation was subsequently validated, with high mutational load associated with prolonged survival in independent cohorts of NSCLC patients treated with Pembrolizumab.(56) Those responders demonstrating the highest mutational burden harbored multiple specific DDR gene mutations (BRCA2, MSH2, POLD1, POLE, RAD51c, and RAD17) that likely intensified tumor immunogenicity through somatic mutation generation. Higher response rates to PD-1 blockade in MMR-deficient tumors (47, 57) as well as in BRCA2-mutated melanoma support this concept.(58)

In the BRCA2-mutated melanomas, pretreatment samples revealed that although the overall somatic mutational loads of anti-PD-1-responsive melanoma tumors were not greater than those of non-responsive tumors, higher mutational loads were significantly associated with prolonged survival after immune checkpoint blockade. These data remain consistent with the concept that neoepitopes derived from somatic non-synonymous mutations may be critical for determining clinical benefits from anti-PD-1 therapy.

Case studies of exceptional responders to immune checkpoint inhibitors provide opportunities to identify potentially useful patient and tumor-specific predictors. In the study of an exceptional responder to anti-PD1 therapy in metastatic bladder cancer, somatic and germline JAK3 mutations were identified in the same allele.(59) Transfection of this mutant allele into cell lines increased PD-L1 expression, which may explain this patient’s response to anti-PD-L1 therapy. In support of this finding, NSCLC tumors with the same mutant allele also exhibited higher rates of PD-L1 positivity than tumors with other JAK3 genotypes. This example highlights the increased information that can be added to our current understanding by integrating gene expression germ-line mutation analysis in further studies.

Response to immunotherapy in the context of tumor-specific neoantigens

Recent technological innovations have made it possible to dissect the immune response to patient-specific neoantigens that arise as a consequence of tumor-specific somatic mutations, and emerging data suggest that recognition of such neoantigens is a major factor in the activity of clinical immunotherapies. These tumor antigens are recognized by T-cells as foreign based on the changes in the peptide sequence compared to the wild-type peptide. While it is likely that highly mutated tumors form neoantigens more readily, the stochastic nature of neoantigen generation calls for a functional validation as not all formed neoantigens will be immunologically relevant. Bioinformatics approaches have been successful in correlating mutational load with predicted neoantigen load.

Although the complete response rate to anti-CTLA-4 therapy is low for patients with metastatic melanoma, durable clinical benefit is consistently observed in approximately 20% of patients undergoing therapy.(60–62) Clinical studies of exceptional responders and small cohorts of melanoma patients attribute this response to NRAS mutation status,(63) total neoantigen load, and a neoantigen-derived consensus tetrapeptide signature thought to be necessary for activation of an antitumor T-cell response.(55) The recent whole exome analysis confirmed that neoantigen loads were significantly correlated with clinical benefit in melanoma patients treated with ipilimumab(38), though the tetrapeptide signature was not confirmed in additional analyses. (38, 58)

Subsequently, it has been suggested that tumors with high mutational frequency may produce sufficient neoantigens to generate anti-tumor immunogenicity, while tumors with low frequency may not.(64) Similar observations have been made in DDR deficient tumors, notably high MSI tumors,(57) BRCA-mutant ovarian tumors(65) and melanoma.(58) While favoring immunogenic neopeptide generation, the duality of genomic instability and mutation accumulation also supports tumor heterogeneity and emergence of less immunogenic novel clones cable of escaping immune surveillance.

A high degree of intratumoral heterogeneity is associated with poor outcome, while sensitivity to immune checkpoint blockade is associated with low heterogeneity and high numbers of clonal neoantigens.(66) Therefore an incongruous role exists for DDR defects in regulating immune checkpoint blockade response. Although DDR deficient tumors demonstrate profoundly unstable genomes and a high mutational burden, these tumors are the most likely to exhibit high intratumoral heterogeneity as a result random mutation induction.(67)

A significant challenge is the determination of those epitopes that will prime a T-cell response. In melanoma following ipilimumab treatment, only two of 448 immunologically relevant epitopes triggered a patient-specific antitumor T-cell response.(68) In NSCLC patients, response to PD-1 blockade was associated with a T-cell response against only a single neoantigen resulting from a single nucleotide variant in HERC1. (56)

Mutational burden and predicted neoantigen load also mold the landscape and functional capacity of the antitumor immune microenvironment. RNA-seq data has correlated tumor-infiltrating cytotoxic T-lymphocytes with a higher immunogenic mutation rate.(69) WGS studies of RNA expression in the tumor microenvironment also revealed that the predicted neoantigen load correlates with the cytolytic activities of intratumoral cytotoxic T-lymphocytes and natural killer cells.

These data are supportive of high mutational burden correlating with increased neoantigen formation and tumor immunogenicity. However, the high attrition rate from a high mutational burden to the small number of neoepitopes that can generate an anti-tumor immune response illustrates the complexity of neoantigen prediction using genomic data in isolation.

Pancreatic Adenocarcinoma

The role of immunotherapy in many “immunogenic” histologies is well established, however, clinical trials testing immune checkpoint inhibitors in pancreatic ductal adenocarcinoma (PDAC) remain to demonstrate clinically significant responses. Yet, as with other targeted therapeutic approaches in this disease, it is becoming apparent that immune therapies will likely require a predictive biomarker based approach. Using a RNAseq based expression analysis of 100 tumors; four classes of PDAC were identified, including a “pancreatic progenitor subtype”.(79) This subtype contained subclasses with active transcriptional networks characteristic of late stages of pancreas development termed Abnormally Differentiated Endocrine and eXocrine (ADEX). One of these subtypes was an “immunogenic subtype” characterized by elevated expression of multiple transcripts derived from immune cells and immune modulatory cytokines. It is suspected that this cognate molecular signature may possess an inherent vulnerability to targeted immunotherapy. Further, gene expression analysis of the other subtype (squamous) identified an immunosuppressive signaling pathway. This pathway included the expression of PD-LI, PD-L2 and IDO-1 within tumor cells. Therefore, GEP of specific tumor subtypes in this disease may allow patient specific stratification for immunotherapies.

Urothelial Carcinoma

In the multi-center, single-arm, phase II trial of PD-L1 blockade with atezolizumab in patients with locally advanced or metastatic urothelial carcinoma, patient response correlated with genetic tumor subtyping.(80) The Cancer Genome Atlas (TCGA) was utilized on paraffin-embedded tumor samples finding higher response rates (34%, p=0·0017) in the luminal cluster II subtype. These tumors were characterized by transcriptional signatures associated with T-effector cell activation. By contrast, luminal cluster I tumors were associated with low CD8+ effector genes expression, lower PD-L1 immune cell or tumor cell expression, and lower responses to atezolizumab. Moreover, the median mutation load was significantly higher in responders than in non-responders assessed through focussed genomic profiling of a 315-gene panel (12·4 vs. 6·4 per megabase, p<0·0001).

Although PD-L1 immune cell status is clearly associated with atezolizumab response, incorporation of TCGA gene expression subtype, mutation load, or both of these novel biomarkers into a model based on PD-L1 immune cell staining significantly improved the association with response. Thus, disease subtype and mutation load do not simply recapitulate the information provided by PD-L1 expression in immune cells, but rather, they provide independent and complementary information.