Editor’s summary
Tumors, especially those with a high mutational burden, express neoantigens that may be recognized by an individual’s T cells. However, identifying and validating these neoantigens can be a tedious and sometimes unsuccessful effort. To address this, Miller et al. developed a platform that combines whole-exome sequencing and transcriptional analysis of tumors with ex vivo stimulation of peripheral blood cells to more efficiently identify neoantigens that have elicited T cell responses in patients. The approach, termed “identify-prioritize-validate” (IPV), was able to identify clinically meaningful neoantigens in patients with different types of cancer, including those with a lower mutational burden, and with different human leukocyte antigen (HLA) types. Thus, the IPV approach demonstrates an advance in identifying neoantigens that may be useful both as a biomarker and as a therapeutic target. —Courtney Malo
Abstract
The clinical impact of tumor-specific neoantigens as both immunotherapeutic targets and biomarkers has been impeded by the lack of efficient methods for their identification and validation from routine samples. We have developed a platform that combines bioinformatic analysis of tumor exomes and transcriptional data with functional testing of autologous peripheral blood mononuclear cells (PBMCs) to simultaneously identify and validate neoantigens recognized by naturally primed CD4+ and CD8+ T cell responses across a range of tumor types and mutational burdens. The method features a human leukocyte antigen (HLA)–agnostic bioinformatic algorithm that prioritizes mutations recognized by patient PBMCs at a greater than 40% positive predictive value followed by a short-term in vitro functional assay, which allows interrogation of 50 to 75 expressed mutations from a single 50-ml blood sample. Neoantigens validated by this method include both driver and passenger mutations, and this method identified neoantigens that would not have been otherwise detected using an in silico prediction approach. These findings reveal an efficient approach to systematically validate clinically actionable neoantigens and the T cell receptors that recognize them and demonstrate that patients across a variety of human cancers have a diverse repertoire of neoantigen-specific T cells.
INTRODUCTION
The genetic instability underlying neoplastic transformation can also provide tumor-specific targets for host T cells (1–4). These include function-altering mutations in so-called “driver” genes, which endow the tumor with growth and survival advantages, as well as tumor-specific variants of “passenger” genes that lack an obvious protumorigenic function (5). In both cases, fragments of the mutated protein can be processed and presented as linear peptides bound to surface human leukocyte antigen (HLA) class I or class II molecules for recognition as neoantigens by a patient’s CD8+ and CD4+ T lymphocytes, respectively (6). Neoantigens have several important features that distinguish them as optimal targets for immunotherapy, including tumor-specific expression and the availability of a high-affinity responding T cell repertoire that is not subject to central (thymic) tolerance (7–9). Accordingly, mounting evidence from a variety of immunotherapy settings, including immune checkpoint blockade (ICB) (10, 11), adoptive cellular therapy (ACT) (12, 13), and personalized cancer vaccination (PCV) (14–16), supports the idea that clinical benefit from these interventions involves immune targeting of neoantigens [reviewed in (17)]. These observations have raised the hope that tumor- and patient-specific immunogenic mutations can be routinely and reliably identified for therapeutic purposes in the majority of cancers. Impediments to this goal are substantial, however, because each of the current methods for neoantigen identification has intrinsic features that could limit its widespread use with routine clinical samples and across the tumor mutational burden spectrum.
Most neoantigen discovery workflows begin with next-generation sequencing (NGS) of matched tumor and normal tissue samples procured during surgery or routine biopsy to identify the expressed nonsynonymous mutations (NSMs) (7). Because only a small fraction of somatic NSMs have been shown to constitute bona fide neoantigens recognized by T cells (3, 18), these are further filtered to identify the subset likely to have intrinsic immunogenicity. This generally involves application of one or more of a growing list of computational algorithms intended to model key steps in the antigen presentation pathway, which, for HLA class I, includes proteasomal processing, transporter associated with antigen processing–mediated transport, and, of central importance in this approach, predicted binding to the patient’s HLA class I alleles (19). HLA class I peptide binding algorithms fall into two broad categories based on the type of data used to train them. In the first, the binding coefficients of an experimental set of peptides containing the mutation within wild-type flanking sequences is prospectively calculated using algorithms trained on the sequence preferences of defined peptides eluted from surface HLA molecules or with known in vitro binding scores empirically derived from experimental data. Mutations that can generate a peptide with a predicted binding score that falls below a given threshold [half-maximal inhibitory concentration (IC50) or rank order] are considered as potential neoantigens. State-of-the-art examples of these computational algorithms include NetMHC (20), MHCFlurry (21), and NetMHCpan (22), with recent updates incorporating mass spectrometry (MS) data from either cell lines (23) or human tumors (24) in an effort to integrate antigen processing data. In general, these computational algorithms excel in identifying peptides that bind to the HLA class I molecules as predicted but show less accuracy in selecting those that are naturally presented at the surface of cells, with fewer than 5% of the roughly 10% of peptides predicted to bind HLA or detectable by either T cell assays or MS (24–27), rendering them unsuitable for cancers with low tumor mutational burdens where the absolute number of expressed mutations would become limiting. Machine learning–aided approaches have improved the accuracy of computational strategies, but their predictive ability in the clinical setting remains to be evaluated (28). Computational algorithms for predicting HLA class II epitopes are comparatively less developed, in part because of limited datasets of validated neoantigens and the more permissive binding features of HLA class II [reviewed in (29)].
An alternative approach to neoantigen discovery is to assess the presence of T cells capable of recognizing predicted HLA class I binders in peripheral blood or tumor-infiltrating lymphocytes (TILs), either physically using peptide-HLA multimers (30) or functionally using in vitro restimulation assays (27). Although these approaches can confirm both presentation of a given neoantigen and the availability of a responding T cell repertoire, they can be technically challenging or require specialized cellular reagents, such as autologous dendritic cells, thus limiting their utility in routine neoantigen screening using nominal clinical samples.
To address these challenges and in an effort to understand the spontaneous T cell response to cancer in patients, we set out to develop an unbiased and HLA-independent strategy for the functional validation of spontaneous neoantigen-specific CD4+ and CD8+ T cell responses that uses NGS-guided selection to nominate NSMs for functional recognition testing in a short-term culture of autologous peripheral blood mononuclear cells (PBMCs), which reports on both T helper cell 1 (TH1) and TH2 cytokines. Using this approach (herein referred to as “IPV” for identify-prioritize-validate), we now show that preexisting CD4+ and CD8+ T cell responses can be identified and the T cell receptors (TCRs) that mediate them can be isolated at rates that are about 10-fold higher than class I HLA binding–based predictive approaches across a spectrum of tumor types and degrees of tumor mutational burden using routinely available clinical samples.
RESULTS
HLA-agnostic selection and functional validation of neoantigens
We sought to develop a functional neoantigen discovery pipeline to measure preexisting neoantigen-specific T cell responses in patients with cancer using standard clinically available biospecimens, such as formalin-fixed paraffin-embedded (FFPE)–preserved tumor and cryopreserved PBMCs. The main features of this workflow are (i) whole-exome sequencing (WES) and RNA sequencing (RNA-seq) of matched tumor and peripheral blood samples, (ii) HLA-agnostic bioinformatic identification of tumor-specific expressed variants and their prioritization as candidate neoantigens, and (iii) functional recognition analysis of candidate mutations in the form of synthetic peptides by autologous PBMCs. Specifically, WES was performed on tumor and normal peripheral blood cells at a read depth of >150× along with RNA-seq on the tumor RNA with a minimum of 5 × 107 reads to identify somatic mutations. Only highly tumor-specific, nonsynonymous, and expressed mutations were considered by filtering for variants with higher variant allele frequency (VAF) in the tumor than the normal WES and variants that were supported by at least one RNA read. The variants were then ranked according to a combination of various parameters, including DNA and RNA VAFs and gene expression as described in Materials and Methods. Depending on the number of patient-derived PBMCs available for testing purposes, a subset of variants was then selected from this ranked list for functional recognition analysis. There, the candidate mutation was represented as pairs of 20-mer peptides with the mutation at position 6 and position 15 within wild-type flanking sequences (fig. S1). Pools of these peptides were incubated with autologous PBMCs in a 14-day coculture, after which viable cells were obtained and analyzed by enzyme-linked immunospot (ELISPOT) assays for interferon-γ (IFN-γ) and interleukin-5 (IL-5) production in response to the individual peptides comprising the pool. As an example of this workflow, we acquired tumor tissue from a patient with metastatic pancreatic ductal adenocarcinoma (PDAC) and identified 636 coding variants, of which 240 met our criteria of being highly specifically expressed in the tumor. Of these 240 candidates, 16 mutations (occurring in a total of 15 genes) were selected for immunogenicity testing on the basis of the criteria described in Materials and Methods and availability of patient PBMCs and were translated into 33 20-mer peptides (data file S1). Peptide pools consisting of 2 to 10 20-mer candidate peptides were then cocultured with autologous PBMCs for 14 days before performing IFN-γ and IL-5 ELISPOT assays using irrelevant peptide and phytohemagglutinin (PHA) as negative and positive controls, respectively. Four of the seven peptide pools tested for this patient elicited autologous T cell responses, including those of both TH1 and TH2 cytokine polarity (Fig. 1A). Deconvolution of the pool responses to single peptides showed that 9 of the 33 neoantigen candidate peptides identified by our pipeline (27%) stimulated T cell responses (Fig. 1B), representing 6 of the 16 mutations selected by our pipeline (37.5%). The observed responses included those against both passenger mutations and pathognomonic drivers, such as KRAS G12>D.
Validation of neoantigen targets for both CD4+ and CD8+ T cell responses
Having confirmed candidate neoantigens as T cell targets by ELISPOT, we next asked whether we could delineate CD4+ versus CD8+ T subset-specific responses using cytokine secretion assays. For this, aliquots of postexpansion PBMCs shown to contain neoantigen-specific T cells by ELISPOT were restimulated on day 17 with the immunogenic mutant neoantigen peptides recognized by each, and cytokine production was measured by cytokine capture assay. Briefly, cells were labeled with both IFN-γ and IL-5 cytokine capture antibodies, along with CD4 and CD8 surface staining, and the amplitude of effector cytokine release from these T cell subsets was determined after a 3-hour incubation with mutant peptide. For example, ELISPOT assays identified functional T cell responses against the driver mutations KRAS G12>V and passenger mutations ZNF500 A426>T, NRXN2 R861>W, and TP53 L265>P (Fig. 2A). Restimulation of T cells from the 14-day coculture with single-mutant peptides on day 17 discerned distinct populations of CD8 and CD4 T cells producing IFN-γ and IL-5 that correlated with the ELISPOT results (Fig. 2B). In addition, we detected T cells polarized to produce TH1 (IFN-γ) and TH2 (IL-5) cytokines, whereas most existing methods of neoantigen discovery are biased toward the identification TH1 responses. Last, among the 13 patients whose samples we analyzed, we identified functional T cell responses against mutations in driver genes such as TP53, APC, KRAS, and MEN1, which represent attractive targets for T cell–directed therapies because of their critical roles in cancer pathogenesis (data file S2). For one of the patients (Hu_048), expression of immunogenic neoantigens was detected in the tumor RNA from spatially distinct sites of disease, suggesting that these neoantigens could serve as ideal therapeutic targets because of their conserved expression (data file S2).
TCR identification and validation
To further validate the responses identified by ELISPOT and cytokine secretion assay (CSA), we sought to confirm that cytokine-producing T cells from these cultures expressed neoantigen-specific TCRs. These were identified through analysis of paired-end sequencing of cDNA libraries from single neoantigen-specific T cells isolated from the day 14 postexpansion cultures on the basis of neoantigen-specific cytokine production or up-regulation of the activation markers 4-1BB (CD137) and CD25. In the first example, 41 single CD4+ T cells from Hu_159 postexpansion PBMCs that produced IFN-γ in response to a pool of neoantigen peptides including KRASG12>V were isolated by CSA and cell sorting. Eight of these (20%) expressed the same αβ TCR complementarity determining region (CDR) 3 clonotype (CAVPDQTGANNLFF and CASSIGPQGKNIQYF, respectively). The full-length TCRs were then cloned into the pMSGV1 retroviral expression vector in a b-P2A-a construct and transduced into primary human T cells that were subsequently tested for recognition of HLA-matched B cells pulsed with either KRAS mutant or wild-type control peptides, and activation of the transduced T cells was assessed by up-regulation of PD-1. Using this approach with overlapping 15-mer peptides, we first identified the minimal peptide recognized by this KRASG12>V-specific TCR (KLVVVGAGGVGKSAL versus KLVVVGAVGVGKSAL) and confirmed that this induces PD-1 up-regulation (Fig. 2C) that can be blocked by antibodies to HLA-DR but not HLA-DQ molecules (Fig. 2D), thereby identifying the restricting class II allele. This was confirmed using a panel of B cell lines that share specific DR alleles with Hu_159 (Fig. 2E), which showed that only target cells expressing HLA-DRB3*03:01 were recognized, indicating that this TCR recognizes amino acids 5 to 19 of the KRASG12>V peptide bound to a heterodimer formed by HLA-DRA1 with HLA-DRB3*03:01. A similar TCR sequencing and functional testing approach was applied to T cells sorted from a separate patient, Hu_250 (Table 1), on the basis of activation-induced marker up-regulation instead of cytokine production (31). For this, day 14 postexpansion PBMCs were restimulated in the presence of a pool of neoantigen peptides, and 129 single T cells that up-regulated CD25 and CD137 (4-1BB) were sorted. This analysis revealed a set of CD25+ CD137+ T cells expressing the same TCR αβ clonotypes (fig. S2). Four of these were expressed and functionally tested for recognition of the neoantigen peptides used in their expansion from blood. One of these (25%) TCRs (TCR 3) expressing the αβ clonotypes SLATGQGDYGYTFGSGT and GNTGGFKTIFGAGTRLF was cloned into the pMSGV1 vector and transduced in primary human T cells, where it was shown to recognize the TAS2R19K258>N peptide from the neoantigen pool (Fig. 2F) by up-regulation of the acute activation marker PD-1. As with the KRASG12>V TCR described above, blocking antibodies and HLA-matched B cell targets were used to identify the HLA restriction as HLA-DRA1/HLA-DRB4*1:03:01 (Fig. 2, G and H). Together, these data demonstrate that the neoantigen-specific responses identified by functional analyses are based on physiologic T cells verifiable at the level of the specificity of the antigen receptors they express and that the IPV platform can allow these to be efficiently isolated and validated on a patient-specific basis….
