Applications of Next-Generation Sequencing in Neoantigen Prediction and Cancer Vaccine Development

Abstract

Next-generation sequencing has changed the face of cancer immunotherapy research by making tumor-specific cancer vaccines a reality. Whole exome sequencing and RNA sequencing combined with bioinformatic pipelines allow the prediction of neoantigen targets for cancer vaccines. In this review, we discuss the preclinical and early clinical evidence for cancer vaccines; describe methods and challenges in neoantigen prediction; and summarize emerging new technologies that will improve neoantigen cancer vaccine development.

Background

Immunotherapy is rapidly evolving as a promising treatment for many types of solid and hematologic malignancies with the potential to provide targeted therapy while minimizing the toxic side effects of traditional chemotherapy. The majority of clinical trials to date, however, focus on inhibiting known immune checkpoints (CTLA-4, PD-1/PD-L1). Immune checkpoint inhibitors such as ipilimumab (anti-CTLA-4) and pembrolizumab (anti-PD1) activate the immune system by blocking inhibitory ligands, which has proven effective in many types of malignancies. Checkpoint inhibitors, however, are not effective for all tumors and can have undesirable side effects related to immune activation and targeting of healthy tissue (Whiteside et al., 2016).

The more we learn about cancer immunology, the more promising the newest form of immunotherapy has become: cancer vaccines. Although vaccines against tumors seem like a grandiose idea, current advances in technology, specifically next-generation sequencing (NGS), have made it an increasing reality. The goal of cancer vaccines is to activate a patient's endogenous lymphocytes to recognize and destroy cancer cells. Tumor-specific proteins are delivered to patients and presented to T cells by antigen presenting cells, causing activation against tumor cells that display these antigens on their cell surface through major histocompatibility complex (MHC) molecules. Promising preclinical animal trials and small cohort clinical trials suggest that cancer vaccines may represent the future of personalized cancer therapy (Sahin et al., 2017; Aurisicchio et al., 2018).

Cancer Vaccine Development Before NGS

There are two main categories of cancer vaccines, standardized or “off the shelf” vaccines that target epitopes previously defined in multiple patients and cancer types (e.g., EGFR, HER2, and VEGF) and personalized vaccines that target tumor-specific epitopes (Morse et al., 2005; Kaumaya and Foy, 2012). Before the development and clinical use of NGS, advances in creating and studying cancer vaccines were limited by cost, production, and a paucity of appropriate epitope targets (Morse et al., 2005; Kudrin, 2012).

Tumor-associated antigens (TAAs) were the main antigen source for vaccine development before the use of NGS. TAAs are peptides that are overexpressed in tumors, but are also present in normal tissues at lower levels (e.g., HER2, p53, survivin, MUC1, and tyrosinase) (Melero et al., 2014; Li et al., 2017). Vaccines against TAAs have been studied in clinical trials for multiple cancer types including melanoma and lung cancer with less than promising results (Vansteenkiste et al., 2013; Butts et al., 2014). This is likely at least partially due to the heterogeneity of tumors requiring more advanced patient selection as well as the development of central thymic immune tolerance.

Impact of NGS on Clinical Cancer Vaccine Development

With the adoption of NGS, the goal is to no longer find the right tumor to fit the vaccine, but rather to create a vaccine that perfectly fits the tumor. Genetic heterogeneity is a core feature of most malignancies and the underlying reason why an optimal cancer treatment for any one individual is so elusive. NGS can uncover tumor-specific alterations with single nucleotide resolution, allowing us to use this characteristic of malignancy to our advantage.

Unlike TAAs, neoantigens are proteins specific to tumor cells and not produced by normal host cells. They arise from somatic DNA alterations specific to each tumor and generally demonstrate binding affinity for MHC molecules. These qualities make them perfect targets for cancer vaccines. The development of patient-specific neoantigen cancer vaccines generally takes place through the following steps: (1)

Perform whole exome and RNA sequencing of tumor and normal cells to identify somatic mutations (neoantigens) and their expression levels.

(2)

Determine patient-specific HLA alleles and the HLA binding affinity of potential neoantigens.

(3)

Filter and select the most promising candidate neoantigens by expression and HLA binding affinity, possibly validating immunogenicity of predicted neoantigens in vitro.

(4)

Create a vaccine with the top neoantigens using a peptide complex, liposomal, or RNA delivery platform.

Many of these steps have been incorporated, streamlined, and packaged as neoantigen prediction pipelines (described hereunder). Not surprisingly, there are many nuances and variables that make this process more challenging. Despite these challenges, however, preclinical and clinical studies for neoantigen vaccine development and administration have been promising.

Preclinical neoantigen cancer vaccine studies

Castle et al. used NGS to identify >900 somatic mutations in a model of murine melanoma (B16F10), finding that >500 of these mutations were expressed using RNA sequencing. Fifty of these mutations were selected for immunogenicity testing based on prediction models and the corresponding peptides were synthesized. These peptides were then injected into mice, and response, as evidenced by T cell reactivity patterns, was seen with 16 out of the 50 predicted neoantigens. Two of the neoantigens were then tested in vivo for antitumoral activity with peptide vaccination and both significantly inhibited tumor growth in the B16F10 mouse model with complete tumor protection in 40% of mice (Castle et al., 2012).

This group subsequently expanded their protocol to three different murine tumor types (melanoma, colon, and breast) with similar results, showing that many nonsynonymous mutations identified by neoantigen prediction pipelines are in fact immunogenic. They also found that the majority of these mutations elicited a response from CD4⁺ T cells, leading to robust antitumoral activity on immunization. From these data, they concluded that bioinformatics pipelines can be successfully used for neoantigen prediction in preclinical models (Kreiter et al., 2015).

Clinical neoantigen cancer vaccine studies

The first evidence of effective neoantigen cancer vaccine use in humans was published simultaneously by Sahin et al. (2017) and Ott et al. (2017). Sahin et al. studied 13 patients with stages III and IV melanoma who received individualized synthetic RNA vaccines. NGS was used to identify nonsynonymous somatic mutations, with RNA sequencing to confirm expression. Ten neoantigens per patient were selected based on predicted affinity for HLA classes I and II. Pre- and postimmunization IFNγ ELISpot analysis in CD4⁺ and CD8⁺ T cells showed activity against 60% of the neoantigens, mostly by CD4⁺ T cells as described in murine models. Overall, patients showed improvement in sustained progression-free survival with a significant reduction in recurrent metastatic events compared with the standard of care. Eight patients with no detectable disease at time of immunization had no evidence of recurrence throughout the follow-up period of 12-23 months (Sahin et al., 2017).

Ott et al. demonstrated both the immunogenicity and antitumoral effects of personalized neoantigen vaccines in a small group of patients with stages III and IV melanoma who underwent surgical resection but no further treatment. First, whole exome and mRNA sequencing was used to identify unique somatic mutations. Potential neoantigens were selected based on HLA binding prediction methods, filtered, and ranked. Patient-specific neoantigen vaccines with up to 20 unique peptides per patient were created and delivered. The vaccines consisted of 13 to 20 peptides with lengths of 15 to 30 amino acids each, administered as five priming and two booster vaccinations.

On 25-month median follow-up, there was no disease recurrence in the four patients with stage III disease. The two patients with stage IV disease showed evidence of recurrence on imaging but subsequently had complete radiographic response after treatment with the checkpoint inhibitor pembrolizumab. Although only a small number of patients were included in this pilot study, these results were impressive when compared with the 6% rate of complete radiographic response for patients with metastatic melanoma treated with pembrolizumab as a first-line therapy (Robert et al., 2015). In addition to an encouraging clinical response, this study also demonstrated CD4⁺ T cell response to 60% of neoantigens and CD8⁺ T cell response to 16% of neoantigens, consistent with previously reported findings (Ott et al., 2017).

Although there is still much to be learned and perfected in the process of developing neoantigen cancer vaccines, NGS has certainly made this idea a reality with very encouraging early clinical studies.

Pipelines for Neoantigen Prediction

The number of nonsynonymous somatic mutations is large for any given tumor, and predicting which mutations will demonstrate strong HLA binding to produce a robust immune response is critical. The process of identifying effective candidate neoantigens, therefore, is complex. Hereunder we outline the individual steps of neoantigen prediction and discuss pipelines designed to integrate and simplify these steps.

Variant detection

The first step of neoantigen prediction is the identification of mutated proteins specific to the tumor cells, commonly referred to as variant detection or somatic mutation calling. One of the most widely used tools for variant detection is MUTECT, part of the Genome Analysis Toolkit produced by the Broad Institute, which can be used for variant detection using whole exome sequencing, whole genome sequencing, or RNA sequencing data (McKenna et al., 2010). Additional programs that serve a similar purpose include VarScan and EBCall (Koboldt et al., 2012; Shiraishi et al., 2013). The main pitfall of this step is the generation of false positive mutations, due to sequencing errors and read misalignments. False positive errors can be reduced by increasing the number of algorithms used to determine potential variants (Ding et al., 2014; Hackl et al., 2016).

HLA typing

Traditionally, HLA typing is performed using serologic methods. With the development of NGS, however, multiple programs are being developed that can predict HLA type using RNA sequencing data. Although initial tools for HLA typing, including HLAminer (Warren et al., 2012) and Seq2HLA (Boegel et al., 2012), were subject to sequencing errors and low accuracy, newer programs such as Polysolver (Shukla et al., 2015) and Optitype (Szolek et al., 2014) have improved their methods leading to improved accuracy and sensitivity. These newer HLA typing algorithms compare RNAseq or whole genome sequencing data with published reference sequences, selecting for alignments with the least number of mismatches to determine HLA type (Szolek et al., 2014; Hackl et al., 2016).

MHC binding affinity

Most tools that predict neoantigen-MHC binding affinity utilize sequence-based methods that focus on the primary protein sequences rather than structured-based methods that apply predictions of 3D protein configuration. Initially these tools used linear techniques, assessing the binding affinity of each individual amino acid at all possible binding positions. With machine learning, however, there are now more advanced nonlinear methods with better performance.

Currently, the most commonly used tools are NetMHC (Lundegaard et al., 2008), which uses allele-specific epitope prediction (better with well-characterized alleles) and NetMHCpan (Nielsen and Andreatta, 2016), which uses pan-specific machine learning methods (Hundal et al., 2016). The newer pan-specific tools do not rely on previously characterized MHC peptide binding data, but rather use integrated machine learning algorithms to predict the binding affinity of known MHC protein sequences. This method has been shown to be more accurate when there is limited MHC peptide binding data available (Hackl et al., 2016).

Integrated pipelines

pVAC-Seq (personalized variant antigens by cancer sequencing) is a comprehensive neoantigen prediction pipeline developed by Hundal et al. (2016) to streamline the computational flow of this complicated process. The required inputs are an annotated list of nonsynonymous somatic mutations and the HLA haplotypes of the patient. The authors use genome modeling system for alignment and variant calling (Griffith et al., 2015) to create a variant list annotated with amino acid changes and transcript sequences. The user may enter HLA haplotypes obtained by clinical assays, or with one of the already described HLA typing tools; the authors recommend HLAminer (Warren et al., 2012) or ATHLATES (Liu et al., 2013). pVAC-Seq then performs epitope prediction using NetMHC (Lundegaard et al., 2008) and filters neoantigen candidates creating a prioritized list of vaccine candidates.

The CloudNeo pipeline is a recently developed cloud-based tool that integrates many of the steps of neoantigen prediction. It requires only a vcf (variant call format) file for nonsynonymous somatic mutations, which can be created using one of the additional tools already described, and an RNAseq bam file for HLA typing. The pipeline then translates the genomic variants into amino acids using variant effect predictor (VEP) (McLaren et al., 2010) and the user can choose HLAminer or Polysolver for HLA typing. Finally, the NetMHCpan tool predicts HLA binding affinity. The output is a list of peptides with MHC binding affinity scores. The users can then determine cutoffs to choose their neoantigens based on the output (Bais et al., 2017).

TIminer (tumor immunology miner), developed by Tappeiner et al., was designed as a more user friendly pipeline as it melds all required interfaces to facilitate not only neoantigen prediction but also characterization of immune infiltrates and quantification of tumor immunogenicity.

TIminer integrates RNAseq data, inputted as a FASTQ file, as well as somatic DNA mutations, inputted as a vcf file. Using RNAseq data, TIminer quantifies gene expression (in transcripts per million) using Kallisto (Bray et al., 2016). HLA typing is then performed with Optitype (Szolek et al., 2014), mutated DNA coding regions are translated with VEP (McLaren et al., 2010), and finally MHC binding affinity is predicted with NetMHCpan (Nielsen and Andreatta, 2016). In addition to the gene expression data and MHC binding predictions, TIminer uses “gene set enrichment analysis” to characterize tumor infiltrating immune cells within the sample and an “immunophenogram” to predict overall immunogenicity of the tumor (Tappeiner et al., 2017).

Tumor-specific neoantigen detector is a pipeline developed to identify neoantigen-producing somatic mutations and tumor-specific mutant proteins. This group compared two methods of neoantigen prediction: protein topology of mutated extracellular membrane proteins and NetMHCpan (Zhou et al., 2017). The pipeline uses the Genome Analysis Toolkit (GATK v 3.5) and MUTECT 2 for variant calling (McKenna et al., 2010; Cibulskis et al., 2013). HLA typing is performed with SOAP-HLA (Li et al., 2009). To determine MHC binding by protein topography, the program TMHMM is used (Sonnhammer et al., 1998); NetMHCpan is employed as a second method.

Although this pipeline is very efficient at processing raw data in a user friendly way, there are two potential downsides. First, it does not integrate RNAseq data to determine the expression level of the identified somatic mutations. Second, the protein topography data, although helpful for better understanding the mutations, do not clearly provide additional information on potential MHC binding compared with NetMHCpan.

INTEGRATE-Neo is a unique pipeline as it evaluates gene fusions rather than missense mutations (Zhang et al., 2017). This pipeline builds upon the group's prior computational tool, INTEGRATE, which uses RNAseq and whole genome sequencing inputs to determine gene fusions (Zhang et al., 2016). The first step of the INTEGRATE-Neo pipeline requires the human reference genome (FASTA format), gene models (GenePred format), and gene fusions generated from INTEGRATE (BEDPE format). This step identifies and excludes gene fusions that do not create a fusion protein. The second step is HLA prediction using HLAminer (Warren et al., 2012), although the users may upload their own HLA data if they prefer a different method or if HLA alleles have been determined clinically. The final step is neoantigen prediction using NetMHC v4.0. Ideally the neoantigen predictions from this pipeline would be combined with those from another pipeline to give a broader more complete array of possible neoantigens based on a variety of mutation types.

The pipeline NeoantigenR was developed to utilize both short- and long-read sequences to detect abnormal splicing variants in tumor cells (Tang and Subha, 2017). Input data for NeoantigenR includes an annotated gene prediction output file in GFF format that can be DNA or RNA sequencing data and short- or long-read sequences. NeoantigenR uses isoform structural comparison to determine alternative splicing events and then predicts neoantigens formed by identifying insertion, deletion, and substitution mutations. NetMHC is used to predict MHC binding affinity.

Neopepsee is a pipeline with an integrated machine learning tool that not only predicts neoantigens but also uses specialized “immunogenicity features” to reduce potential false positives (Kim et al., 2018). This pipeline requires RNAseq data (FASTQ), somatic mutation data (vcf), and HLA typing that can be provided or predicted by the program. Neopepsee predicts neoantigens based on its own unique machine learning algorithms and provides details on immunogenicity-related factors such as IC₅₀, expression level, and immune regulatory function (i.e., PD1 and PD-L1). Notably, this algorithm was validated in data sets with previously proven neoantigens for melanoma, leukemia, and stomach cancer patients, showing superior neoantigen prediction for melanoma and leukemia compared with other methods (Kim et al., 2018).

FRED 2 (framework for epitope detection) is an open-source framework designed to address the issue of the large number of pipelines and tools that each require unique inputs and provide different output formats (Schubert et al., 2016). This versatile framework provides platforms for HLA typing, neoantigen prediction (with tools to evaluate antigen processing), selection, and assembly. It is unique in its flexibility, which allows users to select their preferred prediction tools without having to significantly modify or recreate their data. For example, for HLA typing, OptiType (Szolek et al., 2014), PolySolver (Shukla et al., 2015), seq2HLA (Boegel et al., 2012), and ATHLATES (Liu et al., 2013) are all available. FRED 2 allows for interaction of most available tools used along each step of a pipeline, empowering users to utilize the overall design that best fits their needs.

Each of these pipelines perform similar functions; however, the methods and results vary. It is important that users try multiple pipelines, choose the most appropriate for their own specific purposes, and be wary of the potential pitfalls including neoantigen false positive and false negative predictions. All of these pipelines are publically available (Table 1).

Table 1.

Summary of Neoantigen Prediction Pipelines

Pipeline	Highlights	Tools	Website	Reference
pVAC-Seq	Standard computational pipeline does not integrate RNA expression	HLAminer or ATHLATES, NetMHC	https://github.com/griffithlab/pVAC-Seq	Hundal et al. (2016)
CloudNeo	Cloud-based pipeline that integrates vcf file of somatic mutations, RNAseq data, and HLA typing to predict MHC binding	VEP, HLAminer or Polysolver, NetMHCpan	https://github.com/TheJacksonLaboratory/CloudNeo	Bais et al. (2017)
TIminer	Predicts MHC binding and uses novel tools to characterize TILs and predict immunogenicity	Kallisto, VEP, Optitype, NetMHCpan	http://icbi.i-med.ac.at/software/timiner/timiner.shtml	Tappeiner et al. (2017)
TSNAD	Compares two methods of neoantigen prediction: protein topology of mutation extracellular membrane proteins and NetMHCpan	GATK, MUTECT 2, SOAP-HLA, TMHMM, NetMHCpan	https://github.com/jiujiezz/tsnad	Zhou et al. (2017)
INTEGRATE-Neo	Evaluates gene fusions as sources of somatic mutations	INTEGRATE, HLAminer, NetMHC	https://github.com/ChrisMaherLab/INTEGRATE-Neo	Zhang et al. (2017)
NeoantigenR	Utilizes long-read sequences to improve detection of alternative splicing variants	NetMHC	https://github.com/ICBI/neoantigeR	Tang and Subha (2017)
Neopepsee	Uses a unique machine learning algorithm based on “immunogenicity features” to predict neoantigens	HLAminer, Unique neoantigen predictor	http://sourceforge.net/projects/neopepsee/	Kim et al. (2018)
FRED 2	Versatile framework that allows users to select their own tools and easily transition between tools	Many options	http://fred-2.github.io	Schubert et al. (2016)

FRED 2, framework for epitope detection; GATK, Genome Analysis Toolkit; MHC, major histocompatibility complex; pVAC-Seq, personalized variant antigens by cancer sequencing; TSNAD, tumor-specific neoantigen detector; vcf, variant call format; VEP, variant effect predictor.

Ex Vivo Testing of Neoantigens

A major challenge of creating novel immunotherapies from neoantigens is determining which of the pipeline-predicted neoantigens will be immunogenic in vivo. To optimize effectiveness, it is crucial that neoantigens with the greatest potential for activating T cells are identified and used. With this goal in mind, methods are being developed to test the immunogenicity of predicted neoantigens ex vivo.

In 2016, Stronen et al. developed a method of evaluating T cell reactivity to neoantigens using healthy donor-derived T cells. They studied stage IV melanoma patients, focusing on 20 neoantigens with high predicted binding to HLA-A*02:01. T cells were isolated from healthy donors with the matching HLA-A*02:01 subtype and cocultured with dendritic cells transfected with mRNA of the selected neoantigens. Peptide-MHC staining showed T cell reactivity toward 3 to 5 of the chosen 20 neoantigens in four total donors, with reactivity to 11 unique neoantigens. Of note, only one of these neoantigens was detected by tumor infiltrating lymphocytes from the melanoma sample. The authors also created a flow cytometry-based assay to assess peptide-MHC stability, which has been linked to immunogenicity (Stronen et al., 2016).

Zhang et al. demonstrated successful in vitro testing of immunogenicity using a mouse patient-derived xenograft model of advanced stage breast cancer. The group used a pipeline to predict neoantigens in three patients (between 33 and 74 per patient) and then cultured the patients' autologous T cells with these selected neoantigens for 12 days. Using IFNγ ELISPOT, they found a significant T cell response in two of nine neoantigens for patient 1, and one of eight neoantigens for the other two patients. They then tested the immunogenicity of these T cells in vivo by injecting tumor cells into genetically immune compromised mice. The stimulated T cells were transferred into these tumor-bearing mice, which subsequently showed decreased tumor growth compared with controls (Zhang et al., 2017).

These studies show that we can further refine our neoantigen prediction methods by testing for immunogenicity and T cell activation ex vivo using both donor-derived T cells and autologous T cells. By developing and perfecting these methods, we will be able to create more effective and less costly immune therapies.

Challenges in Neoantigen Prediction and Vaccine Development

Despite the success of multiple groups in designing and testing cancer vaccines in both preclinical and clinical pilot trials, several challenges must be overcome before the widespread adoption of cancer vaccines as an effective clinical therapy. The cost and time required for whole genome sequencing, neoantigen prediction, and vaccine creation, for example, are significant. Based on previous trends, however, we anticipate that it will just be a matter of time before these barriers are overcome by the development of newer, more efficient, and less expensive technologies.

Another significant challenge is that only a very small fraction of predicted neoantigens are immunogenic and able to induce significant T cell reactivity. Corresponding to and further complicating this issue is that our current methods of neoantigen prediction likely miss a large number of potential neoantigens due to short sequences, sequencing errors, and incomplete algorithms. Contributing factors to these challenges are present in every step of the process. For example, whole exome sequencing is prone to amplification bias and technical errors, creating both false positive and false negative somatic mutations. In addition, there are numerous types of mutations other than single nucleotide variants that can create neoantigens, including missense/non-sense, deletions, insertions, duplications, and frameshifts; most of these mutations are currently not accounted for in described neoantigen prediction pipelines.

Finally, methods for testing and obtaining FDA approval for neoantigen cancer vaccines will be challenging, as the specificity of each vaccine is prohibitive to conducting large clinical trials. New methods to confirm safety and efficacy will need to be developed. Notably, chimeric antigen receptor T cell therapy obtained approval after showing impressive results in a trial of only 63 patients providing hope for a similar approval pathway for cancer vaccines (Aurisicchio et al., 2018).

The Future of NGS and Cancer Vaccines

As the landscape of NGS and bioinformatics continues to evolve, solutions to these and other problems are surfacing rapidly. Pipelines for neoantigen prediction are being developed and refined at an exciting pace, and advances in NGS techniques and applications promise to address many of the current challenges in neoantigen prediction and accelerate the development of effective cancer vaccines. In addition, new collaborations are being created to address these challenges, such as the Tumor Neoantigen Selection Alliance, started in 2016, which includes researchers from multiple institutions each evaluating the same cancers with different pipelines and then testing immunogenicity to refine and validate the best prediction models (Editorial, 2017).

Alternative splicing and intron retention

Recent studies are finding that abnormal RNA splicing is much more common in tumor cells than in normal tissue, largely due to mutations in genes related to RNA splicing (Dvinge and Bradley, 2015; Seiler et al., 2018). These alternative splicing events lead to intron retention within tumors, sometimes in patterns that are specific to cancer type, and often independent of mutations in RNA splicing machinery (Dvinge and Bradley, 2015). mRNA transcripts that include introns due to splicing errors are translated and then degraded, leading to peptides that could represent a rich source of potential neoantigens.

Investigators are testing this hypothesis by developing tools to detect alternative splicing events, finding that the potential neoantigen load nearly doubles when including alternative splicing and retained intron transcripts (Kahles et al., 2018; Smart et al., 2018). In fact, Smart et al. (2018) used mass spectrometry to confirm that presentation of these neoantigens on MHC I occurred at similar rates as neoantigens predicted from somatic mutations. Notably, these studies demonstrate that numerous different types of alternative splicing events can be successfully identified with short-read RNA sequencing, meaning that perhaps long-read sequencing is not essential as Tang and Madhavan previously suggested with their development of the NeoantigenR pipeline (Tang and Subha, 2017; Seiler et al., 2018).

Mass spectrometry-based immunopeptidomics

Immunopeptidomics, the study of the binding and presentation of peptides by HLA molecules, is an expanding field that holds great promise for improving the way neoantigens are identified and validated. Rather than relying on in silico algorithms to predict MHC binding, mass spectrometry-based techniques identify MHC peptide complexes in tumor cells, objectively identifying potential neoantigens. These peptides can then be compared with sequencing data and algorithm predictions to refine neoantigen selection (Bassani-Sternberg and Coukos, 2016; Bassani-Sternberg, 2018). The use of mass spectrometry-based immunopeptidomics will decrease the number false positive neoantigens predicted compared with in silico techniques, which can improve the efficiency of ex vivo immunogenicity testing by narrowing the number of potential neoantigens.

Single cell sequencing

Tumor samples from biopsies or surgical specimens represent a variety of cell types including tumor cells, immune cells, and fibroblasts. The wide variety of cell types included in bulk sequencing can confound NGS data. Single cell sequencing techniques address this issue and may lead to higher quality more accurate tumor sequencing data. The isolation of cancer cells from surrounding cells can be done by multiple methods, including micromanipulation, flow-assisted cell sorting, serial dilution, microfluidics, and laser-capture microdissection.

Single cell sequencing does pose some challenges that must be overcome before its use for cancer vaccine development. First, current cell isolation techniques require fresh tissue rather than formalin-fixed paraffin-embedded tissues, which may be impractical in a clinical setting. Second, single cell DNA sequencing involves whole genome amplification (WGA), which introduces errors. The two main methods for WGA are degenerative-oligonucleotide-PCR, which creates very low physical coverage of only about 10%, and multiple displacement amplification, which has excellent coverage at 90% but a very high error rate, especially for false positive mutations.

Allelic drop out is another error type that can cause significant issues when generating single cell sequencing data. This occurs when one allele in a heterozygous mutation is not amplified, leading to the illusion of a homozygous phenotype, and can occur at 10-50% of mutation sites. Although there are some computation models that can correct for false positive errors, to truly control for these errors one would need to perform sequencing on multiple cells, which can be prohibitively costly (Navin, 2014; Alves and Posada, 2018). Nevertheless, single cell sequencing holds tremendous promise in the development of cancer vaccines as these challenges are overcome, as it may eventually be performed on extremely small tumor samples and ideally one day even on circulating tumor cells in the peripheral blood.

CITE-seq

Another novel application of NGS, CITE-seq, uses single cell RNA sequencing combined with oligonucleotide-labeled antibodies to generate simultaneous protein and transcriptome read-outs on a single cell level (Stoeckius et al., 2017). Single cell RNA sequencing data are significantly more advanced and less prone to errors than singe cell DNA sequencing. CITE-seq methods may offer a way to simultaneously isolate and sequence mRNA from tumor cells versus their surrounding immune and stromal cells. In addition, CITE-seq may allow for the development of novel neoantigen pipelines that combine sequencing data with protein expression data, with enormous potential for improving neoantigen prediction. Lastly, the ability of CITE-seq to simultaneously characterize immune cell populations within tumors may help guide vaccine adjuvant administration in the future.

Conclusions

Despite the challenges of neoantigen prediction and cancer vaccine development, the future of neoantigen cancer vaccines is bright, largely due to the development and continued advancement of NGS. As we improve our methods for predicting and testing neoantigens, cancer vaccines may one day become a mainstay of treatment for many types of malignancies.

Footnotes

Author Disclosure Statement

No competing financial interests exist.

References

Alves

, Posada

(2018). Sensitivity to sequencing depth in single-cell cancer genomics. Genome Med, 10:29.

Aurisicchio

, Pallocca

, Ciliberto

, Palombo

(2018). The perfect personalized cancer therapy: cancer vaccines against neoantigens. J Exp Clin Cancer Res, 37:86.

Bais

, Namburi

, Gatti

, et al. (2017). CloudNeo: a cloud pipeline for identifying patient-specific tumor neoantigens. Bioinformatics, 33:3110-3112.

Bassani-Sternberg

(2018). Mass spectrometry based immunopeptidomics for the discovery of cancer neoantigens. Methods Mol Biol, 1719:209-221.

Bassani-Sternberg

, Coukos

(2016). Mass spectrometry-based antigen discovery for cancer immunotherapy. Curr Opin Immunol, 41:9-17.

Boegel

, Lower

, Schafer

, et al. (2012). HLA typing from RNA-Seq sequence reads. Genome Med, 4:102.

Bray

, Pimentel

, Melsted

, Pachter

(2016). Near-optimal probabilistic RNA-seq quantification. Nat Biotechnol, 34:525-527.

Butts

, Socinski

, Mitchell

, et al. (2014). Tecemotide (L-BLP25) versus placebo after chemoradiotherapy for stage III non-small-cell lung cancer (START): a randomised, double-blind, phase 3 trial. Lancet Oncol, 15:59-68.

Castle

, Kreiter

, Diekmann

, et al. (2012). Exploiting the mutanome for tumor vaccination. Cancer Res, 72:1081-1091.

10.

Cibulskis

, Lawrence

, Carter

, et al. (2013). Sensitive detection of somatic point mutations in impure and heterogeneous cancer samples. Nat Biotechnol, 31:213-219.

11.

Ding

, Wendl

, McMichael

, Raphael

(2014). Expanding the computational toolbox for mining cancer genomes. Nat Rev Genet, 15:556-570.

12.

Dvinge

, Bradley

(2015). Widespread intron retention diversifies most cancer transcriptomes. Genome Med, 7:45.

13.

Editorial

(2017). The problem with neoantigen prediction. Nat Biotechnol, 35:97.

14.

Griffith

, Griffith

, Smith

, et al. (2015). Genome modeling system: a knowledge management platform for genomics. PLoS Comput Biol, 11:e1004274.

15.

Hackl

, Charoentong

, Finotello

, Trajanoski

(2016). Computational genomics tools for dissecting tumour-immune cell interactions. Nat Rev Genet, 17:441-458.

16.

Hundal

, Carreno

, Petti

, et al. (2016). pVAC-Seq: a genome-guided in silico approach to identifying tumor neoantigens. Genome Med, 8:11.

17.

Kahles

, Lehmann

, Toussaint

, et al. (2018). Comprehensive analysis of alternative splicing across tumors from 8,705 patients. Cancer Cell, 34:211-224 e216.

18.

Kaumaya

, Foy

(2012). Peptide vaccines and targeting HER and VEGF proteins may offer a potentially new paradigm in cancer immunotherapy. Future Oncol, 8:961-987.

19.

Kim

, Kim

, et al. (2018). Neopepsee: accurate genome-level prediction of neoantigens by harnessing sequence and amino acid immunogenicity information. Ann Oncol, 29:1030-1036.

20.

Koboldt

, Zhang

, Larson

, et al. (2012). VarScan 2: somatic mutation and copy number alteration discovery in cancer by exome sequencing. Genome Res, 22:568-576.

21.

Kreiter

, Vormehr

, van de Roemer

, et al. (2015). Mutant MHC class II epitopes drive therapeutic immune responses to cancer. Nature, 520:692-696.

22.

Kudrin

(2012). Overview of cancer vaccines: considerations for development. Hum Vaccin Immunother, 8:1335-1353.

23.

, Goedegebuure

, Gillanders

(2017). Preclinical and clinical development of neoantigen vaccines. Ann Oncol, 28(suppl_12):xii11-xii17.

24.

, Yu

, Li

, et al. (2009). SOAP2: an improved ultrafast tool for short read alignment. Bioinformatics, 25:1966-1967.

25.

Liu

, Yang

, Duffy

, et al. (2013). ATHLATES: accurate typing of human leukocyte antigen through exome sequencing. Nucleic Acids Res, 41:e142.

26.

Lundegaard

, Lamberth

, Harndahl

, et al. (2008). NetMHC-3.0: accurate web accessible predictions of human, mouse and monkey MHC class I affinities for peptides of length 8-11. Nucleic Acids Res, 36(Web Server issue):W509-512.

27.

McKenna

, Hanna

, Banks

, et al. (2010). The Genome Analysis Toolkit: a MapReduce framework for analyzing next-generation DNA sequencing data. Genome Res, 20:1297-1303.

28.

McLaren

, Pritchard

, Rios

, et al. (2010). Deriving the consequences of genomic variants with the Ensembl API and SNP Effect Predictor. Bioinformatics, 26:2069-2070.

29.

Melero

, Gaudernack

, Gerritsen

, et al. (2014). Therapeutic vaccines for cancer: an overview of clinical trials. Nat Rev Clin Oncol, 11:509-524.

30.

Morse

, Chui

, Hobeika

, et al. (2005). Recent developments in therapeutic cancer vaccines. Nat Clin Pract Oncol, 2:108-113.

31.

Navin

(2014). Cancer genomics: one cell at a time. Genome Biol, 15:452.

32.

Nielsen

, Andreatta

(2016). NetMHCpan-3.0; improved prediction of binding to MHC class I molecules integrating information from multiple receptor and peptide length datasets. Genome Med, 8:33.

33.

Ott

, Hu

, Keskin

, et al. (2017). An immunogenic personal neoantigen vaccine for patients with melanoma. Nature, 547:217-221.

34.

Robert

, Schachter

, Long

, et al. (2015). Pembrolizumab versus Ipilimumab in advanced melanoma. N Engl J Med, 372:2521-2532.

35.

Sahin

, Derhovanessian

, Miller

, et al. (2017). Personalized RNA mutanome vaccines mobilize poly-specific therapeutic immunity against cancer. Nature, 547:222-226.

36.

Schubert

, Walzer

, Brachvogel

, et al. (2016). FRED 2: an immunoinformatics framework for Python. Bioinformatics, 32:2044-2046.

37.

Seiler

, Peng

, Agrawal

, et al. (2018). Somatic mutational landscape of splicing factor genes and their functional consequences across 33 cancer types. Cell Rep, 23:282-296 e284.

38.

Shiraishi

, Sato

, Chiba

, et al. (2013). An empirical Bayesian framework for somatic mutation detection from cancer genome sequencing data. Nucleic Acids Res, 41:e89.

39.

Shukla

, Rooney

, Rajasagi

, et al. (2015). Comprehensive analysis of cancer-associated somatic mutations in class I HLA genes. Nat Biotechnol, 33:1152-1158.

40.

Smart

, Margolis

, Pimentel

, et al. (2018). Intron retention is a source of neoepitopes in cancer. Nat Biotechnol, 36:1056-1058.

41.

Sonnhammer

, von Heijne

, Krogh

(1998). A hidden Markov model for predicting transmembrane helices in protein sequences. Proc Int Conf Intell Syst Mol Biol, 6:175-182.

42.

Stoeckius

, Hafemeister

, Stephenson

, et al. (2017). Simultaneous epitope and transcriptome measurement in single cells. Nat Methods, 14:865-868.

43.

Stronen

, Toebes

, Kelderman

, et al. (2016). Targeting of cancer neoantigens with donor-derived T cell receptor repertoires. Science, 352:1337-1341.

44.

Szolek

, Schubert

, Mohr

, et al. (2014). OptiType: precision HLA typing from next-generation sequencing data. Bioinformatics, 30:3310-3316.

45.

Tang

, Subha (2017). neoantigenR: an annotation based pipeline for tumor neoantigen identification from sequencing data. bioRxiv; doi.org/10.1101/171843.

46.

Tappeiner

, Finotello

, Charoentong

, et al. (2017). TIminer: NGS data mining pipeline for cancer immunology and immunotherapy. Bioinformatics, 33:3140-3141.

47.

Vansteenkiste

, Zielinski

, Linder

, et al. (2013). Adjuvant MAGE-A3 immunotherapy in resected non-small-cell lung cancer: phase II randomized study results. J Clin Oncol, 31:2396-2403.

48.

Warren

, Choe

, Freeman

, et al. (2012). Derivation of HLA types from shotgun sequence datasets. Genome Med, 4:95.

49.

Whiteside

, Demaria

, Rodriguez-Ruiz

, et al. (2016). Emerging opportunities and challenges in cancer immunotherapy. Clin Cancer Res, 22:1845-1855.

50.

Zhang

, Mardis

, Maher

(2017). INTEGRATE-neo: a pipeline for personalized gene fusion neoantigen discovery. Bioinformatics, 33:555-557.

51.

Zhang

, White

, Schmidt

, et al. (2016). INTEGRATE: gene fusion discovery using whole genome and transcriptome data. Genome Res, 26:108-118.

52.

Zhang

, Kim

, Hundal

, et al. (2017). Breast cancer neoantigens can induce CD8(+) T-cell responses and antitumor immunity. Cancer Immunol Res, 5:516-523.

53.

Zhou

, Lyu

, Wu

, et al. (2017). TSNAD: an integrated software for cancer somatic mutation and tumour-specific neoantigen detection. R Soc Open Sci, 4:170050.