Antigen Selection
1. Start with a good open-source antigen prediction pipeline (rather than developing a new pipeline in-house)
Rationale: Predicting tumor antigens from genomics data seems deceptively easy, e.g., just call some variants from whole exome sequencing, look for expression of these in RNA-seq data, predict HLA binding with netMHCpan or MHCFlurry, and rank predictions by binding. That’s what was done in the first papers describing a neoantigen prediction workflow (and is what many groups still do). However, the vast majority of predicted neoantigens are neither presented by HLA nor immunogenic in vivo, so the nuances of which features to capture from the genomics data, filtering heuristics, and ranking heuristics can matter quite a bit. Although there are multiple pipelines reported in the literature, most of these are proprietary and/or underspecified in methods. There are three main open-source neoantigen prediction workflows to choose from, deploy without much friction, and build upon (pVACtools, OpenVax, LENS). The most widely used in clinical trials so far has been pVACtools (GitHub repo, paper), which was developed by Malachi and Obi Griffith’s lab at Washington University in St. Louis and has supported the WashU and Jamie Leandro Foundation vaccine trials. OpenVax (GitHub repo, paper) was developed by Alex Rubinsteyn, Tim O’Donnell, and Julia Kodysh while working in the Hammerbacher lab at Mt. Sinai in New York, supporting the Mt. Sinai trials run through Nina Bhardwaj’s group. LENS (GitHub repo, paper) is our software, developed to allow prediction of tumor antigens from genomic sources beyond small somatic mutations, harmonize RNA expression across genomic sources for antigen prioritization, capture features important for immunogenicity prediction, and be transparent, reproducible, and extensible based on containerization. So, depending on what you want to do, you could leverage the development time, testing, and prior FDA review of any of these pipelines.
Counterpoints: If you have special insight into genomics prediction of tumor antigens in general or for a subclass of antigens, especially if you have novel bioinformatics methods required for antigen prediction, you’d want to incorporate these. In that case, you might still want to build on one of the open-source workflows given they’ve already been reviewed by the FDA and their libraries could speed development time. You may also have collaborator(s) in academia or industry with proprietary pipelines that you’d like to use. If so, I would benchmark these against one or more open-source pipelines first to ensure good performance.
2. Consider including expressional tumor antigens beyond genomically templated sources (e.g. alternative splice variants, endogenous retroviral elements, and cancer/testis antigens) but use only those with minimal expression in medullary thymic epithelial cells
Rationale: Even in tumors with a high mutation rate, only a minority of predicted neoantigens are presented by cancer cells in vivo, and only a minority of those are immunogenic. In some tumor types, somatic mutations are rare. Optimal antigen selection may include expressional antigens as well as antigens derived from somatic mutations. You can use publicly available RNA-seq data from medullary thymic epithelial cells to exclude expressional antigens that are expressed highly in mTECs (which would be less likely to be immunogenic given negative selection in the thymus).
Counterpoint: Using only neoantigens derived from somatic mutations allows confidence in tumor-specificity of the specific antigen peptide(s). Splice variants, CTAs and ERVs may be expressed by normal tissues, potentially causing toxicity if the vaccine elicits potent antigen-specific T cell responses. That said, using publicly available genomics datasets to confirm tumor-specificity may attenuate this risk.
3. Use long-read as well as short-read sequencing if possible to capture larger somatic mutations
Rationale: Long-read whole genome and whole transcriptome sequencing (e.g., Pacific Biosciences or Oxford Nanopore) allow identification of germline and somatic structural variants that cannot be identified by short-read sequencing (e.g., Illumina) alone. Large somatic variants may elicit more immunogenic neoantigens than single-nucleotide variants, as they contain longer stretches of non-self amino acid subsequences (i.e., look more like viruses than self). Our workflow for long read data analysis (Exacto) is here: GitHub repo / blog post.
Counterpoints: Long read sequencing is expensive and requires fresh or frozen tissue for high-molecular weight DNA and RNA extraction. For these reasons, it may be logistically challenging. It may be the case that most large structural variants can be resolved with short read sequencing (unclear as robust benchmarking has not yet been done). No study has yet shown an example of an immunogenic neoantigen that could be detected by long-read sequencing but not short-read sequencing.
4. Validate antigen presentation as best you can using immunopeptidomics studies
Rationale: Only a minority of predicted neoantigens are presented by HLA in vivo. The true positive rate of prediction is probably 2-20% depending on the tumor and genomics/bioinformatics workflow, so antigen selection using genomics alone would yield mostly false positives. Mass spectrometry based immunopeptidomics can be used to confirm HLA presentation of tumor antigen peptides (examples: 1, 2, 3). Another option would be to validate predicted tumor antigens by looking for antigen-specific T cell responses prior to vaccination (presuming they would only be present if the tumor antigens were presented in vivo). The downside of this second option is that you would be measuring the cancer’s immunological history, not necessarily generating evidence that the cancer cells were presenting the antigen(s) at the time of sample collection.
Counterpoints: Like extraction of high-molecular weight DNA and RNA for long-read sequencing, immunoprecipitation of peptide/HLA molecules for immunopeptidomics currently requires fresh or frozen tumor samples as the input material. Sensitivity of shotgun immunopeptidomics studies is lower than you’d like, and targeted immunopeptidomics is prohibitively expensive to screen hundreds of peptides per patient. Even if you integrate immunopeptidomics into your integral biomarker for antigen selection, you will likely need to allow for cases where there are insufficient numbers of validated tumor antigens to choose from. You could use immunopeptidomics for antigen prioritization rather than making it a hard filter on antigen selection.
5. Include validated antigens such that the vaccine antigen set covers all tumor subclones
Rationale: Failure to eliminate all tumor subclones will select for escape variants that are resistant to T cell populations elicited by the vaccine (due to lack of relevant antigen presentation). You can use single cell sequencing studies to define the clonal landscape of the tumor and select an antigen set that covers all subclones. Our software for data analysis for this is here (github, blog post).
Counterpoints: Single cell sequencing requires capturing fresh or viably cryopreserved tumor cells, which may be logistically challenging. Antigen-directed immunotherapies that target a single antigen, or a set of antigens that do not cover all tumor subclones, may still be clinically effective if they induce epitope spreading and/or are part of a combination therapy strategy where other drugs can eliminate the resistant subclones.
6. Select validated antigens where the set covers HLA proteins encoded by alleles on both chromosomes
Rationale: Loss of heterozygosity of HLA (LOHHLA) is somewhat frequent mechanism of resistance to T cell-directed therapies. LOHHLA can happen without much cost in selective fitness of the cancer cells, however given that more extensive loss of HLA can trigger NK cell-mediated cytotoxicity, and a gene coding for a core subunit of RNA polymerase I lies within the HLA Class I loci, loss of both large chromosome segments containing all classical HLA class I loci has a much higher fitness cost. You can hedge against LOHHLA in vaccine design by including antigens presented by HLA proteins encoded by HLA alleles on both chromosomes. The original LOHHLA workflow is here (paper, github), and a variant of this is implemented in LENS.
Counterpoint: The optimal quantitative trade-offs between selecting antigens presented on HLA across both chromosomes and selecting optimal antigens regardless of HLA are unknown.
7. Select as few tumor antigens as possible (subject to the above constraints)
Rationale: Responses to potentially immunogenic antigens may be attenuated by competitive inhibition when multiple antigens are included in the vaccine (i.e., immunodominance amongst the antigen set). Selecting as few antigens as possible may minimize competitive inhibition of immunogenic antigens.
Counterpoints: The optimal quantitative trade-offs between increasing the frequency of immunogenic antigens in the vaccine (by selecting more antigens) and reducing antigen competitive inhibition (by selecting less antigens) are unknown. You could hedge risk here by allowing changes to antigen inclusion over time based on elicitation of antigen-specific T cell responses and tumor coevolution/immune escape.
Vaccine Formulation
8. Use an mRNA/LNP formulation if possible
Rationale: The only randomized clinical trial which has reported a benefit of neoantigen vaccination (in this case given with anti-PD-1 vs anti-PD-1 only) is the Merck/Moderna adjuvant melanoma trial that used an mRNA/LNP vaccine. We have some preclinical data in mice that mRNA/LNP vaccination may be superior to peptide/adjuvant vaccination, with more benchmarking ongoing (although adjuvant data are difficult to interpret translationally due to differences between innate immune signaling in mice vs humans). There are also practical advantages of an mRNA formulation: faster synthesis, no loss of antigens to synthesis failure based on hydrophobicity and cysteine content, and more flexible combination with other therapies (deliverable via LNPs).
Counterpoints: Peptide/adjuvant vaccines have a longer history of clinical trials and may be cheaper and easier to access for now. mRNA and LNP design and manufacturing are encumbered by IP constraints, which may matter if you are looking for investment or commercial returns from your vaccine platform. If you use a peptide vaccine, consider using multiple adjuvants (along with immune checkpoint inhibition) for potentially greater efficacy. DC vaccines may be more potent, but they are expensive in time and money to manufacture and as such are difficult to scale.
Vaccine Dosing
9. Give the vaccine in the context of combination immunotherapy
Rationale: Eliciting tumor antigen-specific T cells in vivo is not enough to cure people. As best you can, you also want to reverse or preempt resistance mechanisms. The best place to start is probably giving the vaccine with PD-1 inhibition (based on results from the Merck/Moderna melanoma study), however additional resistance mechanisms are probably applicable in your tumor of interest, and you may wish to target these as well.
Counterpoints: Increasing complexity of the intervention means increased cost in time and money for the trial. For anti-PD-1 specifically, this may interfere with establishment and maintenance of antigen-specific T cell memory (even as it improves effector function in the short term) (paper).
Correlative Studies
10. Measure everything you can that may be a resistance mechanism to see which ones occur and at what frequency
Rationale: For a therapeutic vaccine to work clinically, you need all of the following biology: (1) Antigen targets that are presented by HLA, (2) Enough peptide/HLA complexes on the cancer cell surface to support effective targeting by T cells, (3) Elicitation of enough T cells specific for target antigen(s) in the body to outcompete the number of tumor cells, (4) Achievement of T cell ingress into tumors, (5) Capacity of elicited T cells to kill cancer cells in the tumor microenvironment (including proper tuning of T cell-intrinsic phenotypes and subversion of intratumoral T cell suppression and evasion), and (6) Capacity to prevent or resolve acquired resistance mechanisms that develop under evolutionary selection pressure induced by your therapy. If a patient does not have a clinical response, then any of the above could’ve failed. It’s important to understand the which failure modes are common, how common they are, and whether they can be intervened upon to improve cure rates. If you have research biopsies pre- and post-therapy along with serial peripheral blood samples, you can evaluate these.
Counterpoints: Correlative studies are expensive, and the space of potential resistance mechanisms is huge, so evaluating them would be exploratory at first. The extent to which active resistance mechanisms are shared among patients is unknown, so the best you may be able to do in a small trial is categorize likely candidates and show they lead to T cell dysfunction and/or evasion in ex vivo studies (rather than causally linking them to lack of clinical response in specific patients). Best would be to embed these studies in a context of iterative refinement of treatment strategies, like the n-of-1 approach described in Sid Sijbrandij’s “Going Founder Mode on Cancer.”



Great post
The audio engineering metaphor is pretty clever for talking about institutional boundries. The idea that feedback happens when systems amplify signals they werent designed for makes sense both technically and as a framework for thinking about role fidelity. I worked around sound systems briefly and the principle of ringing out a room before going live definitely translates to spotting problems before they get amplified in other contexts too.