An Empirical Antigen Selection Method Identifies Neoantigens That Either Elicit Broad Antitumor T-cell Responses or Drive Tumor Growth

The clinical benefit of immune checkpoint inhibitor (ICI) therapies in cancer provides evidence that the immune system can control and even eradicate tumors (1). The effectiveness of ICI therapies relies primarily on T-cell recognition of neoantigens derived from tumor-specific mutations (2–4). Cancer vaccines based on these neoantigens have the potential to target a broad array of targets by increasing the pool of neoantigen-specific CD4⁺ and CD8⁺ T cells, which may limit metastatic tumor escape. Both T-cell subsets have clear importance in controlling tumor growth (5, 6). Moreover, CD4⁺ T-cell support during priming is critical for secondary expansion and effector function of CD8⁺ T-cell responses (7). Successful generation of immunity against cancer neoantigens has the potential to increase the efficacy of ICI therapies while being safe and well tolerated.

The feasibility of personalized neoantigen cancer vaccines was recently demonstrated in two independent clinical trials in which computational methods were used to predict MHC class I antigens from large numbers of neoantigen candidates (8, 9). In these studies, 20% or fewer of the neoantigens selected as MHC class I epitopes were immunogenic for CD8⁺ T cells when measured directly ex vivo after vaccination. Most of the responses were measured in the CD4⁺ T-cell subset despite having been predicted as CD8⁺ T-cell antigens (10). In silico prediction for MHC class II antigen identification presents multiple challenges including polymorphisms in the class II molecules, variations in epitope length, the influence of peptide flanking regions, and challenges in the identification of the correct binding core (11). Even if computational methods correctly predict MHC antigen presentation, antigens selected by this approach may still fail because the patient's T-cell repertoire cannot respond to the antigen in vivo.

To overcome the challenges associated with in silico neoantigen prediction, we developed ATLAS, a robust ex vivo platform that empirically identifies neoantigens to which a subject has already generated an immune response (12). With this HLA-agnostic bioassay, each putative antigen is expressed individually in Escherichia coli, with or without a coexpressed cytolysin to facilitate MHC class I or class II processing, respectively (13). The E. coli library is fed in an ordered array to autologous dendritic cells derived from each patient's peripheral blood and incubated overnight with their T cells. If the T cells recognize their cognate antigen, they will secrete measurable cytokines. ATLAS has previously yielded novel antigens with clinical efficacy when administered as a vaccine against infectious diseases (14–16), but has not been previously applied to patients with cancer, or to immunologically competent murine cancer models.

Here, we report preclinical and clinical results demonstrating that ATLAS accurately identifies neoantigens that drive both stimulatory and inhibitory T-cell responses to cancer, as defined by inflammatory cytokine secretion. Many identified neoantigens did not fit common selection criteria for in silico neoantigen selection. Therapeutic vaccination of mice with ATLAS-identified stimulatory neoantigens resulted in antitumor efficacy as a monotherapy. By contrast, vaccines that included even a single neoantigen associated with inhibitory responses in ATLAS led to accelerated tumor outgrowth and/or abrogation of vaccine-mediated protection. We have termed these inhibitory antigens “Inhibigens”; they are defined as antigens to which inflammatory T-cell cytokine secretion is suppressed to levels below baseline. These studies demonstrate the feasibility of using an ex vivo antigen detection method to formulate patient-specific cancer vaccines and demonstrate the potential importance of distinguishing between stimulatory and inhibitory antigens in a treatment or prevention setting.

A cohort of nine patients with lung cancer was recruited into an observational study to assess preexisting neoantigen-specific T-cell responses using the ATLAS assay. Patient demographics, HLA haplotypes, and prior treatment are summarized in Table 1, Supplementary Fig. S1A and S1B, and Supplementary Table S1, respectively. Tumor sequencing identified between 84 and 971 somatic variants per patient, with a median of 700. In total, 1,015 unique mutations were profiled: 934 variants for CD8⁺ T-cell responses and 525 variants for CD4⁺ T-cell responses (Table 1; Supplementary Table S2). The median number of mutations classified as immunologically recognized stimulatory neoantigens based on the ATLAS assay was low, with 5% and 1% of the screened mutations for CD4⁺ and CD8⁺ T cells, respectively. Surprisingly, many neoantigen-specific T-cell responses were inhibitory, defined as downregulating IFNγ and/or TNFα cytokine secretion to levels significantly below baseline controls. Inhibitory responses were observed to a median of 3% and 1% of mutations for CD4⁺ and CD8⁺ T cells, respectively (Fig. 1A; Table 1).

Using the developer's recommended 2% cutoff (17), only 6% of the epitopes predicted by NetMHCpan and none of the epitopes predicted by NetMHCIIpan elicited a stimulatory T-cell response in the ATLAS assay (Fig. 1B). Moreover, 32% of the neoantigens identified by ATLAS to elicit CD8⁺ T-cell IFNγ and/or TNFα secretion were not predicted by algorithms. In this cohort of patients, 7% of the NetMHCpan-predicted antigens were identified as Inhibigens, demonstrating that current algorithms are unable to distinguish between neoantigens and Inhibigens, as algorithms predict only peptide binding to specific MHC/HLA and are not designed to predict the immune response to the peptide. Therefore, these algorithms may result in the inadvertent prioritization of potentially deleterious targets. Next, common criteria used for in silico predictions were assessed to determine if any could enrich for mutations that elicited patient T-cell responses in ATLAS. Neoantigens and Inhibigens identified using the ATLAS assay were not enriched for known cancer genes (Fig. 1C), or novel sequences (indels, frameshifts, neoORFs) preferentially over short variants [single-nucleotide variants (SNV) Fig. 1D], nor were all ATLAS-identified neoantigens expressed as determined by RNA-sequencing (RNA-seq) analysis, which measures only genes that are actively transcribed at the time of collection and in the specific portion of tumor that is biopsied (Fig. 1E). Furthermore, there was no relationship between transcript abundance and antigenicity (Fig. 1F). The differential predicted affinity of the mutated versus wild-type antigen for the MHC molecule, measured by use of the differential agretopic index (DAI), was not different for either stimulatory neoantigens or Inhibigens (Fig. 1G). Finally, we tested whether reducing the stringency of the NetMHCpan cutoff would lead to predictions consistent with those detected by ATLAS. Although reducing the binding affinity threshold did decrease neoantigens missed by the algorithm, the proportion of Inhibigens also increased (Fig. 1H). Therefore, the ATLAS bioassay empirically identifies neoantigens in peripheral blood of patients with cancer that are not predicted with in silico methods and distinguishes putative antigens based on their ability to stimulate or inhibit T-cell responses ex vivo.

To determine if the neoantigens identified using ATLAS were biologically relevant, the ATLAS ex vivo assay was applied to the B16F10 murine melanoma model. Whole-exome sequencing of B16F10 melanoma tumors resected from C57BL/6 mice identified 1,679 nonsynonymous mutations, exceeding the 962 previously reported (18) and in line with more recent sequencing of B16F10 (Supplementary Table S3; ref. 19). An ATLAS library expressing each individual mutation was constructed and used to screen CD8⁺ T cells from the spleens of tumor-bearing mice, resulting in the identification of 66 neoantigens and 57 Inhibigens (Fig. 2A; Supplementary Table S3). A lead neoantigen candidate, Gal3St1_P64A (galactose-3-O-sulfotransferase; Supplementary Table S2), was selected for further study based on high IFNγ and TNFα secretion in multiple mouse ATLAS screens. Although the Gal3St1_P64A mutation has been previously identified in B16F10 tumors (18, 19), neither immunogenicity nor efficacy has been described with formulations containing this sequence after vaccination.

When a vaccine containing three antigens—Gal3St1_P64A and previously identified M30 (MHC II neoantigen; ref. 18) and TRP2 (MHC I tumor-associated antigen; ref. 20; Supplementary Table S4)—was therapeutically administered with adjuvant (CpG, QS21, and 3D-PHAD) to tumor-bearing mice, tumor growth was either significantly delayed or completely abrogated compared with controls (Fig. 2B). Gal3St1_P64A as a single antigen vaccine conferred some protection (Supplementary Fig. S2), but a combination of all three antigens was necessary for maximal antitumor efficacy, supporting the concept that a multiantigen, CD8⁺ and CD4⁺ broad-spectrum T cell–targeted vaccine is necessary for protection in this aggressive model (21). Of note, the formulation containing the ATLAS-identified Gal3St1_P64A neoantigen was significantly more protective than the published efficacious combination containing M27 (MHC I neoantigen; Fig. 2B; ref. 18). T-cell responses were generated to all vaccine antigens, with potent IFNγ responses measured to Gal3St1_P64A (Fig. 2C). Cross-reactivity with wild-type Gal3St1 was not observed (Supplementary Fig. S3). Vaccinated mice that showed no evidence of tumor growth were rechallenged with B16F10 on day 63. As compared with age-matched naïve controls, mice previously vaccinated with the Gal3St1_P64A vaccine had negligible tumor growth, confirming the generation of durable protective immunity (Fig. 2D).

The detection of Inhibigens by the ATLAS screen suggests potentially deleterious antigen-specific immune modulation by T cells, the biological relevance of which was evaluated by therapeutic vaccination with a pool of 27-mer synthetic long peptides (SLP) spanning four ATLAS-identified Inhibigens (Supplementary Table S4). In tumor-bearing mice, immunization with Inhibigen peptides led to a marked increase in tumor growth kinetics relative to mice that received adjuvant alone (Fig. 3A). Immune responses to Inhibigen peptide immunization, measured by IFNγ enzyme-linked immunoabsorbent spot (ELISpot) assays, were lower than those to stimulatory neoantigens (Supplementary Fig. S4). Flow cytometric analysis of progressing tumors revealed reduced T-cell infiltration, particularly by CD8⁺ T cells, as compared with control tumors (Supplementary Fig. S5). When the pool of Inhibigen peptides was added to a formulation containing M30 and TRP2, protection conferred by T-cell responses to M30 and TRP2 was completely abrogated (Fig. 3B), suggesting that deleterious Inhibigen-specific responses can be dominant to protective responses. After deconvolution of the peptide pool to test each Inhibigen individually with M30 and TRP2, two Inhibigens were identified as primarily responsible for reduced efficacy. In particular, the Inhibigen peptide spanning the MMP9 frameshift mutation (MMP9_FS) completely abrogated protection (Fig. 3C). The TAAR9_C190W Inhibigen also dampened protection but to a lesser extent, whereas ABCA17 and ASB14 Inhibigens had no demonstrable effect on tumor kinetics. The protumor effect of MMP9_FS immunization correlated with a significant decrease in T-cell IFNγ secretion specific for protective M30 and TRP2 antigens administered in the same vaccine pool (Fig. 3D). In addition, CD8⁺ T-cell infiltration into tumors was significantly diminished by MMP9_FS immunization (Fig. 3E and F; Supplementary Fig. S6). Analysis of CD4⁺CD25⁺FOXP3⁺ cells in the spleens of mice therapeutically immunized with a vaccine containing the MMP9_FS SLP did not reveal significant cell number differences relative to mice that received the vaccine in the absence of MMP9_FS (Fig. 3G). In addition, increases in FOXP3⁺ cells were not observed in the tumor microenvironment (TME) of Inhibigen-immunized mice by immunohistochemistry (IHC; Fig. 3F). Combined, these data suggest that differential expansion of conventional regulatory T cells (Treg) did not account for Inhibigen-induced phenotypic differences. Interestingly, a marked decrease in F4/80⁺ and CD11c⁺ cells in the TME was observed in Inhibigen-immunized mice, suggesting a differential effect on macrophage and/or dendritic cells in the tumor (Fig. 3F). Taken together, these data support the conclusion that the ATLAS approach detects both stimulatory and inhibitory antigens, and that including Inhibigens in a vaccine may lead to detrimental effects including decreased efficacy in vivo.

A planned interim analysis was performed on subjects participating in a phase I/IIa clinical trial that was designed to assess the safety and immunogenicity of ATLAS-identified neoantigens in a personalized vaccine for patients with melanoma, head and neck squamous cell cancer, urothelial carcinoma, or non–small cell lung cancer (NSCLC) who were without evidence of disease after initial therapy. Matched tumor and normal tissue from initial resection or biopsy were sequenced to identify the patient's unique mutanome followed by ATLAS screening (Supplementary Table S5). Neoantigens that elicited stimulatory CD4⁺ or CD8⁺ T-cell responses were included in the vaccine, and Inhibigens were excluded. Comprehensive ATLAS screening data across all patients are presented in Supplementary Table S5, and a representative patient screen is shown graphically in Supplementary Fig. S7. In the patient shown, stimulatory neoantigens were identified in both T-cell subsets, with 1 CD4⁺ neoantigen and 15 CD8⁺ neoantigens. In this patient, inhibitory responses were found in only the CD8⁺ T-cell subset, whereas in other subjects, inhibitory responses were found in either one or both T-cell subsets (Table 2; Supplementary Table S5). As in the observational lung cancer cohort, there was no association of immunogenicity in ATLAS with DNA copy number, DAI, or mutation type (short variant vs. neoORF; Supplementary Fig. S8A–S8H). Although there was a trend toward increased RNA expression of neoantigens versus nonantigens or Inhibigens, the correlation was low (Supplementary Fig. S8E). Not surprisingly for patients with cancer, there was significant interpatient variability in ATLAS outcomes with respect to frequency of stimulatory neoantigens and Inhibigens (Table 2). Application of NetMHCpan to patient-specific mutations was unable to predict most ATLAS-identified stimulatory neoantigens. Conversely, a number of Inhibigens identified by ATLAS would have been included in a NetMHCpan-directed vaccine (Supplementary Fig. S8H).

The clinical vaccine product, GEN-009, consists of 4 to 20 SLPs divided into 4 pools, with each pool containing 1 to 5 SLPs. A single pool was administered into each extremity combined with the adjuvant poly-ICLC (Hiltonol, Oncovir, Inc.). The vaccine was administered in a prime-boost schedule of five immunizations (Supplementary Fig. S9). Eight patients were vaccinated. One patient developed recurrent disease during vaccine manufacturing and therefore did not receive the vaccine (patient D). Interestingly, this subject had complete resection for stage IIIA melanoma without further adjuvant therapy but had a predominance of Inhibigens identified in their mutanome (Table 2). Patients for whom a vaccine was not feasible or who dropped out due to early progression were replaced. Supporting manufacturing feasibility, a GEN-009 personalized vaccine was generated for 100% of eligible patients, and the median overall time from tumor collection to vaccination was less than 16 weeks. To date, 37 vaccinations have been administered without a dose-limiting toxicity; adverse events were consistent with those expected from the adjuvant alone and were self-limited. These consisted of grade 1 injection site reactions/pain, fatigue, chills, headache, and influenza-like illness, with a single occurrence of grade 2 pyrexia. There were no clinically significant laboratory abnormalities. No patients have experienced disease recurrence after vaccination with a median follow-up of 34 weeks (range, 15–46 weeks), although one subject was diagnosed with a molecularly distinct second primary tumor while in the treatment period. Clinical follow-up of these patients is ongoing.

T-cell responses to the GEN-009 vaccine were analyzed by dual-color fluorospot assays in prevaccination and postvaccination samples. Pools of four overlapping peptides (OLP; 15 mer overlapping by 11 aa) were synthesized for each neoantigen SLP represented in the vaccine. Total peripheral blood mononuclear cells (PBMC) and sorted CD4⁺ or CD8⁺ T cells were evaluated for responsiveness to each of the individual immunized SLPs as well as vaccine pools. Ex vivo responses were assessed by IFNγ/Granzyme B (GrB) fluorospot assay. For the in vitro stimulation (IVS) assay, cells were expanded for 10 days with OLP pools and then assessed in an IFNγ/TNFα fluorospot assay. Comprehensive immunogenicity data for a single patient are shown in Fig. 4A–C. Modulation of neoantigen-specific CD4⁺ and CD8⁺ T-cell responses after vaccination was apparent by both ex vivo (plots A and B) and IVS (plot C) fluorospot. The responses varied for each immunized peptide, in some cases increasing after immunization and in other cases decreasing for a particular analyte. After vaccination, all patients had CD4⁺ and CD8⁺ T-cell responses to at least one peptide ex vivo (Fig. 4D; Supplementary Table S6). Across the cohort, most subjects had T-cell responses to a proportion of peptides included in their vaccine measurable ex vivo at baseline (Fig. 4D) and after IVS (Fig. 4E). Overall, there were immune responses detected to nearly all (99%) of the administered peptides (Fig. 4F; Supplementary Table S6). Ex vivo neoantigen-specific CD4⁺ and CD8⁺ T-cell responses were identified to 51% and 41% of peptides, respectively. In addition, IVS CD4⁺ and CD8⁺ T-cell responses were observed against 87% and 59% of administered peptides, respectively. Responder analysis (statistical differences in magnitude after vaccination compared with baseline) revealed that on day 50, by the IFNγ assay, the magnitude of PBMC and CD4⁺ T-cell responses was statistically greater than responses at baseline for a median of 88% of SLPs; for CD8⁺ T-cell responses, increases were statistically significant for 63% of the SLPs (Supplementary Table S7). The TNFα results were similar, but with slightly lower medians. Overall, these data show that empirical selection of stimulatory neoantigens with ATLAS results in a greater proportion of immunogenic responses after vaccination than previously reported (8, 9).

Finding therapeutic T-cell targets has been critical to progress against cancer. The identification of relevant tumor-specific neoantigens holds the promise of improving cancer immunotherapy through broad expansion of tumor-expressed and immunogenic targets, but determination of which mutations are relevant for an individual patient has been elusive. Although de novo neoantigen-specific vaccine responses may be important for successful immunotherapy (22), in silico neoantigen selection has yielded poor predictive value and provides little confidence that putative prioritized antigens are relevant to the individual patient's HLA type and T-cell repertoire. Moreover, the success of tumor-infiltrating lymphocyte (TIL) therapies shows that preexisting T cells can be protective if they are successfully expanded and otherwise augmented (reviewed in ref. 23). De novo responses may also carry additional risk of off-tumor toxicity as the T cells generated are novel to the patient. Here, we describe the implementation of an empirical method to identify immunogenic tumor mutations using a biological assay customized for the patient. ATLAS comprehensively screens all tumor-specific mutations and identifies personalized neoantigens to which a patient's T cells have already responded (i.e., preexisting T-cell responses). Importantly, to be detected by the ATLAS bioassay, identified neoantigens must have been effectively processed and presented in the context of the patient's tumor for recall T-cell responses to be generated. The ATLAS assay has successfully identified protective infectious disease antigens for use in candidate vaccines and immunotherapies (24–27). The ATLAS assay is also well suited for high-throughput and comprehensive screening of all somatic mutations in the setting of solid-tumor cancers because ATLAS is HLA-agnostic and inclusive of all genetic and ethnic backgrounds and does not rely on epitope predictions. In addition, ATLAS identifies whether the T-cell response a patient generated to each of their neoantigens is potentially beneficial (stimulatory) or deleterious (inhibitory).

In the retrospective research cohort of nine patients with lung cancer, screening of peripheral blood T cells against respective tumor mutanomes revealed that approximately 10% of mutations were immunogenic (both CD4⁺ and CD8⁺ T-cell subsets), and, of these, 6% were newly discovered Inhibigens. The relative proportion of neoantigens was highly variable between subjects, confirming previous reports (28). By contrast, the observed proportions of neoantigens greatly exceed the neoantigen-specific response frequencies of 1% to 2% identified in TIL from tandem minigenes or in peripheral blood screens of PD-1–sorted T cells. This discrepancy could be due to multiple factors. First, ATLAS libraries are built based on nonsynonymous mutations (relative to germline) identified in DNA, whereas many other screens have filtered out mutations based on RNA expression prior to in silico selection (29). Our data suggest that this approach may remove some bona fide neoantigens, as these analyses only measure RNA expression in the region of tumor biopsied and do not account for tumor heterogeneity. Conversely, it is possible that our screens identify neoantigens that were expressed in the tumor in the past but have since been immunoedited. Next, TIL approaches require sufficient tumor material to facilitate T-cell outgrowth from resected tumor fragments. It is possible that not all neoantigen-specific T cells will migrate out of the tumor and expand during this process, thereby resulting in a perceived lower response breadth. Indeed, a recent publication showed that addition of 4-1BB agonistic antibody to tumor explants resulted in a greater proportion of neoantigen-specific T-cell responses emerging from tumors (30). Third, multimer approaches are limited in allele coverage and are dependent on algorithmic predictions for selection (31). Finally, peptide-based approaches require epitope prediction and exogenous loading of MHC class I or class II epitopes for presentation, whereas the ATLAS approach utilizes native intracellular antigen processing and presentation.

For each of the datasets presented here, when neoantigens identified by ATLAS were compared with predicted epitopes based on NetMHCpan algorithms (17, 32), there was minimal overlap. In the retrospective lung cancer cohort, for example, the MHC I prediction algorithm had a 13% positive predictive value relative to ATLAS outputs. When the algorithm predicted an antigen that was confirmed by ATLAS, approximately half of the time the response was identified as inhibitory, a biological response that cannot currently be predicted by an algorithm. In addition, 32% of the neoantigens identified by ATLAS were not in silico predicted. NetMHCIIpan was even less accurate for MHC II where the algorithm had a 3% positive predictive value (all were characterized by ATLAS as Inhibigens) and a 93% false negative rate. The poor predictive value of epitope prediction algorithms based on affinity is not surprising; indeed, van Rooij and colleagues (33) showed that less than 1% of predicted high-affinity binders from a patient with melanoma were recognized by the patient's T cells. Moreover, a mass spectrometry approach recently revealed that weakly binding neoepitopes could be protective tumor antigens in mouse models (34), confirming that vaccines designed based upon predicted high-affinity binding epitopes may miss many appropriate antigens and include potentially unsuitable targets. In addition, in order to prioritize predicted epitopes for vaccination using algorithms, supplementary selection criteria are commonly used, such as selecting mutations in known cancer genes or those that generate novel sequences such as insertions or deletions, or using criteria such as expression levels, DNA copy number, or DAI (reviewed in ref. 35). In the current study, no enrichment for neoantigens or Inhibigens was observed by these criteria. There was a slight trend for increased RNA expression of neoantigens, but the differences were small with a large overlap in distributions. We acknowledge that only one prediction algorithm was used in this study, and there are others available for public or private use that may yield different outcomes. In addition, multiple groups are trying to improve on existing algorithms by using mass spectrometry and monoallelic cell line approaches (34, 36, 37). Despite progress in the various in silico approaches, our data suggest that the heterogeneity of MHC, the complex biology of antigen processing, presentation, T-cell responsiveness, and the context in which T cells are educated to mutations during cancer initiation and progression are far more complex than can be currently modeled in silico.

The complexity of neoantigen identification is exemplified by the detection of Inhibigens in our study. In ATLAS, these responses are characterized by the reduction of IFNγ and/or TNFα secretion to levels that are statistically below baseline controls. As Inhibigen responses have also been found in screens of nonmutated tumor antigens and pathogen proteins (38, 39), these responses do not appear to be unique to cancer mutations. The biological relevance of Inhibigen-specific responses was demonstrated in the B16F10 melanoma model, where inclusion of ATLAS-defined Inhibigen peptides to an otherwise protective vaccine abrogated efficacy and, in some cases, accelerated tumor progression. Corresponding loss of neoantigen-specific T-cell immunogenicity to protective antigens was observed, as well as reduced tumor infiltration by CD8⁺ T cells, macrophages, and dendritic cells. Vaccine-augmented tumor growth, rather than protection, has been reported in other mouse models and attributed to tolerization of antitumor response (40, 41).

Further investigation is required to reveal how ATLAS-identified Inhibigens exert protumor effects and determine whether the phenotype is a property of the Inhibigen itself, or whether T cells are aberrantly educated to Inhibigens in the context of the tumor. Several possibilities exist: (i) Inhibigen-specific Tregs may be suppressing antitumor immune responses. Tregs are functionally and phenotypically heterogeneous and can act through a number of mechanisms, including production of inhibitory cytokines, cytolysis, metabolic disruption of effector T cells, and/or inhibition of dendritic cell maturation and function (42). Although the broader class of Tregs cannot be fully discounted at this time, simple induction of classic Tregs (CD4⁺CD25⁺FOXP3⁺) is unlikely, as these Tregs were not more abundant in the TME of Inhibigen-immunized mice relative to controls. Moreover, with ATLAS, Inhibigens are commonly identified in the CD8⁺ T-cell subset. (ii) Inhibigens might modulate myeloid-derived cells directly, either by disrupting antigen presentation or by inducing these cells to express immunosuppressive molecules such as indoleamine 2,3-dioxygenase that suppress effector T-cell function (43). Monocyte-derived suppressor cells (MDSC) are also capable of suppressing effector T-cell responses; however, Inhibigens are identified in ATLAS in the context of a defined culture containing uniform monocyte-derived dendritic cells (MDDC) and sorted T cells during a brief overnight exposure. The differentiation of MDDCs to MDSCs in this short time frame is unlikely. (iii) Effector T cells responding to Inhibigens could be driven into a suppressive state. T-cell dysfunction due to persistent antigen stimulation of the T-cell receptor (TCR) is a common feature of cancer, and it is possible that Inhibigens induce tolerance through this mechanism (44). However, it is unclear how T-cell dysfunction of an Inhibigen-specific cell could suppress effector T cells against other antitumor antigens in trans. Alternatively, regulation of downstream TCR signaling through inhibition of protein kinase or phosphatase activity has been shown to disrupt T-cell activation (45), and Inhibigens could regulate these pathways.

To distinguish between these possibilities or uncover novel mechanisms of inhibition, future studies will explore the presence of inhibitory cytokines (e.g., IL10, TGFβ, and IL35) in various cell subsets after in vitro and in vivo exposure to Inhibigens. In addition, it will be critical to map the MMP9_FS minimal inhibitory epitope, which will enable tetramer development and characterization of cytokine profiles and transcriptomic patterns of purified Inhibigen-specific T cells. Inhibigen identification and characterization in additional syngeneic mouse tumor models will also be explored; however, the frequent identification of Inhibigens in many human screens underscores their broad relevance. We propose that T cells responding to Inhibigens, or the antigen-presenting cells (APC) that display them, can affect the local and systemic immune environment, leading to local changes in the TME as well as tumor growth kinetics; this accentuates the importance of utilizing ex vivo assays to define and exclude such antigens from any personalized cancer vaccine or cell therapy.

GEN-009 is a personalized cancer vaccine that includes only ATLAS-identified neoantigens and excludes Inhibigens. Analysis of the initial cohort of enrolled patients in our phase I clinical trial (NCT03633110) showed that generation of GEN-009 vaccine was consistently feasible and well tolerated. All vaccinated patients had detectable T-cell immune responses to the neoantigens in their vaccine, including immune responses overall against 99% of all immunized peptides, which is a remarkable and unprecedented result in the field (8, 9, 46). Broad CD4⁺ and CD8⁺ T-cell responses across multiple neoantigens were generated, including CD8⁺ T-cell responses that were measurable directly ex vivo. Both outcomes are notable in comparison with previously reported first-in-human, in silico–selected neoantigen vaccines (8, 9) in which ex vivo CD8⁺ T-cell responses were undetectable, and the proportion of peptides to which responses were measured was approximately 60% after IVS. These prior neoantigen vaccine studies could neither identify nor exclude Inhibigens with their in silico approaches. Notably, if Inhibigens were included in a cancer vaccine, they would likely appear as undetectable in an IFNγ ELISpot analysis. The GEN-009-101 study continues to evaluate the GEN-009 vaccine, excluding Inhibigens, in combination with standard-of-care immunotherapy regimens.

In summary, the translational and clinical data presented here demonstrate the feasibility of a high-throughput personalized approach to identifying mutation-associated neoantigens in murine models and patients with cancer. Administration of a vaccine optimized via selection of neoantigens with ATLAS should amplify preexisting immunity, create de novo responses to the same neoantigens, and allow expansion of targets via epitope spread. The ATLAS approach has unique advantages as it identifies both stimulatory and novel inhibitory antigens (Inhibigens). We hypothesize that the proportion of Inhibigen- to neoantigen-specific T-cell responses may correlate with clinical prognosis and are currently exploring this in a prospective study. It is possible that these inhibitory T-cell responses are associated with observations of resistance or hyperprogression in some subjects treated with ICI (reviewed in ref. 47); however, our data do not support dysregulation of PD-1⁺ Tregs as was recently reported (48). If the deleterious impact of Inhibigens by therapeutic vaccination in a mouse model translates to humans, then it is crucial to identify and exclude these Inhibigens from cancer immunotherapies. In support of this empirical antigen selection approach, a preplanned interim analysis of the first cohort of patients enrolled in the GEN-009-101 clinical trial confirmed that the stimulatory antigens identified via ATLAS were bona fide immunogenic neoantigens. Overall, Genocea's ATLAS platform provides critical information about a patient's individual neoantigen immunity and could enable cancer immunotherapies that augment a patient's preexisting antitumor response and improve clinical outcomes.

The objectives of these studies were to explore the peripheral T-cell responses to tumor mutations in human subjects and in mouse models. Our hypothesis was that unbiased measurement of T-cell responses to each mutation in a subject's tumor would yield more accurate and biologically relevant information about neoantigen specificities than the in silico methods commonly used for neoantigen prediction. To this end, we recruited patients with cancer for an observational study, explored the phenotype of inhibitory and stimulatory T-cell responses in a mouse model, and finally began a phase I/IIa clinical trial to treat patients with cancer with the GEN-009 immunotherapy. Power calculations were performed for group sizes in the mouse model.

The ATLAS assay has been used extensively for screening of T-cell responses in human subjects, with detailed methods described previously (12, 24, 25, 27, 49, 50).

Genomic DNA from formalin-fixed, paraffin-embedded (FFPE) tumor sections and matched saliva was isolated and subjected to whole-exome capture and paired-end sequencing and RNA-seq using the Nova-seq platform. Sequencing and data analysis were performed at Personalis, which have demonstrated accuracy and precision of their experimental protocols, with a sensitivity of 99% for SNVs (3% false positive rate) and 94% for indels (3% false positive rate). The coverage was 150× for normal and 300× for tumor. Somatic variants were called from the exome reads and reference human genome using hs37d5 and BWA and STAR alignment tools, and variants were called using MuTect, Verdict, and Fusion Catcher. Genomic DNA extracted from B16F10 tumors raised in C57BL/6 mice was sequenced using whole-exome sequencing with HiSeq2500 platform. One hundred base pair (bp) paired-end sequencing was performed at Otogenetics, and sequences were aligned against the mm10 reference genome. The coverage was 120×. The B16F10 tumor was then compared against germline C57BL/6 DNA from mouse tail to identify tumor-specific mutations using Strelka and Verdicts. The original B16F10 cell line (ATCC, passage 6) was also sequenced to confirm presence of these same mutations. Variant annotation was performed using SnpEff (dataset: National Center for Biotechnology Information Sequence Read Archive accession #PRJNA662658).

For the lung cohort and murine model studies, 399-bp DNA fragments encoding each of the identified mutations were synthesized (GENEWIZ or Twist Biosciences) and cloned into a proprietary plasmid (pGEN4) with each fragment ligated in-frame with a DNA sequence encoding the SIINFEKL epitope from chicken ovalbumin. In the GEN-009-001 study, 123-bp DNA fragments encoding each identified mutation were synthesized (GENEWIZ) and cloned into a proprietary vector (pGEN132) containing a red fluorescent protein (Fresno RFP, ATUM Bio) coding sequence in-frame. Clones expressing the first 399 or 123 bp (depending on use of pGEN4 or pGEN132, respectively) of green fluorescent protein from Branchiostoma lanceolatum (blFP-Y3; abbreviated NG) were used as an irrelevant negative control antigen. Sequence-verified clones were transformed into E. coli containing a separate plasmid encoding a cytoplasmic variant of listeriolysin O (cLLO) for screening of CD8⁺ T cells or lacking cLLO for CD4⁺ T cells (13). Expression was induced using IPTG and verified either by Western blot or by detection of RFP fluorescence. Clones and negative controls were fixed with formalin, washed, normalized, and then randomly rearrayed in duplicate in 384-well plates. Plates were stored at −80°C until used (13).

PBMC samples were collected by either venipuncture or apheresis, enriched via density gradient centrifugation, and viably frozen within 8 hours. Frozen PBMCs were thawed and sorted using microbeads (Miltenyi) to enrich for monocytes and to separate the CD4⁺ and CD8⁺ T cells. The monocytes were cultured with recombinant human GM-CSF and recombinant human IL4 for 7 days to differentiate into MDDCs. The CD4⁺ and CD8⁺ T cells were separately expanded nonspecifically for 1 week using anti-CD3 and anti-CD28 microbeads (Invitrogen) and low-dose IL2. Due to cell limitations, not all mutations could be screened against both T-cell subsets for all patients. Under those circumstances, CD8⁺ T cells were prioritized.

For screening of ATLAS libraries, autologous MDDCs were added to individual wells of the arrayed E. coli plates, with a target of 5,000 cells per well. Plates were incubated at 37°C with 5% CO₂ for 2 hours to facilitate antigen uptake and processing. After washing, 80,000 autologous T cells (either CD4⁺ or CD8⁺) were added to each well and incubated overnight. In addition to the negative controls expressing blFP-Y3 (NG), phorbol 12-myristate 13-acetate (PMA)/ionomycin mitogen (81 nmol/L and 1.34 μmol/L, respectively) was added as a positive control for T-cell activity. Supernatants were transferred to new 384-well plates and frozen at −20°C until analyzed for IFNγ and TNFα cytokine levels by Meso Scale Discovery (MSD) assay. Briefly, supernatants and calibrators were added to precoated MSD plates and incubated. After washing, the SULFO-TAG–conjugated detection antibody was added, and after washing, read buffer was added followed by analysis with an MSD instrument.

The murine neoantigen screens were performed as described above, with the exception that only CD8⁺ splenocytes were tested, and the RAW309 Cr.1 macrophage cell line (ATCC TIB-69) was used as APC. The RAW309 cells were incubated overnight and then the ATLAS library was added the next day. The original ATLAS screen interrogated the entire mutanome, and two repeat screens evaluated a subset of 174 clones based on the initial data.

Responses were identified based on mean and SD estimates from cytokine levels of both the negative control (NG) and replicate test wells from MSD plates based on defined standards. Data from MSD plates are first log-transformed (natural log) prior to analysis. Any test wells with responses that do not fall within the linear range of the standard curve are excluded. Given the approximate normal distribution of the data, a span of 2 SDs from the mean represents a 95% confidence interval for nonresponse. Therefore, a threshold of 2 SDs above the mean was used for identification of neoantigens with the expectation of a false positive rate of 5%. Z-transformed data based on calculated mean and SDs were used to identify and visualize antigen calling from MSD plate data.

B16F10 cells were obtained from the ATCC, expanded to six passages, and frozen as a stock for use across mouse studies. Thawed cells were cultured for up to 4 days without passaging prior to mouse injection. Cells at final passage prior to freezing tested negative for Mycoplasma by Lonza MycoAlert detection kit (2018) and negative for all infectious disease by Charles River Research Animal Diagnostic Services (2018). Whole-exome sequencing of the final frozen stock was performed by Otogenetics confirming cell line identity (2018).

Female 6- to 8-week-old C57BL/6 mice were obtained from The Jackson Laboratory for use in mouse studies under the Institutional Animal Care and Use Committee–approved protocol #18-0625-1: Evaluation of the effects of vaccination in mouse tumor models.

SLPs (27 aa) were synthesized by GenScript, New England Peptide, or 21st Century Biochemicals. B16F10 melanoma cells (1 × 10⁵; ATCC CRL-6475) were inoculated subcutaneously into the flanks of 6- to 8-week-old female C57BL/6 mice. Tumor-bearing mice were randomized and then injected s.c. in the tail base on days 3, 10, and 17 with formulations containing 50 μg per SLP, 5 μg CpG ODN1826 (Invivogen), 5 μg 3D-PHAD (Avanti), and 25 μg QS21 (Invivogen). Tumor sizes were measured every 3 days until ≥2,000 mm³ and daily until tumor ≥2,500 mm³, upon which mice were euthanized. For immunogenicity studies, mice were immunized on days 0 and 7 in the absence of tumor and euthanized for splenocyte analysis on day 14.

ELISpot assay for the measurement of IFNγ from mouse T cells pulsed with antigen has been previously described (51, 52). Stimulants of media alone, PMA/ionomycin (200 ng/mL; 1 μg/mL), or OLPs (15 mer overlapping by 11 aa; 1 μg/peptide) were prepared. Stimulants and cells were coincubated for 16 hours at 37°C after which IFNγ secretion was detected using anti-IFNγ capture antibody (10 μg/mL clone AN18; Mabtech) and anti–IFNγ-biotin (1 μg/mL clone R4-6A2; Mabtech). All SFCs were normalized to 1 × 10⁶ splenocytes.

TILs were isolated using enzymatic digestion with collagenase I (100 U/mL; Worthington), collagenase IV (500 U/mL; Sigma), and DNAse (200 μg/mL; Sigma). After digestion, single-cell suspensions were filtered through a 70-μm filter, and lymphocytes were sorted using CD45 microbeads (Biolegend, Clone 30F11, STEMCELL isolation kit). Cells were stained in 5% FBS in PBS. Antibodies included fixable viability dye (BD), CD19 (6D5, Biolegend), NK1.1 (PK136, Biolegend), CD4 (RM4-5, Thermo Fisher Scientific), CD8 (53-6.7, Biolegend), CD25 (PC61, Biolegend), CD3 (17A2, Biolegend), and CD45 (30-F11, Biolegend). Collection was performed on a BD LSRFortessa, and analysis was performed using FlowJo.

For IHC, resected and FFPE tumors were prepared and sectioned by JMD Histology and Histologistics. Slides were deparaffinized and subjected to antigen retrieval in either citrate or EDTA buffer at 100°C for 20 minutes in a TintoRetriever pressure cooker from Bio SB. Slides were blocked with 10% goat serum and incubated with primary antibody overnight at 4°C. Antibodies included CD4 (D7D2Z, 1:50), CD8 (D4W2Z, 1:200), FoxP3 (D6O8R, 1:100), F4/80 (D2S9R, 1:200), and CD11c (D1V9Y, 1:350) from Cell Signaling Technology. All washes were performed with TBS-Tween20. Slides were incubated with polink-1 AP rabbit secondary antibody and subsequently with permanent red chromogen dye from Origene. Tissues were counterstained with hematoxylin and mounted with aqueous mounting medium. Images were acquired on the EVOS FL Auto imaging system (Life Technologies) and prepared in ImageJ (NIH) software.

ATLAS-prioritized neoantigens were manufactured as SLPs (27 aa) under cGMP conditions, formulated in physiologic buffer, and frozen at -20°C. Each personalized vaccine consisted of between 4 and 20 SLPs divided into 4 pools, with each pool containing 1 to 5 SLPs. The specific number of SLPs in each patient's vaccine was dependent on the number of neoantigens identified by ATLAS and the number that could be synthesized and formulated. Peptides were combined with poly-ICLC (Hiltonol) for subcutaneous administration.

GEN-009-101 is a first-in-human study evaluating the safety, immunogenicity, and anticancer activity of the GEN-009 vaccine in patients with solid tumor malignancies. The clinical trial was registered with designation NCT03633110 (https://clinicaltrials.gov/ct2/show/NCT03633110). Part A, for which we now report preliminary results, enrolled patients without evidence of disease after initial therapy. The primary endpoint is safety and immunogenicity of administered peptides. The vaccine was injected s.c. for 5 doses over 6 months on days 1, 22, 43, 85, and 169 (Supplementary Fig. S9). The clinical sites enrolling the study obtained written informed consent from patients via an Institutional Review Board–approved consent form, and the trial was conducted in accordance with recognized ethical guidelines under the supervision of an institutional or central review board.

Part A was open to adults with cutaneous melanoma, NSCLC, squamous cell carcinoma of the head and neck, or urothelial carcinoma, who had completed standard treatment (e.g., surgical resection, adjuvant therapy, and/or radiotherapy) with curative intent. Additional eligibility requirements included Eastern Cooperative Oncology Group (ECOG) performance status of 0 or 1 and adequate hematologic, liver, and kidney function. Exclusion criteria included active autoimmune disease, primary immune deficiency, or immunosuppressive therapies within 28 days prior to study.

All patients signed written informed consent after the nature and possible consequences of the studies were explained, and before the conduct of any protocol-specific procedures. Leukapheresis was performed at baseline and day 50 for the ATLAS process and immunogenicity assessments. In addition, whole blood was collected at additional timepoints for immunogenicity. Disease assessment, CT, or MRI was performed within 4 weeks prior to GEN-009 dosing to rule out disease recurrence. Patients with urothelial carcinoma who did not have a complete cystectomy additionally underwent urine cytology and, if appropriate, cystoscopy to verify disease status.

Disease-free survival is being assessed every 12 weeks through the first year of treatment, and every 26 weeks thereafter. Safety measurements performed throughout the vaccine treatment period included evaluation of adverse events, clinical laboratory tests, physical examination, vital signs, and ECOG status. Adverse events are graded according to the NCI Common Terminology Criteria for Adverse Events v5.0.

PBMCs were stimulated in a 24-well plate at 5 × 10⁶ cells/well with 1 μg each peptide/mL in the presence of IL7. On days 3, 6, and 9, IL2 was added. After 10 days, T cells were sorted for CD4⁺ or CD8⁺ T cells using magnetic bead isolation (Miltenyi) before plating for the fluorospot assay.

96-well precoated human IFNγ/GrB (ex vivo) or IFNγ/TNFα (IVS) fluorospot kits (Mabtech) were used. OLPs (15 mer overlapping by 11 aa) were synthesized for all vaccinated SLPs. For the ex vivo fluorospot, 2 × 10⁵ PBMC or 2 × 10⁵ CD4- or CD8-depleted PBMCs were added to each well with 1 μg/mL each OLP and incubated for 48 hours. For the IVS fluorospot, 2 × 10⁴ CD4⁺ or CD8⁺ T cells, 5 × 10³ lymphocyte-depleted PBMCs, and 1 μg/mL each OLP were added and incubated for 24 hours. Plates were washed and then detection antibody was added, followed by anti–BAM-490 and SA-550 (Mabtech). After washing, fluorescence enhancer was added, and then the plates were read on the AID iSpot Reader System (AID).

After bias correction and failure removal, the amount of cytokine production per well was normalized to the multiple redundant measurements of the negative control neon green (NG). Only cytokine measurements that fell within the linear range of the standard curves run on the same plate were included in analyses. A candidate antigen was defined as stimulatory or inhibitory if the cytokine production was 2 times the median average deviations above or below the median of the negative controls, respectively.

In addition to B16F10 ATLAS clones, clones expressing mouse actin, a housekeeping gene for which mice should not have any T-cell reactivity, were included as negative control antigen arrayed randomly across the plates. For each MSD screening plate and cytokine, a multilevel model was fit to the log-transformed data. The estimated mean response to the negative control was subtracted from the mean response to each candidate neoantigen, and the estimated residual SD was divided by this difference to yield a normalized log cytokine concentration that can be interpreted as SDs above or below the background control. Values greater than 2 were deemed stimulatory responses, and values less than -2 were deemed inhibitory responses.

A response was defined as positive if the mean SFC exceeded the LOD and if the P value between the test and negative control was ≤ 0.05 by the DFR(eq) test (53).

H. Lam reports a patent for PCT/US18/23442 pending. L.K. McNeil reports patents for PCT/US18/23442 and PCT/US20/33277 pending to Genocea Biosciences. R.B. Cohen reports grants and personal fees from Genocea Biosciences during the conduct of the study; and grants and personal fees from HEAT Biologics outside the submitted work. M.L. Johnson reports grants from Genocea Biosciences during the conduct of the study; and personal fees from Otsuka and Astellas Pharma, grants and other from Genentech/Roche, Boehringer Ingelheim, AstraZeneca, Calithera Biosciences, Merck, Loxo, Sanofi, Mirati Therapeutics, Pfizer, Guardant Health, Incyte, AbbVie, Atreca, GlaxoSmithKline, Gritstone Oncology, Janssen, Lilly, Novartis, Amgen, Daiichi Sankyo, EMD Serono, and WindMIL, other from Ribon Therapeutics, Achillies Therapeutics, Association of Community Cancer Centers, Bristol-Myers Squibb, and G1 Therapeutics, and grants from Kadmon, Genmab, Stem CentRx, Checkpoint Therapeutics, Array BioPharma, Regeneron, Hengrui Pharmaceuticals, Lycera, BeiGene, Tarveda Therapeutics, CytomX Therapeutics, Dynavax, Corvus Pharmaceuticals, Genocea Biosciences, Adaptimmune, Syndax, Neovia Oncology, Acerta Pharma, Takeda, Shattuck Labs, Apexigen, Atreca, OncoMed, Immunocore, Jounce Therapeutics, University of Michigan, TCR2 Therapeutics, Arcus Biosciences, Ribon Therapeutics, BerGenBio, Tmunity Therapeutics, and Seven and Eight Biopharmaceuticals outside the submitted work. M.L. Gillison reports other from Genocea Biosciences during the conduct of the study; and personal fees from Gilead, EMD Serono, BMS, Kura, Shattuck Labs, Bayer Healthcare, Roche, BioNTech AG, Merck, and NewLink Corporation and other from BMS, Kura, and Bicara Therapeutics outside the submitted work. M.N. Stein reports grants from Genocea Biosciences during the conduct of the study; and grants from Exelixis, Tmunity, Seattle Genetics, Nektar, Lilly, Bristol Myers Squibb, Harpoon, and Janssen Oncology outside the submitted work. U.N. Vaishampayan reports grants and personal fees from BMS, Inc., and Alkermes, and personal fees from Exelixis, Bayer, Pfizer, Merck, and Onc Live outside the submitted work. A.P. DeCillis reports personal fees from Genocea Biosciences, Exelixis, Osmol, Cybrexa, Codiak, Monopteros, Pyramid, and Evelo outside the submitted work. J.J. Foti reports other from Genocea Biosciences during the conduct of the study. W. Broom reports a patent for PCT/US18/23442 pending; in addition, W. Broom was an employee of Genocea Biosciences at the time of his contribution to this article. E.M. Jaffee reports personal fees from Genocea Biosciences during the conduct of the study; and personal fees from CSTONE, Achilles, Adaptive Biotech, dragonfly, and Parker Institute, grants from Lustgarten Foundation outside the submitted work. K.-K Wong is a founder and equity holder of G1 Therapeutics and has consulting and sponsored research with AstraZeneca, Janssen, Pfizer/Array Biopharma, Novartis, Merck, Zentalis, and Genocea Biosciences as well as sponsored research (only) with Takeda, BMS, Mirati, Alkermes, Merus, Amgen, Ansun Biopharma, Eliven Therapeutics, Tvardi Therapeutics, Delfi Diagnostics, and Dracen Pharmaceuticals. C.G. Drake reports other from Genocea Biosciences during the conduct of the study; and other from Bayer, BMS, Compugen, F-Star, Kleo, Merck, Merck-Serono, Pfizer, Roche/Genentech, Shattuck Labs, Tizona, and Werewolf outside the submitted work; in addition, C.G. Drake had a patent for LAG-3 issued and licensed to BMS. P.M. Carroll reports a patent for PCT/US18/23442 pending. T.A. Davis reports personal fees and other from Genocea Biotherapeutics during the conduct of the study; and personal fees and other from Genocea Biosciences outside the submitted work. J.B. Flechtner reports personal fees and other from Genocea Biosciences outside the submitted work; in addition, J.B. Flechtner had patents for PCT/US18/23442 and PCT/US20/033277 pending to Genocea Biosciences. No disclosures were reported by the other authors.

H. Lam: Conceptualization, data curation, formal analysis, supervision, validation, investigation, visualization, methodology, writing–original draft, writing–review and editing. L.K. McNeil: Formal analysis, supervision, investigation, methodology, writing–original draft, writing–review and editing. H. Starobinets: Conceptualization, resources, data curation, formal analysis, investigation, visualization, methodology, writing–review and editing. V.L. DeVault: Conceptualization, resources, data curation, formal analysis, validation, investigation, visualization, methodology, writing–review and editing. R.B. Cohen: Resources, investigation. P. Twardowski: Resources, investigation. M.L. Johnson: Resources, investigation. M.L. Gillison: Resources, investigation. M.N. Stein: Resources, investigation. U.N. Vaishampayan: Resources, investigation. A.P. DeCillis: Conceptualization, resources, investigation. J.J. Foti: Resources, investigation, methodology. V. Vemulapalli: Resources, data curation, software, formal analysis, validation, visualization, methodology. E. Tjon: Resources, data curation, software, formal analysis, validation, visualization, methodology. K. Ferber: Resources, data curation, software, formal analysis, validation, visualization, methodology. D.B. DeOliveira: Resources, investigation, methodology. W. Broom: Resources, formal analysis, supervision, investigation. P. Agnihotri: Investigation, methodology. E.M. Jaffee: Conceptualization. K.-K. Wong: Conceptualization. C.G. Drake: Conceptualization. P.M. Carroll: Conceptualization. T.A. Davis: Conceptualization, resources, supervision, validation, investigation, methodology, writing–review and editing. J.B. Flechtner: Conceptualization, resources, data curation, formal analysis, supervision, validation, investigation, visualization, methodology, writing–original draft, writing–review and editing.

The authors are grateful to the patients who consented to participate in our observational study and clinical trial. They also thank Genocea's extraordinary clinical operations team (Jessica Price, Kevin Mancini, and Rich Hernandez) and manufacturing team including Narinderjeet Singh and Manish Jain. They also thank the ATLAS team including Louisa Dowal, Tulin Dadali, Crystal Cabral, Michael O'Keeffe, Christopher Warren, James Loizeaux, Mariya Croll, James Perry, Anna Lyubetskaya, Erick Donis, and Matthew Lanchantin; the Translational Immunology team including Mara Shainheit, Tyler Fenske, Megan Croshier, Devin Champagne, Gabriella Santone, and Syukri Shukor; the Computational Biology team including Ning Wu and Huilei Xu; and the Preclinical Development team including Stephanie Rinaldi, Sanmit Adhikari, Osaruese Odeh, Peri Matatia, Simran Singh, Dylan Sheehan, and Cindy Nguyen.