Tumors often co-exist with T cells that recognize somatically mutated peptides presented by cancer cells on major histocompatibility complex I (MHC-I). However, it is unknown why the immune system fails to eliminate immune-recognizable neoplasms before they manifest as frank disease. To understand the determinants of MHC-I peptide immunogenicity in nascent tumors, we tested the ability of thousands of MHC-I ligands to cause tumor subclone rejection in immunocompetent mice by use of a new ‘PresentER’ antigen presentation platform. Surprisingly, we show that immunogenic tumor antigens do not lead to immune-mediated cell rejection when the fraction of cells bearing each antigen (‘clonal fraction’) is low. Moreover, the clonal fraction necessary to lead to rejection of immunogenic tumor subclones depends on the antigen. These data indicate that tumor neoantigen heterogeneity has an underappreciated impact on immune elimination of cancer cells and has implications for the design of immunotherapeutics such as cancer vaccines.

T cells are specialized agents of the immune system that can detect and attack tumors. They spot their target by identifying small pieces of proteins – or antigens – at the surface of diseased cells. In particular, they can recognize the new and abnormal antigens that a cancer cell often displays. Yet, cells that become cancerous and start displaying suspicious antigens can manage to escape T cells and grow into full tumors. Why does the immune system not recognize and kill these early cancers before they get out of control?

One possibility is that T cells do not identify certain antigens carried by cancer cells. To test this, researchers have conducted experiments where they inject a mouse with cancer cells that display a single new antigen. If the animal develops a tumor, then this antigen does not trigger an immune response. However, this method is slow and laborious, because only one antigen can be tested at the time. Instead, Gejman, Chang et al. developed a new technique, PresentER, where a rodent gets injected with a mix of millions cancer cells that each displays a different antigen. This way, many thousands of new antigens can be studied in one go. The tumors are left to grow for several weeks before they are removed and analyzed to see which cells survived and which have been killed by the immune system.

Unexpectedly, the nature of the antigen did not make a big difference. Instead, cancer cells with new antigens could go undetected if they were rare and made up only a small proportion of all the different cancer cells in a tumor. However, the immune system would eliminate the exact same cancer cells when they were the major component of a cancerous lump.

Future research now has to explore exactly how rare cancer cells can hide amongst other cells, and remain invisible to the body. Armed with this knowledge, it might be possible to improve cancer therapy by prompting the immune system to target these emerging threats earlier.

Human cancers bear uniquely distinguishable features on the surface of their cells in the form of neoantigens, which are peptides derived from mutated, foreign or oncofetal proteins that are presented in complex with major histocompatibility complex I (MHC-I) molecules. These short, 8–11 amino acid fragments specifically mark cancer cells and activate potent immune responses that can lead to effective anti-cancer therapy (Tran et al., 2014; Rosenberg and Restifo, 2015; Zacharakis et al., 2018). Yet, the existence of T cells that recognize neoantigens is often not sufficient to eliminate tumors. Karl Hellström first described the coexistence of tumor-specific lymphocytes together with cancer cells in human solid tumors as a paradox 50 years ago (Hellström et al., 1968). In mice, an analogously enigmatic observation has been made that sporadic tumors occurring in aged or carcinogen-treated mice induce strong T cell responses only when transferred into new hosts, thereby preventing engraftment (Heike et al., 1994; Dubey et al., 1997; Shankaran et al., 2001). Thus, tumor-specific T cells can lead to tumor rejection in a new host or when used as a cancer therapy, but somehow immune surveillance is evaded during early tumorigenesis in the host that originally developed an immunogenic tumor. The increased rate of tumor formation in immunocompromised individuals has led to the hypothesis that the immune system can and does eliminate some tumors, particularly virally induced tumors, before they become clinically apparent (Schulz, 2009). We hypothesized that if the immune system can eliminate some early tumors, but not others, perhaps it is because some antigens are more potent at inducing effective T cell responses during early tumorigenesis. Identification and characterization of neoantigens that can induce an effective immune response and clear cancer cells is critical to understanding why and how immunogenic tumors develop in immunocompetent hosts.

Immunogenic peptides have been discovered in animal studies by injection of thousands or millions of tumor cells bearing neoantigens into animals and observing tumor rejection. However, the robust immune activation and tumor rejection in these cases is not analogous to the events of early tumorigenesis in humans, when the number of transformed cells is miniscule. Thus, even though some tumors are immunogenic, it is not clear why the host cannot eliminate them when the tumors first arise. If the biochemical features of neoantigens that lead to effective T cell responses were known, it might be possible to identify which tumors bear immunogenic antigens. Some reports have linked peptide immunogenicity to the biochemical characteristics of amino acid residues at certain positions along the MHC-I ligand (Calis et al., 2013), while others have focused on the difference between the affinity of a wild-type ligand and a mutated ligand (Duan et al., 2014). However, the absence of a large, unbiased data set of known immunogenic and non-immunogenic peptides has stymied the validation of these approaches. Indeed, most known immunogenic antigens are derived from viral proteins; few mutationally-derived neoantigens are confirmed as bonafide immunogenic peptides in mice or humans. Here, we have developed a novel genetic method to express libraries of precisely defined MHC-I ligands in mammalian tumor cells and have used this method to ask questions about MHC peptide immunogenicity in immunocompetent animals during early tumorigenesis.

Using libraries of genetically encoded MHC-I ligands, we tracked the dynamic growth and depletion of thousands of tumor subclones in vivo and noted a striking failure of cancer immunosurveillance that is potentially analogous to the failure of immune surveillance in humans during early tumorigenesis. We demonstrate for the first time that the ability of the naive immune system to surveille a nascent tumor and reject immunogenic subclones is limited by the fraction of cells expressing each unique antigen. Furthermore, we show that these rejection thresholds vary among antigens. Our data are consistent with the observation in humans that patients whose tumors have high numbers of subclones—and thus more subclonal neoantigens—have increased levels of relapse and worse survival than patients with more homogenous tumors (McGranahan et al., 2016; Reuben et al., 2017; Turajlic et al., 2018a). Thus, our data provide an antigen-specific rationale for the impact that tumor heterogeneity has on survival of human patients. According to our findings, antigen-specific immune effects are limited during early tumorigenesis, which has implications for the emergence and outgrowth of immunogenic tumors.

We have developed a reductionist method for encoding diverse peptide/MHC (pMHC) ligands in mammalian cells (termed ‘PresentER’), which was purposefully designed to enable us to ask questions about the immunogenicity of individual peptide epitopes presented by cancer cells at relatively physiologic levels. We used this method to perform high-throughput, pooled screening of MHC-I ligand immunogenicity in wild-type mice (Figure 1A). The PresentER antigen minigene is comprised of an endoplasmic reticulum (ER) signal sequence followed by a short peptide/epitope. Expression of the peptide and its display on MHC-I does not require proteasomal degradation or peptide processing, thus enabling precise definition of the exact epitope displayed to the immune system. As previously described (Gejman, 2018a), Transporter associated with antigen presentation (Tap) deficient cell lines expressing PresentER antigen minigenes lead to surface presentation of the encoded MHC-I peptide, detectable by multiple modalities, including fluorescently labeled antibodies directed to specific MHC-I ligands, mass spectrometry based immunopeptidomics and antigen-specific T cell reactivity. To demonstrate the applicability of PresentER antigen minigenes to study MHC-I ligand immunogenicity, we first asked if cancer cells encoding known immunogenic (mouse Tyrp1/gp75 TAYRYHLL W223A,H224Y (Dyall et al., 1998); mouse Ddx5/p68 SNFVFAGI S551F (Dubey et al., 1997); synthetic SIYRYYGL (Udaka et al., 1996); synthetic VTFVFAGL (Dubey et al., 1997); chicken ovalbumin SIINFEKL) or non-immunogenic (mouse Serpinf1/Pedf MSIIFFLPL (Wang et al., 2006); scrambled chicken ovalbumin FEKIILSN; mouse Ndufa4/dEV8 EQYKFYSV (Holler et al., 2003); mouse Ddx5/p68 SNFVSAGI (Dubey et al., 1997); mouse Tyrp1/gp75 TWHRYHLL (Dyall et al., 1998); Mouse Trp2 SVYDFFVWL) MHC-I ligands would be rejected by wild-type (WT) animals. The C57BL/6 syngeneic, Tap deficient mouse cell line RMA/S was transduced with these antigen minigenes and we first checked that the antigens were expressed at physiologically normal levels on the cell surface. Previously we, and others, had shown that endogenously expressed MHC-I tumor antigens are presented on the cell surface at several hundred to several thousand sites per cell (Dao et al., 2013; Sergeeva et al., 2011; Mathias et al., 2017). Using a radiolabeled TCR mimic antibody (25-D1.16) reactive with SIINFEKL/H-2Kb, we demonstrated that there were on average 3500 binding sites per RMA/S cell expressing PresentER-SIINFEKL, compared with ~90 sites on cells expressing PresentER-MSIIFFLPL, suggesting that the PresentER system allows display of the epitopes in a relatively physiologic range. We injected 5 × 106 cells subcutaneously into WT C57BL6/N mice. Tumors expressing known non-immunogenic peptides grew, while tumors expressing known immunogenic peptides were rejected in some or most animals (Figure 1B). Tap deficient cells were used because peptides derived from proteins that are not directed to the ER (e.g. eGFP) are not presented on MHC as these cells lack the ability to transport peptide from the cytoplasm into the endoplasmic reticulum.

At 7 days after tumor injection, T cells specific for the chicken ovalbumin peptide SIINFEKL could be detected in tumor draining lymph nodes of animals injected with PresentER-SIINFEKL expressing cells (Figure 1C). Haemotoxylin and Eosin (H and E) staining of regressing PresentER-SIINFEKL tumors showed lymphocytic infiltration with hyalin rich fibrin deposits, indicating cell death. By contrast, tumors with PresentER-FEKIILSN (a non-MHC-I binding peptide) were well-vascularized, highly cellular and with trace lymphocytic infiltration (Figure 1D). In order to verify that the mechanism of tumor rejection was indeed T cell dependent, we injected Rag-/- animals with SIINFEKL or FEKIILSN positive tumors and confirmed that SIINFEKL tumors were not rejected (Figure 1E). Taken together, these results indicate that RMA/S tumors expressing PresentER antigen minigenes can recapitulate the known immunogenicity of individual mouse MHC-I ligands in a T cell dependent manner.

To study immunogenicity in vivo at high throughput and complexity in this model, we first wanted to understand if immune responses directed at immunogenic antigens lead to rejection of cells presenting non-immunogenic antigens. If so, the ability to study immunogenicity in a pooled in vivo setting might be compromised. Mice were injected with pairs of tumors expressing an immunogenic and a non-immunogenic antigen minigenes (one minigene-expressing tumor on each flank): SIINFEKL/FEKIILSN, SIYRYYGL/EQYKFYSV and TAYRYHLL/TWHRYHLL. The immunogenic SIINFEKL, SIYRYYGL and TAYRYHLL expressing tumors were rejected, but the non-immunogenic FEKIILSN, EQYKFYSV and TWHRYHLL tumors found on the contralateral flank were not (Figure 2A). This suggests that tumor rejection is local and that an effective immune response to an immunogenic tumor does not affect the growth of a non-immunogenic tumor.

Next, to test if the immune system could identify and kill immunogenic subclones within a largely non-immunogenic tumor, we cultured cells with varying ratios of immunogenic (SIINFEKL) and non-immunogenic (FEKIILSN) cells. These cells were injected subcutaneously into mice and tumor size was monitored. In tumors where PresentER-SIINFEKL cells were greater than 25% of a tumor, tumors were smaller and tumor rejection occurred more frequently, especially when tumors were comprised of ≥50% immunogenic cells (Figure 2B). Next, we wanted to test if immunogenic subclones within a largely non-immunogenic tumor could be eliminated. We cloned mCherry into the PresentER vector and mixed PresentER-SIINFEKL (mCherry) with non-immunogenic PresentER-MSIIFFLPL (eGFP) cells at varying ratios and injected them into congenically marked CD45.1 mice. On day 16, the tumors were harvested and flow cytometry was performed to identify which (CD45.2 positive) tumor cells remained. Remarkably, within a non-immunogenic (eGFP labeled) tumor, immunogenic (mCherry labeled) sub-populations were eliminated (Figure 2C). Thus, in this model, the immune system is capable of recognizing and selectively depleting immunogenic tumor subclones within the context of a largely non-immunogenic tumor.

We hypothesized that a pooled screen in vivo might reveal the determinants of immunogenic MHC-I ligands if a tumor bearing a library of MHC-I antigen were depleted of cells bearing immunogenic peptide-MHC while tumor cells bearing non-immunogenic pMHC were spared (Figure 3A). Using the PresentER system, such a screen could be done on large scale and identify hundreds or thousands of immunogenic antigens at once, in contrast to identifying immunogenic antigens one-by-one. Immunoediting in vivo leading to loss of tumor clones with immunogenic neoantigen-encoding mutations has previously been observed in syngeneic mouse tumor models (Matsushita et al., 2012; DuPage et al., 2012; DuPage et al., 2011), suggesting that this approach might be viable.

We searched the mouse proteome in silico and randomly selected 5,000 8-mer peptides that were predicted by NetMHCPan to bind to the B6 mouse MHC-I allele H-2Kb. We also designed a library of mutated peptides by selecting single amino acid substituents of each peptide in the wild-type peptide library that did not eliminate MHC-I binding (Figure 3B). The MHC-I affinities and properties of the mutated peptides are further described in Figure 3—figure supplement 1 and Supplementary file 2. On average, the peptides in the mutated library have slightly higher affinity for MHC-I than the peptides in the wild type library, but the plurality of mutated peptides are within 100 nM of their non-mutated counterparts (Figure 3—figure supplement 1A–B). The residues that are changed from the wild-type to the mutant library tend to be at positions 4, 6 and 7 (Figure 3—figure supplement 1C). The majority of mutations are isomorphic (polar >polar, hydrophobic >hydrophobic and charged >charged), but ~⅓ of peptides feature a hydrophobic >other substitution (Figure 3—figure supplement 1D). Several known immunogenic and known non-immunogenic control antigen minigenes were included in each library, including some that were described in Figure 1. The libraries of wild type and mutant libraries were separately cloned and introduced into RMA/S cells by transduction at multiplicity of infection <0.3, thereby ensuring that few cells received more than one minigene. Naive C57BL6/N mice were injected with either (a) 5 × 106 cells expressing the wild type library, (b) 5 × 106 cells expressing the mutant library, (c) 5 × 106 mutant library cells plus 1 × 106 wild type RMA/S (‘padded’) or (d) 5 × 106 wild type library cells mixed with 5 × 106 mutant library cells (n = 5 per group). The tumors were allowed to grow for 17 days and then harvested (Figure 3C and Figure 3—figure supplement 2A). Genomic DNA was extracted from all of the tumors, as well as RMA/S library cells frozen on the day of injection (‘pretumor’ samples) and the minigenes encoded by each cell were amplified by PCR and sequenced by Illumina next generation sequencing.

Comparison of minigene abundance in the tumor outgrowth with minigene abundance in the pre-tumor samples led to the surprising finding that no minigenes were robustly depleted during growth in vivo (Figure 3D and Figure 3—figure supplement 2 and Figure 3—source data 1). This result stood in stark contrast to the experiments described above, in which mutant peptide expressing clones were efficiently depleted in immunocompetent mice. Some minigenes that were not abundant in the library (<1/10,000) at the time of injection were depleted or dropped out entirely in the tumor due simply to stochastic drop-out; however, minigenes that were abundant in the library at time of injection maintained their abundance despite in vivo growth of the tumor. Surprisingly, even positive control minigenes encoding strongly immunogenic peptides (orange points) were not depleted in this context. Analysis of minigene abundance in individual animal tumors (as opposed to the average of several tumors) yielded the same conclusion that few, if any, minigenes were reliably depleted (Figure 3—figure supplement 3).

The absolute number of cells at time of injection can be estimated from the relative abundance of each minigene in the pre-tumor samples. The least abundant minigenes were found at 1 per 5 × 106 cells (0.00002%) and the most abundant at ~23×103 per 5 × 106 (~0.5%). The immunogenic controls SIINFEKL, VTFVFAGL and SNFVFAGI were found at 0.005–0.013% of cells before injection into mice, which correspond to between 250 and 650 cells injected out of 5 × 106. The number of cells expressing the immunogenic antigens upon injection is very low and thus analogous to the number of cells that are present during early tumor development, when immunogenic transformed cells first arise. The surprising inability of the immune system to eliminate these demonstrably immunogenic cells may shed light on the limits of immune cell activation and killing during the early stages of tumorigenesis that enable outgrowth and escape of immunogenic cancers.

A possible explanation for failure of immune surveillance in our RMA/S MHC-I minigene library model might be insufficient antigen available in the growing tumor to activate an initially productive anti-tumor immune response in naive mice. If this were true, we hypothesized that T cells from antigen experienced mice might be able to detect and kill immunogenic cells in a highly heterogeneous tumor. Although vaccination with soluble peptides has been shown to generate robust T cell immunity (Ott et al., 2017; Sahin et al., 2017), this is not cost effective at scale. Alternatively, there is precedent for the idea that vaccination with a library of mutated antigens can lead to immunogenic T cell responses that lead to slower tumor growth or clearance (Engelhorn et al., 2006). We decided to vaccinate mice with irradiated tumor cells bearing the library of MHC-I peptides. In order to avoid confounding immunity to the RMA/S cells themselves (Van Hall et al., 2006), we used CRISPR/Cas9 to generate a B6-syngeic Tap2-/- MCA205 fibrosarcoma cell line. A single cell clone of Tap2 knockout MCA205 was selected and Tap2 knockout was validated by RT-PCR and next generation sequencing. Decreased surface MHC-I staining was expected and observed, because the Tap complex is a key chaperone of peptide/MHC-I formation (Figure 4—figure supplement 1).

WT B6 mice were vaccinated three times, once every 6 days, with 1 × 107 irradiated MCA205∆Tap2 cells bearing the wild-type library minigenes (Figure 4A). Splenocytes and draining lymph nodes from three vaccinated and three non-vaccinated mice were harvested at day 18 after the final vaccination and analyzed for the presence of antigen experienced T cells. Five control peptide tetramers were used, three of which are immunogenic and were present in the library (SIINFEKL, SNFVFAGI, VTFVFAGL), one which is not immunogenic but was present in the library (MSIIFFLPL) and one which is immunogenic but not found in the library (SIYRYYGL). Only the immunogenic peptides found in the library showed an increased number of CD44+/tetramer+ CD8 T cells, while the other two peptides did not show significant changes (Figure 4B). Therefore, vaccination with the library yielded detectable T cell populations specific to the immunogenic peptides.

Vaccinated and non-vaccinated mice were then challenged with 5 × 106 RMA/S cells bearing the wild type peptide library. Slower tumor growth was noted in the vaccinated as compared to the non-vaccinated mice, suggesting that a vaccine-related anti-tumor effect may have occurred (Figure 4C). However, neither vaccinated nor unvaccinated animals showed depletion of immunogenic control minigenes in relation to the non-immunogenic control minigenes (Figure 4D–F and Figure 4—source data 1). Thus, although slower tumor growth was noted, antigen-specific immunity was not observed in response to prophylactic vaccination in the library setting, suggesting a possible response to some other, broadly expressed, cellular antigens not represented in the peptide library.

While we have demonstrated that PresentER minigenes peptides can generate effective antigen-specific immunity in bulk RMA/S tumor assays, the same response does not occur in tumors bearing libraries of MHC-I ligands. In order to test if there is a threshold level of tumor cell clonality necessary to effectively activate the immune system, CD45.1 mice were injected with mixtures of immunogenic (labeled with eGFP) and non-immunogenic (labeled with mCherry) RMA/S cells and flow cytometry was performed on reisolated tumors 17 days later (Figure 5A). Relative to their proportion upon engraftment, tumor cells bearing the immunogenic peptides SIINFEKL and TAYRYHLL were depleted when they comprised as little as 1% of the tumor. Below 1%, depletion of cells bearing these two minigenes could not be detected. Depletion of cells bearing the SNFVSAGI peptide could be reliably detected at 50% and in some tumors at 10%, however depletion could not be detected when the SNFVSAGI cells were found at less than 10% of the tumor (Figure 5B). These findings are surprising, as they indicate that immunogenic tumor subclones can persist within a tumor and that rejection or persistence is dependent on tumor cell percentage of the total tumor mass during tumorigenesis, and not immunogenicity of the cell alone.

Failure of the immune system to eliminate immunogenic subclones present at low fractional abundance could either be due to the low quantity of total antigen present in the tumor or to the low percentage of cells bearing each antigen. In order to discriminate between these two possibilities, we increased the amount of tumor injected, thus increasing the total amount of antigen the immune system sees while keeping the percentage of each antigen within the tumor the same. We grew RMA/S library tumors in Rag-/- mice, harvested the tumors at 20 days and retained a portion of the material for sequencing. The rest of the tumor material was transferred into the flank of WT B6 or Rag-/- mice. Each animal received approximately 1 milliliter of tumor fragments (~2.5×108 cells), which represents a 40–100 fold increase in cells expressing each antigen (Figure 5C). After 17 days, the transferred tumors were harvested and sequenced. Once again, as we observed in both naive and vaccinated mice, robust depletion of cells bearing immunogenic peptide minigenes did not occur. Overall minigene abundance was highly correlated between the tumor material transferred and the tumors harvested 17 days later in both WT and Rag-/- mice and in both the wild type and mutant libraries (Figure 5D and Figure 5—source data 1). These results reveal that it is the percentage of tumor cells bearing each antigen within a tumor and not the total quantity of antigen that determines if an effective immune response occurs.

The immune system is capable of recognizing cancer cells as foreign based on the presentation of altered or unusual MHC-I ligands on the surface of tumor cells—a phenomenon that has been leveraged for cancer therapies such as immune checkpoint blockade and adoptive cell transfer (Tran et al., 2014; Rosenberg and Restifo, 2015; Zacharakis et al., 2018). Despite this, the immune system fails to clear immunogenic, clinically detected tumors on its own, as has been paradoxically noted for decades (Hellström et al., 1968). It is not clear at what point in tumorigenesis the immune system begins to initiate (ineffectively) a response to immunogenic MHC-I ligands presented on cancer cells. There is epidemiologic evidence that immunocompromised patients develop more tumors, suggesting that the immune system may prevent some tumors from ever manifesting clinically. However, recent data also suggests that mutations accumulate at high rates in sun-damaged, but otherwise normal, tissue at levels comparable to cancer cells (Martincorena et al., 2015) and that tumors face overall little negative selective pressure (Martincorena et al., 2017). Indeed, immune escape by loss of MHC (Masuda et al., 2007; Cabrera et al., 2000; Gudmundsdóttir et al., 2000; So et al., 2005; Shukla et al., 2015), when it occurs, is a late and subclonal event (McGranahan et al., 2017; Turajlic et al., 2018b). If T cell surveillance were highly effective during early tumor development, negative selective pressure would be notable in the evolutionary trajectories of tumors and HLA loss would be expected to occur early, frequently and in a clonal manner. We interpret the data from human and animal experiments to suggest that routine T cell immunosurveillance of nascent cancer cells does occur, but is limited by unknown factors, and that recognition of tumors as foreign occurs mostly later in tumor development.

If tumor immunosurveillance is sometimes effective in the early growth of a tumor, we hypothesized that it might be due to potent neoantigens expressed by cancer cells. Discovery of the biochemical characteristics of mutationally-derived neoantigens that are immunogenic would be important to clarify why some neoantigens are tolerated by the immune system and others are not. The significance of this question is underscored by the major challenge currently facing cancer immunotherapy: the identification of patients whose tumors bear immunogenic neoantigens—and are thus likely to respond well to immune checkpoint blockade or other immunotherapies—vs those patients who will be unnecessarily exposed to potentially toxic therapies without a chance for efficacy.

Here, we have developed a reductionist approach to determine which MHC-I ligands are immunogenic in immunocompetent mice. We demonstrate that tumors are rejected when an immunogenic pMHC is expressed on all or most injected cancer cells. To our surprise, we have discovered that effective T cell responses are not mounted against highly immunogenic peptides when cells expressing these peptides are a minor fraction of the tumor, which is the case in early developing tumors in humans. Furthermore, the threshold percentage of tumor cells necessary to yield an antigen-specific immunogenic T cell response varies with the antigen. We propose that ineffective T cell responses may be a consequence of intratumoral (or intra-tissue) heterogeneity and that neoantigen clonal fraction is an important, overlooked aspect of MHC-I antigen immunogenicity.

There are many mechanisms by which immune mediated killing of immunogenic cancer cells can fail. Here we report the first evidence that tumor heterogeneity is an explicit factor leading to failure of effective immunity in a tumor and that the maximum level of heterogeneity tolerable before immune escape occurs is dependent on the antigen. Immune checkpoints or active tumor suppression do not explain the observed immune evasion because such mechanisms would be expected to suppress T cell killing irrespective of the fraction of tumor cells that expresses an immunogenic epitope. This is evidenced by robust depletion of the highly immunogenic peptides SIINFEKL and TAYRYHLL when present at ≥1% of the tumor. The mechanism by which T cell responses are restrained when immunogenic MHC-I antigens are present at low frequency is not yet clear. Low levels of antigen presentation may lead to ineffective cross-presentation of antigen in the tumor draining lymph node, thus limiting T cell activation (Spiotto et al., 2002). In animal models, low levels of immunogenic epitopes of oncogenic drivers presented on transformed cells early during tumorigenesis (at the premalignant stage) were shown to induce a program of cellular hypo-responsiveness in tumor-specific (oncogene-specific) CD8 T cells (Willimsky and Blankenstein, 2005; Schietinger et al., 2016). After prophylactic vaccination with irradiated library cells we observed smaller tumors and increased numbers of antigen-specific T cells in splenocytes and draining lymph nodes, suggesting that some antigen-specific T cell proliferation does occur. While we demonstrated sufficient levels of the high affinity, highly immunogenic antigens on the cell surface, tumor associated epitopes of lower affinity, immunogenicity and expression levels should produce an even more pronounced deficit in T cell recognition. We speculate that the mechanism of immune escape in this model is either ineffective T cell activation or failure of activated T cells to identify and kill antigen-positive cells present at low abundance within a sea of cells displaying irrelevant antigens.

The observation that animals prophylactically vaccinated with MCA205∆Tap2 cells are partially protected from outgrowth of RMA/S tumors may be due to several possibilities. For instance, the vaccinations may have activated T cells that recognize some immunogenic peptides in the library, or T cells that recognize mutated peptides expressed by both cell types, or, most intriguingly, T cells specific to T cell epitopes associated with impaired peptide processing (TEIPPs). TEIPPs are T cell epitopes derived from wild type proteins that are not normally presented, unless antigen presentation is altered by loss of Tap, which yields a relative deficit of peptides translocated into the ER. TEIPPs in RMA/S cells can be immunogenic (Van Hall et al., 2006; Doorduijn et al., 2016). Although RMA/S and MCA205∆Tap2 are unlikely to share neoantigens derived from somatic mutations, they may share TEIPPs that contribute to protection against RMA/S tumor growth. A caveat to the vaccination experiments is that differences in the antigen presentation machinery of MCA205∆Tap2 and RMA/S cells may lead to incomplete congruence in the MHC-I peptidome of these two cell lines. These differences may lead to some peptides encoded by the PresentER library (or by the endogenous genome of the cell) to be poorly presented in one cell line but well presented in another cell line.

Although we have focused our efforts on understanding the role of CD8 T cells, other infiltrating cell types are also important for anti-tumor effects. In particular, CD4 T cells have been shown to play important roles in the CD8 T cell responses to tumors (Borst et al., 2018) and to target neoantigens (Ott et al., 2017; Sahin et al., 2017). However, our experimental design does not allow us to assess the role of CD4 T cells in CD8 T cell mediated anti-tumor responses. Thus, while it may be the case that the magnitude or quality of the immune response may differ when both CD8 and CD4 cells recognize a related immunogenic antigen, we are unable to study this with the tools we have developed to perform pooled library experiments because we cannot encode both a specific MHC II and MHC I ligand in a single minigene. Moreover, there is the confounding possibility that some CD4 help is coming from unrelated antigens, such as mutated proteins found in RMA/S or the exogenous proteins introduced into the cells by the PresentER vector. More broadly, the heterogeneity in actual human tumors is characterized by differential infiltration with antigen presenting cells, which can change the sensitivity of the immune system to detect immunogenic antigens present at low clonal fractions, and thus could impact our findings. Although we could not modulate the infiltrating immune cells in our experimental tumor models, we appreciate that this may influence the potency of a subpopulation of immunogenic cells. It would be interesting to understand how various levels of antigen presenting cells contribute to dynamic thresholds of T cell immunogenicity.

The PresentER system we have employed does have some potential biochemical caveats that may impact the interpretation of the data. While we have not observed alternative cleavage patterns in individual peptides we have studied (Gejman, 2018b), in a library setting we cannot test if every minigene is properly yielding its encoded MHC-I ligand. For instance, some encoded peptides may not bind well to MHC-I or may only bind in the presence of an MHC loading chaperone. Some peptides may lead to improper or no cleavage of the signal peptide whereas other peptides may be shortened (e.g. by ERAP1, an ER associated endopeptidase), altered or destroyed in the endoplasmic reticulum.

New therapeutic modalities such as immune checkpoint blockade have shown clinical efficacy in the treatment of human tumors, but the toxicity of the drug regimens has led to many efforts to find biomarkers that predict patient response. Tumors with high mutation burdens—which are more likely to have immunogenic neoantigens (Van Allen et al., 2015; Yarchoan et al., 2017; Snyder et al., 2014)—and tumors that have mismatch repair (MMR) deficiency or microsatellite instability (MSI-H) respond well to checkpoint blockade (Le et al., 2015; Le et al., 2017). In mouse models, syngeneic tumors with MMR gene knock outs accumulate mutations over time and grow more slowly in wild type mice than do tumors without MMR deficiency. Cell lines derived by sub-cloning of MMR deficient lines—thus increasing the clonal fractions of each neoantigen—grow more slowly or are rejected entirely. In all cases, MMR inactivation in tumors leads to better responses to checkpoint blockade than the parental tumors (Germano et al., 2017). Moreover, survival is inversely related to tumor neoantigen clonality in human patients with lung adenocarcinoma (McGranahan et al., 2016; Reuben et al., 2017). Patients whose tumors bear high numbers of subclonal (or branched) neoantigens have increased levels of relapse and worse survival than those patients with more homogenous tumors (McGranahan et al., 2016; Reuben et al., 2017). Response to checkpoint blockade in lung and skin tumors is also associated with lower levels of intratumoral heterogeneity (McGranahan et al., 2016). The combination of these data are highly suggestive that while total neoantigen burden is important for long term survival and response to checkpoint blockade, neoantigen clonality is an additional important factor in mediating tumor regression.

The discovery that intratumoral heterogeneity prevents effective T cell responses has implications for the development of therapeutic cancer vaccines, checkpoint blockade, adoptive T cell therapies, and studies of tumor immunogenicity. In addition, our data may provide a new understanding of mechanisms of immune surveillance and its failure that allows growth and evolution of tumors and subclones. For instance, recent data shows that subclones with very low clonal fraction yet distinct intratumoral functions may be important to tumor survival and growth (Vinci et al., 2018)—suggesting that even low abundance tumor subclones may be important targets for immunotherapy. In general, we may speculate on one mechanism for cancer escape and progression, in which small early cancers generally do not bear immunogenic epitopes derived from their limited number of driver oncogenic proteins and few passenger mutations. Then as the tumors evolve, the neoantigens appearing in subclones do not reach a clonal fraction high enough to breach their antigen-specific immunogenicity thresholds, thereby allowing escape of these otherwise immunogenic clones. This model may help to explain why sun-damaged and aged healthy tissue is replete with mutations, sometimes reaching frequencies seen in human cancers (Martincorena et al., 2015; Risques and Kennedy, 2018). Cumulatively, these findings suggest that effective T cell immunity is restrained in the context of healthy tissue and growing tumors and paint a picture of T cell immunogenicity that is poorly captured by existing models of tissue immunosurveillance.

6-8 week old C57BL6/N mice were purchased from Envigo or Taconic Biosciences. 6–8 week old B6.SJL-Ptprca/BoyAiTac (known as CD45.1 mice) and B6.129S6-Rag2tm1Fwa N12 Mice (known as Rag2 KO) were purchased from Taconic Biosciences. Mice were shaved before subcutaneous engraftment of the indicated number of RMA/S cells in 100 uL PBS. Tumor volumes were calculated using caliper measurements and the standard modified ellipsoid formula: tumor volume = (LxW2 x 0.52 every 2–3 days. Animals were euthanized when tumor volume exceeded 2000 mm3 or if ulceration was noted. Vaccination of animals was performed by subcutaneous injection of 10 × 106 irradiated (20 Gy) MCA205-ΔTap2 cells expressing libraries of minigenes.

For cell surface staining, cells were incubated with appropriate fluorophore-conjugated mAbs for 30 min on ice and washed twice before resuspension in the viability dye DAPI at 1 μg/mL. Flow cytometry data were collected on a LSRfortessa (BD) or an Accuri C6 (BD) and analyzed with FlowJo V10 software. The antibodies used in this study were anti-H-2Kb-APC clone AF6-88.5 (Biolegend 116517), anti-SIINFEKL/H2kb-APC clone 25-D1.16 (Biolegend 141605), anti-CD45.2-APC clone 104 (eBioscience 17-0454-81), anti-CD8a-FITC clone Ly-2 (BD Pharmingen 553030), anti-CD3-PerCP clone 145–2 C11 (BD Pharmingen 553067). The following fluorescently labeled H-2Kb tetramers were obtained through the NIH Tetramer Core Facility: SIINFEKL, SIYRYYGL, MSIIFFLPL, SNFVFAGI and VTFVFAGL. Radioimmuneassay to determine the number of SIINFEKL/H-2Kb molecules on the surface of RMA/S PresentER-SIINFEKL cells was performed as previously described (Dao et al., 2013; Chang et al., 2017).

A guide RNA sequence targeting murine Tap2 (ATGGGGCTGTTGCGCTGAGC) was cloned into the LentiCRISPRv2 (Sanjana et al., 2014) plasmid (Addgene plasmid 52961), a gift from Feng Zhang (Broad Institute, Cambridge, Massachusetts, USA). MCA205 fibrosarcoma cells were transiently transfected using Lipofectamine 2000 (Thermo Fisher Scientific 11668027) following standard manufacturer’s protocols. 24 hr later, successful transfectants were selected using 5 ug/mL Puromycin for 3 days before expansion and single-cell subculture. Genetic ablation of Tap2 was verified by next generation sequencing of the Tap2 loci confirming a frameshift deletion in both alleles, and RT-PCR analysis. Primers for RT-PCR analysis are listed in Supplementary file 1. Reduced cell-surface H-2Kb expression was also verified by flow cytometry.

RMA/S cells expressing PresentER antigen #1 (eGFP) were mixed with RMA/S cells expressing PresentER antigen #2 (mCherry) at defined ratios (e.g. 1:10, 1:100, 1:1000, etc) and validated by flow cytometry immediately before injection into CD45.1 mice. After 17 days, tumors were harvested, cut into small pieces and disaggregated by incubation at 37°C with Liberase TL (Sigma-Aldrich 5401020001), DNAse I (Worthington Biochemical LS002139) and ACK lysis buffer (Thermo Fisher Scientific A1049201). Single cell suspensions of tumor cells were stained with CD45.2 and DAPI and collected on the same flow cytometer, using the same settings and gates as on day 0. The number of eGFP and mCherry cells was calculated based on the same gates used on the day of injection. Non-fluorescent cells were ignored for the purposes of analysis and the percentage of eGFP and mCherry cells in the tumor were normalized to sum to 100%. The fold change in cells expressing each antigen was calculated as: (percentage of normalized eGFP cells in tumor) / (percentage of eGFP cells at D0).

The individual PresentER minigenes specified above were cloned into the PresentER backbone as previously described (Tran et al., 2014). Oligonucleotides were amplified using T3_SfiI and T7_SfiI primers (Supplementary file 1). HEK293T Phoenix-Ampho cells were transfected with each plasmid and, after 24 hr, viral supernatant was harvested every 12 hr. RMA/S were transduced with limiting amounts of viral supernatant at 1000xg for 2 hr at 37°C in six well non-tissue culture treated plates. PresentER minigenes vectors are available on Addgene (102942, 102943, 102944, 102945, 102946).

The peptides included in the libraries were found in Uniprot database of canonical mouse protein sequences (UP000000589). Substrings of unique eight amino acid sequences were collected and affinity to H-2Kb was calculated using NetMHCPan v4.0. 5000 randomly selected peptides with predicted ic50 <500 nM were selected and constitute the ‘wild type peptide library.’ A single random amino acid substitution was made to each member of the wild type library to generate the ‘mutant peptide library.’ Substitutions which generated another wild type peptide were excluded.

Libraries of PresentER minigenes were cloned as previously described (Tran et al., 2014). Library metadata is provided in Supplementary file 2. Oligonucleotide libraries were ordered from CustomArray and amplified with Phusion polymerase using WT and Mutant minigene library-specific primers (Supplementary file 1). Amplicons were digested with SfiI and passed through a MinElute column. The PresentER cassette vector was also digested with SfiI, treated with calf intestinal phosphatase and ligated to the oligonucleotides with T4 ligase. Ligation products were phenol extracted and electroporated into DH5 electrocompetent cells. Electroporated cells were plated and counted. At least 1000x fold more transformants than minigene library members were required to proceed to plasmid DNA extraction (>5×106 colonies). The colonies were scraped off the plate and grown for 3.5 hr in TB +ampicillin at 37°C at 225 rpm. The bacteria were maxiprepped using the Qiagen maxiprep kit and library representation was checked by Illumina sequencing. Library containing retrovirus was produced by transfection of 15 cm plates of HEK293T Phoenix-AMPHO cells with library plasmid DNA. Viral supernatant fluid was collected beginning at 24 hr after transfection and continuing every 12 hr until 72 hr post transfection. Viral supernatant fluid was pooled, concentrated with Clontech Retro-X concentrator and frozen. Concentrated viral supernatant fluid titers were determined by transduction of RMA/S cells. Libraries of minigene expressing RMA/S cells were generated by transduction at an MOI <0.3 (1,000xg for 2 hr at 37°C in six well non-tissue culture treated plates) and selection with puromycin. Library expressing cells were maintained in cultures of >1000 x cells per number of minigenes in the library (i.e. at least 5 × 106).

To verify that minigene representation was not compromised during cloning, all libraries were sequenced from plasmid prior to transduction into mammalian cells. Minigenes were either directly amplified with P5 and barcoded P7 primers and sequenced using a custom, or they were amplified with a nested PCR protocol followed by Illumina library preparation and sequencing at the Integrated Genomics Operation at MSKCC (Supplementary file 1). Genomic DNA was extracted from cultured cells or mechanically disaggregated tumors with the Gentra Puregene kit and minigenes were amplified from genomic DNA. In order to avoid amplification bias due to high PCR cycle numbers or multiple integration events, PCR was performed with the minimum number of cycles possible. The minimum number of PCR cycles was empirically-determined for each batch of samples, but was typically 18–22 cycles when the template was genomic DNA. All steps were performed in PCR hoods or other clean environments with precautions to avoid cross contamination between samples. In order to avoid under-sampling low abundance minigenes, each sample of genomic DNA was amplified using DNA equivalent to >1,000 fold the number of minigenes in the library. Reads were mapped to the PresentER minigene libraries with Bowtie2 using default settings. Reads that did not map to the minigenes in the library were discarded. The number of reads aligning to each minigene was divided by the total number of reads aligning to the library to obtain the relative abundance of each minigene. Quality control to ensure library representation and absence of contamination were performed on every sample. All data analysis was performed in R.

Some PresentER minigenes are available from Addgene (e.g., #102942, #102943, #102945, #102946, #102944) and others are available by request directly from the investigators. Data are available in the following repositories: DOI:10.5281/zenodo.1310902, DOI:10.5281/zenodo.1309836 and DOI:10.5281/zenodo.1308909.

Thank you for submitting your article "Rejection of immunogenic tumor clones is limited by clonal fraction" for consideration by eLife. Your article has been reviewed by two peer reviewers, and the evaluation has been overseen by a Reviewing Editor and Tadatsugu Taniguchi as the Senior Editor. The following individual involved in review of your submission has agreed to reveal his identity: K Melief (Reviewer #1).

Your careful quantitative analysis of the role of tumor heterogeneity (clonal fraction) in effective T cell elimination of cancer adds an important new level of insight into the reasons for failure of immunotherapy even when a tumor possesses adequate neoantigens for immune recognition and effector T cells specific for that antigen exist.

The reviewers have discussed the reviews with one another and the Reviewing Editor has drafted this decision to help you prepare a revised submission.

Summary:

This paper in an ingenious fashion approaches experimentally the important issue of tumor heterogeneity and immunogenicity. To that purpose the authors have used the antigen processing defective (TAP defective) RMA-S cell-line and supplied it with minigene constructs that allow the MHC class I epitope direct access to the ER for MHC class I loading, circumventing the TAP defect in RMA-S. The startling finding is then made through a number of well-designed experiments that small numbers of tumor cells, presenting a given model neo-epitope in a sea of other neo-epitopes escape being noted by the immune system, leading to tumor outgrowth. Not entirely surprisingly, for some epitopes larger fractions of tumor cells go unnoticed than for others, probably reflecting the " quality" of each epitope. While the interpretation offered by the authors in general may be appropriate, additional consideration apply, which they should discuss, because this adds to the relevance of their paper. If true, the conclusions imply a novel parameter strongly influences immunogenicity/targeting of tumor cells and could explain empirical observations on frequent development of heterogenous human tumors.

Essential revisions:

1) The authors have not reckoned with a potential presentation of epitopes by MHC class II and possible effects of CD4 help. It is well known that optimal CD8+ T cell immunity depends bot for effector cell induction and memory on specific CD4+ T cell immunity. Specific help is better than non-specific help, because the CD4+ T cells not only assist in the induction and programming of optimally activated CD8+ T cells, but also improve the intra-tumoral conditions for activity of CD8+ T cells. In the average neo-antigen carrying human tumors both CD4 and CD8 neo-epitopes exist (see Ott et al., 2017 and Sahin et al., 2017 papers quoted in the manuscript). It would be great to probe the rules for immunogenicity ageing in the presence or absence of specific helper epitopes being expressed and presented from the same RMA-S tumor, but I appreciate the fact that this is asking too much at the present time. Nevertheless it is quite possible that smaller fractions of tumor cells would become eliminated in the presence of robust and tumor-specific CD4+ help. The importance of CD4+ help has been discussed recently in Nature Reviews Cancer by Jannie Borst et al., 2018.

2) Tumor heterogeneity may also involve the fact that some tumors are very rich in APC, while others are not, lacking the appropriate chemokine (receptor) expression. This type of heterogeneity might heavily influence the results one might obtain with the system used by the authors. Moreover most tumors lack intrinsic danger signals to activate these DC. Again this is likely to substantially influence the outcome. It will enrich the paper if these aspects are discussed.

3) The observation that vaccination with the MCA 205 TAP-deficient tumor protects non-specifically against RMA-S is not unexpected, because of the existence in RMAS and likely MCA 205 TAP deficient of so-called TEIPP antigens (T cell epitopes associated with impaired peptide processing), first described by Thorbald van Hall et al. in 2006 and lately last revisited in J. Exp. Med. 2018. These epitopes are generated from normal non-mutated household epitopes in a TAP-independent fashion, but being at low numbers, never make it to the cell surface in TAP-competent cells. However, in TAP-deficient cells they see their chance to load MHC class I in the absence of this competition.

4) The authors assume that the antigenic peptide is cleanly produced in the cells by cleavage of the ER translocation signal. Such cleavages are often influenced by surrounding sequences which in this case include the antigenic peptide. So it is possible that what are thought of as perfect antigenic peptides may not always be so. This should at least be discussed.

5) No thought seems to have been given to other ER trimming events such as those caused by the ERAP1, the ER aminopeptidase that is well known to trim antigenic peptides and influence presentation. Selective effects of ERAP1 trimming could change the outcome of actual peptides presented on the cell surface. Because the analysis carried out is of DNA sequences, rather than of the presented peptides discrepancies might be explained by differential processing of antigenic precursors. Again, this needs to be discussed.

6) The original model system was developed in RMA/S cells, but in later experiments included MCA cells that were rendered TAP2-deficient by targeted mutation. Is peptide presentation in two cell types comparable? Especially if there are differences in ERAP1 expression among the two cell types.

7) How the authors compute the relative frequency of minigenes in the library is key to the analysis but it is unclear how this was done. Is it possible that results could be skewed by the number of viral integration events or amplification errors?

The lack of tumor-specific CD4 T cell help in our model system is a critical point and we agree with the comments from the reviewers that this should be discussed in the manuscript. We have added the following statement to the manuscript to address the question of CD4 T cell help and additionally expanded on our thoughts below:

“Our experimental design does not allow us to assess the role of CD4 T cells in CD8 T cell mediated anti-tumor responses. […] Moreover, there is the confounding possibility that some CD4 help is coming from unrelated antigens, such as mutated proteins found in RMA/S or the exogenous proteins introduced into the cells by the PresentER vector.”

The reviewers make the excellent point that the priming activities of helper CD4 T cells may increase the quality of some CD8 antigens. We strongly agree that CD4 cells likely play an important role in increasing the potency of some poorly immunogenic MHC I antigens. However, we think it is unlikely that CD4 help could normalize T cell responses such that all CD8 antigens would then have the same potency. In other words: even if CD4 cells can help some subclonal antigens to yield more potent immune responses, some antigens would still be more potent than others and this would be reflected in differential ability to clear subclonal cells.

Our ability to study MHC II ligands with our present method is limited (although we are considering building a CD4 version of PresentER). We have not worked out whether our signal sequence method efficiently yields MHC II peptides to APCs. In addition, the concept of pooled libraries of precisely defined MHC I antigens cannot yet be combined with libraries of longer MHC II peptides for technical reasons: two libraries would need to be made, one encoding MHC I peptides and the other encoding MHC II peptides, and there is no existing way to ensure that the quantity of the matched MHC I and II peptides are comparable across two libraries). Moreover, MHC II peptides are more difficult to predict computationally than MHC I peptides. We can imagine experiments with individual model antigens to test if the addition of a defined MHC II epitope to the tumor cells would improve T cell rejection of immunogenic subclones. However, a great deal of preliminary experiments would be needed to first show that MHC II peptides were being generated and seen in the RMA/S model we have defined and there are many caveats that would need to be addressed, e.g. the relative strengths of the CD8 and CD4 ligands, the possibility of CD4 help coming from an unrelated antigen, etc. In sum, we think this is a highly complex problem for which there are many experimental unknowns. We have not yet begun this work and we envisage it would take greater than 1 year to construct the system and to be confident in any assessment. As such, this is better suited for a follow-up manuscript.

We agree. This is also an excellent point. We have added to the Discussion the following statement:

“The heterogeneity in actual human tumors is characterized by differential infiltration with antigen presenting cells, which can change the sensitivity of the immune system to detect immunogenic antigens present at low clonal fractions, and thus could impact our findings. […] It would be very interesting to understand how different levels of antigen presenting cells contribute to dynamic thresholds of T cell immunogenicity.”

We agree that TEIPPs deserve a fuller discussion and thank the reviewers for pointing out that vaccination with MCA205 may be semi-protective against RMA/S tumors because of shared TEIPPs. Although we were aware of the existence of TEIPPs, we did not consider the possibility that shared TEIPPs could confer an antigen-specific protective effect. We have added this possible explanation, among others, to the Discussion as follows:

“The observation that animals prophylactically vaccinated with MCA205∆Tap2 cells are partially protected from outgrowth of RMA/S tumors may be due to a number of distinct possibilities. […] Although RMA/S and MCA205∆Tap2 are unlikely to share neoantigens derived from somatic mutations, they may share TEIPPs that may contribute to protection against RMA/S tumor growth.”

Thank you for the points made in #4 and #5. We agree that the caveats and limitations of the PresentER system should be discussed more fully, especially as they apply to libraries of MHC-I antigens. Some of the considerations have been addressed in our first manuscript on the PresentER technology (see Gejman 2018 doi: https://doi.org/10.1101/267047). To be more explicit in the present paper, we have added a paragraph to the manuscript to discuss these points. Of note: we have not yet found evidence of alternative cleavage sites in any of the model antigens we have studied, however our tools to examine this phenomenon are limited (and would require to conduct MHC IP + mass spectrometry to show that cleavage/alteration is occurring). Nevertheless, we agree that some encoded peptides may cause alternative or no cleavage of the signal peptide leading to altered peptide or no free peptide.

The following paragraph is added:

“The PresentER system we have employed does have some potential biochemical caveats that may impact the interpretation of the data. […] Some peptides may lead to improper or no cleavage of the signal peptide whereas other peptides may be shortened (e.g. by ERAP1, an ER associated endopeptidase), altered or destroyed in the endoplasmic reticulum.“

The reviewers are correct that differences in the antigen presentation machinery of MCA205∆Tap2 and RMA/S might affect the peptides presented in the library setting and we have added a sentence to the Discussion to address this:

“A caveat to the vaccination experiments is that differences between the antigen presentation machinery of MCA205∆Tap2 and RMA/S cells may lead to incomplete congruence in the PresentER-encoded MHC-I peptidome of these two cell lines. These differences may lead to some peptides encoded by the library to be poorly presented in one cell line but well presented in another cell line.”

In general, we believe that this caveat would only impact a small fraction of the members of the library. Moreover, we do not think that these data would change the interpretation of the results because MCA205∆Tap2 cells were only used to boost the initial T cell responses in the pooled library setting. Our later data suggests that there may be a fundamental roadblock to identifying immunogenic peptides in pooled library experiments, namely that low frequency antigens do not lead to effective T cell response.

We agree. This is an excellent point. In general, pooled library experiments can be skewed by viral integration events or amplification errors. We have taken precautions at the level of PCR and viral transduction to avoid these issues. We have added several clarifying statements in the Materials and methods and provide more narrative detail here.

In order to avoid amplification bias due to high PCR cycle numbers or multiple integration events, PCR was performed with the minimum number of cycles necessary to see a clear band on an ethidium bromide gel (i.e. distinct from the gDNA smear). The minimum number of PCR cycles was empirically-determined for each batch of samples, but was typically 22 cycles for outer PCR and 18 cycles for inner PCR when the template was genomic DNA. These are relatively low cycle numbers that do not exhaust available dNTPs or primer (which can lead to amplification errors with >30 cycles). All steps were performed in PCR hoods or other clean environments with precautions to avoid cross contamination between samples. The maximum amount of genomic DNA that could be used in each reaction without inhibiting PCR was used in order to minimize amplification biases in each well. In order to avoid under-sampling low abundance minigenes, each sample was amplified using DNA equivalent to >1,000-fold the number of minigenes in the library. For instance, for the 5,000 minigene library samples used in this manuscript, at least 8e6 cell equivalents of gDNA (6µg of gDNA per cell => 48µg of gDNA) was amplified across >40 PCR reactions (we have subsequently optimized this using Q5 polymerase and can now load 10µg of gDNA per reaction). The total amount of gDNA from which to amplify was determined by examining H&E slides of RMA/S tumors and noting that growing tumors were comprised mainly of sheets of transformed cells, indicating that most gDNA comes from tumor/minigene-bearing cells and not stromal cells lacking the minigene construct. If we had noted more stromal infiltration, we would have amplified from a larger number of total cells (i.e. more total gDNA) in order to avoid under-sampling low abundance minigenes.

Biases arising from multiple viral integration events were avoided by transducing target cells at low multiplicities of infection (MOI). The functional titer was determined empirically for each batch of virus by aliquoting+freezing concentrated viral supernatant and transducing RMA/S (or MCA205) with a range of viral supernatant amounts. The amount of viral supernatant corresponding to <⅓ maximal transduction efficiency was used to generate the library, suggesting that on average <10% of cells were doubly infected. >1,000 more cells than minigenes in the library were transduced (usually 2-3,000) and maintained at all times (including during puromycin selection) to ensure that low abundance minigenes did not drop out of the pool.

The quantification of minigenes in the library after an experiment was performed as follows: genomic DNA was extracted from tumors, the minigenes were amplified by nested PCR and sequenced on Illumina instruments. After sequencing, all reads were aligned to the minigenes in the library. The number of reads aligning to each minigene was divided by the total number of reads aligning to the library to obtain the relative abundance of each minigene.