Low neoantigen expression and poor T cell priming underlie early immune escape in cancer

Immune evasion is a hallmark of cancer, and therapies that restore immune surveillance have proven highly effective in cancers with high tumor mutation burden (TMB) (e.g. microsatellite instable (MSI) colorectal cancer (CRC)). Whether low TMB cancers, which are largely refractory to immunotherapy, harbor T cell neoantigens capable of engaging adaptive immunity remains unclear. Here, we show that the majority of microsatellite stable (MSS) CRC harbors predicted high-affinity neoantigens despite low TMB. Unexpectedly, these neoantigens are broadly expressed at lower levels relative to those in MSI CRC, suggesting a potential role of antigen expression in tumor immune surveillance. To test this, we developed a versatile platform for functional interrogation of high expression of multiple high-affinity low poor T cell priming and or therapeutic rescue of T cells with low neoantigen These findings underscore a critical role of neoantigen expression levels in immune evasion and suggest that poor expression or presentation may be a general feature of neoantigens acquired early in tumorigenesis. Finally, poorly expressed neoantigens, commonly excluded in tumor vaccine pipelines, may hold untapped therapeutic potential. tolerogenic cell response characterized by reduced and functionality. These findings provide broader context to previous flank transplant studies that found lower levels of epitope expression 37 or MHC-I binding affinity 38 facilitate tumor immune escape, and a study that described early T cell dysfunction in an SV40 large T-antigen-driven model of liver cancer 39 . Indeed, a general feature of early immune evasion in many tumors may be that T cell dysfunction begins as a tolerogenic program initiated during priming with insufficient antigen stimulation, in addition to a lack of local inflammatory and/or co-stimulatory cues. By extension, it is likely that immune responses against clonal neoantigens in cancer, at least those acquired early in tumorigenesis, are poorly primed and tolerogenic—axiomatic to their failure to restrain tumor outgrowth. We showed that rescuing early priming is sufficient to prevent tumor initiation with 100% efficiency in our model, consistent with the fact that early neoplasia lack the immunosuppressive mechanisms to evade efficiently primed T cell responses. We also showed that poorly primed T cells in our model undergo progressive exhaustion, in line with prevailing literature and suggesting that T cell dysfunction in cancer is a heterogenous state shaped by multiple processes operative early and late in tumorigenesis. Future studies combining lineage tracing and single cell sequencing technologies should help disentangle the contribution of distinct genetic and epigenetic programs to T cell dysfunction in cancer. (two doses) of OVA 250-270 or non-specific vaccination. Significance was assessed by Wilcoxon rank-sum test of percent change in tumor size. blocking solution and incubated at 4 ° C overnight. Anti- b 2M staining is specific to human b 2M in the SIINFEKL tetramer and serves to amplify signal. 30 minutes prior to imaging, 120 µ L of ProLong Diamond Antifade Mountant with DAPI (ThermoFisher) was added to slides. Images were taken at 30X on an Olympus FV1200 Laser Scanning Confocal Microscope and analyzed in ImageJ.


Results
Neoantigen expression level is a critical determinant of immune outcome in a novel orthotopic mouse model of colon cancer. To explore immune dysfunction in cancer, it is critical that preclinical models faithfully recapitulate the tissue microenvironment and genetics of the human disease. Likewise, to study processes underlying T cell dysfunction and immunotherapy resistance, models should enable manipulation of defined antigen-specific T cell responses. To our knowledge, no single model of CRC meets all of these criteria. Therefore, we adapted a technique employing endoscope-guided submucosal injection 13 to induce geneticallydefined tumors in the mouse colon harboring model CD8 + T cell antigens. We first developed an autochthonous model in Apc flox/flox mice initiated by lentivirus expressing Cre-recombinase and the ovalbumin antigen linked to luciferase (LucOS) (Extended Data Fig. 1a), as our group has previously done in models of lung cancer 14 and soft-tissue sarcoma 15 . Injection with LucOS dramatically reduced tumor incidence in a T cell-mediated manner, and tumors that did arise invariably lost antigen expression (Extended Data Fig. 1b- To enforce stable and continuous expression of antigen, we generated CRC organoids with SIINFEKL directly linked to Apc knockdown, an essential event in transformation. Specifically, we transformed normal colon organoids from C57Bl/6 Kras LSL-G12D ; Trp53 flox/flox (KP) mice with adenoviral Cre, followed by lentivirus expressing miR-30 shRNA against Apc (shApc) 16 and SIINFEKL fused to the fluorophore mScarlet (mScarlet SIIN ) (Fig. 1d). Given that relief from Apc knockdown in shApc-transformed tumors results in regression 16 , dependence on shApc provides powerful selection against antigen loss (Extended Data Fig. 1i).
Finally, deletion of Smad4, commonly mutated in CRC, was achieved by CRISPR-Cas9 editing. Selection of organoids harboring complete mutation of all genes was performed following published protocols 17,18 . This resulted in isogenic quadruple-mutant (shAKPS) organoids modeling the four most common genetic mutations in MSS CRC, which are co-mutated with high frequency in metastatic disease and associated with poor prognosis 19 . To investigate the importance of neoantigen expression level, we generated organoids with 400fold range of mScarlet SIIN expression via modifications to the shApc-expressing lentivirus, including placement in reverse orientation to the promoter (EF1α initiates bidirectional transcription), removal of the Kozak consensus sequence, and swapping 25% of codons with rare variants (Fig. 1d-e, Extended Data Fig. 1j).
Importantly, this flexible system is broadly applicable to other cancers via linkage to relevant essential events (e.g. knockdown of Trp53) and is easily adapted to the study of other immune epitopes.
Endoscope-guided transplant of shAKPS organoids without antigen (no SIIN ) efficiently induced tumors ( Fig. 1f-g) and spontaneous metastasis to liver and lung, with histology remarkably similar to human CRC (Extended Data Fig. 1k). In contrast, transplant of the highest expression variant (hi SIIN ) resulted in CD8 + T cellmediated rejection in all animals ( Fig. 1f-g). We also generated shAKPS organoids harboring different epitopes with high affinity for MHC-I, including SIYRYYGL (hi SIY ), VGFNFRTL (hi VGF ), and ITYTWTRL (hi ITY ) (Extended Data Fig. 1l). The latter two are mutant epitopes of Lama4 (G1254V) and Alg8 (A506T) that arose in a methylcholanthrene-induced mouse sarcoma, which were reported to be insufficient for tumor rejection but central to ICB response in a syngeneic flank transplant model 20 . Here, all three epitopes resulted in tumor rejection (Extended Data Fig. 1m), demonstrating that immunogenicity is not idiosyncratic to SIINFEKL but a general feature of high expression of high-affinity epitopes. This also argues that the major genetic features of MSS CRC do not confer cell-autonomous resistance to T cell-mediated killing.
Strikingly, transplant of the lowest expression variant (lo SIIN ) induced tumors and metastases with similar efficiency, histology and infiltration as no SIIN organoids ( Fig. 1f-m). CD8 + T cells were sparse and only modestly increased in lo SIIN tumors, while helper and regulatory T cell infiltration was not significantly different ( Fig. 1j-n, Extended Data Fig. 1k). This is characteristic of the immune "cold" landscape of MSS CRC in humans 21 . Importantly, immune escape in lo SIIN tumors did not result from neoantigen ignorance, as advanced tumors were infiltrated by antigen-experienced (CD44 + ) and specific (SIINFEKL H-2Kb tetramer + ) CD8 + T cells ( Fig. 1o-p). Therefore, MSS tumors may harbor strong neoantigens despite poor T cell infiltration, consistent with our TCGA neoantigen analysis ( Fig. 1a-b), and antigen expression level is a critical determinant of the anti-tumor immune response.

Low neoantigen expression results in impaired T cell effector commitment and early dysfunction.
To investigate why lo SIIN tumors escaped immune rejection, we first compared the kinetics of the antigen-specific T cell response in lo SIIN versus hi SIIN lesions. Low neoantigen expression resulted in both delayed and lower magnitude response (Fig. 2a). Interestingly, this difference was far less pronounced in the caudal and iliac draining lymph nodes (DLNs) (Extended Data Fig. 2a), suggesting that early lo SIIN T cells are either impaired in their ability to traffic to or proliferate within the tumor. The latter is unlikely, however, as lo SIIN and hi SIIN T cells within tumors and DLNs showed no difference in proliferation (Extended Data Fig. 2b). Alternatively, T cells arriving at the tumor may have undergone deletional tolerance 22 , resulting in lower numbers. A critical step in the early maturation of functional T cell responses is effector differentiation, characterized by production of cytolytic granzymes and cytokines, and loss of progenitor potential [23][24][25] . We assessed levels of Granzyme B (GZMB) and TCF1, which is expressed in naïve, memory precursor and memory T cells and associated with stem-like properties [23][24][25] . Consistent with impaired effector differentiation, significantly more antigen-specific T cells from lo SIIN tumors and DLNs were TCF1 + /GZMB -, and significantly fewer were TCF1 -/GZMB + , at 8 days  Fig. 2e). However, this cytokine-proficient population also showed higher TCF1 and lower GZMB (Extended Data Fig. 2f), suggesting similar lack of effector differentiation. Strikingly, the percentage of double-negative (TCF1 -/GZMB -) T cells was greater in lo SIIN versus hi SIIN animals at 8 days (Fig. 2b,g), and became even more pronounced by 14 days (Fig. 2c,h). Absence of TCF1 and GZMB indicates lack of progenitor or effector functionality and is associated with dysfunction. Indeed, by 14 days TCF1 -/GZMBantigen-specific T cells in lo SIIN versus hi SIIN tumors showed higher expression of co-inhibitory receptors PD-1, TIM3, LAG3, and 2B4 (Extended Data Fig. 2g), and an increased fraction co-expressing three or all four (Fig.   2i, Extended Data Fig. 2h). To further interrogate functionality of the lo SIIN response, we performed an in vivo killing assay 26 by transferring SIINFEKL-loaded "target" splenocytes into tumor-bearing mice. Consistent with rejection of hi SIIN organoids, targets were nearly completely eliminated in the DLNs and spleen 6 hours posttransfer in hi SIIN , but not lo SIIN , animals at 8 days (Fig. 2j,k, Extended Data Fig. 2i). Even at the peak of the lo SIIN response (14 days), killing was incomplete (Fig. 2l,m, Extended Data Fig. 2j), with fewer targets killed per antigen-specific T cell (Fig. 2n). Altogether, these results demonstrate that low neoantigen expression drives an immediately dysfunctional T cell response with attenuated magnitude and per cell functionality.

T cells in tumors with low neoantigen expression become progressively dysfunctional. T cell dysfunction
in human cancer is often attributed to upregulation of co-inhibitory receptors, terminal differentiation and loss of effector function, or "exhaustion" 27,28 . More recently, it has been shown that terminally-differentiated T cells in cancer are characterized by low TCF1 and high TIM3 expression, and are regenerated by "progenitor exhausted" T cells that have high TCF1 and low TIM3 expression [29][30][31] . Consistent with progressive dysfunction, terminallydifferentiated T cells (TCF1 -/TIM3 + ) made up a greater proportion of the response at 42 compared to 8 days in lo SIIN tumors, but not in hi SIIN rejected lesions (Fig. 3a,b, Extended Data Fig. 3a). In contrast, progenitors (TCF1 + /TIM3 -) were substantially depleted in lo SIIN tumors by 42 days (Fig. 3a,b, Extended Data Fig. 3b). In addition to TIM3, antigen-specific T cells in lo SIIN tumors at 42 days showed increased expression of PD-1, LAG3, and TIGIT (Fig. 3c). In lo SIIN tumors, T cells negative for TCF1 and triple-positive for TIM3, PD-1, and LAG3 peaked at 42 days (Fig. 3d), consistent with progressive dysfunction. A similar trend was observed in TCF1 + /TIM3progenitor exhausted T cells double-positive for PD-1 and LAG3 (Fig. 3e). Finally, lo SIIN tumors and DLNs at 42 days showed a reduced percentage of TNFα and IFNγ double-positive antigen-specific T cells (Extended Data Fig. 3c-d), indicating loss of effector functionality. Therefore, in addition to immediate dysfunction, T cells responding to low tumor neoantigen expression underwent progressive exhaustion.
T cell priming is limiting in the context of low neoantigen expression and can be rescued by therapeutic vaccination. Impaired effector differentiation and early dysfunction are indicative of poor priming, such as occurs in the absence of CD4 + T cell "help" 32 . While lo SIIN and hi SIIN organoids lack a defined MHC-II-restricted model neoantigen, depletion of CD4 + T cells completely rescued formation of hi SIIN tumors (Fig. 4a).
Therefore, absence of help is unlikely the mechanism of dysfunction in the lo SIIN model. Consistent with the importance of neoantigen expression level in priming, hi SIIN tumor formation was partially rescued in Batf3 -/mice, which lack conventional cross-presenting dendritic cells (Fig. 4a). However, some Batf3 -/animals still rejected tumors, implicating additional antigen-presenting cells as important players in CD8 + T cell priming in the colon. To interrogate initial T cell activation, we performed a series of experiments with lo SIIN tumors in different priming contexts. First, we showed that in vitro-activated TCR-transgenic T cells specific to SIINFEKL (OT-1) were capable of killing lo SIIN organoids when co-cultured (Fig. 4b-c). To optimize priming in vivo, we performed: 1) re-challenge with lo SIIN organoids 28 days after "vaccination" with hi SIIN organoids, 2) co-injection  (Fig. 4j). These results argue that it is therapeutically tractable to vaccinate against poorly expressed neoantigens, and that strict neoantigen expression cutoffs in anti-tumor vaccine pipelines should be reevaluated.
Agonistic CD40 rescues the poorly primed T cell response and enhances immune checkpoint blockade.
We next asked whether more readily deployable antibody-based immunotherapies are efficacious in our model.
Agonistic antibodies against the CD40 receptor (αCD40) enhance priming by potentiating the co-stimulatory function of antigen-presenting cells 33 . αCD40 is efficacious in preclinical mouse models of pancreatic ductal adenocarcinoma (PDAC), particularly when combined with ICB and immunogenic chemotherapy 34 . Recently, a phase Ib clinical trial in PDAC with αCD40 (APX005M), αPD-1 (nivolumab), and gemcitabine/nab-paclitaxel has shown promising early results 35 . This is particularly exciting in light of the low TMB and immunogenicity of PDAC, which, like MSS CRC, is refractory to ICB 36 . Therapeutic combinations with αCD40 may be able to rescue or generate new T cell responses against weak affinity or poorly expressed neoantigens, or against tumorassociated self-antigens that lack high affinity T cell clones due to central tolerance. However, clinical studies in CRC are lacking.
Interestingly, adoptive cell transfer (ACT) of ex vivo-activated OT-1 T cells at 14 days post-transplant significantly delayed tumor growth but only resulted in one complete response (Extended Data Fig. 5c), suggesting that transferred T cells rapidly become dysfunctional.
Despite initially delayed tumor growth in the single ICB arms, no significant difference in final tumor burden was observed at necropsy, suggesting only transient effect in the majority of tumors (Fig. 5i,k). In addition, incidence of metastasis was not significantly decreased in single ICB arms (Fig. 5j,l-n). These results are reminiscent of the poor response to ICB seen in MSS CRC and demonstrate that ICB is only modestly effective at rescuing a poorly primed T cell response. In contrast, single agent αCD40 significantly decreased primary tumor size at endpoint, while combination with ICB significantly reduced tumor size further (Fig. 5i). All treatment arms with αCD40 resulted in significantly reduced rates of metastasis (Fig. 5j), although this could reflect the absence of primary tumors in many of these animals. However, the combined rate of metastasis in animals with progressive disease across all αCD40 arms was still significantly reduced (Extended Data Fig.   5d). Interestingly, while ACT had no effect on reducing primary tumor size at endpoint, it resulted in complete control of metastatic tumor burden (Fig. 5j).

Discussion
The poor response of most CRC to immunotherapy represents a major unmet clinical need. Mouse models have provided invaluable insights into T cell dysfunction in cancer, but none to our knowledge recapitulate essential features of human CRC while facilitating detailed study of antigen-specific T cells. Here, we developed colonoscopy-guided models that enable comparison of functional versus dysfunctional tumorspecific T cell responses in a context highly faithful to the microenvironment, genetics, histopathology, and metastatic progression of the human disease.
We found that tumors from all MSS CRC patients in the TCGA harbored strong predicted neoantigens, but that these were broadly expressed at lower levels compared to those from MSI CRC. This raises the intriguing possibility that poor immunogenicity in MSS CRC and other immune cold cancers is driven by both lower burden and lower expression of neoantigens. Consistent with this notion, our low neoantigen-expressing model, like MSS CRC, demonstrated poor T cell infiltration and response to ICB. Leveraging the defined antigen in our model, we showed that low neoantigen expression precludes productive priming and drives a tolerogenic T cell response characterized by reduced magnitude, effector commitment, and per cell functionality. These findings provide broader context to previous flank transplant studies that found lower levels of epitope expression 37 or MHC-I binding affinity 38 facilitate tumor immune escape, and a study that described early T cell dysfunction in an SV40 large T-antigen-driven model of liver cancer 39 . Indeed, a general feature of early immune evasion in many tumors may be that T cell dysfunction begins as a tolerogenic program initiated during priming with insufficient antigen stimulation, in addition to a lack of local inflammatory and/or costimulatory cues. By extension, it is likely that immune responses against clonal neoantigens in cancer, at least those acquired early in tumorigenesis, are poorly primed and tolerogenic-axiomatic to their failure to restrain tumor outgrowth. We showed that rescuing early priming is sufficient to prevent tumor initiation with 100% efficiency in our model, consistent with the fact that early neoplasia lack the immunosuppressive mechanisms to evade efficiently primed T cell responses. We also showed that poorly primed T cells in our model undergo progressive exhaustion, in line with prevailing literature and suggesting that T cell dysfunction in cancer is a heterogenous state shaped by multiple processes operative early and late in tumorigenesis. Future studies combining lineage tracing and single cell sequencing technologies should help disentangle the contribution of distinct genetic and epigenetic programs to T cell dysfunction in cancer.
Therapeutically targeting priming via αCD40 was highly efficacious in our model, particularly in combination with ICB, resulting in complete responses in the majority of animals. While ICB alone had no effect on the rate of metastasis, αCD40 and ACT almost completely prevented metastases, even in mice with progressive primary disease. Therefore, targeting priming may be especially efficacious against early metastatic lesions that may not be detected at the time of treatment. These results establish the preclinical utility of our model and highlight the therapeutic promise of combined αCD40 and ICB in the treatment of MSS CRC and other immune cold cancers. Furthermore, demonstration that anti-tumor immunity against a poorly expressed neoantigen can be rescued holds promise for therapeutic vaccination, justifying exploration of candidate neoantigens that by current practice are considered too poorly expressed. Finally, the flexible organoid-based system developed here should facilitate future studies of neoantigen expression level in faithful models of cancer, such as in modulating epitope immunodominance, and interplay of CD8 + and CD4 + neoantigen responses.

Lentivirus Production
Lentivirus was produced in HEK-293 cells co-transfected with lentiviral backbone constructs, packaging vectors (psPax2 and pMD2G), and TransIT-LT1 Transfection Reagent (Mirus Bio). The HEK293s and transfection reagents were incubated at 37 °C overnight. The following day, the culture media was changed.
Virus was harvested 48 and 72 hours post transfection and filtered using a 0.45 micron Stericup Quick Release-HV Sterile Vacuum Filtration System (EMD Millipore). Virus was concentrated using an ultracentrifuge spinning at 25,000 rpm for 2 hours at 4 °C. Virus was resuspended in Opti-MEM (ThermoFisher).

Organoid isolation and culture
Normal colon crypts were isolated from wild-type female C57BL/6 mice as previously described 13 . Briefly, colons were isolated, rinsed in PBS, cut into small pieces and crypts dissociated with 5 mM EDTA at 4 °C for 1 hr. Crypts and organoids were cultured in 65 µL domes comprised of 10 µL conditioned L-WRN (for primary crypts and WT organoids) or minimal media (for Apc knockdown organoids) and 55 µL of growth factor reduced phenol-red free Matrigel Matrix (Corning) in 24-well TC treated Olympus plates (Genesee Scientific).
Organoids were passaged when confluent, every two to three days. To passage, Matrigel domes were mechanically dissociated in 500 µL of TrypLE Express (Thermo) per well and by pipetting up and down in a 1 mL pipette 30 times before incubation at 37 °C for 15 minutes. Following incubation, the organoids were pipetted another 30 times before being washed in PBS. Following centrifugation for 1 minute at 2000 rpm, cells were resuspended and plated as described above. Domes were allowed to solidify for 12 minutes in a 37 °C incubator, and 500 µL of appropriate media added.

Organoid infection and transformation
Confluent organoids were dissociated to single cells using Trype-LE Express, diluted and washed with PBS and then spun down in a tabletop centrifuge. They were then resuspended in 1.5 mL of appropriate media (L-WRN media for primary lines, minimal media for transformed lines) with 10 µM of Y-27632 (Sigma-Aldrich).
Organoids were then divided into three separate wells of a 24-well TC-treated plate, and given a range of virus from 10k to 100k TU. The plate was then sealed using Parafilm, spun in a tabletop centrifuge at 600g for 1 hour, and incubated for 5 hours at 37 °C, after which organoids were plated in Matrigel as described above.
Adeno-Cre was used to recombine Kras LSL-G12D and Trp53 flox/flox alleles in wild-type KP organoids, after which organoids were grown in L-WRN media + Y-27632 for 4 days, and then selected for recombination for 1 week using 10 µM NUTLIN-3A (Sigma-Aldrich), which arrests growth of Trp53 wild-type organoids. Complete recombination of Kras and Trp53 alleles was confirmed by PCR allele-specific genotyping using published primers for these alleles 41,42 . Next, organoids were infected with shApc-expressing lentiviruses. To select for stable integration, organoids were switched to minimal media lacking WNT (described above) one week after infection, as organoids with Apc expression are unable to grow in the absence of WNT. Organoids with the lowest multiplicity of infection (~0.5), as determined by mScarlet or GFP fluorescence, were chosen. Finally, these organoids were infected with integration-deficient lentivirus expressing Cas9 (lenti CRISPR v2) 47 and sgRNA against Smad4 18 , and selected by addition of TGFβ, which arrests growth of Smad4-expressing organoids. Integration deficient lentivirus was generated using a D64V mutant psPax2 packaging vector 48 , and absence of integration in organoids was confirmed by absence of Cas9 protein by Western blot, and sensitivity to puromycin killing, following 1 week selection with TGFβ.

Preparation of organoids for injection
Intact organoids were harvested for injections two days after passaging. Organoids were washed in PBS and dissociated using Dispase (Corning) at a ratio of 100 µL per 65 µl Matrigel dome. Matrigel was broken up by scraping and pipetting up and down four times using a 1 mL pipette and incubated at 37 °C for 30 minutes.
Organoids were then harvested into a 15 mL conical, washed with 10 mL PBS, and resuspended in OPTI-MEM (Gibco). Whole organoids were counted and resuspended to a final concentration of 50,000 organoids / mL in OPTI-MEM and 10% Matrigel.

Colonoscopy-guided injections
Lentiviral injection and orthotopic transplant of transformed organoids was performed following the methods described in Roper et al., 2018 13 . First, mice were anesthetized using 2% isoflurane. Once anesthetized, mice were placed in a supine position and the colon was cleaned using a rapid enema of tap water. Gentle abdominal pressure was applied to propel stool out of the colon. Next the colonoscope was inserted and the colon insufflated with air. 100 µL of lentivirus or organoids was taken up into the injection syringe (Hamilton Inc., part number 7656-01), which was then attached to the injection needle (Hamilton, 33 gauge, small Hub RN NDL, 16 inches long, point 4, 45-degree bevel, like part number 7803-05) which was fed through the working channel of the colonoscope. Next, the needle was inserted into the mucosa of the colon at around a 30° angle relative to the colon wall. Next, a "test" injection of around 10 µL was performed. If the "test" injection resulted in a small mucosal bubble, another 40 µL was quickly injected. Successful injections result in large "blebs" within the mucosa. Lentivirus was injected at 20,000 or 100,000 TU / µL. Organoids were injected at 50 intact organoids / µL. Organoids were always prepared for injection at day 2 post-splitting.

Tissue preparation and flow cytometry
The colon draining lymph nodes (caudal and iliac) of each mouse were harvested into a single well of a 24 well TC plate (Corning) that had been scoured in a crosshatch pattern with a scalpel and containing 1 mL RPMI1640 (Corning) with 5% heat-inactivated fetal bovine serum (HI-FBS) (harvest media). Next, the hardplastic end of a 3 mL syringe was used to grind the lymph nodes along the scoured plate. The cells were then During initial optimization of this protocol, intravenous CD45 staining, prior to sacrifice of animals, was performed to differentiate tissue-infiltrating versus circulating T cells. This resulted in staining of less than 1% of total SIINFEKL-specific T cells, and therefore was not routinely performed.
Prior to cell surface stains, live/dead staining was performed in PBS with ghost ef780 ( DLNs and spleens were harvested 6 hours later and processed for flow cytometry as described above. Target and control splenocytes were identified by live/dead staining and CTV labeling intensity, and percent target killing determined relative to the control population. Targets killed per antigen-specific T cell was determined by dividing the total number of targets killed (control minus target splenocytes) by the total number of SIINFEKL tetramer + CD8 + T cells. This metric was meaningful at 14 days when target killing was incomplete in both lo SIIN and hi SIIN animals, but precluded at 8 days by effectively complete target killing in hi SIIN animals.

Peptide stimulation for cytokine staining
Samples were prepared as described above, and prior to surface staining were stained with antibodies for Ly- µM SIINFEKL peptide (Anaspec) for 3 hours at 37 °C. Cells were washed twice in PBS and stained for surface and intracellular markers following the procedure described above.

OT-1 T cell activation
Spleens and LNs from OT-1 mice were harvested into ice cold PBS and mashed through a 100 µM filter using the rubber end of a syringe plunger. Cells were pelleted, red blood cells lysed using ACK Lysing Buffer (Gibco), resuspended in T-cell media (described in previous section) + 10 ng/mL hIL2 (PeproTech) and 1 µM SIINFEKL peptide (Anaspec), counted, and plated in a 24 well TC plate at 1x10 6 cells/mL, 1.5 mL per well.
Stimulation was performed for 24 hours at 37 °C. CD8 + T cells were then purified using the CD8a + T Cell Isolation Kit, mouse (Miltenyi Biotech), and expanded in T cell media + hIL2 with daily splitting. T cells were used for ACT or co-culture assays at day 3 or 4.

OT-1 and organoid co-culture
OT-1 T cells were activated in vitro as described above. Organoids were broken up into single cells as and analyzing particles with size threshold set to >20 pixels 2 , circularity set to 0.1-1.0, and including holes.
Areas of objects were summed and plotted in R.

Three color immunohistochemistry
Tissue samples were fixed in zinc formalin overnight and washed in 70% ethanol before embedding in paraffin.
Tissue slides were dewaxed using a Gemini automatic self-contained slide strainer. Next, slides underwent antigen retrieval through incubation in citrate buffer pH 6 in a pressure cooker set at 125 °C for five minutes.

In situ SIINFEKL tetramer staining and immunofluorescent imaging
Tissue was stained in situ with SIINFEKL tetramer as previously described 49 . Briefly, freshly harvested tissue was manually cut into thin sections with a razor blade and incubated with PE-conjugated SIIINFEKL tetramer (NIH tetramer core) diluted to 2 µg / mL in PBS with 2% HI-FBS overnight at 4 °C. The next day, tissue was washed and lightly fixed in 2% PFA for 2 hours on ice. Tissue was then washed again and incubated overnight in 30% sucrose at 4 °C. The next day, tissue was snap frozen in Tissue Tek OCT (Sakura) using a 2methylbutane cooling bath in a beaker on dry ice. Sections were cut to a thickness of 7 µm on a CryoStar NX70 (ThermoFisher), fixed in acetone at -20 °C for 10 minutes, and stored at -20 °C.
Prior to additional staining, slides were rehydrated in PBS for 10 minutes, and then blocked with 5% bovine serum albumin in PBS for 45 minutes in a humidified chamber. CD8β AF647 (YTS156.7.7, BioLegend) (1:100), and anti-human b2-microglobulin (b2M) PE (2M2, BioLegend) (1:50) were added in a total volume of 300 µL of blocking solution and incubated at 4 °C overnight. Anti-b2M staining is specific to human b2M in the SIINFEKL tetramer and serves to amplify signal. 30 minutes prior to imaging, 120 µL of ProLong Diamond Antifade Mountant with DAPI (ThermoFisher) was added to slides. Images were taken at 30X on an Olympus FV1200 Laser Scanning Confocal Microscope and analyzed in ImageJ.

Colonoscopy imaging
Tumor progression was monitored longitudinally using a Karl Storz colonoscopy system with white light, RFP and GFP fluorescence. This consists of Image 1 H3-Z Spies HD Camera System (part TH100), Image 1 HUB CCU (parts TC200, TC300), 175 Watt D-Light Cold Light Source (part 20133701-1), AIDA HD capture system, and fluorescent filters in the RFP and GFP channels (all from Karl Storz). The endoscope used for imaging was the Hopkins Telescope (Karl Storz, part 64301AA) with operating sheath (Karl Storz, part 64301AA). To consistently measure tumor area, biopsy forceps (Richard Wolf) were fed through the operating sheath and positioned as consistently as possible given two major landmarks: widthwise grooves that appear as concentric semi-circles in the field of view, and a lengthwise groove at the forceps tip. Images were captured upon contact of the forceps with the tumor to maintain consistent distance between camera and tumor across all images.
Tumor area in the field of view was calculated using ImageJ. Imaging and quantification were performed blind to treatment groups.

TCGA and neoantigen prediction data analysis
Colon and rectal adenocarcinoma (COADREAD) RNA-seq Level 2 normalized data and patient information was downloaded from the NCI Genomics Data Commons data portal using the 'TCGAbiolinks' package in R 52 .
Normal tissue, FFPE-fixed samples, and other duplicates were removed prior to analysis. Predicted neoantigens and affinities were retrieved from a previously published pan-cancer analysis 53 .

TCR sequencing
SIINFEKL tetramer-positive CD8 + T cells were prepared and stained as described for single cell RNAseq. Cells were directly sorted into 50 μl lysis buffer with proteinase K, from the Arcturus PicoPure DNA Extraction kit (ThermoFisher), in low binding microcentrifuge tubes (Biotix), and genomic DNA extraction performed following manufacturer instructions. Mouse TCRβ sequencing was performed by Adaptive Biotechnologies. Analysis was performed in R, and Simpson diversity calculated using the 'vegan' package. To account for differences in total numbers of T cells surveyed in samples between groups, unique productive TCR sequences (clones) were randomly down-sampled to match between groups. Down-sampled data is presented in Extended Data Figure 3d-e, although down-sampling did not impact observed trends.

Statistical Analysis
Statistical analysis was performed in R. For statistical assessment of differences in proportionality, Fisher's exact 2x2 test was performed. For continuous data, Wilcoxon rank sum test was performed, with the exception of the organoid and OT-1 co-culture results, which were analyzed with the Student's T test. Multiple comparison correction was performed using Holm's method.