Mismatch repair deficiency is not sufficient to elicit tumor immunogenicity

DNA mismatch repair deficiency (MMRd) is associated with a high tumor mutational burden (TMB) and sensitivity to immune checkpoint blockade (ICB) therapy. Nevertheless, most MMRd tumors do not durably respond to ICB and critical questions remain about immunosurveillance and TMB in these tumors. In the present study, we developed autochthonous mouse models of MMRd lung and colon cancer. Surprisingly, these models did not display increased T cell infiltration or ICB response, which we showed to be the result of substantial intratumor heterogeneity of mutations. Furthermore, we found that immunosurveillance shapes the clonal architecture but not the overall burden of neoantigens, and T cell responses against subclonal neoantigens are blunted. Finally, we showed that clonal, but not subclonal, neoantigen burden predicts ICB response in clinical trials of MMRd gastric and colorectal cancer. These results provide important context for understanding immune evasion in cancers with a high TMB and have major implications for therapies aimed at increasing TMB.

DNA mismatch repair deficiency (MMRd) is associated with a high tumor mutational burden (TMB) and sensitivity to immune checkpoint blockade (ICB) therapy. Nevertheless, most MMRd tumors do not durably respond to ICB and critical questions remain about immunosurveillance and TMB in these tumors. In the present study, we developed autochthonous mouse models of MMRd lung and colon cancer. Surprisingly, these models did not display increased T cell infiltration or ICB response, which we showed to be the result of substantial intratumor heterogeneity of mutations. Furthermore, we found that immunosurveillance shapes the clonal architecture but not the overall burden of neoantigens, and T cell responses against subclonal neoantigens are blunted. Finally, we showed that clonal, but not subclonal, neoantigen burden predicts ICB response in clinical trials of MMRd gastric and colorectal cancer. These results provide important context for understanding immune evasion in cancers with a high TMB and have major implications for therapies aimed at increasing TMB.
Immunotherapy has revolutionized the treatment landscape of many cancers, particularly those with a high tumor mutational burden (TMB) [1][2][3][4][5] . Somatic mutations can generate neoantigens capable of eliciting tumor-specific T cell responses 6,7 , and it is widely believed that increased TMB renders tumors susceptible to immune attack after ICB treatment. Indeed, multiple pan-cancer meta-analyses have shown that TMB is one of the strongest predictors of ICB response [8][9][10] , leading to approval by the US Food and Drug Administration (FDA) of pembrolizumab (anti-PD-1) for all tumors based on high TMB alone. In particular, MMRd is associated with some of the highest TMBs observed in cancer (hypermutation) [11][12][13][14][15] and remarkable response rates to pembrolizumab 2,5 . However, more than half the patients with MMRd tumors do not durably respond and TMB does not stratify responders 2,5 . This underscores a critical need to understand what factors beyond TMB mediate efficacy. There are also conflicting studies suggesting that TMB is an imperfect biomarker of immunotherapy response 16 and argue that FDA approval of pembrolizumab based on TMB alone may be too broad 17 . Specifically, TMB provides limited to no additional predictive value within specific subsets of cancer known to have high rates of response to ICB, like those with MMRd 2,5 .
One factor that may dilute the power of TMB as a biomarker of ICB response is intratumor heterogeneity (ITH) of mutations or, simply defined, the fraction of mutations that are subclonal-present in a minority of tumor cells. Subclonal neoantigens could be deleted with minimal impact on tumor fitness 18 or fail to elicit productive T cell responses 19 . Indeed, it has been observed in human cancer that ITH is Article https://doi.org/10.1038/s41588-023-01499-4 clone were found in the sequencing data of the parental compared with the unrelated control line (1,685 versus 451), albeit at variant allele frequencies (VAFs) below the threshold of mutation calling (Extended Data Fig. 1s,t), as previously observed in hypermutated gliomas 38 .
Altogether, these results establish the utility of our models to recapitulate fundamental mutational processes underlying hypermutation in MMRd human cancer. The lower clonal TMB that we observed is probably due to neutral evolution in the absence of major selective bottlenecks, a feature of ablating MMR concomitantly with tumor initiation in our models. Although these models cannot capture the accumulation of mutations or clonal evolution over decades in human cancer, they are uniquely suited to study the role of ITH in immune dysfunction of cancers with high and low prevalence of MMRd alike. Importantly, ITH is associated with aggressive disease and decreased ICB response in humans [22][23][24][25] , but it remains unclear how ITH impacts the immune response in MMRd cancers specifically.

Sporadic MMRd in the mouse is not immunogenic
Next, we sought to determine the effects of sporadic MMRd on tumorigenesis in our models. Neither KPC nor KPM Msh2 KO models showed a significant difference in overall tumor burden or grade at either timepoint ( To determine the sensitivity of these models to immunotherapy, we first performed preclinical trials with ICB (anti-CTLA-4/anti-PD-1) in the KPM model. We followed similar dosing as established in seminal preclinical studies that preceded the first human clinical trials of anti-CTLA-4 (ref. 39) and anti-PD-1 (ref. 40), although we continued treatment for longer (4 weeks). In addition, we included treatment with the chemotherapeutic combination of oxaliplatin and low-dose cyclophosphamide (Oxa/Cyc) alone and in combination with ICB ( Fig. 2h-i), because Oxa/Cyc has been shown to synergize with ICB 41,42 . To our surprise, no significant differences were observed between KPM and KP mice across all treatment arms, in both longitudinal change (Fig. 2j) and final tumor burden at necropsy ( Fig. 2k and Extended Data Fig. 2j). Consistent with a lack of immunogenicity, we observed no differences in tumor grade or burden at 16 weeks between KPM and KP mice treated continuously with CD4 + and CD8 + T cell-depleting antibodies (αCD4/8) (Extended Data Fig. 2k-p). These results were not unique to the lung, because ICB treatment failed to induce any responses in sgMsh2-targeted colon tumors (Fig. 2l,m and Extended Data Fig. 2q,r). Likewise, there was no significant difference in endpoint size of sgCtlversus sgMsh2-targeted colon tumors after ICB or continuous αCD4/8 treatment (Fig. 2n). Altogether, these data demonstrate that MMRd in these models is not sufficient to increase immunogenicity or sensitivity of tumors to ICB, alone or in combination with chemotherapy, in stark contrast to previous reports in cell-line transplant models 27,28 .

MMRd drives mutational heterogeneity
To assess the clonal composition of mutations in our models, we estimated cancer cell fractions (CCFs) 43 . TMB was predominantly subclonal and most mutations were present in less than a quarter of cells in MMRd tumors from lung and colon (Fig. 3a,b and Extended Data Fig. 3a-c). These mutations also adhered perfectly to a theoretical model of neutral evolution of subclonal mutations in cancer 44 (Fig. 3c,d), consistent with the absence of selective events after tumor initiation in our models. To investigate ITH more deeply, we performed WES on eight single-cell clones derived from an sgMsh2-targeted lung tumor cell line (09-2). Importantly, before subcloning, we restored MMR by re-expressing Msh2 on a bicistronic lentivirus conferring puromycin associated with decreased T cell infiltration 20,21 and poor survival 22,23 , whereas clonal neoantigens are predictive of response to ICB [23][24][25] . This concept has been exemplified in an experimental setting of ultraviolet light B (UVB)-induced hypermutation of melanoma cell lines 26 . It is reasonable to hypothesize that similar mechanisms are operating in MMRd cancers given their constitutive mutational instability. Although previous studies showed that MMRd mutagenesis in vitro renders cell lines immunogenic 27,28 , it is unclear what the impact is of MMR loss in vivo in the presence of immunosurveillance, a process that exquisitely shapes tumor immunogenicity 6,29,30 . To address these questions, we developed autochthonous mouse models of sporadic MMRd lung and colon cancer via targeted ablation of key genes in the MMR complex, MutL homolog 1 (Mlh1) and MutS homologs 2, 3 and 6 (Msh2, Msh3 and Msh6), and performed preclinical trials to determine ICB sensitivity.
We also performed WES on cell lines derived from sgMsh2-targeted lung tumors. Unexpectedly, these showed on average a tenfold greater TMB than sgMsh2-targeted lung tumors ( Fig. 1e and Extended Data Fig. 1k), suggesting low tumor purity (unlikely, because Msh2 KO lung tumors were on average 72% pure), mutagenesis in culture or substantial ITH reduced by clonal selection on plastic. Arguing against extensive mutagenesis in vitro, WES of a single-cell clone revealed more than double the TMB of the parental line, despite a much smaller increase in TMB after 20 passages (Extended Data Fig. 1r). Consistent with high ITH, more sequencing reads supporting mutations private to the single Article https://doi.org/10.1038/s41588-023-01499-4 resistance (Fig. 3e). Of note, these clones (M1-8) maintained MSH2 expression and showed stable mutational and clonal architecture after 20 passages in the presence of puromycin (Extended Data Fig. 3d-f).
Nevertheless, significantly more somatic mutations were called in all clones than in the parental line (Fig. 3f) and these mutations were not broadly shared across clones (Extended Data Fig. 3g    the notion that ITH is underestimated by bulk sequencing methods 38 . Phylogenetic analysis of clones further confirmed a considerable level of ITH (Fig. 3g). Given that ITH in our models arose in vivo, we sought to understand the role of immunoediting 6,29,30 in shaping this process. We performed WES on micro-dissected Msh2 KO lung tumors (weeks [16][17][18][19][20] from animals continuously depleted of T cells (n = 34) or treated with ICB (n = 12). Continuous T cell depletion had no significant effect on overall TMB or tumor neoantigen burden (TNB), consistent with its lack of effect on tumor progression (Extended Data Fig. 2k-p). However, stratifying mutations by CCF revealed a significant increase in clonal (CCF ≥ 0.75), but not subclonal (CCF ≤ 0.5), TMB and TNB (Fig. 3h,i and Extended Data Fig. 3h,i). T cell-depleted tumors also showed a striking difference in CCF distribution and median CCF of expressed neoantigens, with significant enrichment of clonal neoantigens (Fig. 3j,k and Extended Data Fig. 3j). We observed similar results in the colon Msh2 KO model (Extended Data Fig. 3k-n). It is interesting that tumors from ICB-treated animals showed significantly lower subclonal, but not clonal, TMB and TNB, suggesting possible simultaneous loss of subclonal populations and selective expansion of others in this context of ICB nonresponse (Fig. 3h,i and Extended Data Fig. 3h,i). Altogether, these results argue that neoantigens with high, but not low, clonal fraction are negatively selected by the adaptive immune system during tumor progression. By extension, immunosurveillance in mutationally unstable tumors promotes ITH by selectively pruning clonal neoantigens and thereby increasing the relative fraction of subclonal neoantigens. ICB treatment may lower the CCF threshold for T cell-mediated elimination, but probably also drives expansion of less immunogenic subclones that maintain an overall state of ITH in nonresponders.

Intratumor heterogeneity enables immune evasion
To evaluate the impact of ITH, specifically tumor cell clonality, on immunogenicity in our models, we assessed survival of animals after orthotopic transplantation of the lung tumor cell lines and clones described above. The parental MSH2 knockout lines and low TMB control line were similarly nonimmunogenic, showing no difference in disease progression with and without continuous T cell depletion (Fig. 4a). This is consistent with the lack of immunogenicity of tumors induced by chemical carcinogens in immunocompetent animals and the principle of immunoediting 6,29 . Unlike MMRd in the autochthonous model, however, mice transplanted with the parental Msh2 KO line (09-2) or an equal mixture of 09-2-derived Msh2 KO clones (M1-8) and treated with ICB showed 20-30% durable responses and reduced hazard ratios (HRs) (0.34-0.50) over 30 weeks (Fig. 4b). Mice transplanted individually with five otherwise nonimmunogenic Msh2 KO clones (M1, M2, M4, M5 and M6) and treated with ICB showed 30-75% durable responses and even further reduction in HRs (0.05-0.32) over 30 weeks (Fig. 4b). It is interesting that the other three Msh2 KO clones (M3, M7 and M8) were strongly immunogenic even without ICB treatment, efficiently driving disease only with T cell depletion (Fig. 4b). These differences were not the result of tumor intrinsic growth rates or loss of antigen presentation, because all clones showed similar in vitro growth kinetics and readily expressed major histocompatibility complex class I (MHC-I) (H-2K b , H-2D b ) and programmed death-ligand 1 (PD-L1) on stimulation with interferon-γ (IFN-γ; Extended Data Fig. 4a-e). Although M3, M7 and M8 did not show increased TMB/TNB relative to the other clones, they did express on average higher H-2K b /H-2D b with IFN-γ stimulation, particularly the most immunogenic clone, M3, although this was not significant (Extended Data Fig. 4b,c). It is therefore possible that the higher baseline immunogenicity of these clones is due in part to greater surface presentation of neoantigens. Consistent with MMRd-derived neoantigens underlying the immunogenicity and ICB responsiveness of the clones, mice transplanted individually with single-cell clones derived from the low TMB control line (13-1) were uniformly and completely unresponsive to ICB (Extended Data Fig. 4f). Altogether, these results establish a model of MMRd wherein increasing mutational clonality (autochthonous tumors < cell lines < mixture of clones < clones) correlates with immunogenicity and ICB response.
Given that the parental cell line 09-2 arose in an immunocompetent host yet contained highly immunogenic subclones (M3, M7 and M8), we reasoned that tumors evade immunosurveillance not only by selective outgrowth of nonimmunogenic subclones but passively through the failure of immunity to eliminate otherwise immunogenic tumor cells present at a low clonal fraction. To test this hypothesis, we collected lung tumors and metastases from mice orthotopically transplanted with an equal mixture of clones M1-8 and reconstructed their clonal makeup by ultradeep, targeted amplicon sequencing of unique clone-defining SNVs (Methods, Fig. 4c and Extended Data Fig.  4g,h). Despite the immunogenicity of M3, M7 and M8, these clones were detected in 12 of 20 tumors analyzed from immunocompetent animals at clonal fractions ≥1%. Of these 12 animals 5 also developed liver metastases, 3 of which were clonal outgrowths of an immunogenic clone (M8) (Extended Data Fig. 4h). In contrast, none of the immunocompetent animals transplanted with M3, M7 or M8 alone formed metastases. This pattern was not substantially different to that of transplants in continuously T cell-depleted or ICB-treated mice. However, tumors from T cell-depleted mice on average showed significantly lower clonal diversity with greater dominance by individual clones (Fig. 4c,d). In addition, tumors from immunocompetent animals harboring one or more immunogenic clones (M3, M7 or M8) at ≥1% fraction were significantly more heterogeneous than tumors without these clones present (Fig. 4e). These results are in agreement with experiments in the autochthonous models ( Fig. 3h-k and Extended Data Fig. 3h-n) and further support a revised understanding of immunosurveillance as a process that strongly selects against clonal, but not subclonal, neoantigens. Yellow is for normal lung, red for G1, green for G2, blue for G3 and orange for G4. Scale bar, 5 mm. e-g, IHC staining and Aiforia CNN quantification of T cell subsets in KPM and KP tumor-bearing lungs of animals in a and b. Representative IHC staining of lung tumor from an animal in b (left) with Aiforia CNN annotations (right) for CD4 + (green), CD8 + (yellow) and CD4 + FOXP3 + T reg cells (purple) (e). Scale bar, 100 μm. CNN quantification of IHC staining within lung tumors from KPM and KP animals is shown at 5 (f) and 15 weeks (g) post-initiation. h,i, Preclinical trial design in KPM and KP models (h) and treatment arms (i). j, Change in solid lung volume as measured by μCT pretreatment (10 weeks) and post-treatment (14 weeks). k, Lung tumor burden at necropsy (14 weeks) as measured by manual annotation of H&E-stained whole lung sections. l, Brightfield and fluorescent colonoscopy images of sgMsh2targeted colon tumor, representing 16 animals. m, Change in colon tumor size by colonoscopy pre-treatment (20 weeks) and post-treatment (24 weeks) (n = 23 sgMsh2-and 7 sgCtl-targeted animals treated with ICB). n, Colon tumor burden at necropsy (24 weeks) as measured by stereomicroscopy (n = 23 sgMsh2-and 7 sgCtl-targeted animals treated with ICB (αPD-1/CTLA-4) and 10 sgMsh2-targeted animals treated continuously with T cell-depleting antibodies (αCD4/8)). Boxplots display median and interquartile range (IQR; box bounds), with whiskers extending to most extreme points (≤1.5× IQR) and all datapoints. Significance in a, b, f, g, k and n was assessed using Wilcoxon's rank-sum test with correction for multiple comparisons in a, b, f, g and k. P values in n are uncorrected.

Neoantigen clonality tunes the T cell response
To gain further mechanistic understanding of how T cells shape clonal architecture, we first profiled the MHC-I immunopeptidome 45 of M1-8 clones (Supplementary Table 2). Using quantitative tandem mass spectrometry (MS-MS), we identified seven neoepitopes with relative abundance patterns consistent with the associated mutations in the clones. Of these, five neoepitopes (71%) exhibited immunogenicity in normal mice after dendritic cell prime/boost/boost vaccination, by IFN-γ enzyme-linked immune adsorbent spot (ELISpot) and/ or MHC-I:epitope tetramer staining of splenocytes (Extended Data Tumor area (% of total lung area) Fig. 5a-e). One neoepitope unique to the M5 clone, QAYAFLQHL, elicited a higher-magnitude, antigen-specific CD8 + T cell response than the highly immunogenic model neoantigen SIINFEKL, from chicken ovalbumin 46 (Extended Data Fig. 5d). Although MS immunopeptidomics is far from exhaustive 47 , our identification of five bona fide immunogenic neoepitopes supports the notion that neoantigens underlie the enhanced immunogenicity of Msh2 KO clones ( Fig. 4b and Extended Data Fig. 4f). Next, we transplanted the M5 clone into the lungs of syngeneic mice at different clonal fractions, diluted with the other M1-8 clones, and analyzed the M5-specific CD8 + T cell response using flow cytometry with MHC-I:QAYAFLQHL dextramers (Fig. 5a). Consistent with sensitivity of the M5 clone to ICB (Fig. 4b), a robust QAYAFLQHL-specific T cell response was induced in lungs and mediastinal draining lymph nodes (mLNs) of clonally transplanted mice after 2 weeks of ICB treatment. As the clonal fraction and total number of M5 cells were decreased, however, the magnitude of this response also decreased ( Fig. 5b-d and Extended Data Fig. 5f-h). Similar results were obtained in analogous experiments with the M2 clone using MHC-I:AALQNAVTF tetramers to analyze M2-specific CD8 + T cells (Extended Data Fig. 5i-k). Surprisingly, the quality of the QAYAFLQHL-specific T cell response also decreased with lower clonal fraction, with a significantly smaller percentage of QAYAFLQHL-specific T cells expressing the major effector protease of cytotoxic T cells, granzyme B (GZMB) and a significantly greater percentage expressing TCF1 (Fig. 5e-i and Extended Data Fig. 5l-p). Expressed highly in naive and early activated T cells, the transcription factor TCF1 is lost during effector differentiation 48 . We have previously shown that TCF1 + GZMB − tumor-specific T cells are enriched  and characteristic of tolerogenic dysfunction in a model of microsatellite stable colon cancer 49 . It is interesting that transplantation of an equal number of M5 cells (100,000) at clonality versus CCF of 0.5 (with 100,000 M1-4,6-8 cells) resulted in no significant difference in magnitude, GZMB + or TCF + percentage of the T cell response ( Fig. 5d,g,i and Extended Data Fig. 5h,n,p), suggesting that total mass of neoantigen-expressing tumor cells is the primary determinant of quality of the neoantigen-specific T cell response. Altogether, these results provide a compelling rationale for why immunosurveillance fails to delete tumor cells bearing subclonal neoantigens (Fig. 5j).

Clonal TNB predicts ICB response in human MMRd cancer
To explore the translational relevance of our findings, we reanalyzed sequencing data from two clinical trials of anti-PD-1 treatment in advanced MMRd colorectal cancer (CRC; Bortolomeazzi et al. 50 ) and gastric cancer (Kwon et al. 35 ), including 16 and 13 patients, respectively. Predicting neoantigens and the associated CCFs (Supplementary Table 3), we asked whether clonal TNB and ITH are associated with the response to anti-PD-1. Clonal (CCF ≥ 0.75), but not subclonal (CCF ≤ 0.5), TNB was significantly associated with objective response (OR), whereas high ITH index (subclonal to clonal neoantigen ratio) was significantly associated with nonresponse (NR) (Fig. 6a,b and Extended Data Fig. 6a-d).
Total TNB was also significantly associated with OR (Extended Data Fig. 6e-g), probably because tumors in these studies had generally more clonal than subclonal neoantigens. Although clonal TNB was generally correlated with total TNB, two notable outliers, both nonresponders, had high total TNB but very low clonal TNB (Extended Data Fig. 6h).
Clonal, but not subclonal, TNB was also significantly associated with longer progression-free survival (PFS) in combined analysis of the trials, whereas the ITH index was associated with shorter PFS (Fig. 6c-e). Total TNB also correlated with longer PFS, but did not reach the same level of significance as clonal TNB (Extended Data Fig. 6i). Importantly, there was no significant difference in PFS between the two trials (Extended Data Fig. 6j), justifying their combination in these analyses. Overall, these results support the major conclusions from our mouse models and suggest that clonal TNB is a more accurate predictor of ICB response than overall TMB in MMRd cancers. Although the trials that we reanalyzed in the present study are limited by small sample size and will require validation in larger prospective investigations in MMRd cancers, which are currently lacking, our results align with a growing body of literature that supports the superior predictive power of clonal versus total TMB/TNB as a biomarker of ICB response across other cancers 20-26 .

Discussion
In this Article, we report sporadic MMRd mouse models that recapitulate critical features of human lung and colon cancer, including genetics, histopathology and in situ initiation in the relevant tissue microenvironment. Unlike previous studies that demonstrated a role for MMRd and TMB in immunotherapy response, in which mutagenesis occurred in vitro [26][27][28] , our models enable study of mutations continuously acquired in vivo from tumor initiation through advanced disease. Other studies showed immunotherapy efficacy in autochthonous cancer models employing tissue-specific knockout of Msh2 or activation of mutant Pole (Pole P286R ) during embryogenesis 51,52 . These previous models   recapitulate familial cancers like Lynch syndrome in the accumulation of mutations in normal parenchyma preceding transformation (<5% of colorectal cancers (CRCs)), but not sporadic loss of MMR in established tumors (10-15% of CRCs) 14,15 . In contrast, our models, which more closely resemble sporadic MMRd, followed a model of neutral evolutionary dynamics 44 and did not display increased baseline immunogenicity or response to ICB, probably owing to timing of MMR inactivation and the resulting patterns of clonal expansion 53 . We induced MMRd concomitantly with tumor initiation, resulting in mutation accumulation during exponential cellular expansion that is reminiscent of so-called 'born to be bad' colon tumors that follow an early explosive growth trajectory 54 . It is interesting that it was recently shown that MMRd occurring either de novo or induced by temozolomide treatment in advanced glioma led to extensive ITH and poor ICB response 38 . Tumors in our models developed extensive ITH and a high burden of subclonal mutations that was not detected by bulk sequencing analysis, highlighting the importance of standardization of clinical pipelines to estimate TMB. Strategies to robustly assay ITH, such as multi-region or single-cell DNA-sequencing, may enhance the predictive utility of TMB. Overall, the results from our models strongly support a potential role of ITH in the failure of ICB in some patients with MMRd cancer. However, studies in the mouse cannot be generalized to humans and these results will require clinical validation. Our reanalysis of MMRd cancer clinical trials 35,50 showed significant association of clonal neoantigen burden and ITH index with ICB response and lends credibility to our models, but is limited by small sample size and not powered to differentiate any predictive value of subclonal neoantigen burden or ITH index beyond their association with clonal neoantigen burden. Larger prospective clinical studies will be required to definitively establish the role of ITH and its potential utility as a biomarker of ICB response in human MMRd cancers 5,6 . Similar to other mechanistic studies of ITH 19,26 , we found that experimental reconstitution of ITH potentiated immune evasion. Recently, a preclinical study showed that genetic or pharmacological enrichment of MMRd in the context of mixed MMRd/MMR-proficient (MMRp) cell-line transplants potentiated rejection of the MMRp fraction 55 . This must be interpreted with care, however, because both MMRd and MMRp fractions were derived from the same carcinogen (N-nitroso-N-methylurethane)-induced colon carcinoma line, CT26, and probably share a high burden of clonal neoantigens that may underlie rejection of the MMRp fraction. What distinguishes our study from these prior cell line-based studies is that mutagenesis occurred spontaneously entirely in vivo. That some of the subclones we isolated were highly immunogenic on re-transplantation at clonal, but not subclonal, fraction suggests that they were protected from deletion by high ITH in the original tumor. This probably occurs passively due to low cellularity precluding efficient crosspriming and driving early T cell dysfunction or ignorance 19,49 . Indeed, our high-resolution analysis of neoantigen-specific T cells after ICB treatment showed that the magnitude and effector potential of the response are attenuated with decreasing neoantigen clonality. This is in agreement with an orthogonal study that used retroviral neoantigen libraries to manipulate clonal fraction 19 . However, although that study concluded that neoantigen clonal fraction, not total cellularity, is the major determinant of the response, our experiments suggest the converse. Additional studies are needed to resolve these and other complexities, including the interplay of ITH with T cell interclonal dynamics, where the immune response may be deployed against a limited subset of 'dominant' neoantigens 56 , at the expense of recognition of lower affinity, poorly expressed or subclonal neoantigens.
Paradoxically, we found that immunosurveillance may exacerbate ITH by shaping the clonal architecture of tumors while failing to delete most neoantigens. Therefore, we conclude that ITH is shaped by the interplay of tumor intrinsic, positively selective and immunogenic, negatively selective evolutionary pressures. These results provide nuance to our understanding of immunoediting and highlight the power of our models to capture a hallmark of human cancer that is lacking in carcinogen-induced and genetically engineered models alike 57 -the gradual accumulation of mutations over time 36 .
Our results raise important questions related to therapies aimed at deliberately increasing TMB to enhance tumor immunogenicity 27,58 . These strategies will probably fail to elicit meaningful immune engagement. More concerning, collateral mutagenesis may drive more aggressive cancer, therapy resistance or secondary malignancies. Future studies with models that enable temporal control of cooperating tumorigenic events will be helpful in determining the impact of cancer clonal selection on immunosurveillance and immunotherapy response.

Online content
Any methods, additional references, Nature Portfolio reporting summaries, source data, extended data, supplementary information, acknowledgements, peer review information; details of author contributions and competing interests; and statements of data and code availability are available at https://doi.org/10.1038/s41588-023-01499-4.
Publisher's note Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations.

Tumor models
Tumor burden, where measurable, was not allowed to exceed 1 cm 2 and animals showing discomfort or distress were humanely euthanized following the recommendations of the American Veterinary Medical Association. Autochthonous lung tumors in Kras LSL-G12D ; Trp53 flox/flox ; R26 LSL-Cas9 and Kras LSL-G12D ; Trp53 flox/flox ; Msh2 flox/flox mice were induced by intratracheal instillation of 2 × 10 4 transduction units (TU) of lentivirus and 2 × 10 8 plaque-forming units of adenovirus-expressing Cre driven by the alveolar type II cell-specific surfactant protein C promoter (SPC-Cre), respectively, as previously described 31 . Autochthonous colon tumors in R26 Cas9 mice were induced by endoscope-guided submucosal injection in the distal colon, as previously described 33,63 . Two injections at 1.5 × 10 6 TU of lentivirus in 50 μl of Opti-MEM were delivered per mouse. Lentivirus was produced in HEK293 cells (American Type Culture Collection) and concentrated as previously described 31 , and functional titers (Cre activity, mScarlet fluorescence) measured as previously described 64 . Cell lines were orthotopically transplanted by intratracheal instillation of 1 × 10 5 cells in 50 μl of Spinner Modification of Minimal essential Eagle's Medium (SMEM)/5 mM EDTA, followed by a 30-μl rinse with the same medium. Cell lines were established from autochthonous lung tumors by microdissection and mechanical mincing in digestion buffer (Hanks' balanced salt solution with 1 M Hepes, 125 units ml −1 of collagenase type IV (Worthington) and 20 μg ml −1 of DNase (Sigma-Aldrich)), followed by incubation at 37 °C with gentle agitation for 30 min and plating in RPMI + 10% fetal bovine serum (FBS). Lines were plated into 50:50 RPMI/Dulbecco's modified Eagle's medium (DMEM) + 10% FBS at first passage and DMEM + 10% FBS at second passage and thereafter. Cells were taken for WES at the third passage. Msh2-expressing lentivirus was produced as above. Cells were incubated with lentiviral supernatant and 3 d later selected and maintained thereafter on medium with 6 μg ml −1 of puromycin (Thermo Fisher Scientific).
Somatic SNVs and indels were detected using Mutect2, MuSE v.1.0rc 67 , VarDict v.1.8.2 (ref. 68) and Strelka2 v.2.9.2 (ref. 69) against matched normal tails. Mutect2 was run using a panel of normals compiled from the 18 tails analyzed in the present study. Each caller was run independently on each tumor-normal pair and calls were integrated using SomaticCombiner v. 1.03 (ref. 70). For colon tumors, a panel of four normal tails was used to generate the matched normal control for all samples, because these mice were of pure background. We considered SNV mapping to only exonic regions that were detected by Mutect2 and at least one of the other algorithms. To increase accuracy of indel detection, only indels detected by at least two algorithms were considered. Variants mapping to dbSNP (build ID 150) positions were discarded. Mutations identified in tumors from two or more animals or in at least 50% of tumors from the same animal were discarded. No VAF filter was applied. Microsatellites were annotated using SciRoKo v.3.4 (ref. 71), with minimum score = 8, seed length = 8, repeats = 2 and mismatch penalty = 1.

CCF estimation
Somatic copy-number aberrations were detected by integrating output of GATK and FreeBayes v. Briefly, GATK4 Somatic CNV workflow was utilized for normalization of read counts and genome segmentation using the panel of normals from all tails. FreeBayes was used to obtain B-allele frequencies for dbSNP variant sites. PureCN was used to integrate output from GATK and FreeBayes to estimate allele-specific consensus copy-number profile, purity and ploidy of each sample. The ploidy of cell lines was determined experimentally by metaphase spreads and input into PureCN. Finally, the CCF for each SNV and indel was computed using the R (v.4.0.2) package cDriver 43 .

Clonal deconvolution by targeted amplicon sequencing
To identify private somatic SNVs for distinguishing individual clones in the M1-8 mixed clone tumors (Fig. 4c and Extended Data Fig. 4g,h), we compiled all clonal SNVs in copy-number-neutral regions (four copies, as all lines were tetraploid by metaphase spreads). We then checked the BAM files across all other samples for the complete absence of reads supporting the alternative allele (base quality >20, mapping quality >30) using an in-house Python script relying on the Pysam library. Four Article https://doi.org/10.1038/s41588-023-01499-4 private SNVs for each clone and four common SNVs were validated by PCR amplification and Sanger sequencing before proceeding. The 200-to 250-bp regions spanning these SNVs were either individually PCR amplified from samples, gel purified and combined, or amplified in parallel using a multiplexed PCR panel with primers carrying unique molecular indices (CleanPlex UMI Custom Panel, Paragon Genomics). Amplicon libraries were sequenced on the Illumina NovaSeq 6000 S4 platform with150-bp paired-end chemistry.
Reads were aligned to a fasta reference file of all targets (±250 nt upstream/downstream of SNV, GRCm38) using BWA-MEM (v.0.7.17-r1188) 65 , following the GATK Best Practices workflow. Pileups were generated using the mpileup function of bcftools v.1.10.2 (ref. 76) with --min-BQ 30 and a bed file of all SNV coordinates. For Clean-Plex UMI libraries, the following functions in fgbio (v.2.0.1) (https:// github.com/fulcrumgenomics/fgbio) were called to extract unique molecular identifiers (UMIs) and call consensus reads: ExtractUmis-FromBam, GroupReadsByUmi and CallMolecularConsensusReads. Using a customized R (v.4.0.2) script, total and SNV-specific depths at all locations were extracted. All SNVs were supported by more reads than other alternative alleles in the M1-8 clone-equal mixture control, except for M6_2, which was excluded from subsequent analysis. Background PCR/sequencing error for each SNV was estimated using the median observed frequencies of SNVs in all metastases of different clones, which represented truly clonal controls. SNV frequencies were adjusted by subtracting background values. Clonal percentages in ex vivo tumors were estimated by taking the median of private SNV frequencies, multiplying by 4 (SNVs are 1/4n) and dividing by tumor purity-estimated as the median observed/expected ratio of frequencies of the four common SNVs (present in all clones).

Neoantigen prediction and expression
Variant consequence was annotated using Ensembl Variant Effect Predictor (VEP) v.99 (ref. 77) with Wildtype and Downstream plugins, the VEP cache and reference genome for GRCm38, and the following parameters: --symbol, --terms=SO, --cache, --offline, --transcript_version and --pick. The --pick parameter was reordered from default to report the transcript with most extreme consequence for each variant: rank, canonical, appris, tsl, biotype, ccds, length and mane. Neoepitopes were predicted with C57BL/6 mouse MHC-I alleles, H2-K1 (H-2K b ) and H2-D1 (H-2D b ) and variant effect predictions using pVACtools v. 1.5.7 (ref. 78). Mutant peptides were generated for peptides that were 8-11 amino acids and MHC:peptide binding affinity was predicted for all peptide:MHC allele pairs with NetMHC-4.0, NetMHCpan-4.0, SMM v.1.0 and SMMPMBEC v. 1.0 (refs. 79-82). The median value across all affinity predictions was taken as the final measure of binding affinity. Neoantigens were subset to those with median predicted H-2K b /H-2D b affinity ≤500 nM. Where multiple neoantigens were predicted for the same SNV, only that with highest predicted affinity was retained.

Histology and immunohistochemistry
Quantification of lung tumor burden by grade was performed on scans of hematoxylin and eosin (H&E)-stained sections by automated convolutional neural network (CNN)-developed in collaboration with Aiforia Technologies Oy in consultation with veterinarian pathologist R. Bronson. Using semantic multi-class segmentation, the CNN was trained to classify lung parenchyma and adenocarcinoma grades 1-4. For supervised training, selected areas from 93 slides were chosen. The algorithm performed consistently and with high correlation with human graders across multiple validation datasets independent of the training dataset. Algorithm v.NSCLC_v25 was used. Triple staining (CD8a, CD4 and FOXP3) immunohistochemistry (IHC) and CNN quantification (Aiforia) were performed as previously described 49 . CD3 infiltration in single-stain slide scans was measured as percentage of pixels positive for stain (diaminobenzidine) in Aperio ImageScope. The area of positive and negative MSH2 staining was quantified by manual annotation in QuPath v.0.1.2 (ref. 85).

In vivo tumor imaging and quantification
Lung tumor progression was monitored longitudinally by X-ray microcomputed tomography (μCT) using a GE eXplore CT 120 system, as previously described 86 . Solid lung volume (tumor burden) was quantified using a customized MATLAB (MathWorks) script, as previously described 86 . Colon tumor progression was monitored longitudinally using a Karl Storz colonoscopy system with white light and red fluorescent protein fluorescence and biopsy forceps serving as a landmark for objective positioning, as previously described 49 .

Validation of CRISPR-Cas9 editing and estimation of tumor purity
To validate efficiency of gene editing by clustered regularly interspaced short palindromic repeats (CRISPR)-Cas9, 200-to 250-bp regions spanning sgRNA sites in the genome were amplified and deep sequenced (Massachusetts General Hospital DNA Core). Colon tumor purity was estimated using a non-wild-type allele fraction at the sgRNA-targeted site in Apc. Loss of Apc is prerequisite for tumorigenesis in the model and thus an assumption was made that all tumor cells harbor loss-of-function edits at this locus. Tumor purity in sgMsh2-targeted lung tumors was estimated using WES BAM coverage spanning exons of the Trp53 flox allele and flanking genes (Wrap53, Atp1b2), which were retrieved using the bedcov function of SAMtools v.1.10. The ratio of median coverage in flanking exons (Wrap53 exons 0-9, Trp53 exon 11 and Atp1b2 exons 0-6) versus Trp53 exons flanked by Cre loxP sites (exons 2-10) was calculated in tumors and normal tails. This ratio in tumors, representing the extent of Trp53 flox recombination, was then normalized to the median ratio across matched normal tails to estimate tumor purity, with the assumption that all tumor, but not normal, cells underwent complete recombination of Trp53 flox alleles. Efficiency of Msh2 knockout in KP; Msh2 flox/flox lung tumors was similarly estimated by taking the ratio of reads at the exon flanked by loxP sites (exon 12) and surrounding exons, and adjusting this by tumor purity.

In vitro cell-line assays
Serial live cell imaging of cell lines grown in 96-well plates (Corning) and quantification of confluence were performed with an IncuCyte S3 (Sartorius). Eight replicate wells were seeded with 100 cells and imaged every 3 h for ~6 d.

Phylogenetic tree analysis
All somatic SNVs and indels called by the WES analysis pipeline in M1-8 clones and the 09-2 parental cell line were considered in constructing a phylogenetic tree. The R (v.4.0.2) Bioconductor package phangorn (v.2.7.0) was used to construct a tree from a binary presence/absence matrix of mutations across clones and 09-2_par. Specifically, the function prachet was used to calculate the tree using the parsimonious ratchet method and the function acctran was used to calculate branch lengths.
MS-MS was performed on eluted peptides as we have previously described 49 . Tandem mass spectra were searched with Sequest (Thermo Fisher Scientific, v.IseNode in Proteome Discoverer 2.5.0.400). Sequest was set to search the mouse Uniprot database (3 July 2020 version) with 55,650 entries, including common contaminants and green fluorescent protein, Cas9, puromycin and P2A (present in the cell lines) assuming no digestion enzyme, with fragment and parent ion mass tolerances of 0.02 Da and 10.0 p.p.m., respectively. TMTpro was added as a fixed modification on the carboxy and amino termini of peptides. Oxidation of methionine was specified in Sequest as a variable modification. The resulting peptides were filtered to exclude peptides with an isolation interference >30% and p.p.m. error >±3 of the median p.p.m. error of all peptide-spectrum matches. Peptides were further prioritized based on concordance of relative abundance across clones with presence/ absence of the associated somatic mutation.

Dendritic cell vaccination, ELISpot and MHC-I multimer staining
Bone marrow-derived dendritic cells were prepared, activated, loaded with putative neoepitopes and injected intradermally at the base of the tail of healthy C57BL/6 mice, followed by two heterologous boosts, as previously described 45 . A week after the second boost, spleen and peripheral blood were collected for IFN-γ ELISpot and MHC-I:epitope tetramer flow cytometric assays. Red blood cells were first lysed with ACK lysis buffer. IFN-γ ELISpot was performed following the manufacturer's recommendations (ImmunoSpot, Cellular Technology Limited) using 750,000 splenocytes per well. H-2K b tetramers were custom generated as previously described 45 and used at 1:200 dilution. H-2D b tetramers were generated using UV-labile monomers (UVX Flex-T, BioLegend) following the manufacturer's recommendations and used at 1:50 dilution. H-2K b :QAYAFLQHL dextramers were generated using the U-Load Dextramer Kit (Immudex) following the manufacturer's recommendations and used at 1:10 dilution.

Analysis of human MMRd cancer clinical trials
Raw WES reads from Bortolomeazzi et al. 50 and Kwon et al. 35 trials (ClinicalTrial.gov identifiers: NCT02563002 and NCT02589496) were mapped to the reference human genome (GRCh38) using BWA-MEM v.0.7.17-r1188 (ref. 65). Aligned reads were processed as BAMs following the GATK v.4.1.8.0 Best Practices workflow to remove duplicates and recalibrate base quality scores 66. Somatic SNVs and indels were detected using the same pipeline and callers described above for mouse tumors. CCF values were estimated as described above.
Raw RNA-seq reads were mapped to the human reference genome (GRCh38) using STAR v.2.7.1a 83 . STAR was run using the same parameters as described above in the mouse analysis. The function Htseq-count from the Python library HTSeq (v.2.0.1) 89 was used to compute read counts for each gene (Ensembl release GRCh38.90), which were normalized to transcripts per million. Neoantigens were predicted and prioritized as described above in the mouse analysis. Clonal and subclonal neoantigens were classified as CCF ≥ 0.75 and <0.5, respectively.
Clinical responses were binned into two groups: OR, including partial and complete responders and NR, including patients with stable and progressive disease. PFS analysis was performed on a combination of both trials 35,50 with the trial study as a covariate. Importantly, there was no significant difference in PFS between patients from the two trials. Cox's regression was performed in R (v.4.0.2) using the package survival (v.3.4-0) 90  To assess statistical significance, Fisher's exact 2 × 2 test was used on categorical variables and two-tailed Wilcoxon's rank-sum test or Student's t-test (where the assumption of normality was met) was used on continuous variables. HRs were calculated and compared using Cox's proportional hazards regression. Multiple-comparison corrections were performed using Holm's method. No statistical method was used to predetermine sample size. In preclinical trials of lung and colon models, only those animals with apparent tumors by μCT or colonoscopy were recruited. No other data were excluded from analyses. Preclinical trials were randomized and investigators blinded to allocation during dosing, imaging and quantification. No experiments failed to replicate.

Reporting summary
Further information on research design is available in the Nature Portfolio Reporting Summary linked to this article.

Data availability
Raw exome sequencing and RNA-seq data from Bortolomeazzi et al. 50 are available through controlled-access application via the European Genome-Phenome Archive (hosted by the EMBL-EBI and the CRG) under accession no. EGAD00001006165. Raw sequencing data from Kwon et al. 35 were downloaded from the European Nucleotide Archive (ENA) database under primary accession no. PRJEB40416. The sequencing data generated in the present study are available at ENA under primary accession no. PRJEB56609. Raw MS data generated in the present study are available at MassIVE under accession no. MSV000092096. Source data are provided with this paper.