Mapping intratumor heterogeneity across layers for advancing immunotherapy

SUMMARY

Intratumor heterogeneity (ITH) encompasses genetic, epigenetic, transcriptional, proteomic, and immunopeptidomic diversity. Beyond genetic heterogeneity, it is increasingly clear that non-mutational heterogeneity and plasticity generate dynamic cancer cell states with distinct immune visibility. These layers of complexity converge on the immunopeptidome, the repertoire of peptides displayed by major histocompatibility complex molecules through which tumor cells are surveyed by T cells. Variation in antigen processing, presentation, and peptide abundance across cancer clones and cell states yields spatially and temporally distinct immunological niches that shape immune recognition and therapeutic response. Here, we summarize how multidimensional ITH manifests across cancer types and constrains immunotherapy efficacy. We propose that integrating measurements across layers is a promising direction for improving biomarker identification and informing more precise immune-based treatment strategies.

INTRATUMOR HETEROGENEITY: A MULTIDIMENSIONAL PROBLEM

Tumors are complex and dynamic ecosystems composed of diverse malignant and non-malignant cell populations that coexist and co-evolve within the same lesion.^1,2 Cancer cell diversity, broadly referred to as intratumor heterogeneity (ITH), has long been appreciated at the genetic level, where subclonal mutations and copy-number alterations (CNAs) fuel tumor evolution, disease progression, and resistance to therapy.^3,4 However, it is now clear that genetic variation alone cannot fully explain differences in tumor behavior, immune recognition, or clinical outcome.^5–7 Non-mutational plasticity, driven by epigenetic remodeling, transcriptional and enhancer reprogramming, RNA processing and modifications, and rewiring of translation and protein quality control, can generate proteomic diversity and distinct tumor cell states without accompanying genetic change.^8–10 Importantly, these states are shaped not only by altered protein synthesis but also by post-translational modifications and regulated protein turnover, coupling cellular stress and metabolic programs to dynamic changes in functional phenotype and immune visibility.

While the hallmarks of cancer outline the capabilities tumors must achieve to grow, survive, and adapt, not all cancer cells within a given lesion acquire or deploy these capabilities uniformly. Different subpopulations may fulfill distinct subsets of hallmarks or dynamically switch among them over time. These layers of heterogeneity collectively shape how tumors grow, interact with their microenvironment, and respond to immunotherapy (Figure 1).¹² ITH may therefore be considered an enabling feature of cancer, whereby diversity in cancer cell states enhances overall tumor fitness by buffering stress, facilitating adaptation, and promoting survival under therapeutic pressure. Understanding ITH as a multi-layered, functional property of cancer is therefore essential for deciphering mechanisms of immune escape and for developing more effective, precision-guided therapeutic strategies.⁶ In this review, we highlight mechanistically informative examples that reveal how ITH across molecular layers shapes immune recognition and therapeutic vulnerability, with a particular focus on immunotherapy.

Figure 1. ITH as a lens on the hallmarks of cancer.

Redrawn schematic of the hallmarks of cancer¹¹ arranged in a circular wheel. Each colored wedge represents one classical or emerging hallmark. At the center, a tumor mass composed of multiple genetically, epigenetically, and phenotypically distinct clones is shown. A magnifying glass zooms in on the various hallmarks of cancer, illustrating how intratumor heterogeneity (ITH) can lead to variation across genomic, epigenomic, transcriptomic, proteomic, and immunopeptidomic layers of the tumor. This figure highlights a central concept of the review, namely that ITH is not a separate hallmark but a cross-cutting property that shapes the way cancer hallmarks operate across tumor cell populations.

ITH ON THE GENETIC LEVEL

Genetically, ITH reflects the presence of multiple co-evolving tumor clones within a single tumor, shaped by branching evolutionary trajectories rather than linear clonal expansion.¹³ Each clone is characterized by distinct somatic mutation profiles and clonal prevalence. Importantly, the intratumoral distribution of truncal (clonal, shared by all tumor cells) versus branch (subclonal, present in only a subset of tumor cells) mutations, and the resulting neoepitope architecture, profoundly influences immune surveillance.^3,14–18 This principle was experimentally demonstrated by Gejman et al.¹⁹ who showed that clonal neoantigens induce more robust anti-tumor immune responses than subclonal neoantigens.

Extending these findings in an in vivo setting, Wolf et al.²⁰ employed genetically engineered melanoma models to uncouple tumor mutational burden (TMB) from ITH, demonstrating that increasing genetic heterogeneity, while keeping mutational load constant, markedly impaired T cell infiltration, neoantigen recognition, and tumor control. These data establish genetic ITH itself, rather than mutation number alone, as a dominant determinant of anti-tumor immunity. Consistent with this concept, analyses of human tumors have shown that enrichment for clonal neoantigens correlates with improved immune infiltration and response to immune checkpoint blockade (ICB), whereas subclonal neoantigen dominance is associated with immune evasion despite high overall TMB.^3,14 Together, these studies provide a mechanistic framework explaining how tumors with apparently favorable genomic features, such as high mutational burden, can nonetheless limit tumor-wide immune control through branched clonal architecture, which dilutes neoantigen display and can misdirect anti-tumor immune responses toward only a subset of tumor branches. Downstream consequences of this architecture, including constraints at the level of T cell priming and antigenic target selection, are discussed further below. Supporting this view, multiregional analyses of human cancers reveal that immune selection can actively sculpt tumor clonal composition through depletion of immunogenic neoantigens, loss of antigen-presentation capacity, and HLA loss of heterozygosity.¹⁸ Importantly, these findings also suggest that ITH and immune function may relate in both causal directions: on the one hand, increased heterogeneity can impair potent immune control. On the other hand, ineffective immune surveillance may allow the persistence and outgrowth of multiple subclones, leading to high ITH as a consequence of insufficient immune constraint during tumor evolution.

In general, genetic clonal evolution requires the progressive accumulation of mutations in cancer cells, enabled by genomic instability. The deregulation of DNA repair mechanisms constitutes a recurrent event sustaining genomic instability that is able to generate characteristic mutational processes fueling oncogenesis, shape tumor evolution, and critically influence sensitivity to genotoxic therapy and ICB.²¹ Even in the absence of clear defects in DNA repair, tumors are intrinsically characterized by their distinct mutation rate footprint that is conserved among different clones from the same ancestor but diverges markedly across different tumors.²² Notably, the same mutational processes that increase neoantigen burden can simultaneously amplify subclonal diversity, possibly expanding the evolutionary space for immune escape.

Defects in core DNA damage response (DDR) pathways such as mismatch repair (MMR), homologous recombination (HR), nucleotide excision repair, and direct reversal are pervasive across human cancers and underlie hallmark patterns of genomic instability and mutational signatures.^23,24 Molecular alterations in genes like MLH1/MSH2 (MMR) or BRCA1/2 (HR) not only affect tumor initiation and progression but also create vulnerabilities that can be exploited therapeutically, for example, by alkylating agents, PARP inhibitors, and immune checkpoint inhibitors.^25,26

Alteration of DNA repair pathways and the molecular characteristics of the ensuing genomic instability, particularly in MMR-deficient (MMRd) cancers, also modulates tumor microenvironment (TME) composition and ITH, affecting anti-tumor immune surveillance and response to immunotherapy (Figure 2).

Figure 2. DNA damage accumulation is a major determinant of ITH and immune activation.

The accumulation of DNA lesions caused by direct DNA-damaging factors and/or impaired DNA repair capability has the potential to increase genetic heterogeneity over time, inflating ITH. At the same time, the presence of unrepaired DNA can activate an IFN-1 response, potentially increasing the recruitment and activation of immune effectors.

The peculiar features of the TME in DDR-deficient tumors are influenced by several molecular events that are direct consequences of the impaired DNA repair. For instance, the hypermutation deriving from MMR loss involves a remarkably high number of insertion or deletion (indel)/frameshift mutations and is directly associated with the high TMB and expanded neoantigen landscapes observed in MMRd tumors. This, in turn, favors infiltration by activated CD8⁺ T cells and Th1-polarized immune infiltrates.^27–29 Hyperactivation of anti-tumor CD8⁺ T cells is counterbalanced by the upregulation of multiple immune checkpoints, including PD-1, PD-L1, and CTLA-4, explaining why MMRd tumors, especially in the metastatic setting, are not naturally eliminated despite a hostile immune microenvironment.²⁹ Interestingly, while the high prevalence of frameshift mutations in MMRd tumors has the potential to generate multiple non-self neoantigens per event, potentially offsetting the detrimental effect of ITH on T cell engagement, strong anti-tumor immune surveillance does not rely on frameshifts, as it occurs largely in single-nucleotide variant-driven settings, for example, non-small cell lung cancer (NSCLC) and melanoma.³⁰ In addition, though MMRd tumors show a more frequent loss of major histocompatibility complex (MHC) class I, they still respond to ICB, possibly through the contribution of CD4⁺ and γδ T cells.^31,32

Another consequence of DNA repair defects, shared by both MMRd and HR-deficient (HRD) tumors, is the activation of the cyclic GMP-AMP synthase and stimulator of IFN genes (cGAS-STING) pathway and the consequential stimulation of type I interferon (IFN-1) responses.³³ In MMRd tumors, loss of MLH1 permits unchecked Exo1-mediated DNA excision, increasing single-stranded DNA and driving nuclear DNA leakage into the cytosol, which activates cGAS-STING.^34,35 In HRD tumors, PARP1 inhibition amplifies cytosolic DNA accumulation, likewise triggering cGAS-STING and IFN-I-dependent immunity.^36–38

While sustained accumulation of replication errors in MMRd cancers accounts for increased neoantigen burden and IFN-1 signaling, it is also responsible for a layered genetic architecture with truncal mutations/neoantigens overlaid by extensive subclonal diversification, generating pronounced ITH.^39,40 Notably, and counterintuitively, in early-stage colorectal cancer (CRC), MMRd is associated with a favorable prognosis.^26,41,42 Furthermore, despite high ITH, MMRd cancers are among the most consistent responders to ICB across tumor types, with pivotal trials establishing MMR status as a tumor-agnostic predictive biomarker and leading to tissue-agnostic approvals of pembrolizumab. In metastatic colorectal and endometrial cancers, PD-1 +/− CTLA-4 inhibitors (pembrolizumab, dostarlimab, and nivolumab ± ipilimumab) yield high objective response rates and durable progression-free survival compared with historical chemotherapy outcomes, underscoring the exceptional immunogenicity of these lesions.^28,43–46 These observations raise the question of what makes MMRd tumors so remarkably immunoresponsive and what impact ITH has on immune surveillance in MMRd cancers. Preclinical MMRd models offer insights into this puzzle, showing that increased mutational load alone is insufficient to enhance immunogenicity and that, also in hypermutated MMRd cancers, clonal neoantigen fraction favors response to PD-1/PD-L1 blockade.²⁷ Furthermore, analyses of clinical datasets from MMRd colorectal and gastric cancers demonstrate that clonal, but not subclonal, neoantigen burden correlates with response to ICB.⁴⁷ However, despite evidence that colorectal MMRd tumors with higher clonality of immunogenic mutations and clonally expanded T cells respond better to anti-PD1 agents,⁴⁸ it remains unclear whether a high level of genetic ITH effectively impairs immune responses in MMRd cancers.⁴¹

In order to fully ascertain the interaction between the extreme ITH levels caused by the continuous accumulation of mutations and immune sensitivity in MMRd tumors, future prospective studies will need to integrate multi-region genomics, T cell receptor (TCR) sequencing, and spatial profiling to interrogate how ITH and TME architecture modulate response to PD-1/PD-L1 and CTLA-4 inhibitors across solid tumors, including DDR-altered and MMRd subsets. Emerging and ongoing trials stratify patients by TMB, DDR gene alterations, and neoantigen features and explore rational combinations (e.g., DDR inhibitors plus immune checkpoint inhibitors [ICIs]) to reshape clonal architecture and overcome heterogeneity-driven immune escape, aiming to convert more DDR-deficient tumors into durable responders.⁴⁹

The impact of ITH on treatment responses has been extensively established and reported, and equally relevant could be the modulation of ITH by anticancer treatments.⁵⁰ Specifically, targeted therapies exert selective pressure on cancer cells, which may result in branched evolutionary trees and rapid expansion of genetic heterogeneity across different metastatic lesions.^51–53 While targeted agents can actively select pre-existing resistant subclones, it has also been shown that resistance subclones may arise through gradual, multifactorial adaptation to inhibitors, involving the acquisition of multiple cooperating genetic and epigenetic adaptive changes rather than the commonly assumed single mutational event.⁵⁴ Moreover, cells surviving the initial anti-proliferative effect induced by targeted agents, often referred to as drug-tolerant persisters (DTPs), can exploit an evolutionarily conserved mechanism of genetic diversification under stress by deregulating high-fidelity DNA repair mechanisms in favor of low-fidelity ones, thereby inflating their mutation rate and increasing the chance of generating a resistance mutation.^55–58

Treatment with cytotoxic chemotherapy agents can also affect ITH and immune response. For instance, it has long been known that tumors progressing or relapsing after chemotherapy present a higher degree of genetic heterogeneity.⁵⁹ This increase in ITH may also account for the lack of efficacy of ICB in melanomas that have progressed to some alkylating agents such as dacarbazine and in glioblastomas that have progressed to temozolomide (TMZ).^14,60 Conversely, when standard-of-care cytotoxic agents are combined with ICB, the response rate is typically increased, a strategy widely used in NSCLC, gastroesophageal, biliary tract, urothelial, and breast cancers.⁶¹ Importantly, cytotoxic agents could exert their immunogenic activity independently of their effects on cancer genetic evolution, through antigen-independent mechanisms.⁶² The impact of chemotherapy on the generation of hypermutation and immunogenicity was recently explored in CRC and breast cancer models. Combinatorial treatment with cisplatin (CDDP) and TMZ induced an adaptive downregulation of MMR, resulting in chemotherapy-dependent hypermutability and an increase in predicted neoantigens. Treatment with CDDP and TMZ also remodels the innate immune microenvironment and induces long-lasting responses and complete rejections when combined with anti-PD1 therapy.^63,64 These results integrate previous clinical data on immune stimulation with alkylating agents.^65,66 In one of these studies, the induction of clonal hypermutation with TMZ was followed by long-term benefits with immune checkpoint inhibitors in MMR-proficient metastatic CRC patients.⁶⁵

While easy-to-characterize point mutations remain the most studied determinant of genetic heterogeneity, they are not its only source. CNAs, due to chromosomal instability (CIN), are among the most pervasive sources of genetic ITH, encompassing focal amplifications, deletions, and arm-level or whole-chromosome gains/losses.⁶⁷ These alterations diverge progressively during tumor evolution, producing branching phylogenies detectable at single-cell resolution, with direct consequences for therapy resistance through its impact on driver gene dosage.⁶⁸ Interestingly, besides the progressive accumulation of genetic events, ITH can also be rapidly exacerbated by massive genomic events such as chromothripsis and kataegis. Chromothripsis is the catastrophic shattering of one or a few chromosomes in a single mitotic event, followed by error-prone reassembly, enabling a punctuated burst of structural rearrangements (hundreds in a single division) that can simultaneously inactivate tumor suppressors and amplify oncogenes, leading to the sudden appearance of diversified subclones.⁶⁸ Its most established mechanism involves chromosome sequestration in micronuclei, where impaired DNA replication and rupture trigger massive fragmentation.⁶⁹ Kataegis, on the other hand, refers to hypermutated genomic foci characterized by clusters of C>T and C>G substitutions at TpC dinucleotides, primarily attributable to the APOBEC3A/B cytidine deaminase family acting on single-stranded DNA. Unlike passenger mutation noise, kataegis events can be subclonally distributed, occurring at distinct time points in different subclones, thereby directly seeding mutational ITH. Their spatial co-localization with structural rearrangement breakpoints suggests a mechanistic link to replication stress and DNA repair intermediates.⁷⁰ The relationship between these mutational events and immune surveillance is far from being fully understood and is still a matter of discovery. In the case of CIN-driven alterations, both in their gradual (CNAs) and catastrophic (chromothripsis) establishment, their convergence on the formation of micronuclei promotes a persistent stimulation of inflammatory signaling through the activation of the cGAS-STING pathway. Paradoxically, this constitutes both a trigger of innate immune sensing and a driver of immune evasion when STING signaling becomes chronically rewired.⁶⁸ On the other hand, APOBEC-related mutagenic events like kataegis have been typically associated with the generation of subclonally distributed neoantigens with high immunogenicity potential. However, because these neoantigens are subclonal, they are subject to immune-mediated negative selection and clonal pruning, contributing to immune evasion through selective outgrowth of antigen-loss variants.^71,72

Finally, extrachromosomal DNA (ecDNA) has become a progressively better recognized source of ITH, aggravated by its non-Mendelian segregation during cell division.⁷³ This unpredictable segregation relies on the very nature of ecDNA, composed of large (up to megabase-scale) circular DNA elements lacking centromeres. At each division, daughter cells inherit wildly disparate oncogene copy numbers, generating continuous, high-amplitude copy-number heterogeneity and promoting enhancer rewiring by juxtaposing distant regulatory elements within its circular topology, further diversifying transcriptional states across subclones, including the suppression of immune regulation pathways.^74–76

Overall, clonal evolution can be impacted by both cell-intrinsic factors, such as DNA repair defects and catastrophic mutational events, and cell-extrinsic factors, such as exposure to treatments. These factors can increase genetic heterogeneity and potentially affect cancer-immune recognition. In this context, ITH not only creates a moving target for the immune system but also actively shapes the immune microenvironment in ways that favor tumor persistence. Concurrently, the presence of unrepaired DNA intermediates or cytosolic DNA can stimulate an IFN-I response, which may increase the recruitment and activation of immune effectors. This double-edged phenomenon explains why tumor immunogenicity is difficult to predict from a purely genomic perspective (Figure 2).

NONGENETIC ITH

ITH arises not only from genetic diversification but also from nongenetic plasticity. The latter encompasses reversible, context-dependent changes in chromatin state, transcription, proteome, metabolism, morphology, and behavior that unfold on fast timescales and can be heritably stabilized under selection. Nongenetic ITH allows subpopulations to partition “hallmarks” of cancer in complementary ways and to rapidly adapt to therapy without requiring new mutations, thereby shaping tumor evolution, minimal residual disease (MRD), and clinical resistance.

Although this section focuses on the nongenetic mechanisms that contribute to tumor evolution, genetic instability and non-mutational reprogramming are neither alternatives nor independent evolutionary mechanisms. It has become clear that we must account for multiple evolutionary mechanisms and their interactions to ultimately understand, predict, and steer tumor evolution.^77,78 The recurrent observation that specific genetic alterations, such as loss of RB1 or KRAS oncogenic mutations, for instance,^79,80 increase cancer cell plasticity illustrates this.

Non-mutational reprogramming was only recently recognized as an enabling dimension of the cancer hallmarks¹¹ largely because nongenetic diversity was difficult to measure and validate. Advances in technology and computation now clearly show its role in tumor evolution. Single-cell RNA sequencing (scRNA-seq), the most widely used lens on cell-state diversity, reveals coexisting lineage programs, stress responses, cycling states, a continuum of epithelial-to-mesenchymal transition (EMT)-like states, or DTP phenotypes across cancers.

Single-cell chromatin profiling combined with transcriptome (including long-read) sequencing exposed epigenetic gene regulatory networks, enhancer rewiring, and alternative splicing events underpinning plasticity and drug tolerance.⁸¹

Spatial technologies such as multiplexed small-molecule fluorescence in situ hybridization (smFISH) and high-plex spatial proteomics reveal how specific niches, including hypoxic rims, perivascular zones, and immune-inflamed areas, shape cellular state composition and transitions. They resolve gradients and communities of cancer cells co-expressing overlapping gene programs in the 3D space, patterns that are lost in dissociated single-cell datasets.⁸² These methods have been instrumental in highlighting the important role of the TME as a key driver of nongenetic ITH.⁸³ Proteomic and signaling readouts such as cytometry by time-of-flight (CyTOF), imaging mass cytometry, phospho-flow, and targeted panels quantify functional proteins and signaling nodes that may be decoupled from RNA, including antigen-presentation machinery, stress effectors, and apoptotic priming. They are instrumental in interpreting immune evasion and therapy responses. Lineage and fate tracking, using barcoding, mitochondrial lineage tracing, and CRISPR-based scars, help track clonal dynamics through treatment, distinguishing selection on pre-existing states from induced transitions and linking transient DTP states to later fixed resistance.^84–86 Live-cell and functional phenotyping, including time-lapse imaging, reporter systems, organoids, and co-cultures with immune or stromal cells capture switching kinetics, survival under drug, and immune evasion behaviors, thereby connecting states to consequences. When used in the context of longitudinal monitoring during progression or therapy, these modalities provide orthogonal views of cell-state reprogramming and its dynamics.⁸³

These approaches have illuminated the drug-naive state space, where nongenetic mechanisms are already at work. Melanoma as an anchor, with cross-tumor parallels, exemplifies layered nongenetic heterogeneity. A principal axis reflects differentiation along a melanocytic lineage program (MITF-high) versus dedifferentiated/neural crest-like and mesenchymal-like programs (MITF-low). Cells can occupy intermediate states and move along this continuum in response to microenvironmental cues.^83,87,88 Superimposed on this axis, melanoma cells co-express extrinsic “metaprograms” that are shared across lineages: hypoxia and metabolic rewiring, interferon/antigen-presentation programs, cell-cycle/mitotic programs, oxidative and proteotoxic stress responses, and inflammation-driven phenotypes that intersect with immune interactions.^83,87,88 These metaprograms also appear in tumor cells from other tumor types and even non-malignant cells within the TME, underscoring that transcriptional signatures are not lineage-exclusive and are heavily context-driven.^89–91

Notably, despite melanoma’s exceptionally high TMB, single-cell and multi-omic profiling of mouse and human lesions highlighted that its high transcriptomic diversity is not primarily explained by genotype.^83,87,88 This underscores that plasticity is a dominant driver of state diversity even in highly mutated tumors—and is likely to be even more consequential in cancers with lower mutational burden, such as many pediatric and liquid malignancies.

Bolstering the idea that nongenetic ITH is an important hallmark-enabling feature, growing evidence shows that distinct cancer cell states can cooperate to produce tumors with greater fitness. In melanoma, for example, fibronectin produced by mesenchymal-like cells promotes the survival of melanocytic cells.⁹² This illustrates how communication between cancer cell states may act as a driver of tumor progression.

Nongenetic mechanisms are especially important during bottleneck events that demand rapid adaptation for survival. Such events are particularly notable during metastatic dissemination and upon exposure to therapy (Figure 3).

Figure 3. Contributions of nongenetic mechanisms to tumor evolution and therapy resistance.

Nongenetic mechanisms contribute to cancer cell state diversification at various stages of tumor evolution and therapy processes. They play particularly clinically relevant roles during bottleneck events. They allow cells that exhibit increased ability to adapt and survive owing to increased plasticity to emerge and drive processes such as metastasis and drug tolerance. It has recently been proposed that these cells display a phenotypic state that may be referred to as “driver” cell states,⁹³ by analogy to the “oncogenic” driver mutations. EMT, epithelial-to-mesenchymal transition; MET, mesenchymal-to-epithelial transition; MICs, metastasis-initiating cells; DTPs, drug-tolerant persisters; TME, tumor microenvironment.

Metastasis frequently relies on nongenetic plasticity that enables reversible transitions among epithelial, mesenchymal, and stem-like states.⁹⁴ EMT promotes dissemination, while the mesenchymal-to-epithelial transition (MET) can facilitate outgrowth at distant sites.^95–97 Dissemination can occur without a full EMT, with partial EMT, or EMT-independent routes, underscoring plasticity rather than fixed genetic programs.^98,99 Such state switching can be stochastic and reversible, generating metastasis-competent subpopulations (referred to as metastasis-initiating cells, or MICs) without new mutations, and is underpinned by epigenetic reprogramming that rewires enhancers and chromatin to control EMT/MET and other invasive phenotypes.¹⁰⁰ Microenvironmental cues provide potent, extrinsic instruction: transforming growth factor β (TGF-β) primes organ-specific metastasis, such as lung tropism, and hypoxia-driven lysyl oxidase remodels the extracellular matrix to enable dissemination and colonization without altering genotype.¹⁰¹ Dormancy and later awakening are governed by microenvironmental and inflammatory cues, including neutrophil extracellular traps that “reactivate” quiescent disseminated cells, highlighting metastasis control by context-dependent signals.¹⁰² Mechanical forces further shape metastatic competence: YAP/TAZ-driven mechanotransduction reprograms cell fate in response to matrix stiffness and cytoskeletal tension, promoting invasion and survival without genomic change.¹⁰³ Melanoma exemplifies these principles: cells switch between Prrx1^HIGH; MITF^LOW mesenchymal-like/invasive and Prrx1^LOW; and MITF^HIGH proliferative states as cells leave the primary tumors to ultimately end up colonizing secondary metastatic sites, such as the lung.⁸⁷ Collectively, these lines of evidence demonstrate that metastasis is often powered by reversible cellular states, epigenetic remodeling, and microenvironmental and mechanical cues, rather than by additional genetic alterations (Figure 3). Importantly, cooperative behaviors also matter during metastatic dissemination. This is, for instance, illustrated by collective migration and circulating heterotypic tumor cell clusters, which maintain cell-cell junctions and show markedly enhanced metastatic efficiency.¹⁰⁴ This observation further underscores ITH as a key enabling property during this particularly lethal process.

Under targeted therapy (e.g., mitogen-activated protein kinase [MAPK] pathway inhibitors), a subset of cancer cells frequently enters DTP states characterized by slowed proliferation, chromatin remodeling, altered metabolism, and selective stress-response activation.⁵⁵ These DTPs can persist as MRD after apparent radiographic response and later seed clinical relapse (Figure 3). Melanoma DTP cells can follow divergent trajectories from a common tolerant pool, bifurcating into distinct post-drug fates with different vulnerabilities—complicating eradication because a single targeting strategy may hit one branch but spare another.¹⁰⁵ Beyond melanoma, similar patterns emerge across tumor types. Multiple DTP states have been documented in EGFR- and ALK-driven lung cancers, HER2- or endocrine-driven breast cancers, and in prostate cancers undergoing lineage plasticity toward neuroendocrine phenotypes or CRCs.¹⁰⁶ EMT- and partial EMT-like programs recurrently detected in DTP cells confer these cells with enhanced survival capabilities under therapeutic stress and immune evasion properties, often without fixed genetic changes. Likewise, cells can increase their tolerance to targeted therapy by engaging senescence-like, oncofetal, and embryonic-like programs or engaging in differentiation trajectories.⁵⁵ These state transitions are often accompanied by an exit from the cell cycle—one of the key events linked to survival. Newer studies incorporate time-resolved, barcoding, and spatial multi-omics to link therapy-induced trajectories with later, mutation-fixed resistance, supporting a model in which nongenetic adaptation seeds the “soil” for subsequent genetic takeover. Importantly, the emergence of DTP cells within MRD can be a consequence of both the selection of pre-existing tolerant phenotypes (which may have been acquired through genetic or nongenetic mechanisms) and/or the induction of new states through nongenetic reprogramming. This duality explains the rapid emergence of MRD, often accompanied by little to no genetic change. Notably, even in lesions with high mutation load, such as melanoma, the switch from tolerance to resistance can also be driven by nongenetic events.¹⁰⁷

Immune escape intersects tightly with nongenetic plasticity. Cancer cell-intrinsic nongenetic adaptation to ICB includes reversible suppression of antigen presentation and lineage programs, state switching, and stress-induced reprogramming. Melanoma and other tumors can downregulate MHC class I and the antigen-processing cascade via epigenetic repression of the MHC transactivator NLRC5 and interferon-stimulated genes, which diminishes cytotoxic T cell recognition yet can be restored by epigenetic therapy.¹⁰⁸ Dedifferentiation reduces lineage antigen expression,¹⁰⁹ blunts antigen-presentation programs, and correlates with resistance to anti-PD-1.⁸³ Chronic IFN signaling can drive a stable, epigenetically imprinted resistance program that desensitizes tumors to subsequent immune attack, illustrating nongenetic state adaptation under immune pressure.¹¹⁰ Adaptive upregulation and stabilization of PD-L1 on tumor cells further suppresses T cell function without requiring new mutations.¹¹¹ In addition, cancer cell adaptation to endoplasmic reticulum (ER) stress—induced by oncogenic pressure, hypoxia, nutrient limitation, or therapeutic stress—can shape anti-tumor immunity, impairing dendritic cell (DC) maturation and function, altering macrophage phenotype, suppressing CD8⁺ T cell and natural killer (NK) cell activity, and broadly remodeling the tumor immune microenvironment.^112–115

Functional studies (lineage tracing, depletion, and genetic perturbation) show that, although tumors harbor many recurrent cell states, only a minority fuel clinically relevant processes such as growth, metastasis, and therapy resistance. Building on this, the genetic driver/passenger paradigm has been extended to phenotypic cell states as a blueprint for linking specific states to disease evolution.⁹³ In this framework, driver cell states are the primary engines of progression and therapeutic failure, often superseding individual mutations by coordinating proliferative, survival, and adaptive programs. Passenger states are phenotypes that do not measurably contribute to tumor evolution. A third class, trailer states, modulates behavior in a context-dependent manner: they can shape fitness and tissue ecology beyond (epi)genetics, conferring both driver-like advantages and exploitable liabilities. The framework offers a functional lens to prioritize states for intervention and to identify those that truly determine clinical outcomes. Integrated with ITH as a hallmark-enabling property, trailer states assume special importance: while not intrinsically driving evolution, they can support driver states in specific contexts and become critical during bottlenecks (e.g., therapy, immune attack, and dissemination), thereby attaining high clinical relevance.

DYNAMIC INTERPLAY BETWEEN ITH AND ANTI-TUMOR IMMUNE RESPONSES

The immune system can detect cancerous growths from the premalignant stage through terminal disease. Following activation and detection, specialized immune cells can eliminate cancer cells specifically, leading to immune editing. Thus, the degree of ITH and a robust anti-tumor immune response are intertwined throughout tumor progression. Here, we will review the effects of the immunopeptidome on innate immune sensing and T cell priming and, lastly, on immune evasion.

The classical view of anti-tumor immune responses focuses on cytotoxic CD8⁺ T cells eliminating MHC class I⁺ cancer cells in a highly specific manner by detecting 8–10 amino acid-long peptides (subsequently termed antigens) presented on MHC class I molecules (pMHC class I).¹¹⁶ However, for this process to occur effectively, several additional steps are required to induce a functional immune response.¹¹⁷ Anti-tumor immune responses are initiated by the recruitment of innate immune cells to the site of cancer growth. One highly specialized cell subset is DCs, which are professional antigen-presenting cells.¹¹⁸ DCs are constantly surveilling their surroundings for signals of stress, often referred to as danger signals.^119,120 In cancer, these danger signals are often directly derived from cancer cells and include tumor-cell-derived double-stranded DNA, double-stranded RNA, the histone marker HMGB1, or IFN-I.^121–124 Upon sensing these danger signals, DCs will acquire a mature activation state, which will induce a migratory program while antigen uptake, either via phagocytosis or micropinocytosis, is ceased.^125–127 It is important to note here that tumor cells can reduce the amount of danger signals they produce, for instance, by turning off the cGAS/STING pathway, already establishing a heterogeneous environment at the level of innate immune sensing.¹²⁸ Following maturation, DCs will migrate to the closest draining lymph node, where they will cross-present tumor-cell-derived antigens (this process will be reviewed in detail below) to naive T cells.^118,126 In addition to the migratory program, maturation will result in upregulation of co-stimulatory molecules and cytokines.¹²⁹ Should a T cell with its TCR recognize the presented pMHC class I complex on a mature DC, it will trigger T cell activation, which is characterized by clonal expansion, acquisition of effector functions, and T cell fate decisions. In the cancer setting, this is often a fate of T cell exhaustion (recently reviewed in Bhandarkar et al.¹³⁰). Following T cell activation and differentiation, T cells will migrate back to the site of inflammation, driven by chemokine gradients, often facilitated by DCs that remain in the tumor.^131,132 Once the T cell migrates into the TME, it will exert its cytotoxic function and specifically eliminate tumor cells that present the pMHC class I the T cell is specific for.¹¹⁷ As a consequence, a robust T cell response can eliminate antigen-positive cancer cells, whereas antigen-negative or low-antigen cancer cells will escape the anti-tumor immune response.¹³³ This process has been established as cancer immunoediting.¹³⁴ Given that all tumors harbor some degree of ITH, it could be postulated that only antigens shared between all cancer cells can be efficient targets; however, this is a dramatic oversimplification of the continuous co-evolution of the tumor and the immune response. In the following paragraphs, we will review in detail how ITH affects tumor cell recognition and elimination, T cell activation, and T cell infiltration.

As outlined above, cross-presentation of tumor-cell-derived antigens by DCs is critical for the induction of an anti-tumor immune response.^117,118 Amongst all DCs, conventional type-1 DCs (cDC1) are most proficient at cross-presentation.^125,135,136 cDC1 detects tumor-cell-derived antigen via the c-type lectin receptor DNGR-1 (Clec9a), which binds to exposed filamentous actin (F-actin), mediates phagocytosis followed by antigen degradation and subsequent presentation of the antigen on cDC1.^137,138 For this reason, cDC1 is best at presenting cytosol/cytoskeleton-associated antigens.¹³⁹ While cDC1 are most proficient in cross-presentation, conventional type 2 DCs (cDC2) can acquire and cross-present antigens through various modes and substantially enhance anti-tumor immune responses.¹²² Canonically, cDC2 will be more efficient in presenting soluble antigens,^140,141 but cDC2 can also trogocytose pre-formed pMHC class I complexes from tumor cells when exposed to high levels of tumor-cell-derived IFN-I.¹²² Both cDC subsets have been shown to acquire a CCR7+ migratory phenotype, which is required for the translocation of antigen to the lymph node.¹²⁶ Their cross-presentation directly on migratory DCs or via vesicular transfer to lymph node resident cDC1 and cDC2 can mediate efficient T cell priming.¹⁴²

Regardless of the DC subtype and mode of antigen acquisition, it is important for a productive immune response that antigens cross-presented on DCs and antigens directly presented on tumor cells have a large degree of overlap. However, ITH can directly affect which antigens are the most efficiently cross-presented antigens, thus blunting responses against subdominant antigens and facilitating immune evasion. While it has long been established that DC maturation blunts the uptake of new antigens,¹²⁷ recent studies have highlighted that migratory DCs indeed only acquire antigens from one tumor cell.^143,144 Because of this, cross-presenting DC in the draining lymph node will not only reflect the antigens from the tumor but also mirror the heterogeneity of antigen expression.¹⁴⁴ Given that the above-described antigen transfer in the lymph node required synapse formation between two DCs,¹⁴² it is plausible that this process might further increase heterogeneity within the antigen-presenting cells. Thus, the already rate-limiting process of cross-presentation is further complicated by modes of antigen acquisition, cDC maturation status, and resulting effects on T cell responses.^133,142,144 Distribution of antigen presentation to different individual DCs in the draining lymph node has significant consequences for the resulting T cell response, and, for instance, work by the Schreiber and Murphy groups has shown that effective CD8 T cell priming requires CD4⁺ help, and this help is only provided when both MHC class I- and MHC class II-restricted antigens are presented by the same cDC1.^144,145 This has also been reported for CD4-independent CD8 T cell responses; however, it is complicated by competition for MHC binding, establishing antigen dominance hierarchies.^144,145 These rules of mirrored ITH on DCs were able to explain previous clinical and preclinical observations that high ITH results in blunted immunity and a failed response to ICB therapy.^14,20 Further, these data explain that anti-tumor immune responses at baseline are typically restricted to one to two antigens.^146,147 Consistently, therapeutic vaccines encoding highly expressed targets frequently result in de novo induction of antigen-specific responses when an objective clinical response is observed.^146–148 In this setting, immune pressure may be concentrated on only a subset of tumor subclones bearing the relevant antigen, consistent with the branch-specific effects of ITH described above. Strikingly, if followed over time, these responses are sufficiently strong to cause immune evasion and the outgrowth of tumor cell clones lacking the targeted antigens or evading immunity via other modes (Figure 4).^149,150 Importantly, these emerging rules of ITH mirroring on DCs open new rational strategies for selecting antigens for therapeutic vaccines and might enable the use of low-affinity but highly clonal antigens, such as germline variants, which might avoid immune evasion.¹⁵¹

Figure 4. Schematic describing how ITH shapes the tumor-immune interplay.

Tumor cells (purple) able to evade immune attack promote a cold TME, lacking immune cells and immune editing, which in turn further exacerbates ITH. By contrast, non-immune-evasive states with high NeoAg expression (red) allow a high degree of immune editing and low clonality. This immune state is associated with a high degree of T cell infiltration. Intermediate states of immune suppression are sharpened locally by tumor-cell-intrinsic signaling and heterogeneous NeoAg expression and a high degree of ITH. The intermediate state heterogeneous NeoAg expression results in heterogeneous uptake of antigens by DC, which can suppress priming against tumor-derived antigens and facilitate immune evasion due to a more oligoclonal T cell response. Note that Tregs and macrophages are not exclusive to immunosuppressed tumors, and Tregs can also be present in “hot” tumors with an ongoing CTL response. NeoAg, neoantigen; DC, dendritic cell; CTL, cytotoxic T lymphocyte; Tregs, regulatory T cells.

Over the last decade, an increasing amount of data from preclinical and clinical studies highlighted that alterations to signaling pathways directly affect anti-tumor immune responses. While most of these have already been extensively reviewed,¹³³ we will discuss their effect specifically in the context of heterogeneous perturbation of the pathways in tumor cells and focus on the mechanism of action on the immune cells.

One of the first tumor-cell-intrinsic pathways shown to affect anti-tumor immune responses was the activation of the Wnt/beta-catenin (CTNNB1) pathway in melanoma. Mechanistically, it was shown that CTNNB1 activation in mice and humans results in a lack of cross-presenting DCs infiltrating the tumor, thus blunting T cell priming in the lymph node.¹³⁶ Such an immune evasion could be disrupted by a few tumor cells with no or low CTNNB1 expression. However, the lack of DC infiltration into CTNNB1+ tumors further blunts infiltration of effector T cells into the TME, allowing the tumor to hide in plain sight.¹³² This mode of immune evasion has been shown clinically in melanoma and ovarian cancer, where CTNNB1-positive clones were selected during ICB treatment, establishing a “cold” immune-evasive tumor.^152,153 Similarly to the CTNNB1 pathway, KEAP1 alterations, COX1/2 signaling, and MYC amplification likewise show reduced DC infiltration. Thus, similar modes of immune evasion might be at play in heterogeneous tumors (Figure 4, cold tumor).^154–158

Besides a lack of DC or effector T cell infiltration, tumor-cell-intrinsic signaling can cause active recruitment of immune inhibitory cell populations, which subsequently actively suppress the cytotoxic anti-tumor immune response.¹⁵⁹ These immune inhibitory populations include regulatory T cells, immunosuppressive macrophages, and other myeloid-derived suppressor cells (MDSCs).^133,159,160 The immune inhibitory effect on the TME and CD8⁺ T cells can be as diverse as the suppressive cell populations and range from direct interaction with T cells and over suppression of DCs to inhibition of the vasculature.¹⁶¹ Tumor-cell-intrinsic pathways that are frequently expressed at very heterogeneous levels are TGFb, VEGF, and Sox2, while IDH1/2, EGFR, and p53 mutations are thought to be more clonal drivers (Figure 4, immune excluded).^162–171

Lastly, immune cells, specifically T cells, are highly sensitive to the surrounding levels of nutrients and metabolites.¹⁷² Well-studied examples include adenosine accumulation and tryptophan depletion, both of which can impair T cell function.^173,174 Adenosine levels are regulated by CD39/CD73, while tryptophan depletion is driven by indoleamine 2,3-dioxygenase 1 (IDO1), and these pathways can be highly localized to tumor cells and/or surrounding macrophages. In addition, oxygen levels drastically differ within tumors and are affected by the oxygen consumption of tumor cells.¹⁷⁵ One of the best-understood metabolic programs altered in cancer involves STK11/LKB1 loss, which affects AMP-activated protein kinase (AMPK) and mammalian target of rapamycin (mTOR) signaling, thus affecting both ATP/ADP/AMP/adenosine levels as well as oxygen metabolism.^176–180 Activation of this pathway has been shown to cause T cell exclusion in lung cancer.^176,180 While this well-established specific link between tumor-cell-intrinsic signaling, altered metabolic features, and immune evasion is caused by genetic alterations, many metabolic changes are mediated either epigenetically or in response to environmental pressures, such as new metastatic niches.¹⁸¹ Thus, metabolic heterogeneity is a highly complex feature of ITH that can affect anti-tumor immunity within one lesion but also across different cancer lesions.

While mechanistic studies have revealed how individual tumor-cell-intrinsic processes shape immune recognition, these mechanisms do not operate in isolation in patients. Instead, genetic, epigenetic, proteomic, and immunopeptidomic heterogeneity progressively accumulate and interact over time, together reshaping the TME and giving rise to distinct immunological neighborhoods (Figure 4). When overlaying these sometimes opposing or synergistic effects on the immune system, and, most importantly, considering that all of these can be heterogeneously expressed in the tumor, it becomes clear why any tumor will have a highly heterogeneous TME consisting of various neighborhoods. This concept of complex neighborhoods has been clinically described at various levels.^182,183 These can include immune-active regions, in which T cells recognize antigens and eliminate tumor cells actively, to immunosuppressed areas, in which T cells show poor infiltration or are restrained in their function, all the way to cold areas, which are deprived of T cells due to blunted infiltration or lack of recognizable antigens.^14,182,183 However, clinical studies thus far fall short of resolving the complexity of comprehensively assessing all possible mechanisms at once. The sum of these neighborhoods needs to be considered for therapeutic intervention, as each environment will mediate immune evasion differently.

IMMUNOPEPTIDOME HETEROGENEITY LINKS TUMOR CELL STATES TO IMMUNE RECOGNITION

The ability of the immune system to detect malignant cells ultimately depends on the repertoire of peptides displayed by MHC complexes on the tumor cell surface, collectively known as the immunopeptidome. This repertoire is not a static catalog but a dynamic molecular snapshot shaped by recent protein synthesis, processing, and degradation within each cell. As such, it serves as a direct functional readout of tumor cell state, integrating the consequences of oncogenic signaling, inflammatory cues, metabolic rewiring, genomic instability, and stress-adaptive programs.^184–186 Consequently, immune recognition is continuously negotiated against a shifting antigenic landscape rather than a fixed set of targets. Importantly, not all presented antigens necessarily promote productive anti-tumor immunity. The qualitative properties of certain peptides, particularly those presented in an MHC class II context, can influence CD4 T cell polarization and, in some settings, may favor dysfunctional or regulatory immune programs, providing an additional route by which antigenic heterogeneity may undermine effective tumor control.^187,188

Accordingly, immunopeptidome heterogeneity should be viewed not as an independent layer of diversity but as the integrative surface readout through which tumor-cell-intrinsic heterogeneity is perceived by the immune system (Figure 5). In this context, antigenic diversity can be influenced by genetic ITH both directly, via neoepitope-encoding mutations, and indirectly, via perturbations in trans-acting regulators of RNA processing, translation fidelity, or antigen processing/presentation (e.g., proteases, TAP/tapasin, ERAP1/2, and proteasome/immunoproteasome subunits), that alter which peptides are generated and how efficiently they are displayed.^{144,189–199} Consistent with the impact of trans-acting processing factors on peptide display, alterations in the protease CTSS were shown to remodel the immunopeptidome as quantified by MS-based immunopeptidomics, with accompanying effects on CD8⁺ T cell infiltration.²⁰⁰ Similarly, loss of a single translation-fidelity regulator (TYW2) was shown to induce aberrant pMHC class I presentation by immunopeptidomics and to enhance CD8⁺ T cell infiltration and sensitivity to checkpoint blockade, illustrating how perturbation of a trans-acting translation regulator can reshape both the ligandome and anti-tumor immunity.^198,201 In addition, variability in transcriptional and post-transcriptional cellular programs can diversify the identity and abundance of presented peptides even among neighboring tumor cells, including within genetically similar clones. Immunopeptidome heterogeneity operates at two levels (Figure 6): within individual cells (diversity and abundance hierarchies among co-presented peptides) and between cells (differences in peptide identity and/or relative abundance across neighboring tumor cells). Between-cell differences may reflect clonal architecture (truncal versus branch mutations), trans-acting variation in processing/presentation, or dynamic cell-state fluctuations over time. As a result, immune recognition is inherently spatially and temporally heterogeneous, providing a mechanistic basis for the emergence of distinct immunological neighborhoods within the same tumor lesion.

Figure 5. Multi-layer ITH progressively reshapes the TME and response to treatment.

Schematic representation of a solid tumor illustrating coexisting and interacting layers of ITH. Genetically distinct subclones (different colors) coexist with epigenetic, transcriptomic, and phenotypic diversification, expanding proteomic variability and the repertoire of the immunopeptidome. Heterogeneous immunopeptidomes can also arise among genetically similar clones, reflecting nongenetic variability, including transcriptional states and post-translational regulation, as well as differences in the secretome. These tumor-intrinsic layers are embedded within a complex microenvironment composed of immunosuppressive immune cells and fibroblast-like stromal cells (pink), which both influence and are shaped by tumor cell states. Together, overlapping genomic, epigenomic, transcriptomic, proteomic, and microenvironmental heterogeneity generate a spatially diverse and context-dependent antigenic landscape that supports immune recognition as well as immune escape.

Figure 6. ITH shapes the immunopeptidome at two levels.

Schematic tumor cross-section illustrating how genetic, epigenetic, and proteomic ITH translates into heterogeneity of the immunopeptidome. Each cell in the central tumor mass carries a distinct combination of colored surface dots, representing different peptide-MHC complexes, including potential neoantigens. The upper inset zooms in on a single cell to highlight within-cell immunopeptidome diversity, showing that individual tumor cells simultaneously present multiple peptides of varying abundance. The lower inset illustrates between-cell immunopeptidome heterogeneity, whereby neighboring tumor cells differ in the identity and relative composition of presented peptides. In some cases, these differences reflect quantitative rather than qualitative variation, arising from differential MHC expression levels driven by genetic or epigenetic regulation. Together, these two dimensions (within-cell diversity and between-cell variability) increase the complexity of the tumors’ antigenic landscape and modulate how tumor cells are perceived by the immune system. This figure emphasizes that immunopeptidome heterogeneity constitutes a critical interface through which molecular ITH influences immune recognition and the feasibility of antigen-directed immunotherapies.

CONSTRAINTS OF MUTATION-DERIVED ANTIGENS IN HETEROGENEOUS TUMORS

Somatic mutations generate tumor-specific antigens (TSAs) through the production of neoantigenic peptides that are absent from normal tissues and therefore escape central tolerance. However, the effectiveness of mutation-derived antigens as immune targets is fundamentally constrained by ITH. Most nonsynonymous mutations arise as passenger events and are confined to branch (subclonal) populations, limiting their spatial distribution and utility as immune targets.^3,12,14,202 Targeting such subclonal antigens may transiently eliminate sensitive populations while selectively favoring the persistence of antigen-low/negative clones, thereby reshaping tumor evolution under immune pressure.^{189,191,203–205}

By contrast, driver-derived neoantigens, which can arise from oncogenic driver mutations, are often more widely shared across tumor cells and, in some cases, across patients, motivating efforts to target them clinically.^{14,116,202,203} However, their therapeutic potential remains constrained by antigen processing and presentation. Driver proteins may be expressed at low levels, may exhibit long half-lives that limit proteasomal processing, or may generate peptides with poor MHC class I binding or TCR engagement.^202,205 Strategies aimed at enhancing the degradation or processing of selected proteins, for example, by pharmacologically promoting proteasomal targeting, illustrate that antigen availability it-self is a tunable variable rather than a fixed property of the genome.^206–208 Nevertheless, such approaches underscore the need for rigorous prioritization frameworks that explicitly account for immunopeptidome heterogeneity when selecting mutation-derived targets.

PRIORITIZING ANTIGENS UNDER IMMUNOPEPTIDOME HETEROGENEITY

Recognizing that only a subset of mutation-derived peptides is durably and broadly presented across heterogeneous tumor cell populations has motivated the development of antigen prioritization strategies that move beyond genomic prediction alone.

Along these lines, Niknafs et al.²⁰⁹ introduced the concept of “persistent TMB” (pTMB), emphasizing that mutations that are less susceptible to being lost during tumor evolution may represent more stable, actionable immunotherapeutic targets. This perspective highlights that the clinical relevance of mutational burden depends not only on how many mutations are present but also on their persistence within the tumor genome.

In addition, mutational processes can bias which codons are preferentially altered and thus which amino acid substitutions are enriched, shaping the resulting neoantigen repertoire in an HLA-dependent manner (i.e., the same substitution landscape can yield different predicted neoantigen burdens across HLA backgrounds). Consistent with this concept, recent work proposes that many nucleotide-level mutational signatures collapse into a small number of dominant amino acid substitution signatures with distinct predicted immunogenic potential.²¹⁰

Data-driven approaches integrating immunopeptidomics with genomic and transcriptomic data have demonstrated that direct measurement of peptide presentation provides critical information that cannot be inferred from sequence data alone, including peptide abundance, allele-specific presentation biases, and the functional integrity of the antigen processing and presentation machinery.^{189,191,205,211} Across patient cohorts, such data-driven strategies can also reveal recurrent, naturally presented epitopes from common driver mutations, an important complement to personalized targeting when ITH limits the durability of passenger-derived neoantigens.²¹¹

In the context of ITH, such integrative strategies are particularly valuable, as they enrich for antigens that persist across divergent tumor cell states and microenvironmental contexts.

Importantly, prioritization frameworks that incorporate empirical evidence of peptide presentation implicitly filter against antigens that are lost, diluted, or rendered invisible as tumors evolve.^{204,212–214} This is especially relevant in tumors where immune pressure or therapy induces adaptive remodeling of antigen-processing pathways, leading to spatially restricted or temporally transient antigen display. By focusing on peptides that are reproducibly detected at the tumor cell surface, these data-driven strategies help identify targets that are more likely to support sustained immune recognition despite ongoing tumor evolution.¹¹⁶ However, even highly ranked antigens identified in this manner should not be assumed to be permanently stable, as antigen presentation itself remains subject to modulation by changes in cell state, inflammation, and therapeutic intervention. Importantly, all such prioritization efforts presuppose intact antigen processing and presentation machinery. Tumors that acquire HLA loss of heterozygosity or other defects in the presentation pathway may selectively abrogate neoantigen display, rendering even genomically persistent mutations immunologically silent.

EVOLVED CLONALITY AS AN OPPORTUNITY FOR ANTIGEN TARGETING

While many mutations are subclonal from their inception, tumor evolution can also give rise to a subset of targets characterized by acquired clonality. Under selective pressures imposed by therapy, immune surveillance, or microenvironmental constraints, specific genetic alterations may become enriched and recurrently observed across substantial fractions of the tumor mass. In this setting, mutations that were initially rare can, in some cases, become broadly shared across tumor clones and metastatic lesions, suggesting a potential window of opportunity for antigen-directed interventions.^18,215–217

Crucially, the potential relevance of such targets lies not in their mutational origin but in their apparent enrichment through functional selection across heterogeneous tumor cell populations. When viewed through the lens of ITH, such enrichment may arise because resistance or late-stage progression repeatedly favors similar functional solutions across otherwise distinct subclones, thereby, in some cases, increasing the proportion of tumor cells carrying the same potentially actionable alteration rather than implying full clonal fixation. This conceptual framing highlights the potential utility of targeting antigens associated with late, functionally selected tumor states, particularly in clinical settings where early truncal neoantigens are absent, weakly immunogenic, or no longer presented.

Nevertheless, evolved clonality should be considered as a dynamic and context-dependent property, rather than a permanent one. Continued tumor adaptation may eventually diminish antigen presentation through loss of expression, alterations in processing, or selection of antigen-negative variants.^{12,216,218–220} Thus, even antigens arising from putatively convergent evolution processes must be evaluated cautiously within the broader context of immunopeptidome plasticity and temporal heterogeneity.^221–223

NON-GENOMICALLY ENCODED ANTIGENS MAY BYPASS GENETIC HETEROGENEITY

An alternative strategy to mitigate the constraints imposed by genetic ITH is to target antigenic peptides that do not arise directly from DNA sequence alterations. Tumor cells frequently exhibit dysregulation at the levels of RNA processing, translation, and protein quality control, leading to the generation of aberrant peptides that can be perceived as non-self by the immune system. These include peptides derived from alternative splicing, translational frameshifts, altered decoding fidelity, and noncanonical proteolytic processing.^{185,192,198,224–226} For example, beyond the previously described suppressive effects on T cells, IFNγ-induced IDO1-mediated tryptophan depletion can trigger translational recoding, including ribosomal frameshifting with the generation of trans-frame peptides and promoting tryptophan-to-phenylalanine (W>F) amino acid substitution that yields “substitutant” peptides, thereby directly linking this metabolic axis to non-genomically encoded antigen generation and presentation.^192,197 Furthermore, therapies that perturb RNA or protein quality control pathways (for example, nonsense-mediated decay or chaperone/proteostasis networks) could further amplify the generation and display of such aberrant peptides, effectively adding a pharmacologically “programmable” antigenic layer onto genetically heterogeneous tumors.^193,227

A key conceptual advantage of such antigens is that their generation can be driven by shared cellular states—such as inflammatory stress, metabolic imbalance, or pharmacologic perturbation—rather than by genetic heterogeneity. As a result, non-genomically encoded peptides may be produced across large fractions of tumor cells, effectively creating antigenic clonality at the level of presentation despite underlying genetic diversity.^193,228,229 Moreover, because the mechanisms that generate these peptides are often state-dependent and mechanistically constrained, the resulting antigenic repertoires may be predictable and, in some cases, deliberately inducible.

However, rigorous demonstration of clonality for non-genomically encoded antigens remains technically challenging. Establishing uniform expression and presentation, especially for endogenous peptides, will likely require single-cell or spatially resolved immunopeptidomic approaches, which are still in their infancy. Moreover, their state dependence raises concerns about transient expression and variable pMHC class I display. Whether these mechanisms preferentially yield “high-quality” targets (i.e., sufficiently abundant, stably displayed, and capable of eliciting potent T cell responses) or whether many will be too context-restricted to validate and exploit therapeutically remains an open question. A related hurdle is cancer specificity. Because similar stress-adaptive programs can occur in non-malignant inflamed or perturbed tissues, it may be difficult to pinpoint truly cancer-selective noncanonical peptides and distinguish them from broadly stress-induced peptides when prioritizing therapeutic targets.

Additional non-genomic sources of antigenic diversity may include microbial-derived peptides presented by tumor HLA molecules, as studies have demonstrated that tumor cells can present peptides derived from tumor-associated or intracellular microbes.^11,230–234 These observations raise the intriguing possibility that the tumor-associated microbiome may contribute a distinct, non-human antigenic layer to tumor immune recognition; however, the extent to which such microbial antigens are consistently presented across tumor regions or stable over time remains an important open question.

Nonetheless, the conceptual framework highlights how manipulating tumor cell states, rather than genomes, can reshape immune recognition in the face of extensive ITH.

ANTIGEN ABUNDANCE AS A QUANTITATIVE AXIS OF IMMUNOPEPTIDOME HETEROGENEITY

Beyond differences in antigen identity, ITH also manifests as variation in the amount of antigen presented at the tumor cell surface. For a given antigen, pMHC abundance can vary substantially across tumor regions and over time.²⁰⁵ Indeed, antigen abundance is influenced by multiple cell-intrinsic parameters, including transcriptional output, translational efficiency, protein stability, and the activity of the antigen processing and presentation machinery. As these parameters are themselves subject to nongenetic plasticity, antigen quantity represents an independent and highly dynamic dimension of immunopeptidome heterogeneity.

Quantitative differences in antigen presentation have important functional consequences for T cell responses. Low levels of pMHC class I complexes may fall below the threshold required for effective priming or sustained effector function, particularly in chronically inflamed tumor environments, consistent with the rate-limited nature of DC-mediated cross-presentation and priming discussed above. Under such conditions, T cells can enter dysfunctional or exhausted states despite the presence of clonally expressed antigens, contributing to immune failure without complete loss of antigenicity.²¹⁴ This phenomenon provides a mechanistic explanation for why tumors can remain poorly controlled even when targetable antigens are theoretically present across all tumor cells.

Therapeutic strategies that modulate antigen abundance therefore represent an additional avenue to reshape immune recognition in heterogeneous tumors. Approaches aimed at selectively increasing the degradation of specific proteins can enhance the presentation of otherwise weakly displayed antigens, while more global interventions, such as IFNγ signaling, selected chemotherapies, radiation, oncolytic viruses, or epigenetic and cell-cycle-targeting drugs, can increase HLA/immunoproteasome function and reshape the displayed repertoire.^{192,196,197,235–237} Notably, such interventions may simultaneously increase antigen quantity and alter antigen quality, further amplifying heterogeneity in immune recognition across tumor cell populations. Together, these observations underscore that effective antigen-directed immunotherapy must account not only for which antigens are presented, but also for how much of each antigen is displayed in different tumor cell states.

PREDICTING RESPONSE TO IMMUNOTHERAPY USING ITH-AWARE AI FRAMEWORKS

Although ICB therapies have revolutionized cancer treatment, response rates remain highly variable, with only 20%–40% of patients achieving durable clinical benefit across most tumor types. This wide range in clinical outcomes reflects the complex interplay between tumor clonal architecture, nongenetic variations, and the TME, including the immune compartment.

Attempts to develop biomarkers of response, such as PD-L1 expression and TMB, led to only modest predictive power, underscoring the multifactorial and context-dependent nature of response to ICB. In fact, increasing evidence indicates that—beside cancer cells themselves—almost every cell type present within the TME can contribute to an immunosuppressive environment. Consequently, to capture this complex scenario, hundreds of machine learning and artificial intelligence (AI)-based models have been developed to predict response using diverse cellular, molecular, and clinical inputs, a rapidly expanding literature that has been comprehensively reviewed else-where.²³⁸ Here, we focus specifically on models that leverage descriptors of ITH and the organization of the TME to inform response prediction to ICB.

As discussed above, ITH and tumor clonal structure have profound implications for therapeutic resistance, as minor resistant subclones can rapidly expand under immune or pharmacologic selective pressure. Although tumors with high mutational burden tend to generate more neoantigens and may therefore be more immunogenic, extensive clonal heterogeneity can promote immunoediting and selection of immune-evasive variants. Moreover, spatial heterogeneity in neoantigen distribution may limit effective T cell recognition, often coinciding with regional differences in immune infiltration and checkpoint molecule expression. Together, these features critically shape the immune state of the TME and modulate ICB efficacy.

The advent of single-cell RNA-seq has enabled direct interrogation of tumor and immune heterogeneity at cellular resolution. AI-driven clustering and trajectory inference (e.g., single-cell variational inference with partition-based graph abstraction [scVI/PAGA] with RNA velocity and fate mapping)^239,240 can resolve transitional cell states that link microenvironmental niches to evolving immune phenotypes and clonal dynamics. Emerging spatial transcriptomics and multiplexed imaging technologies now allow mapping of the clonal composition and immune architecture of the TME at an unprecedented scale. These platforms generate highly complex datasets that call for sophisticated computational tools to identify cellular neighborhoods with distinct molecular profiles and to quantify spatial relationships between tumor clones and immune cell populations.¹⁸³ Spatial analyses have consistently demonstrated that the physical organization of tumor subclones and their proximity to immune cells strongly influence response, with immune-excluded phenotypes associated with particularly poor clinical outcome.^241,242 Collectively, these compendia of datasets provide fertile ground for AI-based predictive modeling that moves beyond traditional bulk bioinformatics analyses. However, many of these high-resolution profiling approaches remain costly and technically demanding, limiting their immediate scalability in routine oncology practice. By contrast, histology-based deep-learning models derived from standard hematoxylin and eosin (H&E) slides offer greater potential for broad clinical implementation, as they leverage already available pathology workflows. Accordingly, future efforts may benefit from parallel development of mechanistically rich discovery models and cost-efficient deployment models.

Recent advances in response prediction have increasingly focused on two main determinants emphasized throughout this review: (1) tumor clonal architecture and (2) the cellular and spatial composition of the TME. Clonality-aware frameworks have refined ICB prediction beyond the FDA-approved TMB biomarker by explicitly modeling ITH and tumor evolutionary structure. For example, NeoPrecis integrates neoantigen immunogenicity with tumor phylogenetic features and demonstrates significant associations with durable clinical benefit, achieving 11%–20% relative improvements in area under the receiver operating characteristic curve (AUC) compared with TMB alone in melanoma and NSCLC cohorts.²⁴³ Similarly, analyses of clonal versus subclonal neoantigen burden and expression have robustly stratified responders and non-responders across independent cohorts, establishing clonal neoantigen burden as a dominant predictive feature. Notably, findings initially derived from mouse models were shown to extend to human cancer patients.¹⁴⁴

More recently, a multi-modal genomic model reported that high ITH, whole-genome ploidy alterations, and extensive subclonal copy-number changes are associated with markedly shorter progression-free and overall survival. Notably, three of seven patients predicted to be intrinsically resistant to anti-PD-1 monotherapy responded to an immune checkpoint combination therapy, an encouraging but preliminary observation given the limited cohort size.²⁴⁴ Together, these studies highlight the need for refined, spatially informed measures of immune competence in tumors characterized by highly fragmented clonal landscapes.

A complementary line of research has focused on how TME architecture modulates response. Graph-based representations, where individual cells are modeled as nodes connected by spatial proximity edges, have proven particularly effective in identifying clinically relevant TME configurations. Such analyses have defined tumor-immune interfaces, immune-excluded regions, and immune-desert phenotypes that are strongly associated with patient outcomes and response.²⁴⁵

Building on these concepts, recent TME-based multifactorial predictors of response have been developed. One study combining imaging mass cytometry and spatial transcriptomics in lung cancer identified a resistance-associated cell-type signature from spatial proteomics and a complementary cell-to-gene resistance signature from spatial transcriptomics, both of which achieved significant hazard ratios across three independent patient cohorts.²⁴⁶ In another study, joint modeling of spatial transcriptomics and proteomics in hepatocellular carcinoma identified extracellular matrix-rich resistance niches that exclude immune cells. Using a graph neural network framework, the authors reported an outstanding response prediction performance,²⁴⁷ which, if further validated independently, would suggest substantial promise for spatially explicit AI models. Collectively, these findings demonstrate that AI frameworks that integrate evolutionary and spatial dimensions may indeed advance mechanism-informed immunotherapy biomarkers. Importantly, AI models developed in this space can be broadly divided into two categories: (1) discovery-oriented frameworks aimed at uncovering biological determinants of response and generating mechanistic hypotheses and (2) clinically oriented predictors intended for deployment in real-world decision-making. While the former often leverages high-dimensional multi-omics data to dissect ITH and tumor-immune interactions, the latter must satisfy stricter requirements regarding robustness, scalability, interpretability, and regulatory validation.

While experimental approaches for high-resolution TME characterization have been transformative, they remain costly and labor-intensive, restricting their application so far to small cohorts. This limitation hampers large-scale association studies linking spatial ITH and TME architecture to clinical outcomes. Recent advances in AI-driven digital pathology have begun to address this gap by inferring ITH and TME organization directly from routine H&E-stained histopathology slides. This approach leverages widely available clinical specimens, avoids the need for molecular profiling, and enables retrospective analysis of large pathology archives accumulated over decades of clinical practice. In sum, the ability to use such methods to study large patient cohorts in a ready manner has transformative potential.

One example is HistoTME, a weakly supervised deep-learning framework that infers cell-type-specific TME molecular signatures directly from whole-slide images in NSCLC. In an external cohort of 652 first-line ICI-treated patients, inferred microenvironment signatures predicted treatment response with an AUC of 0.75.²⁴⁸ Another strategy first infers tumor transcriptomic profiles directly from H&E images and then applies established transcriptomics-based response frameworks to the inferred data. In the recent ENLIGHT-DeepPT study, this approach achieved an overall odds ratio of 2.28 for identifying true responders across five independent clinical cohorts spanning multiple treatments and cancer types, without training on these test cohorts.²⁴⁹ Deep-learning convolutional neural networks have also demonstrated that histology and genomic data can be jointly modeled to develop integrated image-omic prognostic frameworks across cancer types.²⁵⁰

Taken together, these and numerous other studies highlight a broader trend: deep-learning analysis of H&E-stained slides has yielded notable advances in ICB response prediction across numerous cancer types, with reporting AUC values around ~0.75, outperforming conventional biomarkers (e.g., for a recent one see Rakaee et al.²⁵¹). Further integration of multi-omics data (with or without digital pathology) using ensemble learning and deep neural networks may yield even more powerful predictive models.²⁵² Importantly, these developments offer an exciting new avenue to democratize access to high-resolution tumor characterization and AI-based biomarkers, particularly in under-resourced regions and low- and middle-income countries.²⁵³

Beyond prediction, AI approaches can provide mechanistic insights into the biological determinants of response and resistance to ICB. Interpretable machine learning strategies, including feature attribution methods such as Shapley additive explanations (SHAP) values and attention-based architectures, have begun to connect model predictions with specific genomic, spatial, or immune features, thereby facilitating biological validation and hypothesis generation.^254–256 While transformer-based and other deep architectures can capture complex, non-linear interactions within heterogeneous datasets, their added value for mechanistic interpretation remains an area of active investigation.

In the context of antigen heterogeneity discussed above, AI-based frameworks are increasingly being applied to neoantigen prioritization and vaccine design. Machine learning models integrating peptide-MHC binding affinity, antigen-processing features, clonality, expression levels, and immunogenicity metrics can rank candidate neoantigens more effectively than sequence-based approaches alone.^189,204,212 In tumors characterized by high ITH, such models may assist in identifying antigens with persistent clonal representation or predicting combinations of subclonal targets that minimize immune escape. As personalized mRNA and peptide vaccine strategies advance clinically, ITH-aware computational antigen selection may become an essential component of rational vaccine design.

Several challenges remain. Beyond predictive performance, translation of ITH-aware AI models into clinical practice requires demonstration of added value over existing biomarkers such as PD-L1 staining, microsatellite instability (MSI) status, and TMB. For regulatory adoption, models must show consistent performance across independent, multi-institutional cohorts, maintain stability under variations in sample processing, and ideally achieve clinically meaningful improvements in discrimination or treatment stratification. Importantly, while metrics such as AUC are widely used in academic studies, modest improvements in AUC alone are unlikely to justify clinical adoption unless they translate into actionable patient reclassification or therapeutic decision changes. Greater emphasis should therefore be placed on clinically interpretable measures (including odds ratios, hazard ratios for progression-free or overall survival, positive and negative predictive values, and precision-recall characteristics) that more directly reflect treatment benefit and risk in real-world oncology practice. Against this backdrop, several barriers must be addressed: (1) model interpretability continues to lag behind predictive performance, limiting clinical acceptance and calling for the development of more explainable AI frameworks; (2) rigorous prospective validation across diverse, multi-institutional cohorts representing varied ancestries, healthcare systems, and real-world clinical workflows is essential to ensure generalizability; (3) the computational infrastructure needed to train and deploy AI models remains substantial and is often unavailable in resource-limited settings; (4) standardization of data acquisition, preprocessing, and quality control pipelines is critical for clinical deployment; and (5) legal and ethical considerations, including algorithmic fairness, data privacy, governance of prospective patient data collection, and equitable access, must be addressed.

Thus, by integrating multi-scale biological data spanning genomics, transcriptomics, spatial organization, and histopathology, AI models are advancing both the mechanistic understanding of the TME and the prediction of ICB outcomes. The recent advances in digital pathology are particularly exciting. As these approaches mature and are validated in prospective clinical studies, they are likely to become an integral component of precision oncology, laying the basis for more comprehensive and effective precision treatments.

CONCLUDING REMARKS: TOWARD AN INTEGRATED FRAMEWORK OF ITH FOR NEXT-GENERATION IMMUNOTHERAPIES

Deciphering how ITH shapes tumor immunity reframes immunotherapy failure as a problem of incomplete “target footprint” and context-dependent immune suppression rather than a single missing biomarker. Across genetic, epigenetic, proteomic, metabolic, and spatial layers, heterogeneous tumors partition immune visibility and immune-evasion programs into coexisting subclones and cell states, creating immunological neighborhoods that are differentially permissive to priming, infiltration, and killing. Improving response to immunotherapy, therefore, requires strategies that (1) broaden antigenic coverage across tumor phylogenies, (2) stabilize or amplify antigen presentation in otherwise “invisible” states, and (3) disrupt neighborhood-level resistance programs that exclude or disable immune effectors. In addition to these approaches, mechanisms capable of extending immune pressure beyond the initially targeted antigen—such as immunogenic cell death, bystander killing, and epitope spreading—offer a complementary route to overcome antigenic or spatial barriers by enabling indirect or propagated killing of tumor cells that were not directly recognized by the initiating immune response. Practically, this argues for ITH-aware target selection (prioritizing empirically presented, persistent, and/or clonal antigens when feasible), multi-epitope approaches designed to cover major branches rather than a single dominant target, and deliberate pairing of ICB with interventions that remodel antigen processing/presentation and immune access (e.g., temporally controlled use of radiation, selected chemotherapies, interferon-pathway modulation, epigenetic agents, and microenvironment-modifying therapies that address exclusionary programs such as TGF-β/VEGF-driven states). Because nongenetic plasticity rapidly generates DTP reservoirs and MRD, combination strategies should not merely “add a second drug” but aim to shape trajectories—prevent entry into tolerant attractors, shorten dwell time in resistant states, or exploit liabilities of slow-cycling/stress-adapted programs (hypoxia adaptation, oxidative stress buffering, autophagy dependence, and lipid metabolism). A central implication of this review is that the immunopeptidome is the functional interface through which these layers become actionable for immunity, and interventions that increase the abundance, breadth, or stability of presented peptides may be as important as expanding functional T cell numbers.

Turning these principles into reliable therapeutic design will require better experimental and computational models that explicitly represent defined clonal/antigenic architectures, state plasticity, and spatial structure. Well-controlled in vivo systems—such as clone-mix tumors with known phylogenies and engineered, quantifiable antigen landscapes—are essential to test how antigen clonality, antigen abundance, and neighborhood context interact to determine priming efficiency, effector function, epitope spreading, and escape. These models should be paired with longitudinal, spatially resolved measurements (multi-region genomics, single-cell and spatial transcriptomics/proteomics, TCR tracking, and—where possible—direct immunopeptidomics) to connect mechanistic perturbations to shifting pMHC landscapes and immune dynamics over time. In parallel, ITH-aware AI/ML frameworks can help unify these data, but their translational value will depend on whether they deliver robust, externally validated patient reclassification beyond established markers (MSI/MMRd, PD-L1, and TMB) and whether model features map onto testable mechanisms that can be therapeutically manipulated.

Key knowledge gaps now define an actionable agenda for the field. First, what degree and type of heterogeneity (genetic versus state driven, antigen identity versus antigen abundance, and spatial versus temporal) most strongly limits durable immune control, and in which tumor contexts? Second, how should vaccines and T cell-based therapies be composed under ITH: should they prioritize clonal targets, deliberately cover phylogenetic branches with selected subclonal targets, or leverage state-dependent/non-genomically encoded peptides to create “presentation-level clonality”—and what are the rules that govern escape under each strategy? Third, how do therapies intended to suppress plasticity (epigenetic, stress-pathway, and metabolic interventions) reshape antigen processing/presentation and DC-mediated cross-presentation in vivo, and when should they be scheduled relative to checkpoint blockade to maximize priming and prevent exhaustion? Fourth, how can we measure immunopeptidome heterogeneity and neighborhood structure in situ at sufficient resolution to guide treatment—ideally with scalable assays suitable for clinical deployment? Finally, can we prospectively predict and then pre-empt evolutionary routes of immune escape (HLA loss, antigen-processing defects, and neighborhood-driven exclusion) using integrated longitudinal sampling and mechanistically constrained models? Addressing these questions will be essential to move from descriptive heterogeneity to rational immunotherapy design that anticipates tumor diversity, expands effective antigenic coverage, and delivers more durable responses across patients.