Modeling human tumor evolution in mice

Summary

Cancer develops through an evolutionary process driven by mutation and selective pressures imposed by the microenvironment. To be translationally relevant, experimental models must recapitulate these clinical complexities. In this Primer, we consider how different mouse-modeling strategies can be aligned with human tumor evolution, focusing on mutation order, the emergence and competition of initially rare mutant clones within genetically mosaic tissues, and microenvironmental context. By comparing conventional genetically engineered mouse models (GEMMs), transplant-based models, and emerging tools such as somatic editing and the tandem arrayed regulator (TAR) allele system, we offer a guide for selecting the most appropriate mouse model to address specific evolutionary questions.

Graphical abstract

Highlights

• •Evaluates mouse models based on how faithfully they mirror human tumour evolutionary trajectories.
• •Defines key principles for 'evolution-aligned' modeling, including precise temporal control and cell-of-origin specificity.
• •Offers a strategic guide for selecting the most appropriate mouse models to answer specific questions for in vivo cancer studies.

This Primer explores the capabilities and limitations of mouse models to study tumor evolution and introduces how emerging technologies can overcome limitations in modeling precise temporal control and cell-type-specificity.

Introduction

Cancer is an evolutionary disease. Most adult tumors are clonally derived from a common ancestral cell, but their development reflects a dynamic competition among initiated clones, shaped by years of tissue-specific selective pressures. Factors such as tissue architecture, including stem-cell niche organization, cellular competition and cooperation, and extracellular matrix composition; immune surveillance, aging, and environmental exposure dictate the selective landscape, driving the formation of complex clonal structures. By the time of clinical detection, this process has typically unfolded over decades, resulting in intratumoral heterogeneity which then complicates treatment^1,2 (Figure 1A).

Figure 1.

Timescales of tumor evolution in humans and genetically engineered mouse models (GEMMs)

(A) Human tumors typically emerge over years to decades, a period in which cells experience fluctuating selective pressures, including inflammatory signals, stromal remodeling, metabolic stress, immune surveillance, and age-related changes; all of which shape the tumor ecosystem. Tumorigenesis proceeds through sequential stages including an initiating mutation, clonal expansion, and subclonal diversification (e.g., acquisition of KRAS activation or loss of TP53), ultimately resulting in a heterogenous tumor.

(B) Conventional GEMMs develop tumors over days to weeks following rapid induction of driver mutations (e.g., Cre-mediated activation or deletion of Apc, Kras or Trp53). This accelerated and often simultaneous genetic alteration enables efficient and reproducible tumor formation, while compressing the timeframe for gradual clonal evolution and prolonged exposure to diverse selective pressures.

(C) Evolution-aligned temporal GEMMs use low-frequency recombination to introduce driver mutations stepwise, enabling clonal expansion and competition under ongoing selective pressures, and ultimately generating heterogeneous tumors.

Experimental model systems are therefore essential for dissecting how randomly acquired mutations interact with selective pressures during tumor evolution, in a controlled manner and for studying how tumors change over time through sequential analyses at defined stages of disease progression. Mouse models, and particularly genetically engineered mouse models (GEMMs), have transformed our understanding of oncogenic signaling, tumor suppressor function, and therapeutic vulnerabilities.³ For example, incorporating combinations of driver alterations identified through human tumor sequencing (such as APC, KRAS, TP53 and NOTCH in colorectal cancer, or KRAS, TP53 and SMAD4 in pancreatic cancer) has enabled investigators to define how clinically relevant mutations cooperate to initiate tumors, shape disease progression, influence tumor phenotype and uncover genotype-specific therapeutic opportunities. While these models may not always replicate the same evolutionary and environmental trajectories through which those genotypes arise in human disease (e.g., exogenous insults, aging, systemic inflammation), GEMMs remain effective tools for studies investigating tumor initiation and progression in vivo (Figure 1).

In this Primer, we aim to provide an overview of how selected mouse models can provide a powerful platform for studying aspects of tumor development and evolutionary processes in vivo, for some cancers. In particular, genetic tractability allows specific biological hypotheses to be tested under controlled conditions. Nonetheless, no single modeling strategy fully captures the continuum of tumor development, and each imposes distinct constraints on evolutionary dynamics (summarized in Table 1). We also introduce a novel Tandem Arrayed Regulator (TAR) allele system, which we are developing and validating as part of the Medical Research Council (MRC)- Funded National Mouse Genetics Network (NMGN) (https://nmgn.mrc.ukri.org/clusters/cancer). This allele system is designed to offer finer temporal and pathway control to model multistep carcinogenesis.

Table 1.

Comparing model strengths and limitations

Model type	Strengths	Limitations	Best used for
Conventional GEMM	•Highly reproducible across cohorts. •Tumors emerge from the correct lineage & developmental timepoints. •Preserves the native microenvironment & intact immune context.	•Mutations are introduced in many cells simultaneously, limiting clonal competition. •Limited temporal control over when mutations occur. •Less suited for modeling late-stage evolution & metastasis without additional hits. •Breeding multiple alleles is slow & expensive.	•Studying tumor initiation & early tumorigenesis. •Mechanistic studies of oncogenic pathways. •Investigating interactions with the immune microenvironment.
Cell/Organoid Transplant	•Rapid generation of experimental cohorts. •Compatible with ex vivo CRISPR editing. •Maintains 3D architecture. •Intact immune system in syngeneic hosts.	•Transplantation bypasses natural initiation & early bottlenecks. •Does not capture early evolution. •Microenvironment can be unnatural, particularly when placed subcutaneously.	•Defining mutation order, clonal competition & evolutionary dynamics. •Modeling late-stage progression & metastatic behavior (in orthotopic settings). •Examining immune-tumour interactions, therapy response & drug resistance (in syngeneic setting).
Patient-Derived Xenografts (PDXs)	•Maintains the native 3D architecture & cellular heterogeneity of the patient tumor. •Preserves subclonal structure & captures patient-specific evolution.	•Requires immunodeficient hosts, limiting immune-related studies. •Does not model tumour initiation or early stages of evolution. •Resource-intensive & slow to establish.	•Personalized medicine, preclinical drug testing & resistance studies. •Biomarker discovery & translational applications. •Studying selection pressures under therapy.
Experimental Metastasis(e.g., intrasplenic, tail vein)	•Ensures synchronous seeding across cohorts. •Produces rapid, reliable colonization in secondary organs. •Technically simpler than orthotopic surgery & scales well.	•Bypasses primary tumor formation, local invasion & early dissemination steps. •Does not model spontaneous metastasis or full evolutionary trajectory.	•Studying organotropism & metastatic colonization. •Investigating the effects of drugs that target late metastatic steps. •Mechanistic studies of extravasation, dormancy, & niche formation.
Sequential (TAR) Allele	•Multi-modal temporal control allows independent genetic deletion, transcriptional activation & protein degradation. •Flexible sequential modeling with stepwise mutation accumulation. •A single polycistronic locus reduces complex breeding.	•Requires optimisation of dosing & administration routes for the different inducing agents. •High logistical complexity, requires precise coordination of three distinct systemic inducers.	•Mapping evolutionary dependencies & dissecting mutation order. •Defining “points of no return”, testing reversibility of oncogenic signaling at different stages of tumor development. •Longitudinal modeling from early-stage initiation to late-stage progression within the same model.

How traditional GEMMs fit in the evolutionary framework

If we prioritize evolutionary accuracy, the standard GEMM framework needs to be reconsidered. For over 45 years, genetic mouse models have underpinned cancer research and have been essential for defining mechanisms of carcinogenesis. A historical overview is not within the scope of this Primer but has been elegantly and extensively described elsewhere.⁴ In brief, the earliest GEMMs were simplified, single oncogene drivers or deletion of common tumor suppressor genes (e.g., Emu-MYC, MMTV-PyMT, Trp53^−/−). Whilst these models were essential in proving the necessity for second hits and identifying permissible cell types, they often fail to recapitulate human tumor histopathology and lack precise control over the targeted cell population, although some continue to remain valuable in the mouse models armory.

Early conceptual models, including the adenoma-to-carcinoma sequence described by Bert Vogelstein, laid the foundation for the classic view that tumorigenesis progresses through an ordered accumulation of defined genetic lesions.⁵ This includes sequential alterations in APC, KRAS, TP53 and additional pathways that link genetic events to histopathological stages. While we now know that this linearity is not universal across tumor types, this model established several core principles, namely, that tumors advance through discrete stages where clonal expansions are driven by fitness-enhancing mutations, and additional alterations accumulate on pre-existing mutated backgrounds. However, this “stepwise” dogma is increasingly challenged by data showing substantial inter and intra tumor heterogeneity, variable mutation orders and branched evolutionary trajectories (Figure 1).

Recent genomic analyses have revealed a far more chaotic reality where mutations do not always follow a predictable, linear path. For example, the “Big Bang” model proposes that an early burst of diversification can generate substantial heterogeneity which may be established early and then followed by relative evolutionary stasis.⁶ Rather than representing mutually exclusive frameworks, such models define a spectrum of evolutionary dynamics, ranging from Vogelstein’s stepwise progression to the “Big Bang” model’s rapid, early diversification. Most cancers fall somewhere in between, where periods of selection are interrupted by stretches of neutral drift. Ultimately, these models underscore that tumor development is dynamic, temporally structured, and shaped by both selection and neutral drift, with important implications on how it can be modeled experimentally. Replicating these unpredictable evolutionary patterns in vivo requires tools that let investigators dictate when and in what order mutations occur. This challenge is addressed by emerging multi-modal systems like the TAR allele, discussed below. The ability to stretch or compress the timeline between genetic hits allows investigators to experimentally simulate the spectrum between slow, stepwise progression and rapid, multi-hit clonal bursts.

Evolving understanding of human cancer has highlighted subtypes within cancers, and therefore models that represent this spectrum are required. Modern complex GEMMs address this by using conditional alleles with expression controlled by methods such as promoter driven recombinases, inducible elements, and viral delivery of recombinases (Figure 2), enabling mutations to be activated in defined cell populations and at specific timepoints in the lifespan of the animal. This spatial and temporal control helps overcome limitations of earlier GEMMs in which mutations were activated broadly or constitutively, often with non-target cell phenotypes, allowing investigators to more accurately model distinct tumor subtypes, control mutation activation and better mimic the timing of tumor initiation. These models have a distinct advantage whereby the tumors spontaneously arise in the correct tissue within a competent immune system and microenvironment and often represent the full spectrum of disease, from initiation to metastasis. Alleles can be engineered to recapitulate human-relevant genetic events, such as conditional knock-in of oncogenic Kras (e.g., LSL-Kras^G12D), and combining these alleles (through breeding programs) in complex GEMMs allows us to dissect how tumors develop, evolve and depend on certain signaling pathways.⁹

Figure 2.

Inducible tools for temporal and conditional control of gene expression and protein stability

(A) Recombinases can be expressed in a promoter specific manner or packaged and delivered to the cell by a tropic virus. Recombinase can be fused with a ligand binding domain from the Estrogen receptor (ER) and upon binding to 4-hydroxy-tamoxifen translocate to the nucleus. Recombinase recognition sites (RRS) can be bound and cleaved by recombinases; with this strategy used to disrupt/delete target genes or to activate expression of modified genes preceded by an RRS-STOP-RRS translational stop cassette.

(B) rtTA3 is a third generation modified reverse tetracycline-inducible transactivator which, in response to doxycycline, binds specifically to Tet-response elements (TREs) engineered into the promoter of desired target genes to drive expression in a dose-dependent manner.⁷

(C) OsTir1F74G (TIR1) is an auxin-inducible E3 ligase that recognizes synthetically engineered degron motifs (AID; Auxin Inducible Degron) on target proteins, driving rapid and highly specific protein degradation upon administration of the auxin analogue, 5-phenyl-1H-indole-3-acetic acid (IAA).⁸

Accurate genetic models of colorectal cancer

The adenoma-to-carcinoma model postulates that cancers evolve through sequential acquisition of genetic events that advance disease, epitomized in colorectal cancer (CRC).⁵ GEMMs of CRC comprehensively cover many of the major genetic drivers of human disease, and a suite of models has been shown to align with human tumors across the spectrum of Consensus Molecular Subtypes (CMS) used to stratify CRC subtypes.^10,11 The CMS framework categorizes CRC into four groups based on distinct biological and transcriptional profiles, CMS1 (immune-activated/microsatellite unstable), CMS2 (canonical/WNT-driven), CMS3 (metabolic/KRAS-mutant), and CMS4 (mesenchymal/highly stromal). By introducing combinations of clinically relevant driver mutations, GEMMs have helped define how specific genetic alterations influence tumor subtype, disease progression, and biological behavior. For example, Apc loss activates WNT signaling and drives adenoma formation, recapitulating key features of canonical CMS2 disease, while Kras activation promotes progression and metabolic reprogramming associated with CMS3 tumors. Additional alterations such as Trp53 loss, impacts disease progression. A disruption of TGF-beta signaling accelerates progression and increases aggressiveness, contributing to features associated with advanced and mesenchymal-like disease states. These models provide a tractable system for linking genotype to phenotype and investigating how layering combinations of driver mutations contribute to clinically relevant disease trajectories.

Tumor phenotypes can be further refined with models of activated Kras and Trp53 deletion developing fewer but more invasive tumors, while inclusion of an activated Notch allele drives a serrated morphology and metastatic dissemination, reflecting aggressive serrated CRC subtypes and poor prognosis molecular subgroups observed in some patients.¹² These models predominantly give rise to small intestinal tumors, reflecting differences in the murine intestinal biology and the distribution of recombinase activity. To mitigate this anatomical limitation, local gene induction in the distal colon (injection of 4-Hydroxy-Tamoxifen into colon sub-mucosa to activate recombinases) or colon-specific Cre drivers can be used to generate colonic tumors.¹³ Although these approaches involve experimental manipulation; they have circumvented a technical limitation with previous tool availability and hence improved anatomical concordance with human CRC. A key advancement has also been the validation of selecting appropriate driver mutation such as combining activating Braf mutations with deletion of Alk5,¹⁴ and loss of Trp53¹⁵ which has recapitulated several key features of right-sided CRC, including proximal tumor location and molecular characteristics associated with BRAF-mutant disease,¹⁶ accurately modelling an important clinical subtype.

Alignment with human disease continues to feed back into model development. Pathway-derived subtyping highlights that although two human subtypes are well recapitulated across several CRC GEMMs, a third subtype remains underrepresented, underscoring the need to develop or refine GEMMs that capture this biology for mechanistic and therapeutic studies.¹⁷ Comprehensive phenotyping of models further facilitates mechanistic discovery, enabling targetable vulnerabilities identified in human data sets to be genetically validated across CRC GEMMs.¹⁸ Through using well-characterized GEMMs it has been possible to unpick mechanistic aspects of clonal evolution,¹⁹ invasive disease,¹² stem cell plasticity²⁰ and therapeutic evolution,¹⁵ to enhance understanding of cancer dynamics and aspects of CRC tumor evolution.

The KPC model of pancreatic cancer

Pancreatic cancer is a particularly lethal cancer, with less than 15% patient survival at 5 years, even though the predominate genetic drivers are known.²¹ By targeting these patient-relevant mutations (e.g., KRAS activation and TP53 loss) to pancreas cells using a Pdx1-Cre driver, for instance, in GEMMs, it is possible to reliably generate pancreatic ductal adenocarcinoma (PDAC).²² This “KPC” model is now the gold standard to study the entire evolution of pancreatic cancer from early/pre-neoplastic lesions, primary tumor growth, through invasion, and eventual metastasis. Importantly, this model also exhibits tumor heterogeneity, stromal remodeling, and resistance to conventional chemotherapy, akin to the human disease setting.²³

The KPC model is an invaluable resource to study vulnerabilities in PDAC. As an example, the introduction of an inducible Kras allele confirmed the essential nature of RAS signaling in PDAC.²⁴ The use of RAS inhibitors in the KPC model further confirmed the necessity of this signaling node in PDAC,^23,25 accurately predictive of how these inhibitors have shown some promise in early clinical trials, although resistance does develop.²⁶ The dual recombinase iterations of this model, utilizing both Flp and Cre, permit temporal control by enabling oncogenic Kras activation and Trp53 deletion to be induced at separate time points. Initial Flp-recombinase-mediated activation of mutant Kras establishes precursor lesions, while subsequent tamoxifen-induced Cre-recombinase activation permits later deletion of Trp53. This sequential genetic manipulation verified that Trp53 loss, following activation of oncogenic Kras, drives progression to invasive PDAC. Additionally, recombinase expression in distinct cell lineages, has also provided a system to specifically modulate genes in the tumor microenvironment during tumorigenesis.^27,28

New patient-relevant genetic models of liver cancer

Alternative systems that do not rely on promoter-driven recombinases are available, for instance, viral delivery of vectors or CRISPR guides (Figure 2). These methods are useful when there is no pre-existing cell targeted recombinase alleles and where agents can be delivered to the tissue of interest. Their usage has been typified by a recent paper generating a suite of liver cancer GEMMs. The delivery of an adeno-associated virus (AAV8), containing a Cre recombinase expressing vector controlled by a hepatocyte specific promoter (Thyroxine-binding globulin), combined with a variety of conditional alleles, allows the study of drivers of hepatocellular carcinoma. Of all genetic combinations generated most developed tumors (e.g., Ctnnb1^(exon3)/WT;R26^{LSL-MYC/LSL-MYC}, R26^{LSL-MYC/LSL-MYC};Trp53^R172H/WT;Pten^fl/fl) with varying latencies, whilst some did not (e.g., R26^{LSL-MYC/LSL-MYC};Trp53^R172H/WT).²⁹ At a transcriptional and histopathological level, these GEMMs aligned with patient tumors, providing a basis for translating therapy into targeted patient subsets. Indeed, these GEMMs were predictive of therapy response with some, but not all cohorts showing a similar response trajectory to standard of care as is seen in patients. By recapitulating clinically aligned liver cancer models, these can extend beyond tracking tumor growth and can provide a preclinical platform for evaluating therapeutic responses in defined patient populations with a view to expediting (potentially repurposed) therapies to clinic.²⁹

These are just a few examples which highlight the utility of well-annotated GEMMs and how best they can be deployed. Importantly, they provide a strong shared resource within the scientific community ensuring the correct model can be applied to the appropriate scientific questions. A major advantage of these models is the fact that they develop in an immunocompetent host. This is exemplified by recent work using disease-positioned GEMMs to explore immunotherapeutic strategies, where both colorectal and pancreatic models have provided insight into how immune-modulating interventions engage with, and are constrained by, native tumor–immune interactions.^30,31

Modeling advanced and metastatic disease using transplantation models

Orthotopic models pertain to the implantation of tumor cells or tissue fragments into their anatomically correct location within the mouse to mimic the clinical setting of tumor growth and progression. This should be compared with the standard approach of transplanting into ectopic sites such as the subcutaneous space. While ectopic tumors easily grow and can be measured, these models generally invoke an artificial setting with an irrelevant tissue microenvironment (summarized in Table 1).

Orthotopic models can involve transplantation of human material into immuno-compromised mice, typically as cell line derived xenografts (CDX) or patient derived xenografts (PDX), or mouse material into a strain matched recipient mouse (syngeneic). CDXs offer a high throughput method to study human tumor cells. However, they lack the capacity to investigate the impact of the immune system. PDXs, from patient tumor material, preserve aspects of the original tumor though lack a complete immune system, and human stroma can be replaced by mouse stroma over serial passages. Syngeneic models preserve appropriate stromal, immune, and vascular interactions, providing a more physiologically relevant system to study tumor behavior and therapeutic response; albeit human specific targets may not translate.

Given the limitations of traditional xenograft systems (Table 1), particularly in recapitulating tumor architecture and signaling dynamics, many researchers have replaced cell suspensions with three-dimensional organoids. Organoids, generated from primary human and murine tumors, retain strong fidelity to their parental tumors at both transcriptomic and proteomic levels. Foundational work from Clevers and colleagues,³² established adult stem cell–derived organoid culture systems, which have been subsequently adapted by Tuveson, and collaborators³³ to model PDAC, showing that orthotopically transplanted neoplastic organoids can recapitulate the full spectrum of PDAC progression. Organoid based co-culture and transplantation approaches have served as a platform to study cell-to-cell crosstalk.³⁴ These systems are now widely applied for various tumor types and used both for in vitro drug screening and as defined, transplantable units for in vivo modeling.

CRC organoids have been successfully used for syngeneic orthotopic transplant studies as they engraft into immunocompetent hosts, preserving an intact tumor–immune microenvironment. Injecting CRC organoids directly into the large bowel or caecal wall retains the influence of the native intestinal microenvironment, including the local microbiota and tissue-specific stromal and immune interactions, resulting in tumor behavior that reflects human CRC biology. This approach has already been used as a tractable model to uncover new mechanisms underlying drug resistance¹⁵ and spontaneous metastasis.³⁵

While GEMMs and orthotopic models effectively reproduce primary disease, there are only a few examples which display a robust and realistic metastatic phenotype where the cancer spreads to distant organs.¹² Given that liver metastasis remains a major cause of mortality for colorectal cancer patients, the field has turned to additional models to interrogate late-stage disease, and metastatic dissemination.

Experimental metastasis models, in which tumor cells are introduced directly into the vasculature or a target organ, are widely used to study specific stages of metastatic progression. For example, intravenous injection of tumor cells predominately lead to lung colonization, while cells introduced by the intracaudal arterial route enable efficient bone metastasis modeling. Tumor cells can also be seeded directly into a number of organs which may provide a controlled way to study metastatic colonization, such as with intrasplenic injection of colorectal cancer cells which robustly produces liver metastasis.³⁶ Although this approach bypasses the early metastatic stages such as local invasion, intravasation, and survival in the systemic circulation, it does allow researchers to focus specifically on hepatic outgrowth, stromal education, and therapeutic response under highly standardized conditions.

These transplantation models have been instrumental in demonstrating plasticity of metastasis as tumors progress from early seeding to large outgrowths,¹⁵ as well as the intrinsic plasticity that drives therapy resistance.³⁷ However, because deliberate introduction of cells circumvents the initial bottlenecks and education that result from tumor–host interactions early on in the tissue of origin, the experimental metastatic models cannot recapitulate the selective pressures that shape natural metastatic competence.

Resection models, where a primary tumor is surgically removed, not only mimic the patient journey but can also be used to study minimal residual disease. Again, using CRC as a paradigm, transplantation of mouse CRC organoids into the base of the caecum, before subsequent primary tumor removal, allowed researchers to track metastatic liver recurrence over time.³⁸ Resection of primary tumors has also been extensively utilized to model spontaneous metastasis in breast cancer models.³⁹ These models have the additional advantage of being able to study the effects of neoadjuvant (pretumor removal) versus adjuvant (post-tumor removal) therapeutic intervention, revealing how the timing of chemical intervention dictates recurrence and long-term survival.³⁹

Limitations of current GEMMs and remaining gaps

Although we highlight above the many advantages of GEMMs when used appropriately (also in Table 1), developing and characterizing these complex models requires a multi-disciplinary team, large resources, and appropriate infrastructure. Significant gaps do remain. As discussed, most complex models recombine alleles in a “one-hit” manner, differing from the progressive acquisition of mutations that occur in human carcinogenesis (Figure 1).

Furthermore, evidence suggests that early initiating lesions may have distinct mutational requirements, which become less important at later stages⁴⁰ and that multi-hit models influence tumor clonal growth.⁴¹ Traditional GEMMs also often succumb to primary tumor burden before metastasis spontaneously occurs, and the difficulty of surgically removing certain primary tumors in mice, a standard treatment approach in human solid cancers, means that researchers are often forced to use different, albeit complementary models, to interrogate primary and metastatic disease. Importantly, these divergences do not invalidate GEMMs; rather, they highlight how model design influences evolutionary outcomes.

If evolutionary alignment is the goal, choosing between a traditional multi-strain layout and a consolidated platform comes down to the genetic complexity your disease model demands. These operational trade-offs, advantages, and specific use cases are outlined for each modeling strategy in Table 1. For experiments relying on a single, clean genetic switch, conventional GEMMs are still the most practical choice. However, when a study requires discrete, sequential steps, like tracking a nascent lesion as it evolves into a metastatic outgrowth, layering individual inducible lines (such as distinct recombinase (Cre/Flp) and Tet-inducible drivers) exponentially increases genetic complexity. Managing these traditional crossbreeding schemes drastically increases animal numbers, operational costs, and the overall time required to establish a single, viable experimental cohort. Consolidating these drivers into a single multi-modal allele reduces the logistical problems of stacking individual strains. This platform also provides the precise, step-by-step control needed to manipulate mutational order and delay metastatic drivers, effectively solving the primary tumor burden bottleneck and allowing researchers to capture late-stage disease biology.

Somatic modeling approaches

While germline-derived GEMMs remain essential for studying cancer initiation and progression, somatic induction strategies provide a valuable complementary approach by allowing targeted genetic alterations to be introduced with temporal and spatial precision within adult tissues. Improvement with techniques such as viral delivery, electroporation⁴² or nanoparticle-mediated transfer of recombinases or genome-editing tools has enabled tumor-initiating mutations in a restricted founder population of cells, thus introducing genetic change in a controlled manner.^43,44 These advances have made it possible to model cancer initiation in situ, capturing aspects of tissue context and timing that are sometimes difficult to access using germline systems alone.

Work from Tyler Jacks laboratory, for example, has shown that viral delivery of CRISPR/Cas9 into the adult lung can somatically disrupt multiple tumor-suppressor genes, generating lung tumors directly from edited founder cells.⁴⁵ These alterations are introduced into a limited number of cells and generate genetically mosaic tissues, in which initially rare mutant clones compete with surrounding wild-type cells, more closely mimicking early tumor evolution. Importantly, restricting these alterations to somatic founder cells keeps the surrounding tumor microenvironment (TME) wild type. In contrast, traditional GEMMs can carry the engineered alleles in a heterozygous configuration throughout the stromal and immune compartments, which may inadvertently impact on native selective pressures. However, both direct somatic delivery and recombinase-based germline systems typically permit only a single event to be triggered at a time. Accurately modeling human disease requires the ability to manipulate multiple events at different times within the same cell. This highlights the need for systems capable of imposing multi-step, precise, sequential, and reversible manipulation of multiple pathways in the cell population.

This evolutionary disconnect has practical consequences for how tumors initiate, diversify, metastasize, and respond to therapy. If we are to faithfully model human tumor biology, therefore, how a tumor evolves is just as important as the mutations it carries at endpoint. Consequently, evolutionary fidelity must become a critical design principle. The challenge is not simply to engineer the correct combination of driver alleles, but to model the selective pressures, temporal spacing, and cellular contexts that define human tumor evolution. A limitation that we propose the TAR allele can go towards addressing.

Model selection is hypothesis driven

Model selection is not a “one-size-fits all” approach and should be determined by the hypothesis being investigated. It may be that pre-existing GEMMs can be utilized for new studies; or it may be that further development of GEMMs, or alternative models may be required. Different model systems (as discussed above) capture distinct aspects of tumor biology, and therefore selection should consider the specific evolutionary stage of disease, and biological question being asked (Figure 3).

Figure 3.

Generalized framework for decision making in cancer modeling utilizing mouse models

Pragmatic selection of the appropriate model to interrogate the scientific question is required. Consideration in terms of the availability, accessibility and capability of working with the model must be factored in during the decision-making process. Therefore, this figure represents a guide to assist in experimental design. Diamonds represent questions, boxes denote applicable models/systems best suited for the investigated tumor stage, lines represent the decision flow based on Yes/No responses.

Early-stage disease

Early-stage disease is most robustly investigated in GEMMs studied at early time-points after oncogene activation on the background of normal tissue, with the capacity to compare chosen genotypes. The possibility of discovery lead approaches is now more feasible with the advancement of somatic mutation technologies (CRISPR guide screens or base-editors) which can accelerate the study of early processes, although empirical validation of the targets would be required.

Increasingly, somatic genome modeling approaches provide complementary approaches to GEMMs by enabling interrogation of candidate driver events directly within relevant tissues and cellular contexts.

Localized and advanced disease

This stage of tumorigenesis can be effectively modeled in GEMMs and somatic models, with aging immunocompetent cohorts, and the correct activation of genetic events, and allows for comparison of genotypes when appropriate alleles are combined. Whilst orthotopic transplantation of cancer cells offers good models for investigating advanced disease stages they are limited by virtue of bypassing initiating events, due to the direct delivery of competent cancer cells. Nonetheless, in vitro manipulation of cell lines provides a method to modulate the genetic landscape of tumor cells, avoiding the requirement for additional allelic crosses, particularly useful to unpick disease mechanisms. These can also be made traceable through the use of reporter constructs and in vivo imaging techniques (fluorescence/bioluminescence).

Distant metastasis

Some GEMMs are proficient in developing spontaneous and clinically relevant metastasis with time; however, each model requires extensive validation to confirm. Therefore, researchers have often leveraged more scientifically tractable models using cell lines/organoids for transplantation. Organ specific metastasis can generally be controlled by modulating injection sites, for example intra-splenic transplant of CRC cells form liver metastasis, whereas tail vein transplants generally result in lung metastasis (e.g., breast/prostate cancer cells). Arguably these transplants effectively model seeding potential of cell lines in target organs, and do not recapitulate the selective pressures of primary tumor growth and escape from the primary site. Alternatively, “inject-resect” systems can model some of the selective pressures of primary tumor growth and the evasion/escape processes required by cancer cells to seed to metastatic sites through venous and lymphatic routes.

Availability of models, expertise, and factors such as cost often dictate choice. However, an increasing variety of well annotated and validated models provides appropriate tools for scientific investigation.

Despite these advances, accurate modeling of tumor evolution remains challenging, particularly when considering the temporal sequence of genetic events and the selective pressures that shape tumor development.

Regulated sequential pharmacogenetic manipulation in vivo

As discussed above, capturing the selective pressures and temporal spacing that define human tumor evolution is essential, because mutation order strongly influences both tumor phenotype and clonal behavior. For example, introducing mutations simultaneously, or in the wrong order, can produce disease trajectories that differ entirely from those seen in patients. Sequential mutation modeling in mice is currently possible, but only on a limited scale, as the breeding required to intercross multiple regulator alleles for separate events quickly become prohibitively expensive and uses a vast number of animals. We are developing more flexible approaches that reduce the logistical burden of traditional breeding while still incorporating several key features into the models. These include temporal control, cell-of-origin specificity, heterogeneity and clonal dynamics, and the influence of microenvironmental and immune pressures. Together, these form the guiding principles for evolution-aligned modeling that are outlined in Table 2.

Table 2.

Key principles for evolution aligned modeling

Key principle	Evolutionary relevance	Modeling requirements
Temporal Control	Human cancers arise through the sequential acquisition of mutations (multistep carcinogenesis). Mutation order can influence tumor development and clonal succession.	Inducible systems that allow staged activation of oncogenic alleles or temporally separated loss of tumor suppressors will enable investigation of mutation order and clonal succession.
Cell-of-Origin Specificity	The identity and differentiation state of the initiating cell can influence tumor evolution. Initiation in different stem or progenitor cell populations may affect tumor latency, heterogeneity, and metastatic potential.	Precise targeting of defined cell populations, combined with temporal control, enables investigation of how cellular context shapes tumor evolution.
Heterogeneity andClonal Dynamics	Tumors comprise diverse subclones that compete, cooperate, expand, or become extinct over time.	Models should capture intratumor heterogeneity and enable investigation of clonal evolution over time.
Microenvironmentaland Immune Pressures	Tumor evolution is shaped by immune surveillance, stromal interactions, aging, and treatment-related selective pressures.	Models that enable tumor development and progression to be studied over time while retaining immune competence. Integrating tumor-intrinsic genetics with host context is essential for modeling therapy response and resistance.

The tandem arrayed regulator system

To address the need for multi-step manipulation within a single cell population, we are developing the first Tandem Arrayed Regulator (TAR) alleles (Figure 4). We assembled a polycistronic array of well-characterized, drug-inducible regulators linked by 2A slip sequences and inserted them into a safe harbor locus, either Rosa26 or Col1A1 (Figure 4A). The design of the array allows for tiered levels of control. Initially, an upstream Frt-STOP-Frt cassette prevents expression. Providing Flp-recombinase, either via a separate tissue-specific transgene or viral delivery for activation, removes this “brake”. Additionally, a CAAGS enhancer that can be removed by DRE recombinase precedes the array. This allows researchers to adjust the strength of expression once the system is activated, preventing potential toxicity from overexpressing the regulators. Upon activation, a single transcript is produced. The 2A slip sequences ensure ribosomal skipping, releasing four distinct, independently activatable proteins:

• 1.OsTir1^F74G: An auxin-inducible E3 ligase that enables rapid and highly specific degradation of degron-tagged proteins upon administration of the auxin analogue 5-Ph-IAA (5-phenyl-1H-indole-3-acetic acid).^8,46
• 2.rtTA3: A third generation modified reverse tetracycline-inducible transactivator that binds tet-response element (TRE)-containing promoters in the presence of doxycycline to drive gene expression in a dose-dependent manner.⁷
• 3.CreER^T2: A tamoxifen-regulated CRE recombinase that translocates to the nucleus upon binding 4-hydroxy-tamoxifen (4-OHT), where it excises DNA flanked by LoxP sites.
• 4.iRFP: A near-infrared fluorescent protein that marks TAR-expressing cells for longitudinal tracking.⁴⁷

Figure 4.

Schematic representation of the TAR allele and its application in cancer modeling

(A) Linear representation of the TAR allele integration into multiple inducible systems, including Flp-mediated tissue-restricted activation, auxin-dependent protein degradation, doxycycline-inducible gene or shRNA expression, and tamoxifen-inducible recombination. Flp activation also drives infrared deep-tissue reporters for non-invasive monitoring. These elements will allow precise temporal and spatial control of gene function in vivo.

(B) Application of the TAR allele in cancer modeling. 1. Generate mice carrying the TAR allele (A) together with the required combination of engineered alleles. 2. Initiate tumorigenesis and TAR activation via tissue-restricted or virally delivered Flp recombination. 3. Allow neoplastic phenotype to develop over an appropriate time frame. 4. Induce further tumor progression through deletion of candidate tumor suppressor gene (TSG) via tamoxifen-dependent activation of CreER^T2 driving recombination of floxed TSG allele (see Figure 2A). 5. Characterize the impact of TSG deletion on tumor phenotype. 6. Upregulate expression of candidate tumor promoting gene (TPG) via doxycycline administration to induce rtTA-dependent transcription of TRE-TPG gene (see Figure 2B). 7. Characterize the impact of controlled TPG upregulation on tumor progression. 8. Evaluate candidate therapeutic target protein via auxin-dependent activation of OsTir1F74G, leading to rapid degradation of degron-tagged protein (see Figure 2C). 9. Characterize the impact of target protein degradation. 10. Evaluation of selected clinical endpoints.

The TAR alleles allow the three regulatory modalities, CreER^T2 (a genetic modifier), tTA3 (a transcriptional activator), and OsTir1^F74G (a selective protein degrader) to be used in any sequence or combination, enabling precise, staged alterations within the same cell population.

Stacking these regulatory elements into a single polycistronic locus effectively reduces the driver breeding bottlenecks seen with other sequential modeling systems, though it’s in vivo implementation introduces several distinct methodological considerations (summarized comprehensively in Table 3). While it simplifies the driver machinery, significant crossbreeding is still required to introduce the necessary downstream target or conditional (LoxP flanked) alleles. Driving three independent regulatory modalities from a single construct also introduces baseline variability to monitor. When working with these large polycistronic arrays, it is practical to screen for potential structural instability, expression silencing across generations, uneven lineage expression, or insertional mutagenesis, which can occasionally occur even within “safe harbor” loci. Monitoring for background recombinase leakage and verifying that expression levels maintain a physiologically relevant gene dosage are necessary for accurately capturing patient-relevant tumor biology. As local tissue microenvironments further dictate small molecule ligand penetration, running a pilot validation phase is essential to fine-tune compound concentrations, induction schedules, and actual target gene expression for the specific tissue of interest. There is also the consideration of available companion alleles (e.g., AID-degron tagged alleles, and Frt-flanked driver alleles) for each tumor model. To ensure accessibility, the validated TAR allele mice will be archived and made available to the wider scientific community via the Mary Lyon Centre Harwell repository, which acts as the centralized resource hub for the MRC National Mouse Genetics Network.

Table 3.

Methodological advantages and limitations of the tandem arrayed regulator (TAR) allele system

Advantages (pros)	Limitations (cons)
Models Sequential Progression: Allows step-by-step manipulations rather than “all-at-once” delivery, more accurately reflecting human disease progression.	Genetic Complexity: Still requires extensive crossbreeding to introduce multiple downstream target alleles.
Deterministic Control: Gives investigators direct control over the timing of multiple distinct mutations, avoiding stochastic tumor development.	Resource and Time Requirements: Increases baseline housing costs and extends in vivo experimental timelines.
Direct Mechanistic Insight: Isolates the active drivers of tumor progression or regression without relying on inferred pathways.	Variable Expression/Delivery: Susceptible to heterogeneous transgene expression or uneven drug/ligand penetration across different cell and tissue types.
High Genotypic Fidelity: Closer genotype-to-phenotype recapitulation than conventional, static models.	Baseline Human Data: May require some a-priori knowledge of the suspected mutational order seen in patient cohorts.
Sequence Versatility: Full user control to reorder or customize the sequence of manipulations.	Companion Allele Availability: Flp-inducible and degron-tagged alleles must be custom-made if suitable companion lines are not already established.
Reversible Elements: Two out of the three conditional enzymatic components are biologically reversible.	–

Exemplar use of the TAR allele: Evolution-aligned modeling of lung cancer

We are currently using the TAR allele to model Non-Small Cell Lung Cancer (NSCLC) evolution in a temporally controlled manner. Traditional models of this disease typically have Kras activating mutations and delete Trp53 simultaneously, generating an aggressive disease from the outset.⁴⁸ Although efficient, these models bypass the initial stages of tumor development. Using the TAR allele system, we can separate these genetic events, to better reflect the natural disease trajectory.

As detailed in Figure 4B, the process is initiated by delivering a Flp-expressing virus to the lung to activate both the TAR allele and an Frt-STOP-Frt-Kras^G12D allele. This establishes an early stage activated KRAS-driven neoplastic environment. Then administering Tamoxifen triggers CreER^T2 activity, deleting Trp53, only after the primary tumor has been initiated and allowed to develop. Finally, Doxycycline or Auxin can be introduced to activate additional perturbations and allow investigation of later-stage requirements of signaling pathways, or to model therapeutic interventions in an advanced, heterogeneous tumor. Introducing each event in sequence shifts the model from simply carrying the right set of mutations, to recreating the evolutionary pressures that shape human malignancy.

Conclusion

This Primer highlights that the suitability of any mouse model depends on how closely its evolutionary trajectory mirrors human disease. Traditional GEMMs remain essential for studies that require an intact immune system and native tissue context, whereas somatic induction methods can restore early selective bottlenecks and clonal dynamics that single-step systems can compress. Emerging, more sophisticated platforms, such as TAR alleles, offer the finer temporal and pathway control needed for accurate modeling of multistep carcinogenesis, though they typically require more validation and longer experimental timelines (Summarized in Table 1).

Across all model types, the choice of cell-of-origin, timing of genetic events, and induction method will determine a study’s physiological relevance. Aligning these factors with the biological question ensures the system captures the selective pressures driving human tumor evolution. Rather than just replicating mutations, the goal is to select the model that most closely mirrors the evolutionary course seen in patients.

Acknowledgments

The authors thank Catherine Winchester (CRUK Scotland Institute) for reading and reviewing the manuscript. We apologize to the many colleagues whose work was omitted due to space restrictions for this focused Primer Article. The TAR allele design was carried out in collaboration with colleagues through the NMGN, specifically Dr Lydia Teboul, Dr Ben Davis, Dr Andrew Wood, and Dr Douglas Strathdee. The figures were created with the BioRender software at BioRender.com under licence to K Blyth. DMW, SLE & the generation of the TAR Allele were funded by the MRC NMGN Cancer Cluster (MC_PC_21042). L.E.M. and K.B. are funded by Cancer Research UK core funding to K.B. (A29799). D.J.M. acknowledges funding from Cancer Research UK (DRCNPG-Jun22\100007).

Author contributions

D.M.W., conceptualization, writing – original draft, and review & editing; L.E.M., conceptualization, writing – original draft, and review and editing; S.L.E., writing – original draft and review & editing; P.D.D., conceptualization, writing – original draft, and review & editing; D.J.M., conceptualization and writing – review and editing; K.B., conceptualization, and writing – review and editing.

Declaration of interests

The authors declare no competing interests.