New advances in the use of patient-derived organoids and mouse models for assessing heterogeneity among metastatic colorectal cancers.

Background
Colorectal cancer (CRC) is the third most commonly diagnosed malignancy and the second leading cause of cancer-related death worldwide, contributing substantially to the global cancer burden and posing a major threat to human health
1
. The liver is the most common site of distant metastasis in CRC patients. For patients with colorectal liver metastasis (LM), the 5-year overall survival rate remains below 30%, and LM is a predominant cause of CRC-related mortality
2
. The current standard of care for LM involves surgical resection, when feasible, combined with systemic chemotherapy. However, clinical observations reveal that a subset of patients experience disease progression despite receiving chemotherapy regimens recommended by the National Comprehensive Cancer Network (NCCN), such as FOLFOX or FOLFIRI. Moreover, recurrence often develops after adjuvant chemotherapy and liver metastasectomy, suggesting that some metastatic tumors may be intrinsically resistant to standard treatment, whereas others are sensitive. These discrepancies in therapeutic responses reflect the underlying heterogeneity of LM and contribute to the varied clinical outcomes observed among patients with mCRC.
Tumor heterogeneity is a key feature of cancer and manifests at multiple levels, including genetic, epigenetic, transcriptomic, histological, and microenvironmental differences both between and within tumors. Accurately modeling such heterogeneity is crucial for mechanistic studies and for developing precision therapies. Among various preclinical platforms, patient-derived organoids (PDOs) have emerged as powerful in vitro models. PDOs are three-dimensional epithelial structures derived from patient tumor tissues that retain the architecture, molecular features, and functional properties of the original tumors. These models preserve the intratumoral heterogeneity of CRC and allow rapid, scalable, and high-fidelity expansion in vitro
3
. PDOs enable functional assessment of patient-specific drug sensitivity and have shown obvious concordance with treatment response–associated clinical features, thereby supporting their potential utility in treatment stratification rather than direct clinical outcome prediction
4
. In particular, LM-derived PDO models have been shown to mirror the treatment responses of the original tumors and to reflect intertumoral and interpatient heterogeneity
5
. However, their predictive performance can be context-dependent, underscoring the need to interpret in vitro responses alongside clinical factors such as microenvironmental context, pharmacokinetics, pharmacodynamics, and assay standardization.
Complementing PDOs, patient-derived xenograft (PDX) models serve as in vivo platforms for studying CRC. By implanting patient tumor fragments into immunocompromised or humanized mice, PDX models replicate tumor–stroma and tumor–vasculature interactions within a physiologically relevant microenvironment. PDX models faithfully preserve the histological and molecular characteristics of primary tumors, including spatial and cellular heterogeneity, and have been widely used to test drug efficacy, explore resistance mechanisms, and validate therapeutic targets
6
.
In summary, PDO and PDX models provide insights into the biological basis of therapeutic resistance and can aid in the development of individualized treatment strategies for patients with mCRC. They represent complementary and translationally relevant tools for dissecting tumor heterogeneity and optimizing therapy for mCRC. Nevertheless, we still need to continuously develop our research models and address their shortcomings to better simulate heterogeneous mCRC diseases. In this review, we summarize recent advances in the field of mCRC driven by PDO and PDX models and discuss potential future directions for the application of these technologies in this area.

Heterogeneity of mCRC
The heterogeneity of mCRC has been consistently observed. A classic example is that lower rectal cancers are more prone to lung metastasis than upper rectal cancers are, which may be attributed to differences in the vascular drainage of the rectum
7
. Typically, venous return from the upper rectum occurs through the superior rectal vein and the inferior mesenteric vein, ultimately draining into the portal venous system. Tumor emboli disseminated via the bloodstream thus tend to remain in the liver, resulting in LM. In contrast, the lower rectum is drained by the inferior rectal vein, which empties into the internal iliac vein. Consequently, tumor emboli originating from lower rectal cancers can enter the common iliac vein
8
and subsequently the inferior vena cava, reaching the lungs via the right heart and leading to pulmonary metastasis. This phenomenon is well explained by the anatomical and mechanical theory of tumor metastasis proposed by James Ewing
9
. Since Paget proposed the “seed and soil” hypothesis
10
, the critical role of the “soil”—the tumor microenvironment (TME)—has been increasingly emphasized in metastatic cancers and is reflected in the heterogeneity of mCRC. The liver and lungs provide distinct TMEs, particularly in terms of the tumor immune microenvironment (TIME), for mCRC
11
. The liver serves as a homing and expansion site for myeloid-derived suppressor cells, together with Kupffer cells, to form an immunosuppressive microenvironment
12
, which substantially overlaps with the TIME of LM, promoting immune evasion and thereby accelerating the formation of metastatic lesions in advanced CRC
13
. In contrast, the TIME of pulmonary metastases exhibits a distinct inflammatory phenotype characterized by high levels of immune cell infiltration and increased expression of immune checkpoint molecules. These features confer greater immunogenicity to lung metastases than to LMs, potentially indicating a greater likelihood of benefiting from immune checkpoint inhibitor therapy
14
.
These findings underscore that mCRC is not a single entity. Its heterogeneity can be observed across multiple levels—from macroscopic to microscopic scales and from population-level characteristics to cellular features. A comparative study involving 17,641 patients revealed a difference in the incidence of LM between left-sided colon cancer (LCC) and right-sided colon cancer (RCC), with LCC being more common
15
. This further suggests that the location of the primary tumor influences the likelihood of metastases occurring in specific distant organs. Although venous drainage from both LCC and RCC flows into the liver, the observed difference in the incidence of LM indicates that the heterogeneity of metastatic lesions cannot be fully explained by anatomical factors alone. A series of studies have demonstrated that the proximal colon (including RCC) and distal colon (including LCC) differ in several biological characteristics, such as embryonic origin, epigenetic profiles, immunological features, and gut microbiota composition16,17. These findings indicate that some intrinsic factors may influence the initiation and progression of mCRC. Distinct histological subtypes of primary CRC further contribute to the heterogeneity of the resulting metastatic lesions. For example, mucinous adenocarcinoma (MAC) differs from adenocarcinoma (AC) in multiple phenotypic aspects. One notable difference lies in mucin expression levels
18
, which substantially influence their distinct patterns of tumor invasion and metastasis
19
. While AC tends to metastasize to the liver, MAC and signet-ring cell carcinoma (SRCC) are more prone to peritoneal dissemination, with the latter exhibiting an even greater propensity for metastasis
20
. Moreover, among these three subtypes, SRCC has the greatest metastatic tendency and is more frequently associated with metastases to multiple and rare sites, such as bone, the pancreas, and the heart
20
. In addition, multiorgan metastases in mCRC often follow distinct combination patterns. For example, ovarian metastases frequently co-occur with peritoneal metastases (PMs), whereas brain metastases are significantly more likely to be accompanied by lung metastases than by LMs
21
. The former phenomenon can be explained by the anatomical continuity between the ovaries and the peritoneum in females, which facilitates the simultaneous colonization of both sites during intraperitoneal tumor dissemination
21
. A reasonable hypothesis to explain the latter phenomenon is that LMs originate from tumor cells with limited motility, whereas lung metastases may arise from more stem-like and motile tumor cell subpopulations. As a result, these cells have a greater propensity for secondary dissemination to other distant organs
21
.
Moreover, the differences in metastatic patterns described above are ultimately driven by deeper underlying molecular mechanisms. On the basis of its genomic, epigenomic, and transcriptomic heterogeneity, CRC can be classified into four consensus molecular subtypes (CMSs)
22
. These distinct CMS categories are associated with varying metastatic tendencies and prognoses in patients with mCRC
23
. The observation that specific CMS subtypes are enriched in different segments of the colon elucidates the molecular basis of the heterogeneity between LCC and RCC—CMS1 and CMS3 are more commonly found in right-sided tumors, whereas CMS2 is predominantly observed in left-sided tumors
24
. Treatment-related heterogeneity in CRC, such as differences in drug sensitivity, has been confirmed in preclinical models such as PDXs to be associated with CMS subtypes
25
. However, there is currently a lack of high-quality research on the connection between the CMS and metastatic behavior through PDO or PDX models, which may be a hot topic of current research. Furthermore, the discovery of intratumoral heterogeneity in CRC at the single-cell level has paved the way for in-depth investigations into single-cell genomics, epigenomics, and transcriptomics
26
. Moreover, these results indicate that organoid models derived from single cells may serve as a reliable platform for studying CRC heterogeneity
26
. By integrating molecular mechanisms with drug-response assays and directly linking clonal architecture and molecular states to functional phenotypes, PDO and PDX platforms provide experimental resolution and a mechanical understanding of mCRC heterogeneity. Unlike purely observational clinical approaches, these patient-derived preclinical systems enable causal, lesion-matched testing that has revealed mechanisms such as subclonal/branched evolution as a functional driver of heterogeneous responses to therapy, metastasis-associated transcriptional rewiring (e.g., stemness programs in LM), and spatially variable receptor signaling states (e.g., EGFR/HER2 heterodimer variability) underlying differential sensitivity/resistance to targeted agents5,27,28.
For patients with mCRC, tumor heterogeneity manifests as differences in survival outcomes. Although the overall incidence of LM is greater in LCC, patients with LM originating from RCC have poorer survival outcomes than those with LM from LCC, potentially because of the greater number of metastatic lesions and the involvement of more hepatic segments
29
. Similarly, patients with multiorgan metastases have significantly poorer survival than those with metastases confined to a single organ
30
. Moreover, tumors exploit their heterogeneity to undergo continuous genomic evolution under the selective pressure of various therapeutic interventions
31
. This evolutionary process originates from the polyclonal genome within the tumor and is ultimately associated with treatment resistance and poor clinical outcomes
32
. Traditional models based on established tumor cell lines are becoming increasingly disconnected from clinical reality. Robust tools capable of recapitulating the heterogeneity of mCRC and capturing treatment responses associated with such heterogeneity are urgently needed. Such tools would provide deeper mechanistic insights into the role of tumor heterogeneity in mCRC and eventually contribute to more precise diagnostic strategies and therapeutic approaches for affected patients.

Heterogeneity identified in murine models
Mouse models have long been essential tools for studying CRC. Genetically engineered mouse models (GEMMs) and PDX models represent two powerful and indispensable platforms in cancer research, both of which have been widely applied in the CRC field
33
. With the increasing identification of CRC-related gene mutations, there is a growing demand for mouse models to study specific molecular mechanisms. Traditional carcinogen-induced models are inadequate for research purposes because of their lack of tumor invasiveness and lack of colon specificity
33
. In contrast, the development of GEMMs based on the Cre–loxP system has allowed researchers to generate mice with colon- and rectum-specific mutations
34
. Through various transgenic strategies35–37, this system enables the bacteriophage P1-derived recombinase Cre to be specifically expressed in the mouse colon, where it recombines and excises target genes flanked by LoxP sites. These approaches have greatly facilitated the elucidation of the complex molecular mechanisms underlying CRC initiation and progression. GEMMs have also been employed in the study of mCRC. Hung et al.
38
were the first to develop a novel GEMM that successfully induced sporadic colonic tumors in mice, recapitulating the full progression from adenoma to carcinoma in situ and ultimately to metastatic cancer. This model closely mirrors the natural course of human CRC development. Using a GEMM platform, Chanrion et al.
39
further demonstrated that activation of the Notch signaling pathway combined with loss of p53 is associated with the development of mCRC.
However, individual GEMMs typically harbor a limited set of genetic alterations, which differ substantially from the diverse mutational landscape observed in human tumors. As a result, each GEMM has unavoidable limitations in modeling CRC tumor heterogeneity. As previously discussed, a wide array of molecular alterations contributes to mCRC, and a single GEMM is often insufficient to recapitulate the complex molecular crosstalk involved. Consequently, discrepancies may still exist between the resulting tumor phenotypes and those observed in human disease. Similarly, in mCRC, the effects of gene mutations such as those in KRAS, p53, and BRAF vary across populations. The association between a given mutation and the occurrence of distant metastasis in CRC is heterogeneous among different racial groups, suggesting that genetic background also plays a significant role in mCRC heterogeneity
40
. However, owing to long-term inbreeding and stabilization, GEMMs possess a uniform genetic background, which limits their ability to reflect the complex and diverse characteristics of mCRC observed in real-world settings—despite ensuring high reproducibility in experimental studies. In addition, the construction of GEMMs involves extensive genetic screening and breeding processes, requiring considerable numbers of mice and prolonged time periods. Given the increasing emphasis on animal welfare, GEMMs may not represent the optimal choice for studying mCRC.
Unlike GEMMs, PDX models were developed after the advent of immunodeficient mice. Before the challenge of immune rejection was overcome, the application of xenograft models was highly limited. This changed in the 1960s with the development of nude mouse41,42 and SCID mouse
43
models. These immunodeficient mouse models represented major breakthroughs and provided a foundational platform for the successful establishment of xenografts. Researchers have progressively optimized the genotypes of these mouse strains to develop more profoundly immunodeficient models, such as the NSG mouse model
44
. These improvements have largely addressed the issues of “immune leakiness” and high natural killer (NK) cell activity observed in earlier immunodeficient mouse models
45
(Table 1). Profound immunodeficiency ensures a better success rate for xenograft models, and the optimal approach for model construction is to implant patient-derived tumor tissue directly into immunodeficient mice rather than into cell lines. Although cell line–derived xenograft (CDX) models are easier to establish because of the stability of the cell lines, they are limited by a lack of molecular heterogeneity
46
. Moreover, owing to long-term adaptation to in vitro culture conditions, CRC cell lines often acquire phenotypes such as enhanced invasiveness, which may not accurately reflect the characteristics of patient-derived CRC cells
47
. Therefore, CDX models have limited applicability for studying the heterogeneity of human CRC or for guiding individualized clinical treatment, especially when the cell lines used in the CDX model are insufficient
46
. In contrast, extensive studies have demonstrated that PDX models can faithfully recapitulate the pathological features, molecular heterogeneity, and clinical phenotypes of tumors from selected patients. Even after multiple passages, PDX tumors maintain histological characteristics—such as glandular architecture and mucin production—that are consistent with those of the original tumor, along with the preservation of intratumoral genetic heterogeneity48,49. PDX models demonstrate a high degree of concordance with patients’ clinical responses to treatment, including both positive and negative therapeutic outcomes. Remarkably, even when patients have undergone multiple regimens of therapy, the originally established PDX models are still capable of predicting responses to subsequent treatments
49
. PDX models have also proven effective in recapitulating the CMS-based heterogeneity of CRC, further reinforcing the associations among CMS classification, tumor heterogeneity, and individualized therapeutic strategies
25
. These findings highlight high-throughput drug screening as another important application of the PDX platform
25
.
Common strategies for constructing mCRC xenograft models include ectopic transplantation, orthotopic transplantation to metastatic target organs, and orthotopic implantation into the colorectum. The most widely used ectopic transplantation method involves implanting tumor tissue subcutaneously into immunodeficient mice. This approach is broadly applicable to a variety of solid tumor types. Subcutaneous implantation provides only a fundamental TME for the engrafted tissue. However, the procedure is technically simple, causes minimal surgical stress to the mice, and results in superficially located tumors that are easily accessible for observation and tumor volume measurement
50
. Orthotopic transplantation into metastatic target organs involves implanting tumor tissue into common metastatic sites of CRC, such as the liver
51
and peritoneum
52
, to model LM and PM. The major advantage of this approach is its ability to closely replicate a patient’s individual pattern of metastatic organ involvement, providing distinct TMEs that reflect organ-specific metastases, thereby enabling the model to more closely resemble the actual disease condition. Nevertheless, this method is technically demanding and imposes greater surgical stress on the mice. In addition, tumor growth is not readily observable and typically requires indirect monitoring through bioluminescence imaging or small-animal imaging techniques
50
. Orthotopic implantation typically involves placing tumor tissue beneath the serosal layer of the cecum53–55. This model mimics the natural progression of human CRC from primary tumor development to metastasis and mimics interactions between the tumor and the intestine-specific stromal microenvironment
55
. Nevertheless, this approach has several notable limitations. Tumor growth and metastatic progression are difficult to observe and track in real time
50
; in some cases, animals must be sacrificed for dissection to assess metastasis. This procedure is technically challenging, particularly for establishing metastatic models, which often require long-term maintenance for several months and incur high costs. In addition, other implantation sites have been explored for tumor modeling. Bouvet et al.
56
reported that CD11c-positive splenic cells protect CRC cells injected into the spleen and facilitate their survival. These splenic immune cells were shown to migrate with tumor cells via the portal vein to the liver, where they contributed to the establishment and maintenance of LM. In contrast, direct injection of tumor cells into the portal vein has proven to be an unreliable method for constructing LM models.
Xenograft models have long contributed to advancing our understanding of mCRC heterogeneity. As early as 1995, Kuo et al.
57
used orthotopic CDX models developed by implanting different CRC cell lines into the cecum of mice to investigate how variations in the hepatic colonization capacity of these cell lines influence the formation of LM. In recent years, with the development of PDX models, increasing attention has been given to investigating mCRC heterogeneity using patient-derived tissues. Cho et al.
27
employed whole-exome sequencing (WES) to perform phylogenetic and subclonal analyses and demonstrated that PDX models are capable of faithfully recapitulating the branched evolutionary patterns of patient tumors at the genomic, transcriptomic, and DNA methylation epigenomic levels. Furthermore, by integrating PDX-based drug sensitivity assessments, they revealed that subclonal heterogeneity within tumors is a key contributor to the therapeutic heterogeneity observed in multiorgan mCRC. At approximately the same time, Wang et al.
28
also utilized PDX models to investigate the relationships between specific genetic alterations—such as KRAS mutations, HER2 amplification, and FGFR2 amplification—and the heterogeneity of drug sensitivity in mCRC. In addition, they explored differences between primary tumors and LM in terms of genetic loci and chromosomal copy number variations. Recent advancements in the field include the application of PDX models in targeted therapy research, particularly in targeting epidermal growth factor receptor (EGFR) and human epidermal growth factor receptor 2 (HER2). Gupta et al.
58
described a potential mechanism of EGFR-mediated resistance to trastuzumab deruxtecan (T-DXd), an antibody–drug conjugate that targets HER2, in HER2-overexpressing mCRC. They reported that in CRC, when EGFR is overexpressed, it can form heterodimers with HER2, thereby preventing the formation of HER2/HER2 homodimers. This action interferes with T-DXd internalization and subsequently impairs its therapeutic efficacy. The researchers observed intratumoral variability in EGFR expression and differences in the abundance of EGFR/HER2 heterodimers within PDX tumors derived from a CRC patient. This finding serves as another compelling example of how intratumoral heterogeneity in CRC can contribute to differential therapeutic responses.
PDX models have been widely applied in the field of mCRC, and their scientific validity and experimental reproducibility have been extensively demonstrated. These models have enabled the comprehensive representation of mCRC heterogeneity across multiple levels, including clinical phenotypes and molecular profiles. As a powerful tool for studying mCRC heterogeneity, PDX models continue to hold a prominent position in the research domain.

Heterogeneity identified through PDO models
As emerging and highly promising in vitro preclinical models, organoids have a relatively shorter history than PDX models. Professor Hans Clevers and his team have been widely recognized as pioneers in the development of these models, with many landmark discoveries and innovations attributed to their work. The birth of organoid technology has generally been traced back to 2009, marked by the development of a self-organizing intestinal organoid model by the Clevers group, which was based on murine leucine-rich repeat-containing G-protein-coupled receptor 5 (LGR5)-positive intestinal stem cells
59
. This model was the first in vitro system in which a single intestinal stem cell, without reliance on mesenchymal or stromal support cells, could differentiate into a structured intestinal epithelium composed of multiple cell types suspended in an artificial extracellular matrix (ECM) and exhibiting a crypt–villus structure. The establishment of this model helped researchers dynamically and systematically recapitulate the cytological and histological characteristics of the intestine in vitro, from single cells to multicellular tissue structures. The application of this system to CRC research was also pioneered by the Clevers group. In 2011, they optimized the formulation of growth factors, hormones, and vitamins in the organoid culture medium, allowing both normal colonic epithelial cells and CRC cells derived from patient biopsies to form organoids and be stably propagated over the long term
60
. This milestone marked the beginning of PDO models. Originating from human biopsy tissues, rather than murine animals, PDO models offer a highly relevant platform for faithfully modeling the biology of human CRC. The Clevers group subsequently demonstrated that CRC-derived PDO models, even after in vitro culture and serial passaging, faithfully retained the histological, genomic, and transcriptomic features of the original patient tumors
61
, confirming that PDOs can serve as representative models of these tumors. Furthermore, they established a high-throughput drug sensitivity screening platform based on PDO models and showed that variations in drug response were associated with alterations in molecular signaling pathways
61
.
The success of the PDO model has provided an unprecedented opportunity for more precise investigations of CRC, leading to a surge of research across various domains in a short period. In the context of tumor heterogeneity, subclonal populations observed in PDO models closely resemble those identified in PDX models
61
, suggesting that PDO models also represent a robust and efficient platform for studying tumor heterogeneity. They offer valuable insights into the factors driving tumor evolution and heterogeneity, as well as their impact on drug response and patient prognosis. In the field of mCRC, PDO models derived from metastatic tumors were established around the same time. Sequencing analyses demonstrated that these mCRC PDO models exhibited excellent representativeness in terms of tumor driver mutations and DNA copy number alterations
62
. Notably, the work of Vlachogiannis et al.
63
provided compelling evidence that PDOs can achieve high sensitivity, specificity, positive predictive value, and negative predictive value in predicting patient responses to targeted therapies and chemotherapeutic agents. Collectively, these studies highlight the use of organoids as a feasible, rapid, and personalized platform for in vitro drug screening in mCRC. As mentioned above, differences in histology, gene mutations, and molecular pathways in mCRC are fundamentally driven by intratumoral heterogeneity. The variation among tumor cells within the primary lesion leads to the formation of heterogeneous metastatic sites following the dissemination of distinct cellular subpopulations. Organoid models constitute an ideal platform that recapitulates this single-cell-level diversity in vitro. Roerink et al.
26
provided evidence for this concept. By sampling multiple regions of the same CRC tumor lesion from a single patient and establishing separate PDO models from each region, the authors identified heterogeneity at the genomic, methylation, and transcriptomic levels among the models. Notably, these PDOs also exhibited differential responses to drug treatment. These findings clearly demonstrated that PDO models are capable of faithfully preserving the intratumoral heterogeneity of CRC at the single-cell level. Building upon the extraordinary work described above, organoids have gradually gained recognition in recent years as excellent preclinical models
64
. The ability of tumor organoids to facilitate antitumor drug screening and predict individualized treatment responses has attracted substantial attention, with researchers increasingly shifting their focus from traditional 2D cancer cell cultures and PDX models to PDO models.
Remarkable progress has also been made in the application of PDO models in the mCRC field. Narasimhan et al.
65
focused on CRC PM and established a “peritonoid” model using PDO technology. By integrating medium-throughput drug screening with WES, they identified both population-level commonalities and individualized heterogeneity in drug sensitivities and genetic profiles among PM patients. In a separate study, Mo et al.
5
creatively established a cohort of mCRC patients with synchronous LM and constructed a large-scale PDO biobank by generating matched organoid models from both primary CRC and LM tissues from the same patients. Through integrative analyses, including histopathology, genomics, and single-cell transcriptomics, the study revealed a high degree of concordance between primary CRC tumors and LMs in terms of histological phenotypes and driver mutations. However, the authors also noted an upregulation of stemness-associated gene pathways in LMs compared with primary lesions, suggesting that LMs may possess greater proliferative capacity and plasticity. In both single-agent and combination chemotherapy drug sensitivity assays, the pronounced heterogeneity among patients underscores the importance of personalized treatment strategies in clinical practice. Moreover, the relatively low level of heterogeneity observed between primary tumors and matched LMs from the same patient suggests that metastatic lesions tend to inherit the sensitive or resistant status of the primary tumor with respect to chemotherapies. Bruun et al.
66
established a “pharmacotranscriptomic heterogeneity” research platform based on PDO modeling. By enrolling a cohort of mCRC patients with multiple LMs and comparing different metastatic lesions from the same individual, the authors reported that in the vast majority of patients, multiple metastatic lesions exhibited consistent responses to a panel of 40 clinically relevant chemotherapeutic and targeted therapy agents, with only a minority showing heterogeneity among lesions within a patient. However, significant heterogeneity among patients in terms of drug sensitivity was still observed and strongly correlated with differences in transcriptomic profiles. These findings suggest that transcriptomic heterogeneity may be a key driver of pharmacological heterogeneity in mCRC.
The above studies strongly highlight the heterogeneity among mCRC patients, reinforcing the need for personalized treatment strategies for these patients. However, the presence of heterogeneity among different lesions within the same patient remains inconclusive and warrants further investigation. In accordance with the work by Chen et al.
67
on patterns of metastasis and genomic evolution in mCRC, three distinct evolutionary modes of metastasis have been proposed: the slow branch-off mode, the faster sequential mode, and the rapidly disseminating diaspora mode. In branch-off mode, tumor cells undergo multidirectional evolution over time, and different evolutionary subclones give rise to separate metastatic lesions. This mode has the highest likelihood of heterogeneity among different metastatic lesions. In contrast, both the sequential and diaspora modes typically involve metastatic lesions derived from the same cluster, resulting in no heterogeneity among metastases. Therefore, when metastases develop a prolonged time after the formation of the primary tumor or when different metastatic lesions emerge at widely separated time points, heterogeneity may still be present. These findings indicate that the mechanisms underlying the initiation and progression of mCRC are incompletely understood and that further investigation is needed.
In summary, PDO models represent excellent tools for studying the heterogeneity of mCRC. They accurately recapitulate the histological, genetic, and transcriptomic heterogeneity of tumors, and their drug response profiles derived from sensitivity assays provide valuable guidance for personalized treatment strategies. Furthermore, owing to their rapid expansion and short culture duration, compared with PDX models, PDO models offer considerable advantages for high-throughput drug screening
68
. Recent regulatory initiatives have encouraged the development of alternative preclinical platforms that may reduce reliance on animal models. The US Food and Drug Administration (FDA) announced plans to gradually phase out the requirements for animal testing in the drug development process
69
. In its statement, the FDA noted that data derived from organoid-based testing may be more direct, obtained more rapidly, and better aligned with current standards for animal welfare. In this context, PDOs offer advantages in terms of scalability, speed, and patient specificity. However, PDOs are more likely to complement PDX models, particularly for addressing questions involving pharmacokinetics, tumor–stroma interactions, and systemic treatment responses.

Limitations of PDX and PDO models
Each of the above models presents inherent limitations despite their respective advantages (Table 2). PDO models lack components of the TIME. A proposed solution to this limitation is the development of coculture systems that integrate PDOs with immune cells
70
. Dijkstra et al.
71
established and validated a coculture platform using autologous tumor organoids and peripheral blood lymphocytes, which enables the enrichment and expansion of tumor-reactive T cells from the peripheral blood of patients with mismatch repair-deficient CRC and non-small-cell lung cancer. By reconstructing tumor–immune cell interactions in vitro, this platform provides strong support for the development of personalized immunotherapy and the identification of novel immunotherapeutic targets. A recent advancement in the mCRC field was reported by Subtil et al.
72
, who developed a coculture model of dendritic cells (DCs) and mCRC organoids, demonstrating phenotypic changes in DCs derived from peripheral blood following interaction with tumor cells. These studies have focused primarily on mechanistic investigations. Beyond these advances, no other notable breakthroughs with significant clinical relevance have yet been reported in the mCRC field.
The predictive accuracy of PDO-based drug screening is still influenced by multiple biological and technical factors. In addition to the TIME, the absence of other native TME components, including stromal and vascular elements, may lead to discrepancies when evaluating therapies such as anti-EGFR therapy, which rely on cell‒cell interactions or paracrine signaling. Fortunately, the technology of organoid vascularization supported by induced pluripotent stem cells (iPSCs) has emerged and is likely applicable to gut organoids
75
, which may partially solve this problem. In addition, PDO systems do not capture pharmacokinetic and pharmacodynamic parameters that critically shape in vivo drug exposure and efficacy, which raises doubts about whether the results of drug sensitivity tests accurately reflect the situation in the body. Variability in culture conditions, operation protocols, and assay protocols across laboratories further contributes to interstudy reproducibility challenges and emphasizes the need for standardized workflows and quality control measures, including primary tumor tissue processing, organoid culture condition control, and drug testing result interpretation
76
. In addition, in the scenario in which the drug sensitivity results of PDOs guide the postsurgery chemotherapy regimens of patients, the timeliness of the drug sensitivity results of PDOs is very important because patients generally need to start the first treatment approximately 1 month after surgery, and shortening the experimental timeline of PDO drug sensitivity testing as much as possible is also an issue that needs to be solved.
One of the major limitations of PDX models is their inability to evaluate the efficacy of cancer immunotherapies. These models are established in immunodeficient mice, resulting in a tumor stroma that lacks immune cell infiltration and a tumor vasculature devoid of functional circulating immune cell populations capable of interacting with tumor cells. Even in partially immunocompetent models, such as nude mice, the immune system remains substantially different from that of cancer patients, limiting the translational relevance of immunotherapy studies conducted in such models. A promising solution is the development of humanized PDX mouse models. Human CD34+ hematopoietic stem cells—typically derived from umbilical cord blood—are injected into immunodeficient mice lacking T cells, B cells, and NK cells. This approach enables the reconstitution of human lymphoid and myeloid hematopoiesis in the host mouse
77
. Capasso et al.
78
demonstrated that these humanized mice can support the growth of CRC-derived PDX tumors and tolerate anti-programmed death 1 (PD-1) therapy. They found that anti-PD-1 treatment significantly suppressed tumor growth in microsatellite instability-high (MSI-H) models, with greater efficacy than in microsatellite stable (MSS) PDX models—closely mirroring clinical responses observed in human patients. These findings suggest that humanized PDX mouse models represent promising platforms for evaluating CRC immunotherapy responses. However, a critical concern remains: does the immune system reconstructed from CD34+ hematopoietic stem cells truly recapitulate the patient’s own immune system? At present, there is no definitive answer. Since the immune system in humanized mouse models is usually derived from a homogeneous source, the immune heterogeneity observed among different cancer patients cannot be recapitulated in the model, whereas obtaining CD34+ hematopoietic stem cells directly from tumor patients remains challenging
77
. In addition, mouse models are inherently unable to provide implanted human immune cells with the same environment as the human body, which can lead to a series of problems, ultimately affecting the effects of human immune cells in mice and making it impossible to study tumor–immune interactions79,80. This presents a potential barrier to further investigating interpatient heterogeneity for personalized immunotherapy response prediction for mCRC patients using humanized PDX models.
Compared with each other, PDX and PDO models provide complementary strengths for mCRC research. PDX models preserve the original tissue architecture, molecular features, and tumor heterogeneity and provide a vascularized in vivo microenvironment that more closely resembles that of human tumors, but they require a long establishment time (typically 4–8 months), are costly, and are generally unsuitable for immunotherapy evaluation
6
. PDO models, in contrast, can be rapidly expanded while retaining patient-specific features and tumor heterogeneity, making them well suited for high-throughput drug screening; however, they lack immune and vascular components and therefore often require immune coculture strategies or microfluidic support to better approximate the tumor microenvironmental context
64
.

Future applications
In addition to coculture PDO models and humanized PDX models, sequential or combined culture approaches that integrate PDO and PDX models (“PDO+PDX models” for short) may yield hybrid models such as patient-derived organoid xenograft (PDOX) or patient-derived xenograft organoid (PDXO) models
50
. In a PDOX model, the organoid is first cultured and then transplanted into immunodeficient mice by injection to construct a xenograft model and continue experiments that cannot be performed in vitro. A PDXO model would be established in the opposite manner. First, a xenograft model, such as a mouse subcutaneous tumor, is established. After successful establishment, some samples are extracted from the tumors in the mice, after which the samples are digested to obtain tumor cells. The tumor cells are then used to construct an organoid model, which is used to complete the verification of the in vitro model. These models leverage the respective strengths of both systems—allowing faster expansion of patient-derived tumor tissues and making them well suited for the rapid establishment of clinically relevant disease model biobanks and therapeutic screening
73
. These models also exhibit high concordance with the paired PDX and PDO models at the genomic, histopathological, and pharmacological levels
74
. In addition, these hybrid models retain tumor heterogeneity
73
and recapitulate tumor phenotypes within a vascularized in vivo microenvironment, thereby addressing the limitations of conventional PDOs, which lack vascular components. These hybrid models have already been successfully applied in mCRC research to predict treatment sensitivity in patients undergoing chemotherapy
81
.
The development of microfluidic technologies has greatly expanded the scope and applicability of PDO models. Challenges related to PDO vascularization have been further addressed through the use of microfluidic systems, in which self-assembled vasculature is integrated with microfluidic channels
82
. This configuration enables immune cells, inflammatory factors, and therapeutic agents to be perfused into the interior of PDOs through a vascularized ECM, thereby more faithfully recapitulating disease model features such as inflammation. Another major application of microfluidics is the organoid-on-a-chip (OoC) platform, which moves organoids beyond static culture at the bottom of culture wells and allows them to be cocultured with tissues from multiple organs or even with microorganisms such as bacteria
83
. This approach provides an in vitro window into multiorgan interactions and multifactorial cooperative or sequential processes, opening new avenues for modeling and investigating complex diseases.
Looking ahead, the integration of PDX and PDO models with spatial transcriptomics, single-cell sequencing, and machine learning is expected to substantially expand their analytical and translational capacity. The integration of machine learning and computational modeling into PDX platforms enables real-time, multiomic data analysis while preserving tumor heterogeneity with high fidelity, thereby potentially enhancing the predictive accuracy of PDX models for therapeutic responses
84
. In parallel, single-cell sequencing and spatial transcriptomics are being increasingly applied to PDO models. Single-cell sequencing offers exceptional resolution for dissecting cellular heterogeneity within organoids, particularly the diversity among distinct tumor cell subpopulations, and is well suited for studying organoid differentiation dynamics, disease mechanisms, and drug modes of action
85
. When combined with spatial transcriptomics, single-cell approaches can further elucidate the spatial organization and intercellular interaction networks of different cellular populations within organoids
85
. Collectively, the convergence of PDO/PDX platforms with spatial transcriptomics and AI-driven analytics represents a critical future direction toward resolving spatial and temporal tumor heterogeneity and narrowing the gap between experimental modeling and clinical decision-making.
Among various metastatic sites, the liver is the most extensively studied site in the field of mCRC. Research focusing on peritoneal, ovarian, pulmonary, and multiorgan metastasis and other rare metastatic sites using PDO- or PDX-based models remains limited. These areas represent vital research fields for future investigations into mCRC heterogeneity and may continuously guide personalized treatment strategies for patients with such metastatic patterns.

Conclusion
Metastatic colorectal cancer (mCRC) is a highly heterogeneous disease, with heterogeneity observed across multiple levels—from populations to individual cells and from macroscopic to microscopic scales. Its complexity spans genetic, transcriptomic, histological, and microenvironmental dimensions and has continued to attract the attention of researchers, with GEMM, PDO, and PDX preclinical models extensively utilized to yield novel insights (Table 3). Selecting the optimal research model to recapitulate specific heterogeneity is essential for understanding disease progression and optimizing therapeutic strategies (Table 4). Over the past decade, PDX and PDO models have emerged as powerful and complementary platforms for capturing mCRC heterogeneity. These models enable functional investigations of intralesional, interlesional, and interpatient variability; facilitate high-throughput drug screening; and support personalized treatment response prediction. Recent innovations—including organoid coculture systems, humanized PDX models, and PDX+PDO models, and OoC technologies—have further expanded their applications by addressing key limitations such as the absence of immune components or vascular structures. Among the most promising next steps is the development of standardized immune-enabled PDO coculture platforms that incorporate autologous or defined immune components to better capture microenvironment-dependent vulnerabilities and to functionally evaluate immunotherapies and rational combinations. Importantly, these models have not yet fully bridged the gap between experimental prediction and routine clinical implementation. Looking forward, the most impactful directions include (1) standardized immune-enabled and microenvironment-integrated modeling strategies, (2) rigorous quality control and benchmarking frameworks, and (3) prospective clinical validation coupled with multiomic and computational approaches. Together, these advances are expected to strengthen the translational utility of patient-derived models and support more context-aware precision strategies for modeling and treating mCRC.