Harnessing PDX and PDX 2.0: the next-generation paradigm for precision oncology and translational breakthroughs

Abstract

Cancer research has achieved remarkable breakthroughs over the past decades with the aid of patient-derived xenograft (PDX) models. However, the limitations of conventional PDX models in hindering clinical translation have become increasingly apparent. In 2025, the National Institutes of Health (NIH) announced a funding shift away from exclusive reliance on animal models without justified integration of novel alternative methods (NAMs), human-relevant modeling approaches. Nevertheless, PDX models cannot be fully replaced currently due to the lingering immaturity and uncertainties of NAMs technologies, indicating that a complete non-animal research paradigm will require sustained methodological development. Therefore, developing an innovative, optimized PDX model to navigate this transitional phase remains the holy grail of preclinical cancer research. Herein, we propose PDX 2.0, a novel conceptual framework that advances conventional PDX models via the systematic integration of NAMs and complementary technologies, thereby enabling more efficient and precise cancer research. This review first delineates the core determinants, major applications, and critical limitations of traditional PDX models, then defines the conceptual architecture and distinctive characteristics of PDX 2.0. We further highlight emerging applications of this framework in high-throughput drug screening, biomarker discovery, and adaptive therapeutic evaluation, positioning PDX 2.0 as a critical evolution of PDX-based research to better support clinically actionable precision oncology.

Introduction

The patient-derived xenograft (PDX) models are preclinical research platforms using immunodeficient mice to implant cancerous tissues and cells from tumor patients, thereby obtaining a series of characteristics similar to the primary tumor. It has elucidated partial mechanisms of tumors and developed drugs based on the differences in the molecular characteristics of malignant tumors to promote the development of personalized medicine (precision medicine) [1, 2]. The most significant feature of the PDX model is that it can retain the heterogeneity of the primary tumor to some extent, including the molecular and phenotypic information, which is conducive to personalized and predictive research, development of drugs, and biomarker analysis in the clinical process of cancer [3–5]. Given that intratumor heterogeneity (ITH) is a critical factor in drug resistance and cancer treatment failure [6], the associated research and development of chemotherapy, molecularly targeted therapy, and tumor immunotherapy provides updated solutions for cancer treatment, especially for the heterogeneity caused by genetic and non-genetic variation of tumors [7]. With advances of molecular annotation technologies, the role of xenograft models in genuinely reflecting the biological features of tumors and evaluating drug efficacy is immediate [8].

As the limitations of conventional PDX models become increasingly apparent, including time lag, microenvironmental shifts, subclone selection, metastasis-related discrepancies, immunodeficiency, labor intensity, and other factors [1, 9–13], their translational utility remains constrained. At the same time, breakthroughs in key modeling technologies have propelled novel alternative methods (NAMs) into prominence [14]. NAMs encompass in silico, in chemico, and in vitro approaches designed to reduce or eliminate animal use, lower costs, increase throughput, and accelerate experimental turnaround. In 2025, the National Institutes of Health (NIH) announced that grants based solely on animal models will be discontinued unless such models are scientifically justified and appropriately integrated with NAMs [15, 16]. However, current NAMs remain immature and unstable for addressing complex biological questions, with insufficiently validated benchmarks and comprehensive training [17, 18], which underscores that a fully non-animal research paradigm remains a distant goal. In this critical transitional period, we propose PDX 2.0, a transformative paradigm that integrates human-based approaches and artificial intelligence (AI)-driven innovation to address these challenges and redefine precision oncology. This paradigm enables iterative treatment optimization aligned with tumor evolution by integrating multi-omics data, AI-based drug response prediction, patient-specific treatment strategies, and biomedical image analysis, among others. Such capabilities are critical in an era where cancer therapies increasingly rely on real-time molecular insights to accelerate the translation of tailored treatments into clinical practice.

In this review, we first present the development of PDX 2.0 models, the key factors for successful establishment, and the main applications of conventional PDX models. More importantly, we highlight the major limitations of traditional PDX models that hinder clinical translation and summarize the corresponding solutions and technological breakthroughs. Building on this foundation, we introduce the innovative concept of PDX 2.0, along with its defining features. Finally, we discuss the core applications of PDX 2.0 across major cancer types and outline future directions for its development.

Development history of PDX and PDX 2.0 models

The concept of “model grafting” represents a crucial innovation in research methodologies, encapsulating the scientific community’s pursuit of the optimal animal object. Since the first xenotransplantation experiments were conducted, nude mice emerged as the first generation of immunodeficient recipient mice to overcome immune rejection. Their application in cancer research marked the formal establishment of PDX models [19]. In the latter half of the twentieth century, the iterative development of diverse immunodeficient mouse strains with distinct advantages propelled PDX models into an era of technological breakthroughs [20–22]. Since the early twenty-first century, the continuous advancement of more sophisticated host mouse strains has coincided with PDX models entering a new phase of large-scale application across diverse areas of cancer research [23]. Necessarily, in 2014, Europe established the EurOPDX Data Portal to share PDX mouse model data information, reach unified standards, and avoid unnecessary waste of resources [24]. The American Institute for Cancer Research (AICR) announced it would stop using the NCI-60 cell line, a 20-year research model that contributed to screening over 100,000 compounds in 2016 [25]. Since that time, the applications of the PDX models have seen a novel rise in prevalence within the research community. In 2017, the National Cancer Institute (NCI) launched PDXNet [26], its official PDX network database, to further enhance the accessibility and utilization of PDX models. Over the past two decades, given the importance of comprehensive annotation, robust logistics, and data resource contributions, numerous large-scale PDX biobanks have been progressively established in many countries or regions (Table 1).

Table 1.

Current large resources repositories of PDX

Name	Website link	Established year	Cancer type	Models
PDXNet	https://www.pdxnetwork.org/	2017	Pan-cancer	(i) PDX
Patient-Derived Models Repository	https://dctd.cancer.gov/drug-discovery-development/reagents-materials/pdmr	2012	Pan-cancer	(i) PDX (ii) Patient-derived in vitro tumor cell cultures (PDCs) (iii) Patient-derived organoids (PDOs) (iv) Cancer associated fibroblasts (CAFs)
Champions oncology	https://www.championsoncology.com/	2007	Pan-cancer	(i) PDX (ii) PDX Tissue Microarrays (TMAs) (iii) humanized PDX (iv) Cell-line derived xenograft (v) 3D cancer model (vi) Hematological VitroScreens (vii) Syngeneic models
MMCHdb	https://tumor.informatics.jax.org/mtbwi/index.do	1998	Pan-cancer	(i) Spontaneous and induced tumors in mice (ii) Genetically defined mouse models of cancer (inbred, hybrid, mutant, and genetically engineered strains of mice) (iii) PDX
Center for patient-derived models (CPDM)	https://www.dana-farber.org/research/integrative-research/center-for-patient-derived-models	2016	Pan-cancer	(i) PDX (ii) Organoids (iii) Spheroids (iv) 2D cultures
Washington University PDX development and Trialcenter (WU-PDTC)	https://pdx.wustl.edu/	2017	Pan-cancer	(i) PDX
Children’s Oncology Group (COG) cell culture and xenograft repository	https://www.cccells.org	2000	Pediatric cancer	(i) Cell lines (ii) PDX
Jackson laboratory	https://www.jax.org/	1929	Pan-cancer	(i) PDX (ii) Humanized mouses
Crown bioscience	https://www.crownbio.com/	2006	Pan-cancer	(i) PDX (ii) PDO (iii) Cell-line derived xenografts (iv) Syngeneic Models (v) Humanized GEMMs (vi) Tumor homografts (vii)Humanized Models (HSC & PBMC) (viii) Orthotopic/Metastatic & Imaging-enabled Models
Princess Margaret Living Biobank	https://pmlivingbiobank.uhnresearch.ca/	2017	Pan-cancer	(i) PDX (ii) PDO

With the rapid emergence of NAMs, both the NIH and the Food and Drug Administration (FDA) highlighted their importance in partially replacing animal experiments in 2023 and 2024. In 2025, the NIH announced that grants exclusively supporting animal models would no longer be provided. PDX 2.0 has been proposed as an integrative model that combines the respective strengths of traditional PDX and NAMs, offering a comprehensive framework for advancing translational cancer research. The timeline of key events in the development of PDX and PDX 2.0 is depicted in Fig. 1.

Fig. 1.

The timeline of key events in the developments of PDX and PDX 2.0. Since the first laboratory animal engraftment, the PDX model has undergone four stages. “Early Exploration and Preparation” showcases the attempts to construct early tumor animal models and the birth of nude mice. “Model Breakthroughs and Technological Innovations” showcases significant innovations in immunodeficient mice for PDX models. “Widespread Adoption and Development” showcases the mature application of PDX models in various cancer research studies and the emergence of PDX repositories. “PDX 2.0 era: integrative modeling paradigm” showcases the integration of conventional PDX models and NAMs in PDX 2.0 as an emerging and evolving modeling paradigm. SCID, Severe Combined Immunodeficiency; NOD/SCID, Non-Obese Diabetic/Severe Combined Immunodeficiency; NCI, National Cancer Institute; NIH, National Institutes of Health; NAMs, Novel Alternative Methods

Establishment of the PDX model

To avoid tumor engraftment rejection in mouse models, the key to establishing PDX models resides in the use of immunodeficient mice. Since the introduction of nude mice, progressive advancements in immunodeficient mouse models have enabled the development of optimized systems tailored to specific experimental requirements. The general process for creating a PDX model involves initially taking tumor cells or tissues obtained from patient surgical tumor specimens or biopsy samples (including circulating tumor cells, CTCs) with subsequent essential processing. Then, for varied research purposes, interstitial cells (e.g., fibroblasts), immune cells (e.g., tumor-infiltrating lymphocytes (TILs)), peripheral blood mononuclear cells (PBMCs), and basal membrane matrix components (i.e., Matrigel) that influence the primary tumor phenotype are co-transplanted with tumor tissue into selected immunodeficient mice under anesthesia via distinct routes [27]. Following the establishment of the first-generation (F1) PDX models, tumors are harvested upon reaching a predefined volume and re-transplanted into subsequent murine generations to generate F2, F3, and later PDX models, which are subsequently used for a wide range of research application [27]. Adherence to the principles of the 3Rs (Replacement, Reduction, and Refinement) is essential throughout the PDX modeling process. This includes the tumor screening, the use of trained staff, aseptic techniques, peri-operative care in the pre-engraftment phase, and systematic monitoring during the post-engraftment phase, involving ethology, health status, tumor-associated pathological signs, and severity assessment [28, 29].

In summary, we have proposed four factors that affect the success rate of PDX model establishment: (i) characteristics of the donor tumor itself, (ii) recipient mouse strain, (iii) selection of engraftment site, (iv) post-establishment analysis and evaluation (Fig. 2).

Fig. 2.

The various factors about the establishment of PDX models. The key factors that potentially impact the construction of PDX models have been delineated, including the characteristics of the donor tumor, the species of the recipient mice, the selection of engraftment way, and evaluation of the fidelity of xenograft tumors after successful construction

Characteristics of the donor tumor

Phenotypic features of the donor tumor

The success rate of engraftment when establishing PDX models is closely associated with tumor phenotype, with reflecting the significant variations in underlying driver factors, including tumor (sub)type, the tumor’s aggressiveness, tumor cell content in the tissue, and whether the tumor is primary or metastatic [23, 30]. A single-center study establishing 171 PDX models for gastrointestinal tumors revealed that primary tumor site (esophagus, stomach, colorectum), histology (adenocarcinoma, squamous cell carcinoma, signet ring cell carcinoma, others), and differentiation (poor, moderate, well) were significantly associated with PDX engraftment success (P < 0.05) [31]. The relationship between engraftment rate and tumor molecular biomarkers varies across cancer types, with Ki-67 serving as one illustrative example linked to higher engraftment success in aggressive subtypes [31–34].

Other features

A retrospective analysis of the osteosarcoma-derived PDX models showed that the model establishment success rate was related to whether the patient received chemotherapy. The success rate of the non-chemotherapy group (47.1%) was much higher than that of the chemotherapy group (17.4%) [35]. Consistently, in the development of an ovarian cancer PDX biobank, tumors that had been pretreated before engraftment exhibited significantly lower establishment rates [36]. This improved engraftment success may be attributed to the suppressive effects of chemotherapy on tumor progression, consistent with the aforementioned hypothesis. Moreover, in colorectal carcinoma studies, the method of tumor acquisition and the time required for engraftment may significantly affect the success rate of establishing PDX models [37]. Other features of the donor tumor influencing the success rate of engraftment also include sterility, size, and the freshness of human tumor samples [4, 23, 27, 30].

Recipient mouse strain

With respect to the species of recipient mice, the tumor model established by a single cell line cannot faithfully replicate the genomic and phenotypic characteristics of the donor tumor due to the lack of the native survival microenvironment of the tumor.

Nude mice, harboring a Foxn1 gene mutation that causes thymus degeneration, exhibit reduced T-cell numbers and compromised humoral and cellular immunity, serving as early immunodeficient models [38]. They were later replaced by Severe Combined Immunodeficiency (SCID) mice, which carry a Prkdc gene defect impairing V(D)J recombination and DNA double-strand break repair, resulting in a complete absence of functional T and B lymphocytes. SCID mice provide a more profoundly immunocompromised microenvironment, which enhances tumor growth support [21]. By backcrossing SCID mutations in the Non-Obese Diabetic (NOD)/Lt murine strain [22], scientists produced NOD/SCID mice, which have multiple defects in both adaptive and innate immunologic function and are suitable recipients for human hematopoietic stem cell engraftment. Based on NOD/SCID mice, to suppress residual natural killer (NK) activity and to improve transplantation efficiency, IL-2 receptor subunit gamma (IL-2Rγ) intracellular domain loss (NOG), IL-2Rγ intact loss (NSG), and Jak3-deficient mice (NOJ) have been generated [39–41]. BALB/c mice are among the most widely used inbred strains in immunology and cancer research, owing to their high susceptibility to infectious and neoplastic diseases. In the early twenty-first century, BALB/c mice and their derivative strains—BALB/c Rag-2null/IL2Rγnull (BRG) and Rag-2null/Jak3null (BRJ) mice—were widely adopted in immunology and cancer research, largely due to their unique phagocytic properties [42, 43]. In general, increasing levels of immunodeficiency are associated with higher engraftment rates, with the approximate hierarchy being BRG/BRJ > NSG > NOD/SCID > SCID > nude mice [3, 23].

Notably, in addition to classical mouse models, other animal species have also been employed to establish PDX models (Table 2).

Table 2.

Comparison of different preclinical models

Model		Construction difficulty	Genetic heterogeneity	TME fidelity	High-throughput drug screening capability	Advantages	Challenges
Mouse	PDX models	+ +	+ + +	+ +	+ +	(i) Preserve genetic & phenotypic heterogeneity of donor tumor (ii) Partially retain the TME (iii) Improve the predictivity of clinical drug responses	(i) Long latency and high cost establishment (ii) Murine stromal substitution (iii) Immunodeficiency background (iv) Limited modeling of metastasis (v) Tumor subclone depletion during engraftment and passaging
	Genetically-engineered mouse model (GEMM) [44–48]	+ + +	+ ~ + +	+	+	(i) Enable dissection of oncogenic driver-phenotype relationships (ii) Maintain native tumor–microenvironment co-evolution (iii) Model multistep tumor progression and metastasis	(i) Limited representation of human tumor heterogeneity (ii) Complex and time-consuming to construct (iii) Species-specific biological differences
	Immune-humanized PDX (iHu-PDX) models [49, 50]	+ + + +	+ + + ~ + + + +	+ + + +	+ ~ + +	(i) Superior standard for evaluating immunotherapies (ii) Studies of tumor-immune interactions	(i) Complex and expensive (ii) Inherited limitations of humanized mice (e.g., GVHD) (iii) Human-mouse physiological mismatches
Chick embryo chorioallantoic membrane (CAM) models [51–53]		+ +	+ ~ + +	+	+ + +	(i) Allow high-throughput and cost-effective experimentation (ii) Bypass mature adaptive immune rejection (iii) Facilitate real-time imaging and manipulation	(i) Species differences and immunodeficiency background (ii) Limited experimental duration (iii) Restricted modeling of advanced tumor
Pig models [54–58]		+ + + +	+ + + ~ + + + +	+ + + +	+	(i) Exhibit high anatomical and physiological similarity (ii) Enable advanced surgical and interventional studies (iii) Accommodate longitudinal studies	(i) High cost and logistical complexity (ii) Long generation time and slow experimental timelines (iii) Complex regulatory and ethical oversight
Larvae zebrafish models [59–61]		+	+ ~ + +	+ +	+ + +	(i) Allow high-throughput and cost-effective experimentation (ii) Permit xenograft without immune rejection (iii) Support optical transparency and real-time imaging	(i) Short transplant window (ii) Species differences and immunodeficiency background (iii) Short experimental window

Selection of engraftment site

Orthotopic engraftment

Patient-derived orthotopic xenograft models enable the tumor to develop in an environment that closely resembles the patient’s physiology, more faithfully recapitulates the native tumor microenvironment, including neoplastic progression, proliferation, invasion, metastasis, tumor–stroma interactions observed in clinical settings, as well as clinically relevant drug responses [62, 63]. At the same time, orthotopic models generally necessitate labor-intensive procedures, require intricate surgical techniques, incur higher costs, and frequently demand imaging modalities for monitoring tumor progression and response to therapy [64]. Orthotopic engraftment can improve engraftment success rates for brain cancers and metastatic tumors [27] and should be prioritized for metastasis-associated research [65].

Ectopic engraftment

1. Subcutaneous engraftment: This is the most commonly used approach due to its easy accessibility, straightforward engraftment process, and ease of monitoring [65]. Based on research objectives, the choice between subcutaneous and orthotopic engraftment should be strategically evaluated. A recent single-cell transcriptomic study of medulloblastoma demonstrated that subcutaneous PDX models accurately recapitulate the overall molecular characteristics of primary tumors to a degree comparable with intracranial PDX models, and may serve as a practical first-line modeling approach [66]. This preference appears to be context-dependent, influenced by tumor type and specific analytical endpoints.

2. Other approaches: Ascitic engraftment is common in ovarian and peritoneal cancers [67], a process closely associated with tumor progression. Hematologic cancers such as leukemia and lymphoma are readily modeled via tail vein injection, while breast cancer models can be effectively established through fat pad implantation (a subcutaneous variant) [68, 69].

Notably, the lack of standardized criteria for 'successful engraftment' remains a notable knowledge gap in PDX methodology [70]. Researchers may adapt this definition to specific research objectives and experimental constraints to ensure practical feasibility.

Post-establishment analysis and evaluation

Given that PDX models demonstrate divergent capacities in recapitulating the histopathological and molecular heterogeneity of donor tumors across experimental contexts, the verification of these characteristics post-establishment is critical for evaluating the fidelity of PDX models and providing essential biological annotation for downstream research.

Phenotypic heterogeneity

Tumor phenotypic information is dynamically shaped by the interplay of tumor-intrinsic drivers (genetic mutation and epigenetic modification) and tumor-extrinsic factors (tumor microenvironment (TME), metabolic reprogramming). To evaluate the concordance of donor tumor characteristics, researchers assess key gross macroscopic and histologic features, which provide an initial assessment of PDX model fidelity. These typically include architectural patterns (e.g., tumor-stroma relationship, epithelial morphology, microvascular invasion), cytologic features (e.g., nuclear atypia, pleomorphism, mitotic activity), and secondary changes (e.g., necrosis). These features are primarily observed using hematoxylin and eosin (H&E) staining and immunohistochemistry (IHC) for specific biomarkers [36, 71]. Meanwhile, the increasing application of machine learning approaches has transcended manual interpretation in PDX data analysis, enabling automated, efficient, and precise data processing, which enhances the applicability and reliability of PDX-derived data for cancer research [72, 73].

Molecular heterogeneity

Comprehensive biological annotation at the molecular level across large PDX panels is essential to assess model fidelity, understand activating oncogenic events, and discover druggable targets that yield critical biological insights and directly guide clinical actionability. The utilization of PDX models, combined with multi-omics technologies including transcriptomics, genomics, epigenomics, proteomics, and other emerging omics modalities, has become pivotal in deciphering tumor drug resistance mechanisms and enabling iterative optimization of molecularly targeted therapies in response to evolving tumor heterogeneity [74, 75]. A study comprehensively characterized the genomic features of PDX models based on current big PDX repositories across 25 cancer types, together with whole-genome duplications (WGDs), oncogenic driver mutations, and gene fusions to identify numerous key oncogenic drivers and potential therapeutic targets consistent with findings from The Cancer Genome Atlas (TCGA studies [76]. Additional examples include proteomic analyses of non-small cell lung cancer (NSCLC) PDX models that identified two distinct proteotypes associated with multiple druggable targets [77], and multi-generational pancreatic ductal adenocarcinoma (PDAC) PDX models characterized using RNA-seq (RNA sequencing) and 4D-DIA mass spectrometry, revealing differentially expressed genes and proteins that offer insights into tumor evolution and potential therapeutic strategies [78]. In metabolomic analysis, Aparna D. Rao et al. [79] used [U-¹³C] glucose tracing and liquid chromatography–mass spectrometry (LC–MS) to analyze 23 melanoma PDX models and their donor tumors, revealing both conservation and divergence in metabolic phenotypes. Metabolic conservation was observed in host-shaped metabolic drift affecting glycolytic and pyrimidine pathways, whereas divergence manifested as stable, tumor-originating metabolic fingerprints that were retained across passages, particularly in mitochondrial metabolites.

Furthermore, single-cell multiomics profiling of PDX models at cellular resolution enables refined tumor cell subtyping, characterization of therapy-induced dynamic adaptations, identification of predictive biomarkers and novel therapeutic strategies [75, 80]. For example, children and adults with high-risk CRLF2-rearranged, JAK pathway–activated Philadelphia chromosome–like acute lymphoblastic leukemia (Ph-like ALL) exhibit heterogeneous responses to TKI therapies. Through scRNA‑seq and snATAC‑seq on Ph‑like ALL cells obtained from PDX models after treatment with ruxolitinib or dasatinib, a novel AP-1–regulated senescent stem/progenitor-like subpopulation was identified. This dormant leukemic subset was selectively eliminated by dual pharmacologic inhibition of BCL-2 using the senolytic agent venetoclax, in combination with JAK/STAT inhibition by ruxolitinib or SRC/ABL inhibition by dasatinib [81].

Therapeutic analysis

This highlights the preserved drug sensitivity/resistance profiles in PDX models, which closely mimic the responsiveness of the donor tumors to pharmacological agents. The establishment of initial PDX models provides a foundation for investigating pharmacological responses [82]. Therapeutic investigations encompassing both standard regimens and novel agents also serve as an essential indicator for evaluating the success of PDX model establishment. Numerous studies have demonstrated that the response of newly established PDX models to drugs is intricately linked to the mutational status and frequency of specific genes. Accurate assessment of therapeutic efficacy depends on the availability of well-designed drug perturbation data [83]. Furthermore, integrative machine learning algorithms applied to PDX models—exemplified by the CeSta classifier [84], a stacked ensemble architecture—achieve superior predictive accuracy for drug response and identify interpretable biomarkers. Having been trained on multiomic data derived from extensive CRC PDX collections, CeSta successfully modeled cetuximab sensitivity and identified key transcriptional biomarkers, which outperformed conventional predictive methods. The categorization of therapeutic response in PDX models based on initial tumor volume change—classified as < 100% (responder), 100–300% (partial responder), or > 300% (non-responder) according to Response Evaluation Criteria in Solid Tumors (RECIST)—may serve as a reference [75].

The PDX Minimal Information (PDX-MI) standard defines a unified reporting framework for PDX models, comprising four core modules (clinical information, model creation, model quality assurance, and model study) and an associated metadata category. This standardized framework aims to enhance data interoperability, support resource sharing, and improve reproducibility in preclinical cancer research [85].

Applications of the PDX model

PDX models have been applied across a wide range of clinical research contexts. Here, four key application areas are summarized to highlight their most critical roles.

Biomarker identification and drug screening

PDX models are primarily used to support the development of anticancer agents and the personalization of oncological therapies. Comprehensive molecular annotation to identify predictive biomarkers, together with high-throughput drug screening to inform therapeutic selection, are prerequisites for defining optimal treatment strategies in cancer management [86]. Building upon this foundation, the molecular profiling of PDX models through multiomics analysis and its correlation with therapeutic outcomes enhance the identification of predictive biomarkers and elucidate the underlying molecular mechanisms of drug resistance crucial, thereby informing the development of novel therapeutic strategies [71, 87, 88]. PDX models exhibit a relatively high degree of predictability regarding pharmacodynamic responses, which substantiates their utility in prioritizing potential therapeutic indications and assessing the efficacy of individual anti-cancer interventions for specific patient cohorts, supporting its prognostic and predictive utility [89]. In practice, PDX-based pharmacological screening is commonly applied in two contexts. When drug mechanisms are well characterized, PDX models can be used to identify optimal therapeutic regimens, with subsequent validation in clinical trials to inform standardized treatment guidelines. When pharmacological mechanisms remain incompletely understood, PDX models enable preliminary efficacy assessment prior to clinical validation [90]. However, the establishment complexity, time requirements, and risk of engraftment failure limit the scalability of PDX models for high-throughput drug screening.

To address these constraints, several attempted solutions have been developed. The “one animal per model per treatment” (1 × 1 × 1) protocol, which leverages large-scale PDX collections to support high-throughput drug screening (HTDS) [91, 92]. In addition, OncoVee MiniPDX [93] provides a capsule-based implant system that enables rapid in vivo assessment of drug sensitivity. Complementary approaches, such as intratumoral drug-delivery microdevices for treatment response prediction [94], and the integration of machine learning algorithms (e.g., the CeSta classifier [84]) further enhance the efficiency and interpretability of PDX-based screening studies. combining PDX models with PDOs and drug-resistant patient-derived cells establishes a multidimensional preclinical platform that supports high-throughput screening and rigorous validation, facilitating the identification of drug response profiles in treatment-resistant cancers [95].

PDX preclinical trials and precision medicine

It is acknowledged that due to the inherent intratumoral and intertumoral heterogeneity, together with the dynamic temporal evolution, no universally applicable or static therapeutic regimen can treat all patient populations. In the management of malignant tumors, pronounced interpatient variability exists in responses to first-line therapies, ranging from high sensitivity to complete resistance. Even the response magnitude of the patients exhibiting chemosensitivity is variable, underscoring the need for individualized pharmacotherapeutic strategies to optimize clinical outcomes. PDX models faithfully recapitulate donor tumor heterogeneity and demonstrate relatively high predictive value for therapeutic efficacy, particularly in the evaluation of hypothesis-driven combination therapies. This process enables the alignment of critical pharmacological parameters—including drug class, compatibility, and dosing—to identify the most effective treatment regimen, whose efficacy can be validated in corresponding PDX models prior to clinical application. Consequently, PDX models provide a pivotal, evidence-based decision-making platform within the precision oncology framework [1, 96]. Furthermore, preclinical population-based studies established by transplanting diverse human tumor samples into immunodeficient mice can mitigate the influence of uncharacterized heterogeneity on experimental outcomes, thereby laying a robust foundation for subsequent drug development efforts [97].

Patient disease trajectories vary according to the specific molecular alterations inherent to individual tumors. Accordingly, precision medicine increasingly relies on comprehensive biomarker annotation to elucidate resistance mechanisms, enabling patients to undergo extensive tumor profiling and receive matched therapeutic interventions [98]. As demonstrated by Wong et al. in the Zero Childhood Cancer Program, at least one reportable molecular aberration was identified in 93.7% of patients with high-risk pediatric cancers, with 71.4% harboring potentially actionable therapeutic targets. The comprehensive molecular profiling not only uncovered driver mutations, gene fusions, copy number variants (CNVs), and aberrant gene expression patterns, but also identified resistance mechanisms, diagnostic reclassification, and germline cancer predisposition variants [99]. Together, these advances facilitate truly personalized treatment based on systematic tumor molecular characterization and informed clinical interpretation, thereby translating molecular discoveries into actionable therapeutic decisions. Because PDX models are typically generated from therapy-resistant tumors, they offer a powerful platform for elucidating drug resistance mechanisms. Leveraging extensive PDX repositories, researchers can systematically evaluate novel therapeutic strategies using a direct “one-to-one” correspondence between each PDX model and its donor tumor [100–102]. The demonstrated capacities of PDX models represent a sustainable opportunity to develop innovative therapeutic strategies and to drive further clinical trials aimed at advancing personalized cancer medicine, as exemplified by in pancreatic cancer [103, 104], NSCLC [105] and other genetically defined tumors [106].

For rare cancers characterized by pronounced intertumoral heterogeneity, extensive PDX collections generated via serial passaging to overcome limited tumor specimen availability and to support downstream drug perturbation studies for therapeutic discovery and personalization. This approach is exemplified by a high-throughput functional drug screen demonstrating that combined inhibition of the antiapoptotic protein Bcl-xL with TOPO1, HDAC inhibitors, and napabucasin enhances efficacy in fibrolamellar carcinoma (FLC), a rare liver cancer subtype [107]. Moreover, highly aggressive rare sarcomas—including metastatic soft tissue sarcoma, mesenchymal chondrosarcoma, osteosarcoma, dedifferentiated liposarcoma, and pseudomyxoma peritonei—can similarly benefit from PDX-based platforms for treatment development and validation [108–112]. The NCI Patient-Derived Models Repository (PDMR) has established over 500 PDX models spanning 10 categories of rare cancers [113], providing valuable resources for therapeutic target identification, novel drug discovery, and preclinical safety and efficacy evaluation. A summary of PDX models driving clinical trials is presented in Fig. 3.

Fig. 3.

The main applications workflow of PDX models. The essential applications and workflow of PDX models in clinical research are outlined as follows: a, Tumor samples from patients with drug-sensitive or resistant cancer are used to establish PDX models, which are then passaging and randomized for drug efficacy assessment. b, Experimental results guide multi-omics analyses at cellular and molecular levels to identify new molecular biomarkers. c, Preclinical trials involve the development of new drugs or combinations, with comprehensive molecular profiling on PDX models of individual patients for precision therapy prototyping. d, PDX models are employed alongside clinical trials to supply supplementary data. Following drug approval, clinical dosing is informed based on real-world patient scenarios. Finally, the developed interventions can be directly translated to guide treatment for patients sampled at an early cancer stage, whereas samples obtained at advanced stages may inform clinical guidelines for future patients with matching molecular features. F0, tumor specimen procured via biopsy or surgery for subsequent engraftment; F1, the first generation PDX models following engraftment into immunodeficient mice, F2, the second generation, and so on; SEN, sensitive to medications, REN, resistance to medications; Re, regimen; Q&D, research and development

Co-clinical trial

In cancer clinical research, certain sporadic tumors require prolonged drug development cycles owing to limited biological characterization, and overall clinical trial success rates remain modest. In this context, the integration of PDX models offers a promising strategy to enhance the efficiency and translational relevance of clinical trials. Given their capacity to faithfully recapitulate tumor biology and predict clinical outcomes, the deployment of PDX models prior to or in parallel with Phase I/II trials is increasingly recommended. This co-clinical approach enables real-time evaluation of therapeutic efficacy in both patients and corresponding models, thereby facilitating the identification of predictive biomarkers and targetable mechanisms of adaptive resistance [114]. By contrast, omitting preclinical animal studies or proceeding directly to patient testing without complementary validation from parallel PDX modeling may be time-consuming and carries a significantly higher clinical risk. The result of the “one-time therapeutic strategy per patient” design in many clinical trials allows limited tolerance for error, often constraining opportunities for extended clinical observation [71, 115]. To support co-clinical trial initiatives, the establishment of integrated infrastructures—including mouse hospitals, comparative pathology platforms, and bioinformatics pipelines—have been developed as a strategic priority [116].

The translational value of this approach has been illustrated in a zanidatamab Phase I study, in which a cohort of HER2-positive PDX models spanning multiple tumor types was established in parallel with patient enrollment. The overall co-clinical Trial of PDX models encompasses two components: (i) to validate zanidatamab activity in HER2-expressing xenografts and demonstrated concordance with patient responses of pretreatment or post-progression biopsies, (ii) to identify MET and MYC amplification as potential mechanisms of acquired resistance and supporting rational combination strategies with MET inhibitors [117]. For rare cancers, in which large-scale clinical trials are often infeasible, PDX models can further complement clinical studies by providing predictive efficacy data and serving as synthetic controls alongside real-world evidence. In the Phase II STATICE trial, co-clinical evaluation using HER2-expressing uterine carcinosarcoma (UCS) PDX models showed response rates comparable to those observed in patients treated with trastuzumab deruxtecan (67% versus 70%) demonstrating a reasonably accurate prediction of clinical efficacy [118].

The principal applications of PDX models discussed above are summarized in Table 3. Notably, experimental procedures are often tailored to specific research objectives, incorporating additional elements as necessary.

Table 3.

Summary of PDX—related clinical trials from 2021to 2025

Study Title	Study Status	Conditions	Interventions	Study Type	NCT Number
Patient-Derived Xenograft (PDX) Modeling in Adult Patients With Metastatic or Recurrent Sarcoma	TERMINATED	Sarcoma	OTHER: PDX drug sensitivity testing	INTERVENTIONAL	NCT02720796
Development of Patient-Derived Xenografts (PDX) in Patients With Breast Cancer	RECRUITING	Breast Cancer/Residual	PROCEDURE: chemotherapy or endocrine therapy for breast cancer	OBSERVATIONAL	NCT04703244
A Platform of Patient-Derived Xenografts (PDX) and 2D/3D Cell Cultures of Soft Tissue Sarcomas (STS)	RECRUITING	Soft Tissue Sarcoma/Xenograft Model/2D/3D Cell Cultures	OTHER: tumor biopsy	INTERVENTIONAL	NCT02910895
Pembrolizumab for Metastatic NSCLC Patients Expressing PD-L1 Who Have Their Own PDX	COMPLETED	Metastatic Non-Small Cell Lung Carcinoma	DRUG: Pembrolizumab Injection	INTERVENTIONAL	NCT03134456
Evaluation of Combined Sensitising and Hypomethylating Therapy Outcomes in AML PDX	RECRUITING	AML—Acute Myeloid Leukemia	NA	OBSERVATIONAL	NCT06782971
Reconstitution of a Human Immune System in a Patient Derived Xenograft (PDX) Model of Genitourinary (GU) Cancers	TERMINATED	Genito Urinary Cancer/Bladder Cancer/Kidney Cancer/Prostate Cancer	PROCEDURE: Bone marrow biopsy/PROCEDURE: Tumor biopsy	OBSERVATIONAL	NCT03134027
Personalized Mini-PDX for Metastatic CRPC	UNKNOWN	Prostate Cancer	OTHER: MiniPDX Group	INTERVENTIONAL	NCT03786848
Patient-Derived Xenografts in Personalizing Treatment for Patients With Relapsed/Refractory Mantle Cell Lymphoma	COMPLETED	Recurrent Mantle Cell Lymphoma/Refractory Mantle Cell Lymphoma	OTHER: Best Practice/DRUG: Ibrutinib/OTHER: Patient Derived Xenograft/OTHER: Personalized Medicine	INTERVENTIONAL	NCT03219047
PDX Models From EGFR Mutant Tumors	WITHDRAWN	Non Small Cell Lung Cancer	NA	OBSERVATIONAL	NCT03872440
Personalized Patient-Derived Xenograft (pPDX) Modeling to Test Drug Response in Matching Host	RECRUITING	Colorectal Neoplasms/Colorectal Cancer/Breast Cancer/Breast Neoplasms/Ovarian Cancer/Ovarian Neoplasm	OTHER: Molecular Profiling & In Vivo drug testing in pPDX and organoid cultures	OBSERVATIONAL	NCT02732860
Patient-derived Xenograft Models of Tumor From Patients With Head and Neck Cancer	RECRUITING	Squamous Cell Carcinoma of the Head and Neck	PROCEDURE: Local biopsy in the tumor	OBSERVATIONAL	NCT02572778
Xenografts Development from Surgical Tumor Samples of Patients with Triple Negative or Luminal B Breast Cancer	RECRUITING	Breast Cancer Female	BIOLOGICAL: blood samples collection	INTERVENTIONAL	NCT04133077
Estrogen Receptor-Positive Breast Cancer Patient-Derived Xenografts	COMPLETED	Breast Cancer	NA	OBSERVATIONAL	NCT02752893
Hyper-Personalized Medicine Using Patient-Derived Xenografts (PDXovo) for Metastatic Solid Tumors	UNKNOWN	Kidney Neoplasm/Carcinoma, Renal Cell/Metastatic Solid Tumor	NA	OBSERVATIONAL	NCT04602702
Patient-Derived Breast Cancer Xenografts	TERMINATED	Metastatic Breast Cancer	NA	OBSERVATIONAL	NCT01750164
Lung Cancer Organoids and Patient-Derived Tumor Xenografts	RECRUITING	Lung Cancer	Other: Tissue and blood	OBSERVATIONAL	NCT05092009

Cancer nanomedicine

Nanomedicine is an innovative field that entails the development, modification, and manipulation of substances at the nanoscale to diagnose, treat, and prevent diseases. Compared to conventional antitumor drugs, nanotherapeutics offer advantages including high drug payload, targeted delivery, controlled release, and compatibility with localized treatment modalities (e.g., radiotherapy, ultrasound, hyperthermia). Collectively, these features can improve drug bioavailability, prolong circulating drug concentrations, and enhance the efficacy of combination therapies [119]. Nevertheless, fundamental barriers hinder the clinical translation of extensive nanoparticle-based research, with many candidates failing to progress from Phase I to Phase II clinical trials [120, 121]. These barriers primarily arise from two sources: firstly, limitations in the innovation of nanomaterial applications, such as constrained development of the physicochemical properties of nanoparticles, and insufficient understanding of nano-bio interactions affecting nanoparticle delivery, distribution, metabolism, and excretion; and secondly, practical hurdles in the clinical translation, encompassing preclinical evaluation, trial design, regulatory approval, and commercialization, necessitating integrated efforts across multiple stakeholders [122].

For the first instance, rigorous experimental design—including appropriate engraftment site selection, randomized controlled studies, and predefined “go/no-go” decision criteria—should be aligned with specific scientific objectives [123]. A critical yet often underappreciated contributor to this discordance is the complex interplay between nanotherapeutics and the tumor immune microenvironment. Nano-immuno interactions, encompassing modulation of immune signaling pathways, polarization of tumor-associated macrophages, and therapy-induced metabolic and epigenetic reprogramming, can profoundly influence therapeutic efficacy and treatment response [119, 124]. In this context, the development of humanized mouse models carrying PDX tumors (i.e., immune-humanized patient-derived xenograft, iHu-PDX), which aim to reconstitute functional human immune systems, represents an important direction for advancing nanomedicine research. Additional factors, such as age- and sex-related biases, microbiome interactions, and limited in vivo passaging, also warrant careful consideration to enhance translational relevance [125], More broadly, PDX-based platforms support robust assessment of nanoparticle biodistribution, intratumoral penetration, and therapeutic efficacy under clinically relevant conditions. Longitudinal analyses within these systems can further inform resistance mechanisms and biomarker discovery, thereby guiding the rational design of next-generation nanomedicines and supporting clinical trial development.

Challenges of PDX and evolution of PDX 2.0

With the advent of deep sequencing, increasing evidence has raised concerns regarding the fidelity of PDX models in preserving the biological characteristics of donor tumors. At the same time, emerging technologies and expanding research demands further complicate their translational application. In response, technological advancements are transforming conventional PDX systems into a next-generation framework—PDX 2.0—characterized by enhanced precision, integrative capability, and data-driven analysis. This review outlines the major challenges confronting PDX models and proposes corresponding strategic solutions to guide future investigations (Fig. 4).

Fig. 4.

Challenges and opportunities of conventional PDX model. The most important challenges and the latest research directions in PDX construction and application are summarized, along with the corresponding solutions that can be attempted. TME, tumor microenvironment; MSCs, human mesenchymal stem cells; GEMM, genetically engineered mouse model; TIL, tumor infiltrating lymphocyte; PBMC, peripheral blood mononuclear cell; CD34 + HSC, CD34 + hematopoietic stem cell; PDOX, patient-derived organoid xenograft; PDXO, patient-derived xenograft organoid; OOC, organ-on-a-chip; OrgOC, organoid-on-chip; CDX, cell line-derived xenograft

Time lag of PDX establishment

When deployed in parallel with clinical trials, PDX studies can provide clinically relevant functional insights. Nonetheless, the establishment and expansion of PDX models typically require more than six months (approximately 2–8 months), causing a delay that is misaligned with clinical treatment timelines. This temporal disconnect precludes timely efficacy assessment for advanced cancers and may lead to missed optimal therapeutic windows [9, 96]. Moreover, incorporating PDX-validated treatment plans into broader clinical guidelines would further exacerbate this constraint. As such, PDX models may not be applicable for guiding treatment for the original patient but could facilitate the identification of precision oncology strategies for subsequent patients with similar molecular traits (Fig. 3).

In contrast, tumor cell lines cultured in vitro proliferate rapidly and lose dependence on the original tissue architecture. Owing to their cost-effectiveness and operational simplicity, they are well suited for mechanistic studies and high-throughput pharmacological screening, thereby serving as a preliminary validation platform to reduce the experimental burden associated with PDX studies. The HTS384 NCI60 platform, a next-phase high-throughput screening tool, employs 384-well plates with a 3-day drug exposure period and a CellTiter-Glo luminescent endpoint. It has screened more than three dozen groups of targeted therapies from over 1,000 FDA-approved anticancer agents, demonstrating high reproducibility and comparability across independent screens [126]. Additionally, the larval zebrafish xenograft model allows for subcellular monitoring of tumor behavior just five days post-transplant, offering valuable insights into the early responses without the need for immunosuppression and requiring only minimal tissue samples. Accordingly, zebrafish serve as an efficient host for tumor xenografts, substantially shortening experimental cycles of PDX models while meeting clinical animal model standards [127]. Furthermore, as previously mentioned, the miniPDX platform—a rapid functional screening tool that arrays patient-derived tumor cells in hollow fiber capsules for subcutaneous implantation in mice—achieves complete drug-response evaluation within 7 days, with sensitivity, specificity, and both positive and negative predictive values exceeding 80% [93].

Issues in the PDX engraftment and passaging

Substitution of tumor stroma

Components of the TME, including the extracellular matrix (ECM), secreted cytokines, immune cells, physicochemical conditions, and stromal cells such as endothelial cells, pericytes, and CAFs, are increasingly recognized as critical regulators of tumor growth, initiation, cancer stem cell maintenance, and metastasis, thereby influencing responsiveness to specific therapies [128]. Generally, from F2 onwards, the human tumor stroma is progressively replaced by murine stroma; concurrently, the three-dimensional architecture of the stroma and adjacent tissues degrades throughout the passages [129, 130]. Alterations in the host microenvironment, as well as the variable quantity of mouse stroma contamination across passages, also contribute to copy number changes between patient tumors and corresponding PDX tumors [131]. Moreover, environmental stress adaptation caused by stroma replacement can influence the proportions of PDX tumor subtypes. Due to stromal transition in PDX, there is a decline in aberrantly differentiated endocrine exocrine (ADEX) and immunogenic PDX models subtypes in PDAC classified by Bailey gene signatures, with no overlapping genetic features across other subtypes [132]. Additionally, this replacement could result in a more antagonistic signaling barrier existing in the context of interspecies chimerism. Research demonstrated that this signaling barrier mediated by the activation of the NF-κB signaling pathway in human cells, which are often eliminated through contact-dependent interspecies cell competition [133]. Beyond biochemical signaling, this potential incompatibility in murine-human cell communication is also reflected in structural barriers. Ballard et al. [134] identified impaired cell adhesion, specifically the inability to form robust E-cadherin junctions, as a key reason why human cells are physically extruded from developing murine embryos. These findings suggest that stromal alterations confer unique molecular and cellular characteristics to PDX models, creating a microenvironment distinct from that of the primary tumor.

Hence, solutions to these challenges fall into two main categories. (i) Establishing TME and restoring tumor-stroma interactions: Selection of NSG mice for PDX models can help sustain the human stromal component [135]. Genetically engineered mice that preserve human components are used to mimic better the human immune milieu [136, 137]. (ii) Advanced bioinformatics tools isolate species-specific signals from mixed sequencing data to profile TME cell traits separately. Host–graft read separation tools such as XenofilteR [138] and Xenome [139] assign mouse- and human-derived reads in xenograft samples, thereby removing cross-species contamination. Moreover, deconvolution tools such as CIBERSORTx [140] and RCTD [141] can estimate the relative abundances of stromal and other cell types from bulk or spatial transcriptomic data. Additionally, it has been reported that intravital imaging of the dynamic TME using the zebrafish model to analyze tumor-stromal crosstalk and dynamic adaptation represents a promising avenue [142]. Notably, the analysis of PDX tumors introduces unique complexities compared to human samples. The absence of patient-matched germline DNA and the contamination from murine stromal cells necessitate tumor-only variant calling algorithms and stringent filters to remove mouse-derived reads and ensure analytical fidelity [76]. It is recommended to establish an evaluation framework to categorize and grade the contributions of stromal characteristics across cancer cells, while tracking immune and non-immune dynamics in the TME throughout treatment.

Clone selection and genetic alteration

Clonal selection is a universal phenomenon during the engraftment and serial propagation of PDX models, resulting from genomic evolution (i.e., the genetic alteration of the tumor genome over time under environmental selective pressures) [143]. Furthermore, the continuous accumulation of copy number alterations (CNAs) during this process and their increasing divergence from the primary tumor genome could influence the model’s drug response [144]. It is widely considered that this clonal selection is reproducible and deterministic, representing a non-random process driven by fitness advantages, whereby distinct PDX models from the same tumor may converge to define common subclones with the highest adaptive fitness [143–145]. However, the direction of clonal selection in PDX models may not fully recapitulate that of the donor tumor, as clonal selection can alter the representation of driver mutations in these models. This divergence may be attributable to alleviating previous immune constraints due to propagation in immunodeficient mice [145]. The absence of previous immune constraints, including innate immunity deficits, including NK cell deficiency and low inflammasome activity, as well as adaptive immune impairment, such as reduced HLA gene expression, may enable immunogenic subclones C_pdx that were suppressed in the primary tumor to become dominant in PDX models. What’s more, during serial engraftment, various subclones from a single primary tumor can form distant PDX lineages (intra-sample subclonality [76, 146]), reflecting the stochastic nature of sampling within spatially heterogeneous tumors and leading to lineage-specific drift, which enhance the purity of PDX models during the passaging process [131]. These observations warrant caution in relying on single-region PDX models for pharmacodynamic analysis and the importance of parallel molecular target identification and sequencing throughout the experimental timeline [131, 132, 146, 147].

Conversely, the specific PDX subclones emerging from clonal selection enable dynamic tracking of the progression of clinical resistant subclones from latency to dominance, thereby informing target selection and drug development. Based on the evolution of PDX models during establishment and passaging, Playa-Albinyana, H. et al. delineated the progression of chronic lymphocytic leukemia (CLL) toward Richter transformation (RT)—a resistant form of diffuse large B-cell lymphoma—characterized by a “latency, expansion, dominance” sequence. Using resultant RT-PDX models with a high oxidative phosphorylation (OXPHOS) profile, they demonstrated that the OXPHOS inhibitor IACS-010759 overcomes venetoclax resistance in CLL [148].

Passaging PDX models raises concerns regarding the stability of genetic information, which can vary depending on the cancer type, sequencing methods, sample handling, and the definition of "consistency" and "inconsistency", and more importantly, the spatial heterogeneity of the donor tumor and contamination by mouse sequences [10, 131, 144, 149]. Some hallmark cancer events, such as genes located within CNAs, differed between early and late passage PDX models in many studies [150, 151], whereas late-passaged (F6) PDX models demonstrated a relatively stable retention of morphological, genomic, and metabolic characteristics [36, 79]. It’s reiterated that multi-region sampling for generating multi-lineage PDX models is critical for capturing diverse tumor subclones and better recapitulating intra-tumor heterogeneity (spatial heterogeneity) while improving the success rate of model establishment [36, 132, 146].

Problems with metastasis

Metastasis refers to the dissemination of tumor cells from the primary site to distant organs, representing one of the adverse behaviors of malignant neoplasms [152]. Approaches for introducing preclinical models of metastasis can be categorized into three types: (i) ectopic and orthotopic transplantation of primary tumors, (ii) injection into circulation, and (iii) direct transplantation into metastatic sites [153]. Ectopic transplantation (e.g., subcutaneous) permits tumor growth but often fails to fully recreate the native TME [154]. By contrast, orthotopic transplantation more faithfully reproduces the metastatic cascade but is often associated with a low incidence of metastasis and fails to model the early stages of cellular transformation and tumorigenesis observed in autochthonous tumors [13]. Injection into circulation (e.g., tail vein [155], intracardiac [156], or intrasplenic [157] injection) facilitates the study of metastatic dormancy and colonization, particularly with respect to organotropism. Direct transplantation into metastatic sites (e.g., intracranial [158] or intrahepatic [159] injection) allows focused analysis of metastatic outgrowth and colonization mechanisms. PDX models can be used to investigate general metastasis principles and evaluate drug responses for testing potential treatments anti-metastatic treatments [154, 160, 161]. Nevertheless, PDX models carry inherent limitations, including tumor stroma replacement, clone selection, genetic alteration during engraftment and passaging, and the absence of a functional immune tumor microenvironment, all of which may hinder in-depth mechanistic investigations into specific aspects of metastasis and therapeutic response.

No single preclinical model is universally suitable for studying tumor metastasis, and the choice of appropriate model should therefore be guided by its specific strengths and limitations relative to the research objective. GEMM-based autochthonous models driven by defined oncogenic alterations can recapitulate the entire process of spontaneous metastasis, in contrast to orthotopic PDX models, which fail to model the early stages of metastatic initiation. Conversely, cancer cell lines and tumor organoids can be introduced into the circulation or directly implanted into metastatic sites to mimic later phases of metastasis, including hematogenous dissemination and metastatic outgrowth [13].

Integrated considerations for integrated models

In clinical research, a single model rarely satisfies all experimental requirements; therefore, combining complementary models and selecting them based on their respective strengths and limitations is essential.

Humanized mice and immune-humanized PDX (iHu-PDX) models

The absence of a functional immune microenvironment due to the use of immunodeficient mice, fundamentally limits the utility of PDX models for investigating tumor–immune interactions and evaluating immunotherapies, representing a major constraint [2]. Humanized mice, also referred to as human hematolymphoid chimeric mice or human immune system (HIS) models, are extensively utilized in immuno-oncology research through the co-engrafting human tumors and immune components into immunodeficient mouse hosts [12]. Humanized mice are primarily categorized into three types: (i) human peripheral blood leukocyte (Hu-PBL) SCID mice involve the engraftment of peripheral blood mononuclear cells (PBMCs) into NOG or NSG mice and are primarily used to evaluate T cell–based immunotherapies. However, these models frequently develop acute graft-versus-host disease (GVHD) mediated by human T cells, leading to a limited experimental window and potential model mortality [162]. Strategies to mitigate GVHD have led to the development of modified models, including NSG-MHC-DKO and NOG-dKO mice with MHC class I and II inactivation [163], HUAMICE models expressing human HLA-A2 and HLA-DR1 [164], NOG-hIL-4-Tg mice that convert T cell CD8⁺ to CD4⁺ T cells via IL-4 overexpression [165], Rag1/IL2rg/CD47 triple-knockout (TKO) mouse strain [166], and NSG-Tg (CMV-IL3, CSF2, KITLG) 1Eav/MloySzJ (NSGS) mice reconstituted with cord blood mononuclear cells (CB-MNCs) [167]. (ii) Human SCID repopulating cell (Hu-SRC) mice are generated by transplanting CD34⁺ hematopoietic stem and progenitor cells (HSPCs) from cord blood (CB), bone marrow (BM), or fetal liver (FL), enabling long-term and more complete hematopoietic system reconstitution. Nevertheless, the significant challenges Hu-SRC murine model faced include impaired development of human MHC (HLA)-restricted T cells, limited antibody class switching in B cells, and insufficient reconstitution of human innate immune lineages [12, 162]. Transgenic expression of human cytokines in immunodeficient mice, as exemplified by NSG-SGM3 and MISTRG models, represents an effective strategy to reconstitute human innate immunity [168]. The NOGW model derived from NOG mice with a c-kit W41 mutation improves long-term HSC maintenance, and its advanced derivative NOGW-EXL expresses human IL-3 and GM-CSF to support multilineage reconstitution [169]. (iii) BLT humanized mice are generated by engrafting human fetal liver–derived CD3⁺ HSPCs and autologous fetal thymus tissue into SCID mice to facilitate HLA-restricted T cell maturation. Limitations of BLT mice include limited fetal tissue availability, surgical complexity, and frequent GVHD-like wasting syndromes in most strains [162]. Stem cell transplantation conducted at an early circadian phase could reduce acute GVHD severity and improve survival rates [170].

Collectively, the above challenges and targeted technical advances significantly expanded the applicability of humanized mice in tumor immunology research. Expending into the field of immuno-oncology research, integrating humanized mice with PDX platforms has enabled the development of iHu-PDX models, which provide a powerful system for investigating immune targets and evaluating immunotherapies in precision oncology. A representative study by Scherer et al. [171] developed the first iHu-PDX model of endocrine-resistant ER⁺ breast cancer harboring a natural ESR1 mutation. This model was established by engrafting CD34⁺ hematopoietic cells into NSG-SGM3 mice followed by implantation of estrogen-independent HCI-013EI PDX tumors, effectively recapitulating the clinical immune TME and endocrine-resistant phenotype. Similarly, one study established iHu-PDX models of melanoma and confirmed that an immunologically "cold" TME is a primary driver of resistance to anti-PD-1 therapy [172]. Separately, evaluation of anti-PD-1 response in iHu-PDX models is performed by assessing the extent of CD3⁺ T cells infiltration [173]. In these studies, patient-derived tumors were implanted into pre-established humanized mice, allowing tumor evolution under immune selective pressure while reducing the risk of GVHD. Also, iHu-PDX models have also been generated by introducing human immune components directly into established PDX tumors [174–176]. Overall, iHu-PDX models integrate the complementary strengths of humanized immune systems and patient-derived tumors, enabling more physiologically relevant investigation of tumor–immune interactions, dissection of immunotherapy resistance mechanisms, and support for precision oncology–oriented therapeutic stratification, thereby representing a promising and evolving platform for future translational cancer research.

PDX-Derived Cells (PDCs)

The derivation of cell lines from PDX tumors, known as PDX-derived cells (PDCs), provides a cost-effective, rapid, and high-throughput platform for drug sensitivity screening and biomarker validation. The paired “PDX–PDC” platform enables a closed-loop preclinical workflow, facilitating in vitro drug screening followed by in vivo efficacy validation [177]. However, PDCs have inherent limitations, including potential loss of tumor heterogeneity and the absence of TME recapitulation in 2D culture. In HER2⁺ CRC, PDX-derived cells (xeno-cells, XL) demonstrated cetuximab resistance mediated by RAS/BRAF mutations or ERBB2 amplification, while also establishing lapatinib and trastuzumab combination efficacy [177]. In another study of cholangiocarcinoma (CCA), a panel of 69 anticancer agents was screened using PDCs, identifying proteasome inhibitors such as bortezomib and carfilzomib as selectively cytotoxic in PTEN-deficient CCA PDCs. This screening further revealed the “PTEN–AKT–FOXO1–BACH1/MAFF–proteasome” axis as a potential therapeutic target, with the antitumor efficacy of bortezomib subsequently validated in PDX models [178]. PDC also serves as a model system for recapitulating metastatic features and investigating metastatic mechanisms. In both intravenous and orthotopic models, researchers demonstrated that the organotropism of osteosarcoma PDCs—particularly towards lung and bone—closely mirrors clinical patterns and further confirmed the preclinical efficacy of the CDK inhibitor dinaciclib in suppressing metastasis via the MYC pathway [179].

Patient-Derived Tumor Organoids (PDOs), Patient-Derived Xenograft Organoid (PDXO) and Patient-Derived Organoid Xenograft (PDOX)

Organoids are stem cell–derived three-dimensional (3D) in vitro cultures that self-organize to resemble key structural and functional features of their native organs [180, 181]. PDOs are three-dimensional culture systems generated from isolated tumor sample cells or from engineered cancer stem cells [182, 183]. PDOs overcome limitations of both 2D cell cultures that lack a tumor microenvironment and exhibit genetic drift, as well as animal models with inherent species disparities and prolonged experimental cycles [183, 184]. Consequently, PDOs have been widely applied in high-throughput drug screening, target discovery, elucidation of drug resistance mechanisms, and prediction of therapeutic responses, thereby advancing precision oncology [183, 185, 186]. However, PDOs remain limited by restrictive culture conditions, incomplete modeling of the immune tumor microenvironment, relatively low establishment efficiency, and prolonged culture duration. Ongoing efforts aim to overcome these challenges through multicellular organoid systems, cancer organoids 2.0, 3D and 4D bioprinting, microfluidic organ-on-a-chip platforms, and AI–assisted approaches [184, 187, 188].

Integrating organoid and PDX technologies has yielded complementary hybrid platforms with enhanced translational potential. PDOX provides a robust platform for partial reconstruction of the TME, effectively mitigating culture-induced deviations in cancer cell states common in PDOs, while offering higher establishment success rates and significantly reduced time and financial costs compared to traditional PDX models. On the other hand, PDXO models preserve key molecular and phenotypic characteristics of the original tumors while offering improved scalability and cost-effectiveness, thereby facilitating higher-throughput drug testing [180, 189, 190]. A series of studies have progressively employed PDOX models to validate model fidelity [191], elucidate mechanisms of tumorigenesis and metastasis [192, 193], and conduct in vivo drug response assays [186, 194]. Simultaneously, multiple studies using the PDXO platform facilitate drug screening, the identification of novel biomarkers and therapeutic targets, and the advancement of precision oncology [195–197]. A summary of PDX-associated integrated models is provided in Table 4.

Table 4.

Preclinical multi-models combined with PDX models

Model		Construction method	Key Characteristics	Advantage	Major Limitations	Primary Applications
iHu-PDX	Immune human peripheral blood leukocyte (Hu-PBL) SCID mice	Injection of human peripheral blood mononuclear cells (PBMCs) into adult immunodeficient mice	Evaluate T cell-related immunotherapies	(i) Simple and rapid reconstruction of human immunity (ii) Broad tissue source availability (iii) Stable human T cell engraftment (iv) Low cost	(i) High incidence of GVHD limits study window (ii) The immune system primarily consists of T cells; lacks other lineages (iii) No HLA restriction	(i) Evaluation of human-specific immunotherapies (ii) Oncolytic virus and vaccine research
	Human SCID repopulating cell (Hu-SRC) mice	Engraftment of human CD34+ hematopoietic stem cells (HSCs) into pre-conditioned (sub-lethally irradiated) neonatal or adult immunodeficient mice	Long-lasting and complete hematopoietic system reconstitution	(i) More complete and long-lasting human immune system reconstitution (multiple lineages) (ii) Long-term experimental window	(i) Expensive and complex construction (ii) Difficult to obtain high-quality CD34+ HSCs (iii) T cells are mouse MHC-restricted; poor B cell response and antibody production	(i) Evaluation of human-specific immunotherapies (ii) Study of human hematopoiesis and immune function (iii) Infection research (such as typhoid, HIV)
	Human bone marrow, liver, thymus (BLT) mice	Engraftment human fetal liver CD34+ HSPCs and autologous fetal thymus tissue into SCID mice	Facilitate HLA-restricted T cell maturation	(i) Generation of human MHC-restricted T cells (ii) Development of human innate immune cells and primary immune organs (iii) Superior model for human adaptive immune responses	(i) More complex and expensive to build (ii) Ethical concerns (iii) High variability and morbidity	(i) Long-term analysis of human immune development & function (ii) Infection Immunity and Vaccine Development (such as HIV) (iii) Transplant immunology and tolerance studies
PDCs		Cell lines derived from PDX tumors	2D in vitro culture system	(i) Facilitates high-throughput drug screening (ii) Enables mechanistic studies	(i) Loss of tumor heterogeneity (ii) Lack of TME (iii) Genetic drift during long-term culture	(i) High-throughput drug screening (ii) Biomarker validation (iii) Mechanistic studies of drug resistance
PDOs		3D in vitro cultures derived from patient tumor samples	Recapitulate the histology and molecular features of the original tumor	(i) High tumor fidelity (ii) Scalable for high-throughput applications (iii) High success rate and short cycle	(i) Constrained culture conditions (ii) Incomplete modeling of the immune TME (iii) Low establishment efficiency	(i) High-throughput drug screening (ii) Novel target discovery (iii) Analysis of drug resistance mechanisms (iv) Predictive biomarker identification
PDOX		Implantation of in vitro-expanded PDOs into immunodeficient mice	Reconstitutes an in vivo microenvironment	(i) Supplement the in vivo microenvironment for PDOs (ii) Higher establishment success rates vs. traditional PDX (iii) Reversion to the original cancer cell state by reconstituting TME (iv) Modified genetically more easily	(i) Murine host microenvironment (ii) Time-consuming and costly	(i) In vivo validation of drug response
PDXO		Derivation of organoid cultures from established PDX tissue	Retains fidelity to the original PDX or patient tumor	(i) Combines the advantages of PDX and PDOs (ii) Provision of a renewable resource for screening	(i) Inherits limitations of PDX and PDOs (ii) Secondary selection in vitro	(i) High-throughput drug screening (ii) Biomarker and resistance mechanism discovery

Other technological considerations in PDX research

Applications and challenges of combining PDX models with Artificial Intelligence (AI)

In the future, innovations in AI-driven computational methods will undoubtedly transform the role of PDX models in delivering precision oncology solutions. AI has enabled multimodal progress across disciplines, including precision medicine, to require the analysis and integration of patients’ heterogeneous data to delineate and predict optimal care paths for future [198]. AI-based computational algorithms are poised to revolutionize the role of PDX models in precision oncology across multiple domains, including identification of novel druggable targets, prediction of treatment response, and biomedical image analysis [198–200]. More importantly, for drug discovery and preclinical trial optimization, AI-based computational algorithms could predict drug candidates by processing knowledge graphs (a structured network organizing entities and their relationships), allowing for rigorous PDX model examination. Meanwhile, machine learning algorithms can be trained to develop more robust and generalizable AI models by integrating data obtained with PDX models with multimodal clinical datasets, enabling enhanced biomedical image analysis, network-based biomarker discovery, or other domains. For example, an intelligent mechanistic framework with strong drug prediction capabilities was established. OncoTarget and OncoTreat [199] represent two complementary RNA-based approaches for personalized cancer therapy prediction (more than 30 therapeutic agents), which were validated in PDX models transplanted across seven aggressive, treatment-refractory cancers. OncoTarget utilizes the virtual inference of protein enrichment repertoire (VIPER) algorithm to quantitatively analyze tumor protein activity profiles and identify aberrantly activated master regulators (MRs) based on stringent statistical thresholds. In contrast, OncoTreat analyzes large-scale perturbational RNA Sequencing (RNASeq) profiles from patient and PDX tumors to identify compounds that invert tumor-specific tumor-checkpoint modules (TCMs) activity. Furthermore, Brian S. Whiter et al [200]. established a comprehensive pan-cancer repository of > 1,000 PDX H&E histopathological images integrated with multi-mics data (including genomic and transcriptomic profiles), detailed pathological annotations, and extensive clinical metadata. This resource enabled the successful implementation of three deep learning applications. These computational advances are promising for advancing PDX-based therapeutic research by developing image-derived predictive algorithms with potential clinical translation.

Current challenges in the application of AI in clinical oncology encompass the increasing complexity of oncology data streams, which fuels the demand for training large-scale models across various stages of cancer care; potential marginalization biases in clinical outputs with AI; problems related to model scale and cost; challenges in data privacy and governance; and novel obstacles encountered when expanding AI applications to a broader range of medical issues [201–203]. Regrettably, the high-dimensional data above have not fully integrated patient-specific insights derived from preclinical trials involving PDX models. Although inpatient data directly associated with patients may have been obtained post-PDX model testing, the need for ongoing iteration of genome-therapy mappings and the critical requirement to assess further patient heterogeneity necessitate the inclusion of personalized data streams from PDX models in laboratories as specific reference indices in machine learning training datasets. This consideration will significantly enhance the comprehensiveness of clinical decision-making and risk mitigation. Although it may exacerbate certain challenges, such as data complexity and the need for robust analytics, this strategy is crucial for refining the precision of clinical diagnostics and patient care.

Age and sex influence in PDX models

Aging initially suppresses cancer through cellular senescence, but accumulated senescent cells ultimately promote tumorigenesis by creating a pro-inflammatory and genomically unstable microenvironment [204, 205]. He et al. did not elucidate the impact of younger age on the consistency of driver gene mutations in PDX models with machine leaning (ML) algorithms, yet advocated for establishing these models using 8-week-old mice [206]. The 6-to 8-week-old mice are analogous to ~ 20-year-old humans, with aging mice potentially influencing tumor metastasis and pharmacological responsiveness [142]. Importantly, the variability of age-specific models for reflecting human biological tissue systems should be incorporated into the considerations of PDX models when modeling different tumors [205].

The effect of gender disparity on the development of different cancers (e.g., prostate cancer, ovarian cancer) encompasses factors such as physiology, genetics, and environment [207]. Sex-specific mice exemplify the complexity in modeling human cancer regulatory mechanisms. For example, male iKAP mice with KRAS*-expressing colorectal cancer have metastatic frequency and survival consistent with humans compared to females [208]. The genes that escape from X chromosome inactivation appear inconsistent in female mice and humans [209]. Cross-sex tumor research based on PDX models is something that needs to be considered in the context of proposing that GEMM can effectively overcome gender-biased responses [125].

Prospective definition and advantages of PDX 2.0

Conventional PDX models with inherent limitations, including species‑specific differences, high costs, and prolonged study timelines, impede their clinical translatability. While NAMs offer numerous advantages, they also suffer from technical immaturity and poor interpretability in addressing complex biological questions [17]. This highlights the need for combinatorial models in which traditional PDXs are necessarily augmented with NAMs to accelerate innovation and deliver transformative therapies. Accordingly, we propose PDX 2.0 as an integrated platform that leverages NAMs and other complementary technologies to overcome the limitations of conventional PDX models and enhance translational fidelity, serving as a transitional framework toward non-animal research paradigms. Based on recent breakthrough findings, PDX 2.0 comprises two core components: human-based experimental approaches and AI-driven computational models (Fig. 5).

Fig. 5.

The prospective definition and framework of PDX 2.0. PDX 2.0 represents a multidimensional reengineering of traditional PDX models through cutting-edge technologies, while simultaneously evolving novel research paradigms via four key dimensions empowered by AI-driven deep learning, thereby advancing PDX models in oncology research. Anticipated future research frameworks integrating PDX 2.0 with AI-enhanced data technologies comprise three optimized phases: 1) Initial AI-powered predictive profiling in silico (e.g., compound screening and toxicity analysis; 2) Intermediate high-throughput validation using cell lines/organoids in vitro; 3) Terminal systemic and organ-specific experimental validation through murine or higher-order animal models, which enables automated data processing/analysis, reduces reliance on animal models, and enhances research efficiency, cost-effectiveness, and result reliability

PDX 2.0 with human-based approaches

PDX 2.0 with in vitro models

• PDC: Although PDC lines are intrinsically limited in recapitulating key features of TME, including dynamic cell–cell and cell-extracellular matrix interactions, heterogeneous PDC cultures preserve core histopathological and genetic landscapes of the parental tumors, supporting their use in early-phase high-throughput cytotoxicity and drug efficacy screening as well as tumor response profiling.

• PDXO: By preserving cancer stem cells (CSCs), PDXO provides a highly scalable, rapid, and high-throughput in vitro system for large-scale drug screening, from which candidate agents can be advanced to matched in vivo PDX models. This enables the refinement and validation of therapeutic strategies within a paired in vitro–in vivo predictive platform, which accelerates drug development while reducing preclinical costs.

• PDOX: PDOX models substantially address the limitations of conventional organoid culture systems in recapitulating key aspects of the in vivo tumor microenvironment, while improving engraftment success rates [189], expanding the scope of translational research, and enabling the evaluation of a broader range of therapeutic modalities.

Beyond PDXO/PDOX, tissue chips (organ-on-a-chip platforms) that enable more controlled simulation of tissue biology, when integrated with PDX models, allow in-depth investigation of how specific physicochemical factors influence precision cancer therapy. An hepatocellular carcinoma (HCC) PDX chip—a dual-gradient system that supports drug screening while segregating oxygen gradient–dependent HCC cell populations—was cross-validated using a multi-site PDX chip model, in which HCC-containing chips were implanted at distinct anatomical locations exhibiting different oxygen gradients, thereby reinforcing personalized therapeutic strategies for HCC based on oxygen heterogeneity [210].

PDX 2.0 with in chemico methods

In chemico methods enable experiments performed on the interrogation of molecular functional states, biochemical activities, and interaction networks. At this level, advanced genetic engineering approaches, particularly CRISPR–Cas9 biotechnology, applied to PDX models within a pathophysiologically relevant tumor context, can partially alleviate in vitro culture–induced selective pressures and enable the functional interrogation of microenvironment-dependent targets while preserving integrated phenotypic states. This, in turn, enables the establishment of genetically editable tumor models for dissecting mechanisms of drug resistance and disease biology [211–213].

PDX 2.0 with iHu-PDX and naturalized models

In immuno-oncology research, iHu-PDX models, generated by engrafting human tumor tissues together with a reconstituted human immune system into immunodeficient mice, represent a promising platform. Despite persistent limitations, including GVHD, incomplete development of innate immune lineages, and insufficient HLA compatibility, ongoing methodological advances enable iHu-PDX models to reconstitute key immune microenvironmental features and complement patient-specific tumor–immune interactions. These properties provide a valuable preclinical system for elucidating immunotherapy mechanisms and informing personalized immune-based treatment strategies.

Notably, PDX models, as laboratory mouse-based experimental systems, are maintained under highly standardized husbandry conditions, which can bias the fidelity with which human immune phenotypes are recapitulated. Without compromising experimental objectives or inducing adverse outcomes, efforts to naturalize laboratory mice [214], including manipulation of biotic/abiotic factors or rewilding strategies, are expected to improve the immunophenotypic alignment of PDX 2.0 models with wild mammals and humans, thereby enhancing their translational relevance for immunology research.

PDX 2.0 with other innovative advances or considerations

As previously discussed, to overcome the limitations of conventional PDX models, several innovative strategies, such as miniPDX and larval zebrafish xenografts for rapid functional screening, multi-region sampling to capture intra-tumor heterogeneity, and integrated multi-model approaches combining GEMMs, PDX models, and organoids to represent tumor progression from early stages to metastasis, aim to improve the translational fidelity of preclinical models.

PDX 2.0 with AI-driven computational models

AI is reshaping precision oncology across the prevention and diagnosis to the optimization of current therapies and the development of novel treatment [215]. Within this context, AI is also emerging as a key driver of PDX 2.0, enabling advances along several interrelated dimensions.

First, AI substantially enhances the capacity to process, integrate, and interpret the rapidly expanding volume of PDX-derived data. These include automated extraction of pathological features, cell-type classification, and multimodal data association analyses that link histopathology with molecular readouts [72], as well as multi-omics–based prognostic stratification and more comprehensive mechanistic insights [216]. Second, AI-based algorithms facilitate the identification of cancer vulnerabilities and treatment response biomarkers within PDX models [217]. Third, AI frameworks trained on drug perturbation data generated from PDX models support the development of more accurate drug response prediction models [218–220], particularly when aligned with matched primary tumor expression profiles across datasets [221]. Ultimately, AI facilitates the optimization of PDX experimental design, thereby reducing both costs and technical barriers to adoption. This includes the in silico design and virtual screening of candidate compounds [222], as well as strategies to minimize redundant PDX sampling while improving translational relevance [220]. In parallel, these algorithms, trained on PDX models, can be validated using real-world data (RWDs) and iteratively refined through continual learning, thereby improving model interpretability and transparency while ensuring appropriate attention to data privacy and security.

Overall, PDX 2.0 represents an integrative modeling paradigm developed to address distinct, deeper layers of cancer-specific biological questions. No single PDX 2.0 model can fully capture the full spectrum of biological variability; rather, increasingly optimized and clinically relevant predictive PDX 2.0 platforms are expected to emerge. Accordingly, rigorous standards, rational model selection, and the potential for integrative use across the complementary PDX 2.0–PDX 2.0 system will be carefully considered in practical research settings. Finally, a comparative overview of PDX and PDX 2.0, including improvement metrics, evaluation criteria, and applications, is provided (Table 5).

Table 5.

Comparative Overview of Conventional PDX and PDX 2.0

Characteristic	Conventional PDX	PDX 2.0
Model Definition	Static in vivo model: Patient tumor tissue engrafted into immunodeficient mice	Integrative Paradigm: A combinatorial framework of traditional PDX complemented by NAMs to enhance translational fidelity and overcome traditional PDX limitations
Core Components	Single in vivo mouse model with patient tumor tissue	Two core components: (i) Human-based experimental approaches (PDC, PDXO, PDOX, tissue chips, CRISPR-Cas9, iHu-PDX, naturalized models) (ii) AI-driven computational models (data integration, biomarker identification, drug response prediction, experimental design optimization
Key Characteristics	(i) Preserves histopathological and genetic features of original tumors (ii) Maintains tumor heterogeneity and some TME components (iii) Requires immunodeficient mice for engraftment	(i) Multi-modal integration of in vitro, in chemico, in vivo, and in silico approaches (ii) Enhanced TME and immune microenvironment recapitulation (iii) Reduced species-specific differences through humanized systems (iv) Accelerated research timelines with parallel in vitro/in silico screening
Improvement Metrics	-	(i) Translational Relevance: Preservation of human-specific TME components and Immune microenvironment similarity to patient tumors (ii) Efficiency Gains: Model establishment time, Cost per experiment and Throughput capacity (iii) Throughput: higher screening capacity via in vitro/in silico parallelization (iv) Predictive Power: Drug response prediction accuracy and biomarker identification reliability (v) Data Integration: enhanced multi-omics data fusion for comprehensive mechanistic insight
Evaluation Criteria	(i) Tumor take rate and growth kinetics (ii) Histopathological similarity to original tumor (iii) Genetic stability across passages (iv) Drug response correlation with patient outcomes	(i) Experimental Validation: Cross-model comparison in tumor functional assays and drug response concordance (ii) Computational Validation: AI model benchmarking, predictive accuracy metrics, data integration quality (iii) Clinical Validation: Real-world data correlation, co-clinical trial integration, longitudinal validation (iv) Quantitative Metrics Validation: reduction in failure rates, immunological concordance score, translational efficiency ratio
applications	(i) Biomarker identification and Drug screening (ii) PDX preclinical trials and Precision Medicine (iii) Co-clinical Trial (iv) Cancer nanomedicine	(i) Precision Biomarker Discovery and High-Throughput Drug Screening System (ii) Refined Preclinical Trials and Whole-Cycle Management of Precision Medicine (iii) Precision Development and Multi-Dimensional efficacy Evaluation of treatment (iv) Exclusive Evaluation and Mechanism Analysis of Immunotherapy

From PDX to PDX 2.0 in cancer research

For key malignant tumors across major organ systems, we elaborate the transformative role and advanced applications of the PDX 2.0 platform, supported by representative case studies. Applications about target discovery, drug development, and clinical translation are also examined in the context of PDX models.

The respiratory system-lung cancer

Lung cancer is a malignant neoplasm that originates within the lung parenchyma or bronchus. It remains a primary global health challenge; in 2022, there were over 2,400,000 recorded incidents and 1,800,000 deaths. Currently, lung cancer accounts for 12.4% of all cancer morbidity and 18.7% of mortality, ranking as the leading cause of cancer-related death across both genders [223]. The disease is broadly categorized into two types: non-small cell lung cancer (NSCLC) and small cell lung cancer (SCLC), representing 85% and 15% of cases, respectively [224].

Recent advances in sophisticated genome editing technologies, such as inducible CRISPR-Cas9 systems, have transformed PDX models from high-fidelity tumor platforms into precisely programmable in vivo research systems, ushering in the era of PDX 2.0, which enables the direct dissection of gene function and drug resistance mechanisms. The pSpCTRE platform is an all-in-one doxycycline-inducible CRISPR–Cas9 system that enables efficient in vivo genome editing in PDX models, incorporating a truncated CD4⁺ T-cell surface selection marker to ensure stringent transcriptional control and enrichment of transduced cells without intervening in vitro culture. Utilizing this technology in lung adenocarcinoma PDX models, the researchers effectively disrupted essential genes to interrogate genetic dependencies and introduced EGFR T790M and C797S mutations via templated homology-directed repair. This approach demonstrated that the C797S mutation alone can confer acquired resistance to the combination therapy of osimertinib and crizotinib [225].

Machine learning and deep learning approaches have progressively shifted PDX data analysis from a descriptive state to a predictive, high-precision, and automated mode of processing, thereby enhancing the efficiency and accuracy of PDX 2.0 for precision lung cancer treatment. For instance, Juppet et al. developed the SCC, a deep learning-based tool that automates species-specific cell classification in histological sections with up to 96% accuracy. The SCC significantly improves automation and precision by prioritizing contextual features among cells over cell-intrinsic morphological characteristics [73]. Furthermore, PDXGEM, a bioinformatics pipeline employing a random forest algorithm, constructs predictive gene expression models based on PDX gene expression and drug response data. This system effectively forecasts clinical responses to anticancer agents like erlotinib [226].

Acquired cross-resistance remains a critical barrier to treating chemotherapy-refractory SCLC. Integrated analysis of 51 PDX models using high-depth whole-genome sequencing (WGS) and RNA-seq, followed by exposure to a panel of clinical treatment regimens, revealed that focal amplification of MYC paralogs (MYC, MYCN, MYCL) serves as a key driver of acquired cross-resistance. This finding was corroborated through longitudinal studies of over 30 chemotherapy-refractory SCLC PDX lines [227]. Regarding novel drug validation, Sen, T. et al. established a panel of PDX‐derived SCLC cell lines, including LX110 (SCLC-A), LX33 (SCLC-N), and LX1322 (SCLC-P). While the drug lurbinectedin demonstrated inhibitory effects on cell lines other than the SCLC-A subtype, subsequent in vivo verification in corresponding PDX models revealed that the SCLC-A subtype exhibited drug sensitivity results discordant with the in vitro findings [228], emphasizing the necessity of integrating in vivo and in vitro models. And the drug has now entered Phase III clinical trial (NCT05153239).

In immunotherapy, integrating humanized mouse models with PDX to form iHu-PDX models, which advances the validation of novel combination therapeutic strategies. For example, Pan et al. established a humanized NSCLC-PDX model by subcutaneously implanting tumor tissues from NSCLC patients into immunodeficient NTG mice, followed by intravenous injection of autologous CD4⁺ T cell-depleted PBMCs. In this iHu-PDX model, they demonstrated that the combination of percutaneous thermal ablation (PTA) and the adoptive transfer of in vitro differentiated human Th9 cells significantly suppressed tumor growth compared to either monotherapy. Subsequent mechanistic analyses of tumor tissues and blood samples revealed that the combination therapy enhanced antitumor efficacy by increasing tumor-infiltrating Th9 cells and boosting CD8⁺ T cell function, a process primarily mediated through PTA-induced IL-1β release and the subsequent activation of the STAT1/IRF1 pathway in Th9 cells [88]. In addition, other researchers have utilized humanized PBMC-PDX models to evaluate the efficacy of Bifidobacterium longum in combination with the immunotherapeutic agent pembrolizumab. Although no synergistic antitumor effect was observed, the iHu-PDX models established a precedent for bacterial-related research in oncology [229]. Moreover, iHu-PDX models also offer valuable insights into tumor-immune interactions. For instance, a study employing humanized NSCLC PDX systems characterized tumors with varying levels of lymphocyte infiltration, categorizing them as “immune-hot” or “immune-cold” and distinct expression profiles of immune-related markers like PD-1 [230].

The digestive system-liver cancer

Liver cancer, predominantly HCC, was responsible for an estimated 865,000 new cases and over 757,000 deaths globally in 2022, representing 4.3% of total cancer diagnoses (ranking 6th) and 7.8% of cancer-related mortality (ranking 3rd) [223]. By leveraging large-scale HCC PDX cohorts, researchers have integrated clinical characteristics, biomolecular information, and drug-response data from PDX models to elucidate key resistance mechanisms and create publicly available resources for drug development. A primary example is the PDXliver database, which consists of 116 Chinese HCC PDX models; this platform not only maps the associations between the molecular profiles and drug sensitivities of the corresponding HCC cases but also serves as an open-access platform to expand and optimize the HCC PDX model repository [231]. Similarly, A cohort comprising 253 patients and over 100 PDX models revealed the role of MAP3K1 in sorafenib resistance [232].

Advancements in machine learning-based dependency mapping approaches have empowered PDX models to not only recapitulate tumor lineage specificity and genetic dependencies of liver cancer but also effectively predict responses to targeted therapies. Shi et al. constructed a translational cancer dependency map to address the limitation of conventional cell line-based dependency maps in linking to patient clinical data. By leveraging gene essentiality prediction models from DEPMAP, they built the PDXE_DEPMAP profile (comprising 191 PDX tumors). This resource demonstrated that unsupervised clustering of PDX models recapitulated patient-relevant lineage dependencies, complementing the findings from TCGA_DEPMAP. Furthermore, a unique strength of PDXE_DEPMAP is its ability to correlate predicted gene essentiality with drug response data from PDX models, successfully predicting the therapeutic response for 80% of the 15 targeted therapies tested. This capability fills a critical gap in The Cancer Genome Atlas (TCGA), which lacks systematic treatment response data, thereby accelerating the validation of tumor vulnerabilities and informing more effective clinical strategies for targeted therapies [217].

Due to the high construction costs of PDX models, their application in high-throughput drug screening is limited. By strategically combining the distinct advantages of 2D cultures, organoids, and PDX models, researchers can establish a streamlined workflow from high-throughput screening that could enable PDX 2.0 to accelerate precision oncology in liver cancer. For example, Sun et al. employed a PDO adhesion-promoting strategy to process patient-derived tumor tissues, optimizing the culture of PDCs compared to the conventional single-cell suspension method directly applied to patient tumor samples, which reduced the culture time to two weeks. Subsequently, 80 compounds were screened using the resulting PDC models. With the WGS analysis performed on the resistance PDC, they conclude that the abnormal activation of HDAC2 and EGFR leads to drug resistance. And the screened drug HDACi Romidepsin can inhibit HDAC1/2 and multiple tyrosine kinases, including AKT2, MET, exhibiting superior antitumor efficacy compared to sorafenib in PDX models [95]. Additionally, the patient-derived organoids-based xenografted liver metastases (PDOX-LM) model has proven effective in predicting treatment sensitivity while retaining the histologic and genetic features of the donor liver metastases. And the PDOX-LM model addresses key limitations of traditional PDX models by enabling efficient establishment, reliable liver-specific metastasis without concurrent lung involvement, and dynamic biomarker monitoring for metastatic progression [233].

In the aspect of drug validation, B7-H4 expression was evaluated in hepatocellular carcinoma; notably, the combination of AZD8205 (a B7-H4–targeted agent) with AZD5305 (a PARP1 inhibitor) demonstrated synergistic antitumor efficacy in PDX models [234]. The result is an ongoing phase I/IIa clinical study (NCT05123482). Another example of target validation involves the inhibition of αvβ3 integrin, which further blocks the POSTN/TGFβ1 positive feedback pathway [235]. Upon administration of the αvβ3 antagonist cilengitide—either in combination with lenvatinib or used alone—in PDX models, superior tumor reduction was observed [236]. In immunotherapy, elevated expression of phosphoglycerate mutase 1 (PGAM1) was identified to correlate with poor prognosis and an immunosuppressive microenvironment. Pharmacological inhibition of PGAM1 by KH3 induces ferroptosis and increases CD8⁺T-cell infiltration, thereby enhancing the efficacy of anti-PD-1 therapy in PDX models [237]. Moreover, Yin et al. constructed iHu-PDX models by transplanting PBMCs into third-generation pre-established HCC-PDX models. In these models, a co-targeting nanoplatform, D/F@MRL (MMP-2-responsive hybrid liposomes co-loading digoxin and PD-L1-degrading nanofibers), was tested to demonstrate to superior antitumor efficacy compared to the D@MRL and F@MRL groups, with results of increased cell populations such as IFN-γ⁺ CD8⁺ T cells and activated dendritic cells (DCs), along with enhanced release of immune cytokines. [238]. These findings suggest that the treatment regimen effectively remodeled the tumor immune microenvironment.

The digestive system-colorectal cancer

Colorectal cancer (CRC), which encompasses colon cancer and rectal cancer (including the anus), remains one of the most prevalent malignancies globally, ranking third in incidence and the second in mortality. Geographically, Europe reports the highest number of cases, with a notably higher incidence observed in males than in females [239]. A study established XENTURION [240], a large-scale resource of matched PDX and PDX-derived tumoroid (PDXT) models (a concept analogous to PDXO in PDX 2.0) from metastatic CRC patients. Multi-omics characterization demonstrated high molecular fidelity between PDXTs and their parental PDXs, confirming that PDXTs reliably predict in vivo responses to cetuximab in PDXs. Furthermore, researchers leveraged this resource to identify adaptive resistance targets involved in autophagy, like NUAK2, ULK1, and HDACs, whose inhibition synergized with cetuximab to enhance antitumor efficacy. By openly sharing data and models, XENTURION provides a standardized platform for preclinical research in metastatic CRC and establishes an integrated “PDXT (PDXO)-based screening to PDX-based validation” pipeline, offering critical support for precision medicine and drug discovery in CRC fields.

The innovation and integration of new computational technologies enable refined analysis of observational data from PDX models, supporting the development of PDX 2.0 for more precise and reliable outcome assessments in personalized treatment strategies. Beyond the SCC [73], KuLGaP represents a sophisticated computational framework that integrates machine learning (Gaussian Processes), information theory (Kullback–Leibler divergence, KL divergence), and classical statistics [241]. KuLGaP quantifies treatment response by modeling tumor growth curves in treatment and control arms using Gaussian processes and computing the KL divergence between them. This approach explicitly accounts for inter-replicate variability and leverages full longitudinal data, overcoming limitations of conventional methods like mRECIST, which rely solely on endpoint measurements or ignore control groups. In a comprehensive evaluation across 329 PDX models, including colorectal cancer, KuLGaP demonstrated superior selectivity and significantly reduced false-positive rates, establishing it as a robust tool for assessing drug efficacy in preclinical oncology.

For CRC, iHu-PDX models preserve patient tumor heterogeneity while reconstructing a humanized immune system. These models enable the integration of tumor growth response and immune response, such as changes in CD8⁺ T cells and immune factors like CXCL12 and TGF-β, to investigate immunotherapies [242] and other novel treatments [243] that modulate the immune microenvironment. This approach provides deeper mechanistic insights into drug action and enables more comprehensive and accurate prediction of treatment responses.

For metastatic colorectal cancer (mCRC), the anti-CEACAM5 antibody–drug conjugate (ADC) Precemtabart tocentecan (Precem-TcT, formerly M9140) has demonstrated potent antitumor activity in PDX models with high CEACAM5 expression and in a PDX model derived from a patient who had progressed on irinotecan therapy, suggesting potential efficacy in irinotecan-resistant settings. In a phase I trial (NCT05464030) involving heavily pretreated patients with irinotecan-refractory mCRC, Precem-TcT showed promising preliminary efficacy at doses ≥ 2.4 mg/kg, with a confirmed objective response rate (ORR) of 8.8%, a disease control rate (DCR) of 58.8%, and a median progression-free survival (mPFS) of 6.7 months [244]. Furthermore, FDA-approved immune checkpoint inhibitors like Pembrolizumab, Nivolumab, and Ipilimumab, remain the primary immunotherapy choice, particularly for MSI-H/MMR-D mCRC [245]. To better evaluate those treatments, PDX models established in humanized cord blood BALB/c-Rag2^null Il2rγ^null SIRαNOD (Hu-CB-BRGS) mice offer significant inherent advantages in the test of preclinical immunotherapy, tumor immunogenicity, and immune response evaluation [246].

The digestive system-pancreatic cancer

Pancreatic ductal adenocarcinoma (PDAC) accounts for approximately 90% of pancreatic cancers. Most PDAC originate from ductal intraepithelial neoplasia, while a smaller subset arises from cystic tumors [247]. Despite the high mortality rate associated with pancreatic cancer, guideline-recommended targeted therapies and immunotherapies currently cover only a small fraction of patients. Cytotoxic chemotherapy, with (m)FOLFIRINOX and gemcitabine-nab-paclitaxel as its pillars, continues to serve as the cornerstone treatment [248]. PDX or their derivative xenograft-derived organoids (XDOs/PDXOs) exhibit high establishment rates in digestive tract tumors, retain key phenotypic and genotypic characteristics, and serve as valid models for assessing the efficacy of inhibitors targeting specific gene mutations [249, 250].

Within the genetic landscape of PDAC, mutations in homologous recombination(HR) gene, including BRCA1, BRCA2, and PALB2, are reported in 15–19% of cases, in contrast to KRAS mutations, which occur in over 90% [251, 252]. For patients with homologous recombination-deficient (HRD) PDAC, PDX models are instrumental in revealing mechanisms of de novo and acquired resistance to PARP inhibitors and platinum-based chemotherapies, as well as in advancing novel drug development [250]. Based on the clinical stratification of PDAC patients (refractory, acquired resistance, long-term responders), 25 PDX models were established from patient samples at different clinical stages (treatment-naive, treatment-exposed). And these were utilized to generate "in vivo acquired resistance" models, which identified key resistance factors including BRCA monoallelic status, secondary mutations, aneuploidy, and basal-like subtype. More importantly, iHu-PDX model was utilized to conduct preclinical efficacy testing of anti-PD-1 immunotherapy in germline BRCA-associated pancreatic cancer [173]. For KRAS inhibitors, as exemplified by the non-covalent agent MRTX1133 that selectively targets the G12D mutant, specificity, potency, and efficacy have been demonstrated in multiple preclinical models, including PDX models [103]. Nevertheless, elucidating acquired resistance mechanisms to KRAS inhibition in PDAC PDX models remains essential. Using clinical samples and multiple preclinical models, including cell lines, PDO, and PDX models, a study identified co-evolving resistance to KRAS inhibition in PDAC, encompassing genetic alterations such as PIK3CA and KRAS mutations and amplifications of KRAS, MYC, MET, EGFR, and CDK6, alongside non-genetic mechanisms including epithelial-to-mesenchymal transition and PI3K pathway activation, leading to proposed combination therapies targeting the partial epithelial-to-mesenchymal transition (EMT) state and utilizing gemcitabine/nab-paclitaxel with the MRTX1133 [104]. Supported by the robust preclinical data, MRTX1133 is now under evaluation in a multicenter, first-in-human phase 1/2 clinical trial (NCT05737706). Furthermore, additional novel agents, such as the covalent KRASG12C inhibitor JDQ443, have completed preclinical testing and entered clinical trials [106], holding promise for advancing PDAC therapy.

In immune and TME modulation, the combination of the vitamin D receptor agonist paricalcitol with gemcitabine and nab-paclitaxel has advanced to clinical trial evaluation for metastatic pancreatic cancer (NCT03520790). The combination of gemcitabine, paricalcitol, and hydroxychloroquine (GPH) triggers AMPK-ULK1 pathway-mediated autophagic death in PDAC cells, reprograms CAFs toward a quiescent phenotype, and reverses immunosuppression by promoting M1 macrophage polarization and CD8⁺ T cell infiltration, as demonstrated in PDX models [253]. A phase II trial investigates the effects of GPH in treating patients with pancreatic cancer (NCT04524702). Furthermore, the elucidation of resistance mechanisms to immunotherapy was critically facilitated by the use of iHu-PDX mouse models. Specifically, it was discovered that EHF deficiency induces the transcription of C-X-C motif chemokine ligand 1 (CXCL1), which enhances the migration of C-X-C motif chemokine receptor 2 + (CXCR2⁺) neutrophils into the tumor. To study this, Xie et al. constructed iHu-PDX models by transplanting CD34⁺ hematopoietic stem cells into NSG mice. These models subsequently demonstrated that blockade of the CXCL1–CXCR2 axis can reverse therapy resistance, and that Nifurtimox, a drug which effectively induces tumoral EHF expression, synergistically enhances the efficacy of combined PD-1 inhibitor and gemcitabine chemotherapy [254].

Machine learning was applied to deconvolute multi-omics data and construct a prognostic model, termed MR-Gradient, to address the pronounced heterogeneity and poor prognostic predictability of PDAC. This study leveraged PDX models to validate its reliability and to elucidate the regulatory roles of SUV39H1/SUV39H2 (H3K9Me3) and KAT2B (H3K9Ac) in the relationship with transcriptomic, epigenomic, and metabolic features to determine PDAC prognosis. The study offers a tool for prognostic prediction and pinpoints potential therapeutic avenues [255].

The urinary system-bladder cancer

Bladder cancer is a common malignant tumor originating primarily from the urothelial cells. Globally, it accounts for 3.1% of the incidence and 2.3% of the mortality among all tumors [223], ranking 9th and 13th in these respective categories.

In bladder cancer metastasis models, PDX tumors are established across multiple anatomical sites to simulate heterogeneous in vivo growth microenvironments. This approach, integrated with multi-omics profiling and functional validations, enables the systematic dissection of metastatic mechanisms and supports multidimensional translational research. For example, Bernardo et al. established orthotopic bladder cancer PDX models at different anatomical locations, and through multi-omics analysis of primary orthotopic tumors and post-metastatic tumors, they demonstrated that the bladder cancer subtypes are not easily altered by changes in the metastatic microenvironment, revealing the unique metastatic characteristics [256]. In another specific aspect, lymphatic metastasis, as a common metastatic pattern of tumors, has been linked to multiple metastasis-associated genetic loci and corresponding treatment approaches. By combining PDX models with technologies such as high-throughput sequencing and RNA purification, Li et al. discovered that the CAF-related long non-coding RNA LINC00665 is associated with lymphatic metastasis of bladder cancer, and blocking this pathway can impair lymphatic metastasis in PDX models [257]. Another site is circNCOR1, an intron-retained circular RNA negatively correlated with lymphatic metastasis. Overexpression of circNCOR1 or inhibition of downstream TGFb signaling can inhibit the growth of metastatic bladder cancer in PDX models [258]. In bladder cancer PDX models that connect with high-throughput sequencing, inhibition of the ubiquitin-conjugating enzyme E2 C (UBE2C) can also suppress tumor growth and reduce lymphangiogenesis, thereby decreasing lymphatic metastasis [259].

In bladder cancer PDX models, targeted therapy and immunotherapy are widely evaluated to investigate therapeutic efficacy and elucidate mechanisms underlying treatment resistance. For instance, in models harboring aberrations of histone lysine demethylase 4 A (KDM4A)—which associates with the Sqle-ROS-JNK/c-Jun signaling axis—administration of the KDM4A inhibitor ML324 resulted in significant tumor growth attenuation [260]. Similarly, the Bloom syndrome protein (BLM) [261], a helicase in the HR repair process has also elevated its lactylation levels following chemotherapy, thereby enhancing DNA HR repair and inducing chemoresistance. And the drug targeting BLM combined with irinotecan successfully alleviated resistance to epirubicin (EPI) in PDX models [262]. The phase I clinical trial further validated the safety of the combination of irinotecan and EPI (NCT06766266). Moreover, the fibroblast growth factor receptor 3 (FGFR3) inhibitor erdafitinib has been shown to enhance the efficacy of ICB in hu-HSC-NOG mice [263], with a Phase I trial underway (NCT05316155). Regarding immunotherapy, iHu-PDX models of bladder cancer have integrated therapeutic evaluation with immune profiling. In these models, a novel DIFP-FA plus US treatment modality demonstrated enhanced antitumor efficacy and synergized effectively with anti-PD-1 therapy by increasing tumor-infiltrating CD4⁺ and CD8⁺ T cells and boosting the secretion of immune cytokines like TNF-α and IFN-γ [264].

The reproductive system-breast cancer

Breast cancer is a malignant tumor originating from the breast ducts or milk-producing lobular units; in 2022, it accounted for about 11.6% of all tumors, and has a mortality of 6.9%, ranking 2nd and 4th in these respective categories [223]. Driven by significant advancements in molecular and genetic profiling, breast cancer is clinically classified into luminal A, luminal B, triple-negative breast cancer (TNBC), and HER2-positive subtypes to facilitate personalized treatment [265]. In preclinical models for breast cancer metastasis research, examples have now emerged where immunodeficient SRG rats (Sprague Dawley Rag2 −/−; Il2rg −/−) have been successfully used to establish breast cancer brain metastasis PDX models [266].

Whether identified through virtual screening or computational repurposing analysis, drug candidates can be validated in PDX and PDXO models. Notably, PDXOs offer significant advantages for preserving the original tumor's complexity and fidelity. For example, a study utilized the Connectivity Map (CMap) Query tool, a platform developed at the Broad Institute, to screen out vinburnine, efavirenz, and ouabain. And the therapeutic potential for tumor attenuation was validated by repurposed drugs in TNBC PDX and PDXO models [267]. Similarly, a generative deep learning approach was employed to design a novel TGFβR1 inhibitor, YH395A. And YH395A significantly suppressed tumor growth in TNBC PDX models and inhibited lung metastasis in an experimental metastasis mouse model [268].

Moreover, the construction of fusion models in PDX 2.0 integrating multiple platforms facilitates the elucidation of drug resistance mechanisms and therapeutic screening in breast cancer. For example, Guillen et al. used paired PDX and PDO models to establish a high-throughput drug screening platform using PDXOs plated in 384-well plates, employing a broad dose range to assess growth promotion, cytotoxicity, and other effects across a 4-day assay. Herein, birinapant was identified as highly effective in half of the TNBC organoids, a finding that was confirmed in corresponding PDX models. Consistently, in a TNBC patient case, PDXO screening of a pretreatment biopsy selected eribulin, which induced pathologic remission in both the PDX and the patient [197]. Recent studies also highlight the synergy of PDO models with the mini-PDX platform, confirming the potential combined efficacy of alpelisib and fulvestrant [269]. Additionally, researchers used the high-throughput characteristics of the PDX-derived TNBC cell line (WHIM12) to confirm the accuracy of the predictions of their developed computer model [270], highlighting the application of PDCs in rapid drug screening and precision therapy.

As drug validation, the tumor-suppressive effect of Zanidatamab (a bispecific human HER2-targeted antibody) was confirmed across a series of PDX models of HER2-expressing tumors, including breast cancer [271]. A phase II clinical trials of this drug are currently recruiting (NCT06695845). Similarly, Kurani et al. used the drug EPZ-5676 to inhibit DOT1L, a key CSCs regulator in TNBC, and significantly suppressed tumor growth in TNBC PDX models [272], following a Phase I trial in pediatric leukemia (NCT02141828). Moreover, the expression of CD47 in HER2-positive breast cancer is associated with drug resistance [273]. Therefore, researchers constructed an anti-CD47/HER2 bispecific antibody (BsAb), IMM2902, which overcame trastuzumab (a humanized monoclonal antibody targeting HER2) resistance in PDX models and humanized models without inducing off-target toxicity [274]. In a study investigating the interaction between breast cancer and immunity, researchers transplanted cultured 4th-generation TNBC PDX tumor cells into Hu-SRC mice to culture iHu-PDX mice. This model revealed the role of tumor necrosis factor receptor superfamily member 11B (TNFRSF11B) in breast cancer in tumor metastasis and growth [275]. In breast cancer immunotherapy, understanding the immune characteristics of breast cancer is an important means to understand the mechanism of drug resistance and develop new drugs [276]. The iHu-PDX model preserves the immune environment better than the PDX model. Accordingly, in a combination drug validation study, the researchers constructed an iHu-PDX model of HR⁺advanced breast cancer and verified that the combination therapy of PI3Kδ inhibitors and anti-PD-1 monoclonal antibodies overcomes tumor resistance to anti-PD-1 monoclonal antibodies by reshaping the tumor immune microenvironment through the inhibition of Treg proliferation by PI3Kδ inhibitors [277].

The sensory system-melanoma

Melanoma is a highly malignant tumor derived from melanocytes, predominantly manifesting in the skin, although it also occurs in mucosal tissues, the choroid, and other ocular structures. Its pathogenesis is largely driven by a combination of somatic oncogenic aberrations and inherited germline genetic modifiers that dysregulate cellular proliferation and differentiation pathways, frequently as a consequence of cumulative ultraviolet (UV) radiation exposure [278].

Employing pre-trained neural network models to extract data features or accomplish specific tasks automatically enhances the efficiency and accuracy of integrated spatial transcriptomic and imaging data analysis in melanoma PDX samples, thereby supporting research on spatial heterogeneity in PDX models, clinical translation, and the development of multimodal predictive biomarkers. Domanskyi et al. [279] developed Spatial Transcriptomics Quantification (STQ), a Nextflow DSL2-based pipeline for integrated analysis of Visium spatial transcriptomics data and H&E-stained whole-slide images (WSIs) from PDX samples. This pipeline comprises three analytical workflows: dual-species, single-species, and arbitrary-grid, enabling species deconvolution, transcriptome quantification, image and nuclear morphometric feature extraction, and spatial alignment and integration.

For metastatic melanoma research, orthotopic transplantation of metastatic tumor fragments into corresponding anatomical sites enables the construction of PDX models that recapitulate the pathological and molecular features of the donor tumor, thereby ensuring reliability for therapeutic evaluation. For instance, metastatic inguinal lymph node fragments from melanoma patients, which contain tumor cells with high metastatic potential, were orthotopically implanted into the flanks of immunodeficient mice [280]. These cells were subsequently passaged to establish pulmonary metastatic PDX models, which were used to evaluate the efficacy of TP-siRC@tHyNPs against pulmonary metastasis [281].

Simultaneously, a comprehensive large-scale platform equipped with extensive characterization techniques and substantial model testing capacity should be integrated to optimize the clinical management of melanoma patients and formulate guidelines for precision medicine. Maria Romina Girotti et al. completed the implantation of 126 melanoma samples in NOD mice. They performed whole-exome sequencing (WES) and circulating tumor DNA targeted sequencing, revealing potential supposition-driven treatment strategies for BRAF wild-type and anti-inhibitor BRAF mutant tumors mentioned earlier [96].

Establishing iHu-PDX models represents a critical milestone in the "PDX 2.0" era, enabling the evaluation of immunotherapy and combination strategies while predicting mechanisms of inherent or acquired resistance. These models successfully recapitulate the status of TILs within the immune microenvironment, demonstrating that immunotherapy exerts its antitumor efficacy by modulating TILs—specifically by enhancing their cytotoxic capacity and alleviating exhaustion [282]. A specialized iHu-PDX model, developed using transgenic human IL-2 (hIL2-NOG) mice rather than simple co-transplantation, allows for the activation of injected autologous TILs; this model faithfully recapitulates patient-specific responses and has demonstrated particular utility in evaluating anti-PD-1 efficacy in advanced melanoma [283]. Furthermore, the ability of lipid nanoparticle-encapsulated FAS-expressing plasmid to restore Fas expression in melanoma cells, thereby enhancing cytotoxic T lymphocyte (CTL) tumor infiltration and exerting antitumor effects, was also evaluated in iHu-PDX models [284].

The blood and immune system-leukemia

Leukemia is a malignant clonal disease of hematopoietic stem/progenitor cells. Among all cancer types in 2022, leukemia accounted for 2.4% of new cancer cases and 3.1% of cancer-related deaths, respectively [223]. Based on the degree of cell differentiation and the natural disease course, leukemia is classified into two major categories: acute leukemia (AL) and chronic leukemia (CL). In AL, cell differentiation is arrested at an early stage, with a predominance of blast cells and early immature cells. The disease progresses rapidly, with a natural history of only a few months. According to the principal hematopoietic lineage involved, AL can be further categorized into acute lymphoblastic leukemia (ALL) and acute myelogenous leukemia (AML) [285, 286]. Different from other solid tumors, leukemia, as a hematological cancer, is a better choice to establish the PDX models by intravenous injection or direct injection into bone marrow (i.e., intrafemoral injection) [68].

The integration of CRISPR technology with PDX cells (i.e., PDCs in PDX 2.0) has transformed the paradigm of leukemia research: optimized editing strategies like the Split-Cas9 system for PDX cells have elevated PDX models from merely passively mimicking human leukemia to becoming active platforms for functional genetic studies and therapeutic testing. Conversely, PDX models provide the clinically relevant in vivo architecture and microenvironment dependencies necessary for PDCs to ensure the physiological significance of CRISPR screening outcomes. Their combination has not only facilitated the discovery and validation of novel leukemia targets such as SLC5A3, MARCH5, and ADAM10, but also offered new insights into mechanisms of metabolic reprogramming, apoptosis regulation, and leukemia–niche crosstalk. Moreover, this synergy demonstrates immerse potential for clinical translation in validating personalized treatment strategies and drug development [211, 212].

Novel therapeutic agents are increasingly integrated with standard chemotherapy to overcome drug resistance in leukemia by targeting specific epigenetic and molecular drivers. For instance, the interaction between the MENIN protein and the MLL fusion protein sustains a hematopoietic stem cell program that promotes malignant proliferation and blocks differentiation; DSP-5336, an inhibitor of this interaction, significantly enhances the efficacy of standard induction chemotherapy (DNR plus Ara-C) in AML PDX models and shows promise for MLL-rearranged or NPM1-mutated leukemias [87]. Another novel drug, a potent and selective covalent KDM1A inhibitor, ORY-1001, was evaluated to decrease the percentage of 9.3% ALL-T1 cells in the PDX models and prolong murine survival by 4 days on average. Finally, it was proved that ORY-1001 resists leukemia cells' proliferation, inducing blast cells' myeloid differentiation and self-renewal of compromised leukemic stem cell (LSC) [287]. Two clinical trials have been initiated to evaluate the efficacy of DSP-5336 and ORY-1001 for the treatment of relapsed or refractory acute myeloid leukemia (R/R AML) harboring FMS-like tyrosine kinase mutations (NCT05546580, NCT04988555). Asparaginase (ASNase) therapy is also a mainstay for AL [288], but the emergence of resistance has prompted the birth of new drugs. Utilizing CRISPR/Cas9-based loss-of-function screening, investigators pinpointed Bruton’s tyrosine kinase (BTK) as a promising therapeutic target. Subsequent studies in PDX models demonstrated that BTK inhibition via ibrutinib significantly sensitizes cells to ASNase [289].

Future perspectives and conclusions

Over the past century and likely for a protracted duration into the future, humanity will continue to grapple with cancer, ensnared in a cyclical pattern of the development-resistance-innovation of drugs. Within the context of specific and intricate therapeutic regimens, no solitary biomarker can reliably forecast the ultimate therapeutic response.

The prerequisite for refining the cancer mechanistic network and the drug-genotype association framework is cross-referencing newly acquired pivotal cancer-promoting datasets with normative datasets [142, 290]. Despite the establishment of extensive repositories of serially transplantable PDX for various significant tumors and the conduction of a considerable number of drug response assays, PDX cohorts representing under-represented cancer subtypes, patient diversity, racial, and geographical variations are still being tested preclinically in a decentralized manner across laboratories, with a lack of standardized quality control for research outcomes [291]. The advent of PDX 2.0 marks a transformative leap in this domain, integrating PDOX/PDXO, PDC, iHu-PDX AI-driven data analytics, advanced humanized immune models, and single-cell resolution technologies to address these limitations. By preserving tumor-stroma interactions while enabling high-throughput drug screening, PDX 2.0 bridges the gap between in vitro simplicity and in vivo complexity, offering a physiologically relevant platform for therapeutic exploration. Furthermore, AI-driven longitudinal profiling empowers dynamic interrogation of tumor evolution and resistance mechanisms, aligning preclinical studies with real-time clinical insights.

Considering that precision medicine is a long-term evolutionary process entailing the integration of diverse heterogeneous datasets from patients and model organisms, the execution of efficient design and management protocols, synchronized governmental strategies and collaborations, and the advancement of technology and legal frameworks [292], PDX 2.0’s AI-driven analytics will play a pivotal role in deciphering complex biomarker networks and predicting therapeutic responses. We are committed to establishing a publicly accessible data repository focusing on the pivotal role of PDX 2.0 in facilitating the benefits of clinical and preclinical trials for personalized cancer therapies (Fig. 6). The data platform, operating as a containerized analytical workflow, integrates bioinformatics infrastructure with repositories of PDX models, aiming to establish an efficient closed-loop management system that connects hospitals, patients, laboratories, PDX, multi-dimensional datasets, therapeutic regimens, and hospitals. The multi-dimensional datasets include multiple heterogeneous data from individual patients (such as electronic health records, multi-omics profiles, clinicopathological information, and post-treatment outcomes) and the testing feedback data from corresponding PDX models (such as identified biomarkers and treatment response metrics), as well as the correlations between multi-omics data of PDX models and the underlying mechanisms of carcinogenesis and drug resistance, necessitating ongoing updates and refinement.

Fig. 6.

Patient-PDX-Regimen platform: Future vision for PDX 2.0 development. a The integrated PDX data platform for closed-loop targeted personalized therapy amalgamates current PDX repositories with regional biomedical informatics architectures. It initiates from addressing drug resistance in healthcare settings, transitioning seamlessly into experimental research domains. Collaboratively, healthcare providers and research laboratories supply the foundational patient’s biological information, while the platform contributes its PDX insights. In the event of unresolved resistance, the platform is perpetually refined by the laboratories’ innovative findings, assimilating high-dimensional patient datasets with corresponding PDX response data. The platform is destined to cater to more public services, considering the integration of novel technologies and various challenges. b Patient’s fundament and PDX’s contribution serve as symbiotic elements within the Patient-PDX-Regimen platform to expedite clinical translation for the precision of oncological diagnosis and treatment decision-making. ML, machine learning

The high fidelity of PDX models in recapitulating tumor heterogeneity is a principal attribute that has advanced the clinical oncology paradigm across numerous domains. The microbiome-tumor axis represents a critical yet underexplored frontier in oncology. PDX 2.0 systems, with their ability to incorporate patient-specific microbiomes, multi-omic analyses, and AI-powered analytics, provide a groundbreaking approach to address this gap. By engineering PDX models with matched intratumoral and gut microbial communities from patients, researchers can unravel the microbial impacts on tumor progression, immune evasion, and therapeutic resistance, identify microbiome-linked biomarkers, and accelerate the development of microbiota-targeted therapies.

Additionally, PDX 2.0 amplifies this strength by embedding ethical and translational efficiencies, which can minimize animal use through organoid scalability, enhance predictive accuracy via AI-driven analytics, and align preclinical outcomes with clinical timelines. We synthesize and delineate the characteristics and applications of PDX models, utilizing a framework of eight primary cancer systems, and illustrate and discuss cutting-edge research and discrepancies in the context of three therapeutic drug categories. While numerous promising PDX applications for clinical oncology are under development, significant challenges hinder the translation to clinical practice. PDX 2.0 addresses these challenges through technological convergence, promoting interoperability among models, data streams, and clinical workflows.

Through concerted efforts to address these challenges, the establishment of the Patient-PDX matched repository based on multi-dimensional data and computational algorithms emerges as an impending trend in PDX development, which is well within reach. PDX 2.0 stands at the forefront of this evolution, redefining translational oncology by harmonizing innovation, ethics, and precision to deliver transformative cancer care.

Abbreviation

AI

Artificial intelligence

ALL

Acute lymphoblastic leukemia

AML

Acute myelogenous leukemia

BLT

Bone marrow, liver, thymus

BTK

Bruton’s tyrosine kinase

CAFs

Carcinoma-associated fibroblasts

CLL

Chronic lymphocytic leukemia

CNAs

Copy number alterations

CRC

Colorectal cancer CRC

CTCs

Circulating tumor cells

ECM

Extracellular matrix

EGFR

Epidermal growth factor receptor

FDA

Food and Drug Administration

GEMM

Genetically-engineered mouse model

GVHD

Graft-versus-host disease

H&E

Hematoxylin and eosin

HCC

Hepatocellular carcinoma

HER

Human epidermal growth factor receptor

HR

Homologous Recombination

HSPCs

Hematopoietic stem and progenitor cells

HTDS

High-throughput drug screening

Hu-PBL

Human peripheral blood leukocyte

Hu-SRC

Human SCID repopulating cell

ICB

Immune checkpoint blockade

iHu-PDX

Immune-humanized patient-derived xenograft

ITH

Intratumor heterogeneity

mCRC

Metastatic colorectal cancer

MRs

Master regulators

NCI

National Cancer Institute

NIH

National Institutes of Health

NK

Natural killer

NOD

Non-obese diabetic

NSCLC

Non-small cell lung cancer

PBL

Peripheral blood leukocyte

PBMCs

Peripheral blood mononuclear cells

PDAC

Pancreatic ductal adenocarcinoma

PDCs

PDX-derived cells

PDMR

Patient-derived models repository

PDOs

Patient-derived organoids

PDOX

Patient-Derived Organoid Xenograft

PDX

Patient-derived xenograft

PDXO

Patient-Derived Xenograft Organoid

RNASeq

RNA Sequencing

SCC

Single Cell Classifier

SCID

Severe combined immunodeficiency

SCLC

Small cell lung cancer

TCGA

The Cancer Genome Atlas

TILs

Tumor-infiltrating lymphocytes

TME

Tumor microenvironment

TNBC

Triple-negative breast cancer

WES

Whole-exome sequencing

WGD

Whole-genome duplication

WGS

Whole-genome sequencing

Author’s contributions

QC, ZX, PL, and HZ conceived of and designed the study. CJ, HF, and WW wrote the manuscript and drafted the figures. CJ, HF, QC, ZX, PL, HZ, WW, ZL, and NZ revised the manuscript. QC, WW, ZX, and HZ provided the funding support. All authors read, edited, and approved the final manuscript.

Funding

This work was supported by the National Natural Science Foundation of China (NO. 82573181, 82372943, 82303610, 82503442), Chongqing Medical Youth Top Talent Program (NO. YXQN202558), Chongqing Youth Innovation Talent Project (NO. CSTB2025YITP-QCRCX0096), Natural Science Foundation of Chongqing (NO. CSTB2025NSCQ-GPX1144), The Science and Technology Innovation Program of Hunan Province (NO. 2023RC3074), Key R&D Program of Jiangxi Province (NO. 20243BBI91007), Jiangxi Provincial Health Technology Project (NO. 202510008), Beijing Xisike Clinical Oncology Research Foundation (NO. Y-Gilead2024-PT-0070), Jiangxi Provincial Natural Science Foundation (NO. 20252BAC220048), Science Foundation of the AMHT Group (NO. 2025YK06), Science Foundation of Hunan Aerospace Hospital (NO. 2025YJ01).

Data availability

No datasets were generated or analysed during the current study.

Declarations

Ethics approval and consent to participate

Not applicable.

Consent for publication

Not applicable.

Competing interests

The authors declare no competing interests.