PDX models for functional precision medicine and discovery science

Abstract

Precision oncology relies on detailed molecular analysis of how diverse tumors respond to various therapies, with the aim to optimize treatment outcomes for individual patients. Patient-derived xenograft (PDX) models have been key to preclinical validation of precision oncology approaches, enabling the analysis of each tumor’s unique genomic landscape and testing therapies that are predicted to be effective based on specific mutations, gene expression patterns, or signaling abnormalities. To extend these standard precision oncology approaches the field has strived to complement the otherwise static and often descriptive measurements with functional assays, termed functional precision oncology (FPO). By utilizing diverse PDX and PDX-derived models, FPO has gained traction as an effective pre-clinical and clinical tool to more precisely recapitulate patient biology using in vivo and ex vivo functional assays. Here, we explore advances and limitations of PDX and PDX-derived models for precision oncology and FPO. We also examine the future of PDX models for precision oncology in the age of artificial intelligence. Integrating these two disciplines could be the key to fast, accurate, and cost-effective treatment prediction, revolutionizing oncology and providing cancer patients with the most effective, personalized treatments.

Introduction

Precision oncology uses molecular and genetic characteristics of patient tumors to accurately diagnose their disease and guide therapy¹ (Figure 1A). Large-scale genomic analyses of various cancers identified recurrent alterations across tumors and led to the hypothesis that identifying genomic drivers [G] within individual tumors could help tailor more effective therapies for each. The landmark clinical trial National Cancer Institute Molecular Analysis for Therapy Choice (NCI-MATCH)² tested this concept and revealed two key points: the percentage of patients from the trial deriving clinical benefit was low; and tumors with the same mutations can have different responses to the same targeted drugs (BOX 1). Other clinical trials have reported similar results³, highlighting the need to explore alternative methods, such as integrating tumor genotypes or phenotypes with functional evaluation of available drugs to more effectively combat cancer.

Figure 1. Patient-derived xenografts pipeline.

(A) Precision oncology treatment pipeline. The current precision oncology paradigm begins with the collection of a patient’s tumor samples followed by tumor sequencing and/or molecular profiling, data analysis, identification of potential tumor vulnerabilities and selecting drugs predicted to target these vulnerabilities. Patients receive this therapy in the context of standard of care treatments or clinical trials.

(B) The precision oncology research pipeline extends this paradigm and can be used to identify precision oncology treatments for the future In this pipeline, human tumors derived from patients at different disease stages, for example before onset of anticancer therapy (treatment naïve), during treatment (on-treatment), or post-treatment are implanted into animal hosts (for example mice or zebrafish) as patient-derived xenograft (PDX) models for research. Patient and PDX tumors are profiled to gather data on tumor genomic, epigenomic, RNA, protein, or metabolic profiles. This information is utilized to identify genomic signatures and biomarkers for investigational therapy response, and candidate treatments can be tested in PDX models.

BOX 1: Recent developments in precision oncology.

• The Human Genome Project, initiated in 1990, transformed biological and clinical research by providing comprehensive insight into the human genome^171,172. This ambitious project substantially advanced our understanding of diseases and treatment strategies, intertwining genomics with clinical medicine to enhance patient care¹⁷³. Notable successes for cancer treatment include the anti-human epidermal growth factor receptor 2 (HER2) therapy trastuzumab, approved in 1998 for HER2-amplified breast cancer^174–176, and the tyrosine kinase inhibitor imatinib, approved in 2001 for chronic myeloid leukemia¹⁷⁷.

• The Cancer Genome Atlas (TCGA)¹⁷⁸, launched in 2005, represented a landmark effort in genomic sequencing and profiling of more than 10,000 samples from 33 cancer types. By enhancing our understanding of the genetic, epigenetic, transcriptomic, and proteomic features of cancer, this initiative greatly expanded the precision oncology paradigm, leveraging multi-layered data to develop targeted therapies and design clinical trials utilizing those therapies.

• The U.S. federal government’s $200 million investment in the Precision Medicine Initiative in 2015 revitalized that interest, encouraging comprehensive cancer centers nationwide to establish precision oncology programs, identify therapeutic vulnerabilities, and match patients with targeted treatments to optimize treatment efficacy while minimizing toxicity.

• The National Cancer Institute’s Molecular Analysis for Therapy Choice (NCI-MATCH) trial, also launched in 2015, was an unprecedented precision oncology program that performed molecular testing on around 6000 patient tumors and evaluated the efficacy of mutation-matched targeted therapies in various cancers². This study showed that 37.6% of patients had actionable mutations with possible targeted treatment, but only 12.4% of patients were able to receive the identified treatment¹²⁷. With an objective response rate set at a minimum of 16% for each arm of the study (5 or more of 31 patients per arm), 7 of the 27 arms met these criteria. Across all treatment arms of evaluable patients who received treatment, the response rate was 10.3% (79/765).

The combination of financial investment and scientific discovery culminated in landmark achievements in the field of precision medicine, and there are now more than 100 FDA approved biomarker-driven targeted therapies approved for cancer, according to the Personalized Medicine Coalition¹⁸⁰. However, these studies also provide a cautionary tale: despite some notable successes, treatment responses were surprisingly underwhelming when matching targeted therapies to genetic vulnerabilities¹⁷⁹, especially from an individual point of view. The presence of a targetable mutation provides clues that can guide therapeutic options, but this must be considered within the complexity of the tumor’s landscape. Transcriptomics, epigenetic profiling and/or proteomics need to be integrated with molecular profiling, to get a better understanding of the tumor’s biology, which cannot be described by genomics alone. To advance the field, we must move beyond the mutation-centric paradigm, integrating hypothesis-driven functional experiments to identify more efficacious therapies for a broader patient population.

Functional precision oncology (FPO)^3,4 is an approach that goes beyond the descriptive and static analysis of tumors, such as the presence of mutations at a given time, to identify precision therapy. With FPO, specimens derived from patients are used to test tumor behavior in functional assays, with the overarching goal of predicting successful therapies for individual tumors in a personalized way. Establishment of stable, patient-derived models for FPO also enables long-term studies of tumor biology and drug responses. There are many types of patient-derived models that can be used for precision oncology and FPO; we here focus on xenograft systems that enable in vivo experimentation, we refer readers to recent reviews of other types of patient-derived models used for FPO assays^5,6.

Patient-derived xenografts (PDX) are generated by implanting patient tumor samples into immunodeficient hosts (typically mice), and largely conserve the patient’s genomic and phenotypic tumor characteristics^7,8,9. PDX models can be used to advance both conventional precision oncology via preclinical studies (Figure 1B), and FPO (Figure 2), on a more individual level^{3,7,10 11}. First, preclinical studies in PDX can be used to better define which tumors respond to specific drugs, such as testing the hypothesis that tumors bearing a certain oncogenic mutation might respond to drugs targeting the cognizant activated oncogene. This sets the stage for clinical trials using typical precision oncology approaches (Figure 1A). Second, as part of the FPO pipeline, PDX can be used for additional assays, and, in concert with other methods, models, and tools, can predict individual patient prognosis or therapeutic response (Figure 2). PDX models have been invaluable within the FPO pipeline. Multiple studies have shown that successful PDX engraftment in mice by itself is prognostic, as it is associated with recurrence and metastases in patients^3,12–15. Once the patient’s tumor is engrafted as a PDX, drugs can be tested for efficacy, with high concordance with patient responses^8,16. PDX models have evolved over time and have expanded beyond mouse hosts to facilitate various aspects of precision oncology. Alternative hosts include zebrafish, pigs, rats, and the chicken egg chorioallantoic membrane [G] (CAM), each with distinct advantages and disadvantages¹⁷.

Figure 2. The premise of functional precision oncology.

Functional precision oncology (FPO) not only analyzes patient tumors for static features, such as the presence of targetable mutations, but also utilizes patient-derived models, such as CAM PDX, zebrafish PDX, pig PDX, or short or long-term ex vivo PDX derived cultures, to functionally test responses to predicted therapies. Integrating that information with omics analysis and newer machine learning algorithms leads to hypothesis driven evaluation of drug targets, and more informed clinical trials that benefit patients. The overarching goal is to identify vulnerabilities using functional assays, providing more comprehensive, personalized data to guide therapeutic decisions.

In this Review, we summarize the development and current state of PDX use in preclinical precision oncology research. We also highlight the use of PDX in FPO, to optimize treatment strategies, perform drug testing, and uncover biomarkers for drug response/resistance, while acknowledging the limitations and challenges that these models face for FPO. We conclude with a discussion of the future of FPO, including how artificial intelligence techniques can integrate with PDX-based FPO to advance oncology research.

Use of PDX models in precision oncology

A fundamental aspect of precision oncology is utilizing a tumor’s genetic aberrations or other vulnerabilities to inform selection of efficacious therapies. PDX and PDX-derived models representing human tumors, on a non-individualized basis, provide a key resource to bring this research into the precision oncology space, for drug validation and identifying biomarkers of response. We here provide key recent examples across several tumor types where precision oncology approaches have been developed, and where this type of research is emerging to identify new treatments.

PDX models for targeted therapy identification

A common use of PDX models is to test novel treatment strategies that address specific genetic aberrations that are present in different cancer types. A notable example is research targeting the epidermal growth factor receptor (EGFR) and MAPK signaling pathways, which are frequently dysregulated in cancer, with numerous targeted therapies approved by the US Food and Drug Administration (FDA) available. For example, in esophagogastric PDX models with mutations affecting mTOR, RAS or EGFR signaling, treatment with the respective targeted therapies attenuated tumor growth¹⁸. In pancreatic cancer, aberrant activity of MEK1¹⁹ and NPC1-like intracellular cholesterol transporter 1 (NPC1L1)²⁰, which regulates cholesterol uptake, were identified as putative vulnerabilities. Targeting these proteins in PDX models using the MEK1 inhibitor AZD6244 or the cholesterol-lowering drug ezetimibe, respectively, significantly limited primary tumor growth compared to vehicle. PDX models are also used to test new inhibitors. For example, in glioblastoma PDX models the new EGFR inhibitor WSD-0922 was evaluated and demonstrated potent antitumor effects²¹.

Other examples of target evaluation include inhibition of cyclin-dependent kinases (CDKs) and cyclin-related proteins. To test putative therapies, a biobank of adolescent and young adult sarcoma PDX models of varying subtypes, including osteosarcoma and rhabdomyosarcoma, that underwent extensive sequencing and pathway analysis was established. Upon pharmacological inhibition of CDK4 and CDK6 with palbociclib or the BET family of proteins with OTX015 in osteosarcoma PDX models, tumor burden decreased, revealing novel targets for osteosarcoma treatment²². In leiomyosarcoma, profiling of PDX using sequencing and dynamic BH3-profiling [G] revealed new drug candidates and identified CDK inhibition as a potential vulnerability. Treatment of leiomyosarcoma PDX with the CDK7 inhibitor YKL-5–124 significantly attenuated tumor growth²³. Furthermore, in head and neck cancers, molecular characterization of PDX suggested cyclin D1 and cyclin-dependent kinase inhibitor 2A as vulnerabilities, and when PDX models were treated with the CDK4 and CDK6-selective inhibitors palbociclib²⁴ or abemaciclib²⁵, there was significant antitumor activity. These examples demonstrate the value of identifying putative targets in tumors and testing their validity in predicting therapeutic response through functional testing. The convergence of data pointing towards targeting key proteins across multiple cancer types also holds promise that interrogation of tumor biology through PDX can produce biomarker-driven treatment strategies using the same class of drugs across many patients.

Inhibition of other common signaling pathways has been investigated in PDX preclinical models for tumor types that may not have aberrant signaling in the pathways discussed above, but still lack effective targeted therapies. In breast cancer, PDX models were used to show that targeting aberrant p53 expression²⁶, Src family kinases²⁷, or fibroblast growth factor receptor (FGFR)²⁸ with idasanutlin, dasatinib, and futibatinib, respectively, could effectively shrink or stabilize tumors. In glioblastoma PDX models, the MDM2 inhibitor SAR405838²⁹ revealed antitumor activity. In breast-implant associated anaplastic large cell lymphoma, a PDX model and matched PDX-derived cell line was used to identify a Janus kinase 1 (JAK1) mutation, which could be targeted using the pan kinase inhibitor ruxolitinib to reduce tumor burden in PDX models expressing the mutant protein³⁰. Similarly, the therapeutic benefit of targeting novel oncogenic fusions such as MYH9::PDGFRB³¹ and mutations in PDGFRA^31,32 in T-cell lymphoblastic leukemia and lymphoblastic lymphoma was demonstrated in PDX models.

Using PDX to optimize therapeutic combinations

Targeting a single oncogenic protein or pathway rarely induces and sustains complete responses. Combination strategies are often necessary and must be investigated preclinically for both efficacy and tolerability^33–35. Furthermore, new targeted therapies are often tested in combination with standard-of-care treatment regimens, so that approved therapies are not withheld from patients while testing experimental drugs. Preclinical evidence for effective therapy regimens using PDX models can inform clinical trial design, and even lead to fast-track status for approval, as in the case of the EGFR-targeting drug HLX42 for non-small-cell lung cancer (NSCLC)³⁶. Examples of strategies used in PDX to develop or optimize therapeutic combinations include combining new targeted therapies with approved chemotherapy; using more than one investigational therapy simultaneously; combining drugs that are FDA-approved for a different indication; or even discovering biomarkers of response to existing chemotherapy combinations³⁷. These studies can power effective clinical trials by providing a compelling basis for efficacious drug combinations in various patient populations represented by PDX.

As noted previously, targeting EGFR or the MAPK pathway is promising for a growing number of cancers, and integration with additional chemotherapies or targeted agents to optimize tumor responses is being evaluated. In triple-negative breast cancer (TNBC) PDX models, the combination of navitoclax, which inhibits the anti-apoptotic proteins BCL-2 and BCL-xL, with the EGFR inhibitors ABT-414 or ABVV-321, was found to elicit better responses than each inhibitor alone. Co-expression of EGFR, BCL-2 and BCL-xL was suggested as biomarkers for this treatment strategy³⁸. Multiple studies have used colorectal cancer (CRC) PDX models to assess the benefit of adding MAPK pathway inhibitors to chemotherapy and small-molecule inhibitors to target frequent mutation of KRAS, NRAS, and BRAF. Strategies include combining the MAPK inhibitior LSN3074753 with the EGFR inhibitor cextuximab³⁹, combining the MAPK inhibitor encorafenib with cetuximab and each of the chemotherapies 5-FU, ozaliplatin, or irinotecan⁴⁰, and co-administering the MAPK inhibitor trametinib and the CDK4/6 inhibitor palbociclib⁴¹. Each of these three co-treatment strategies showed improved tumor response compared to the benchmark of MAPK inhibition alone^39–41.

Another area of research is identifying treatment combinations to improve the efficacy of drugs that target the human epidermal growth factor receptor 2 (HER2) and HER3 family of receptors. In ovarian cancer PDX models, successful combinations have included targeting HER2 with pertuzumab or trastuzumab, together with carboplatin and paclitaxel⁴². Utilizing a large biobank of cervix cancer PDX and PDX-derived organoids (PDXOs) revealed efficacy with the combination of the EGFR and HER2 inhibitor neratinib and chemotherapies in HER2-mutated models⁴³. Similarly, in HER2-amplified gastric cancer PDX, combination of trastuzumab plus other HER2-blocking drugs such as pertuzumab and lapatinib showed greatly improved tumor control compared to the single agents⁴⁴.

PDXs are also being used to test innovative therapeutic combinations tailored to specific cancer types, particularly in cases where clinical treatment regimens or unmet medical needs dictate unique approaches. The integration of targeted therapies with standard of care has the potential to accelerate clinical testing of new combinations. For example, in esophageal adenocarcinoma the addition of a hedgehog signaling inhibitor (LDE225) to radiation, chemotherapy, or both was tested using PDX models. The combination of radiation and chemotherapy could inhibit or delay tumor growth, which was in some PDX models accompanied by upregulation of hedgehog signaling. Addition of LDE225 delayed tumor growth in these models⁴⁵. Other examples include the combination of gemcitabine with the recombinant methioninase (rMETase), trametinib, or the bacteria-based tumor-targeting therapy Salmonella typhimurium A1-R [G], that were significantly more effective than either agent alone in pancreatic cancer PDX models^46–48. Combination treatment with Salmonella typhimurium A1-R also improved efficacy of the chemotherapeutic drug temozolomide or the BRAF inhibitor vemurafenib in melanoma PDX models⁴⁹. Along the same line, ovarian cancer PDX models were used to demonstrate that combination of the checkpoint kinase 1 (CHK1) and CHK2 inhibitor prexasertib, with the PARP inhibitor olaparib, and the chemotherapeutic drugs lurbinectedin and cisplatin inhibited tumor growth^50,51. Finally, PDX models are invaluable to test combinatorial drug regimens in rarer cancer types, including follicular leukemia⁵², medulloblastoma with chromothripsis⁵³ and renal cell carcinoma⁵⁴. While not all of these examples involve preclinical testing of precision therapies in the classical sense (such as targeting a mutated protein in a given tumor), each study revealed key mechanisms or pathways that inform whether a therapy can be effective and therefore sets the stage for biomarker development and refinement of future precision oncology strategies.

Targeting residual disease and drug resistance

The development of strategies to overcome de novo or acquired drug resistance is of great interest to prolong patient survival. PDX are particularly suitable to study this problem because they contain aberrations that will be encountered in the clinic, and they are developed contemporary with today’s treatment regimens, often derived from patient samples after acquired resistance to certain therapies. Mapping drug resistance mechanisms in PDX can lead researchers to identify treatment strategies that can be effective across a subset of the patient population exhibiting similar biology. Such strategies can also help with patient selection, to avoid giving toxic treatment regimens to individuals whose tumors are predicted to be resistant to the therapy. By studying CRC residual disease through PDX modeling after EGFR blockade with cetuximab, HER2 and HER3 upregulation was found to mediate tumor cell survival after treatment. Targeting these HER family members with pan-HER antibodies that disrupt HER receptor signaling, in combination with cetuximab, reduced tumor size and improved overall survival of PDX mice compared to EGFR inhibition alone⁵⁵. Pharmacologically co-inhibiting EGFR (with cetuximab) and the anti-apoptotic protein BCL-xL (with A-1331852) in CRC PDX and PDXOs was shown to eliminate more residual tumor cells than inhibition of EGFR alone. Mechanistically, cetuximab was found to upregulate the pro-apoptotic protein BIM, which was no longer sequestered by BCL-xL upon inhibition⁵⁶. Efforts in CRC to combat development of resistance to anti-EGFR therapy have also utilized cetuximab-resistant PDX models. Studies identified and targeted increased STAT3 signaling using cetuximab in combination with ruxolitinib⁵⁷, and blocked SRC kinase activity with dasatinib⁵⁸, in order to overcome resistance to EGFR inhibition in CRC PDX models. In NSCLC, PDX models were also used to investigate the third-generation EGFR inhibitor osimertinib in overcoming resistance to prior EGFR inhibitors⁵⁹.

PDX models can also be used to identify potential mechanisms of resistance to targeted therapies. Results of these investigations can provide insight into alternative treatment strategies that circumvent these resistance mechanisms and eradicate tumors. PDX models were used to identify ways to target drug-resistant, residual BRAF-mutant melanoma cells after MAPK inhibition with dabrafenib and trametinib, revealing that resistant cells are characterized by a neural crest stem cell-like transcriptional state. Targeting this transcriptional state by inhibiting both MAPK signaling and either focal adhesion kinase (FAK) with PF562271⁶⁰ or the nuclear retinoid acid receptor RXRG with HX531⁶¹ was able to delay development of resistance and shrink PDX tumor size in vivo. Similarly, BRAF-mutant brain tumor PDXs were used to understand responsiveness of BRAF-V600E-mutant brain tumors to the MEK inhibitor trametinib. It was observed that trametinib blocks MAPK and TORC1 activity; however, resistance quickly occurred. Co-treatment with rapamycin sustained signaling inhibition and tumor shrinkage⁶². Additionally, mutations in NRAS, MEK1, BRAF, PTEN, and amplification of MET are common in melanomas resistant to BRAF inhibition. Diverse PDX models bearing these different mutational backgrounds were all resistant to BRAF inhibition, but all responded to the combined treatment of the MET inhibitor capmatinib, the BRAF inhibitor encorafenib, and the MEK inhibitor binimetinib. Because high MET amplification alone was not sufficient to predict response to capmatinib, BRAF inhibitor-resistant PDX models can be used to improve clinical prediction of therapeutic response to BRAF inhibition in melanoma patients⁶³. As an example of using PDX to investigate drug resistance in hematological malignancy, PDX modeling in lymphoma models of acquired resistance to ibrutinib, a common B-cell targeting therapy, revealed an increase in PI3K pathway signaling in resistant PDXs compared to parental; targeting this pathway using idelalisib in tandem with ibrutinib was able to attenuate the growth of resistant models⁶⁴.

PDX models are also used to identify biomarkers of resistance. In ovarian cancer, for example, gene expression analysis of residual tumors following treatment of BRCA-mutant ovarian PDX models with PARP inhibitors revealed biomarkers for PARP inhibitor response (including CA125, KRAS, and ATK1)⁶⁵, whereas RAD51C methylation predicted resistance⁶⁶. Analysis of melanoma PDX models that were either sensitive or resistant to karonudib, an inhibitor of NUDT1 (also known as MTH1) identified ABCB1 as a marker for karonudib resistance⁶⁷.

Genetic screens in PDX models

PDX are amenable to genetic manipulation for mechanistic studies. PDX-derived cell lines have been created as ‘xeno-cell lines’ that can be utilized alongside PDX for mechanistic analysis. In HER2⁺ CRC, individual transduction of these xeno-lines with mutant BRAF, KRAS, or PIK3CA provided insight into the resistance that these mutations provide to the common therapeutic combination of lapatinib and trastuzumab used clinically for HER2⁺ CRC patients⁶⁸. This revealed the importance of taking each patients’ mutation profile into account when predicting treatment efficacy. In acute leukemias, genetic screening techniques in PDX models, such as use of Cre–ERT2–loxP-inducible RNA interference mediated-gene silencing or CRISPR–Cas9 knockout, have been used to identify genes that are crucial for tumor cell survival^69–71. These genes identified included the sodium/myo-inositol cotransporter SLC5A3, the E3 ligase MARCH5, double homeobox protein 4 (DUX4), and the fusion MLL–AF4^69–71, and provide the basis for development of pharmacological inhibitors that target the proteins that these genes produce, in order to address mutations in patient tumors and provide new treatment strategies. For example, upon silencing MARCH5, the efficacy of venetoclax, a BCL2-targeting drug used for treatment of acute myeloid leukemia, was improved.⁶⁹ This represents a new avenue for venetoclax co-treatment strategies that was previously unknown.

PDX and PDX-derived models for drug screening

An important approach in precision oncology is the use of PDX or PDX-derived models, such as PDXOs or xeno-cell lines, to identify new therapies that are effective in tumors with defined characteristics. This is important for identifying drug responses that are not directly related to known genetic or other vulnerabilities in tumors. Although PDX models are not generally amenable to high-throughput screening, creative solutions have been developed to efficiently screen drugs in PDX and derived models. One example is the OncoVee^® MiniPDX platform that enables evaluating three tumors simultaneously in one mouse. In this system, PDX-derived or primary tumor cells are loaded into individual hollow fiber capsules, which are then implanted into mice that are treated with drugs. Results in the MiniPDX system closely recapitulate those in regular PDX and can reduce time and costs of materials through maximizing each individual animal model⁷². Other unique approaches include drug screening on ex vivo PDX-derived tissue explants⁷³, and use of implantable microdevices that can continuously dose PDX models in vivo⁷⁴. Large scale PDX drug screening using over 300 PDX models has also been conducted using a design analogous to human clinical trials whereby 1 mouse carrying 1 tumor type receives 1 drug. When utilizing very large PDX collections, this 1×1×1 design increases scalability and allows a view of heterogeneous patient responses⁷⁵.

PDX are of particular value in evaluating drug responses of rare cancers where few relevant models exist. PDX models of the rare liver cancer subtype fibrolamellar carcinoma were used to demonstrate that both primary and metastatic PDX models were sensitive to inhibition of topoisomerase I (TOP1), HDACs, and to napabucasin (a drug developed to target cancer stem cells). Response to these drugs was correlated to expression of the anti-apoptotic protein BCL-xL and inhibition of BCL-xL in combination with the above therapies increased response as compared to single drugs⁷⁶. Another example is drug screening in CD19⁺ hepatocellular carcinoma (HCC) PDX, a tumor subtype that has no specified treatment option. This study identified the multi-kinase inhibitor regorafenib to be highly effective in CD19⁺ HCC and demonstrate that, mechanistically, regorafenib restricts mitochondrial respiration to promote apoptosis in CD19⁺ HCC cells⁷⁷.

PDX drug screening has been key to uncovering treatment strategies for primary or drug-resistant rare sarcomas including undifferentiated/unclassified soft tissue sarcoma, leiomyosarcoma, synovial sarcoma, spindle cell sarcoma, osteosarcoma, and pleomorphic liposarcoma^78–85. Discovery of notable novel treatment strategies include pazopanib for synovial sarcoma and osteosarcoma, recombinant methioninase (rMETase) with doxorubicin for spindle cell sarcoma, and trabectedin for pleomorphic liposarcoma^78–85.

PDX are also used to screen for improved therapies that target specific known signaling mutations or resistance mechanisms within tumors. In melanoma, PDX were used to screen for new therapies for BRAF^V600E mutant patient tumors. Patients with BRAF-mutated cancer would typically be prescribed the BRAF inhibitor vemurafenib as a first-line therapy. However, vemurafenib treatment in PDX models did not show anti-tumor effects. Surprisingly, the MEK inhibitor trametinib was most effective, despite the BRAF inhibitor vemurafenib being the obvious choice⁸⁶. These results highlight an opportunity of employing FPO at the forefront of patient care to model and personalize each patient’s treatment beyond a “one size fits all” approach.

Another powerful screening technique is high-throughput ex vivo therapeutic testing using primary PDX-derived cells, PDXOs, or PDX-derived cell lines. Although removed from the in vivo environment, these models allow for quicker experimental turnaround and expanded drug panels, and results can serve as the basis for more selective preclinical validation in their matched PDX in vivo. In medulloblastoma, PDX models were used to find tailored therapies for patients to either eradicate tumors with fewer side effects, or overcome resistance frequently seen to the hedgehog signaling inhibitor sonidegib. Screening of over 4400 drugs using PDX-derived cells, and in vivo PDX validation of promising candidates, revealed that actinomycin D treatment was effective against the most aggressive medulloblastoma subtypes, and that selected resistance to sonidegib induces new vulnerabilities to the nuclear export inhibitor selinexor^34,87. The frequent problem of multi-drug resistance in acute myeloid leukemia (AML) was addressed using AML PDX-derived bone marrow cells isolated from matched parental and acquired-resistant PDX models that underwent dynamic BH3-profiling and were tested with various FDA-approved and investigational targeted therapies. Drug resistance was consistently marked with decreased apoptotic priming within tumor cells, and treatment of models with the SMAC mimetics birinapant and LCL-161, venetoclax, or vemurafenib were able to induce cell death and shrink tumor size in PDX⁸⁸. In gliomas, PDXOs were used to identify sensitivity to temozolomide, a common chemotherapy for brain cancers, as well as erlotinib and palbociclib to target EGFR and CDK4 and CDK6, respectively. Resistance to temozolomide was addressed through additional screening, revealing that the FDA-investigational therapy dianhydrogalactitol was able to overcome resistance⁸⁹.

To improve scalability of drug screening in PDX and PDX-derived models, new and unique high-throughput platforms to efficiently screen hundreds to thousands of drugs and gather data on response and resistance were developed. In T-cell acute lymphoblastic leukemia (T-ALL), a unique stepwise approach was developed subjecting PDX-derived cells to a large panel of therapies to first identify singularly effective drugs, and then drugs that remain effective in the presence of endothelial cells, which are known to modify drug response in T-ALL⁹⁰. Several drugs identified, such as ruxolitinib and irinotecan, were able to kill PDX tumor cells in vivo in the presence of endothelial cells⁹⁰. In hepatocellular carcinoma (HCC), the quadratic phenotypic optimization platform (QPOP) was used to analyze results from a PDXO drug screen of proteosome and kinase inhibitors to select treatment combinations. The combination of the proteasome inhibitor ixazomib and the CDK inhibitor dinaciclib demonstrated synergistic activity in PDX and PDXO models, surprisingly outperforming sorafenib, the current first-line therapy for HCC⁹¹. This combination shows strong clinical promise and demonstrates the power of these screening platforms to bring unexpected therapeutic options to patients. PDX-derived breast cancer brain metastasis tumor cells were used to screen over 300 potential therapies for breast cancer brain metastases, revealing sensitivity to histone deacetylase and proteasome inhibitors across PDX models⁹². High-throughput PDXO drug screening for breast cancer has also been performed using short term stable cultures⁹³ and screening stable PDXO lines from several breast cancer subtypes revealed remarkable sensitivity of an as-yet-undefined subset of TNBCs to the SMAC mimetic birinapant, as a single agent or combined with chemotherapies⁹⁴. Similar approaches were used in CRC, where a combination of PDX and matched PDX-derived cell lines were screened to find new treatment options for primary and metastatic CRC. The tyrosine kinase inhibitors ponatinib and erlotinib, irinotecan (DNA-targeting chemotherapy), bevacizumab (VEGF inhibitor), and cetuximab (EGFR inhibitor) all showed tumor-eradicating effects^33,95,96. Additionally, co-inhibiting CDK2 and CDK9 showed potent antitumor effects^33,95,96.

PDX drug screening, whether done in vivo or ex vivo, benefits from integration of omics into the pipeline, to identify biomarkers for drug response or resistance, and even prospectively predict tumor behavior. In ovarian cancer, PDX models have been used to identify novel therapy regimens that eradicate tumor growth, and retrospectively select biomarkers for response. Screening monotherapies and combination therapies revealed that combined treatment with the PI3K inhibitor buparlisib and the MEK1 and MEK2 inhibitor binimetinib, or single agent treatment with the anti-cadherin 6 antibody HKT288 or olaparib were able to eradicate PDX tumors⁹⁷. Sequencing of responsive tumors revealed biomarkers of response for each treatment. Using these biomarkers, approximately 90% of PDXs were prospectively assigned to each treatment regimen and evaluated for drug response. These selected PDX responded to their designated treatment regimen, and this guided therapy approach even elicited tumor killing response in PDX models resistant to standard-of-care chemotherapies, representing potential clinical treatment options and demonstrating the benefit of individual drug selection for patients based on tumor biomarkers⁹⁷.

PDX modeling beyond mouse hosts

Not all PDX are implanted into mouse hosts. Although mice are a widely used model organism due to conservation of many biological processes and robust genetic engineering techniques, the immunodeficient mice required for PDX experimentation are costly and, as discussed above, do not facilitate high-throughput testing. The advantages and drawbacks of mouse and alternative PDX models are described in Table 1.The utilization of zebrafish PDX (zPDX) offer benefits of faster turnout time, lower cost, and higher throughput drug testing in vivo^17,98,99,100. Various methods of making zPDX directly from patient tumors and with serial transplantation have been developed; these details have been covered in recent reviews^98,101. Detailed characterization and functional testing has shown that zPDX recapitulate response trends observed in patient-derived organoids or mouse PDXs when compared head-to-head^102,103. Therefore, zPDXs can increase feasibility of drug screening because of the ability to efficiently inject tumor cells and screen drugs in many animals¹⁰⁴. For example, zPDX models have been used as precision oncology models to test drug responses and screen a variety of therapies for acute myeloid leukemia¹⁰⁵. Additionally, zPDX models were used to screen a range of targeted therapies and quickly modeled response, partial response, and resistance within 3 days to a panel of targeted therapies using tumor samples from a range of pediatric cancers¹⁰³. This is in contrast to mouse PDX experiments, which typically take weeks to months.

Table 1.

Animal models used for PDX precision oncology.

Host	Advantages	Disadvantages	ref
CAM	Faster screening time (few days); Low cost; Metastatic capability can be assessed;	Potential for inflammatory responses after day 15 Not suited to predict metastasis to other organs Immune-deficient Lacks human stroma and vasculature Short drug testing time	Armstrong et al162 Tsimpaki et al163 Ribatti164 DeBord et al165
Zebrafish	Fast turnout for drug screening (few weeks); Metastatic capability can be assessed;	Not suited to predict metastasis to specific organs; Immune-deficient; Lacks human stroma and vasculature;	Haldi et al166 Marques et al167 Yan et al99 Astone et al17 Fazio et al98
Mouse	Preserves much tumor heterogeneity; Reliable for drug testing, resistance; Toxicity validation and prediction; Wealth of preclinical data; Ability to manipulate host genetics;	High cost, low throughput; Generally immune-deficient; Lacks human stroma and vasculature;	Fiebig et al168 Woo et al7 Liu et al8 Hidalgo et al169
Rat	Bigger tumors can be established; Can serve as a model for some cancer types that have low engraftment rates in mice;	Higher cost for breeding and housing compared to mice; Generally immune-deficient; Lacks human stroma and vasculature; Fewer host strains available;	Noto et al110 Ozaki et al170
Pig	Large tumor size relevant to humans; Similar dielectric, anatomical and physiological properties, organ size to humans	Require bigger space and higher cost compared to rodents for pathogen free housing; Longer time to breed; Generally immunosuppressed;	Hendricks-Wenger et al114 Hoopes et al115

Chicken embryo chorioallantoic membrane (CAM) models can also serve as a host for PDX to enable high-throughput screens that are rapid and cost effective. CAM-PDXs are made by transplanting tumor pieces directly from patients or as serially transplanted tumor tissue directly onto the surface of a normal chicken embryo inside of a fertilized egg. Tumors can then be visualized during drug treatments to assess effects. In bladder cancer, CAM-PDX were used to find new treatment strategies for non-metastatic muscle-invasive urothelial bladder cancer, which are often resistant to neoadjuvant chemotherapy. Screening of kinase inhibitors in CAM-PDXs revealed response to combined inhibition of HER2 and EGFR signaling¹⁰⁶. CAM-PDX have also been used to investigate therapeutic response in renal cancer¹⁰⁷ and for engraftment of breast tumors¹⁰⁸.

Although less frequently utilized, rat and pig PDX models bear advantages, such as growing larger PDX tumors over a longer period of time in longer lived animals¹⁰⁹. The development of immunodeficient Sprague-Dawley Rag2/Il2rg rat models offers the benefit of directly engrafting cancer PDX models that are not very successful in immunocompromised mice; for example, the VCaP prostate cancer PDX model¹¹⁰, or estrogen receptor (ER)⁺ breast cancer PDX, which grow better in rat mammary glands compared to mice¹⁰⁹, probably due to the higher levels of estrogen in rats compared to mice^111,112 and the fact that rats are advantageous over mice in developing estrogen-dependent mammary tumors¹¹³. Pigs represent a good model for testing devices that require closer anatomical similarities to humans. For example, CRISPR/Cas9 technology was used to generate ‘on-demand’ immune deficient pigs as hosts for human pancreatic tumors, to grow large amounts of human tumor that was needed to study electrophysiology of the tissue¹¹³. In another study, Yucatan mini pigs were used following immunosuppression with cyclosporin, prednisone, and mycophenolate as hosts for human glioblastoma tumors, as a model with intracranial volume more similar to humans for studies of tumor resection^114,115.

Use of PDX for functional precision medicine

The use of PDX and PDX-derived models as tools to predict recurrence and inform individualized therapy selection in clinical trials of FPO has gained traction¹⁵. Drug responses in PDX models have closely recapitulated patient response profiles observed in the clinic, leading to improved prospective treatment strategies for patients informed by their PDX or PDXO models^{16,59,94,116–118}. For instance, to inform clinical care for breast cancer patients, matched PDX and PDXO were developed and used to screen and validate potential therapies^15,94. Treatment with the chemotherapy eribulin showed the strongest anti-tumor effect compared to other regimens (such as the combination of doxorubicin, cyclophosphamide and paclitaxel known as AC-T) and proved to be effective in the patient after disease recurrence, ultimately demonstrating the power of harnessing these models to tailor a therapy regimen to the individual^15,94. Similarly, OncoVee^® MiniPDX models of a patient with metastatic duodenal adenocarcinoma were used to test and validate chemotherapeutic regimens that were selected based on RNA-sequencing of tumor biopsies¹¹⁶. To provide quicker turnaround with high accuracy and predictive capabilities, multiple efforts have also aimed to incorporate zPDXs into FPO-based clinical trials in addition to mouse PDX^{72,94,117,119}. In gastric cancer, zPDX models were used to screen chemotherapeutic drugs and to identify the most effective treatment for each individual patient. Patients were subsequently treated with the chemotherapy regimen that was predicted to be most effective based on their PDX models, and their clinical responses closely matched the PDX-based predictions¹¹⁷. Along the same line, a rhabdomyosarcoma zPDX model was also used to test combined olaparib and temozolomide efficacy⁹⁹ and has progressed to an ongoing clinical trial for Ewing Sarcoma (NCT01858168¹⁸¹). Another example is the use of zebrafish to predict responses of ovarian cancer patients to individual chemotherapies¹²⁰, demonstrating the predictive power of alternative host PDX models for FPO.

FPO clinical trials have built upon these platforms to provide patients with precise treatment strategies that are informed by each tumor’s unique biology. PDX and PDXO models have the potential to be used clinically as early detection of recurrence for aggressive cancers with limited diagnostic opportunities through the observed correlation of in vivo patient tumor engraftment success with metastatic potential¹⁵. This enables physicians to have additional time to develop treatment plans for breast cancer recurrence (TOWARDS-II, NCT05464082¹⁸²), expand the treatment options of aggressive tumors (FORESEE, NCT04450706¹⁸³); (ZERO2, NCT05504772¹⁸⁴), or identify therapies for patients with tumors resistant to standard treatments (PPDTPC, NCT04373928¹⁸⁵)^94,121,122.

Limitations and challenges of PDX models

Challenges to establish PDX

Successful PDX establishment for precision oncology studies is complex and requires approval of Institutional Review Boards [G] of both funding agencies and institutions, as well as integration between clinical staff and research teams to facilitate timely acquisition of viable tumor samples. For FPO trials, there are a plethora of complications involving standardization of testing such as compliance with Clinical Laboratory Improvement Amendments regulations, and agreements with insurance companies to cover off-label drug administration. This complex subject has been previously reviewed³.

Communication barriers between oncologists and patients and their families^123,124 also represents a formidable challenge. Oncology clinicians often face the challenge of clearly explaining what precision oncology is, balancing potential benefits, uncertainties, and limitations, to ensure realistic expectations without overpromising outcomes¹²³. Creating PDX models for FPO also requires tumor samples, which can create additional discomfort for patients and/or require more visits to the clinic.

Disparities in biobanking and representation of PDX models

Participation in clinical trials and specimen donation are key to advancing PDX establishment and related research in precision oncology. The majority of biospecimen repositories in the U.S. are overrepresented with tumors from patients of European descent¹²⁵, and minority racial and ethnic groups are underrepresented in precision oncology clinical trials^126,127. Such discrepancies need to be addressed, especially in light of recent studies that revealed that biological differences in tumors from patients of diverse ancestries contribute to worse clinical outcomes for minority patients with CRC¹²⁸ and TNBC¹²⁹. An important new initiative within the NCI-funded PDX Network (PDXNet) is developing and characterizing 200 PDX models for gastric, liver, bladder, and lung cancers of which 60% of tumor specimens from underrepresented patients (NCT04410302¹⁸⁶).

Other barriers to equitable FPO trials include the relatively small number of institutions that implement such studies, which are typically larger academic cancer centers, and socioeconomic factors inhibiting participation, such as travel expenses^130,131. It has been proposed that systemic changes in outreach and increased funding to cover expenses would encourage broader public involvement in cancer clinical trials¹³⁰, thus increasing the biodiversity of specimens and the relevance of the assays to broader groups of people.

Unfortunately, not all patient tumors grow as PDXs and some cancers have especially low take rates [G], representing a barrier to having good representation of all cancer types as PDX models and/or implementing them in FPO trials. Prostate, breast, gastric, liver, kidney, and bladder cancers as well as uveal melanoma are examples of tumor types with low take rates^132,133. In addition, strong representation of all cancer subtypes can be challenging as, biologically, some tumors require very specific niches to grow. As already eluded to, a good example is endocrine-resistant ER⁺ breast cancer as these tumors are difficult to grow as PDX, but supplementing mice with estradiol (E2) improves growth of some ER⁺ breast cancer PDX¹³⁴. This emphasizes that an improved understanding of limiting factors can help to overcome issues of growing cancers with low take rates and improve the representation of PDX models across all cancer types^94,135.

Tumor heterogeneity and evolution

Although PDX models require more time than most cell-culture based assays, their ability to retain genetic and molecular features of patient tumors in vivo provides untapped potential for precision oncology. Obtaining samples from patients at initial diagnosis is often more feasible than relapsed or refractory disease, due to a myriad of factors including the location of the tumor, the complexity involved in obtaining the biopsy, and disease stage. Despite these challenges, using PDX to model longitudinal samples from patients collected at multiple time points of their disease, including primary, recurrent and/or refractory disease, can help us better understand the mechanisms that contribute to inter- and intra-tumoral heterogeneity, tumor evolution over time, and drug resistance.

PDX models that are matched to individual patients at different stages of their disease allow for in vivo drug testing that may include standard of care, experimental regimens and combination therapies that cannot be tested in the patient. PDX models can be used as surrogates for patients’ responses to therapy at different stages of disease and on a parallel time scale. This may reveal new therapeutic possibilities for tumors that are refractory to existing treatments. In cases where matched tumors can grow as PDX, they provide invaluable insight into tumor evolution over time and altered responses to drugs, opening avenues to tailor treatment regimens for future patients. Such longitudinal studies, however elegant, are limited by several factors including the amount of tissue available from surgical specimens, how these tissues are processed, whether the acquired samples are adequately representative of the tumor’s heterogeneity, and compounded effects of potentially low engraftment rates for different tumor types and stages.

Standards of validation

To maximize benefit of PDX for precision oncology, the field requires development of rigorous standards to ensure reliability of results. The recently developed Minimal Information standards for PDX establishment and validation (PDX-MI), designed to be shared across institutions to ensure reliable and reproducible data¹³⁶, is an excellent example. PDX-MI includes four general modules that govern PDX mouse model generation and use in a comprehensive manner. They cover the information being documented about patient and tumor characteristics, details of how the PDX model was created, specifics of the tissue origin and means of validation, and data on the genomic characterization and response or resistance to treatment. Adopting similar standards for how drugs are tested (dosing regimens in mice compared to humans; size of tumors being treated, length of treatment, etc.) will increase rigor and reproducibility of results. Indeed, the field is moving in this direction; to bridge the gap between information gained from preclinical PDX studies and clinical trials, recent work published by the NCI PDXNet consortium presented recommendations that deal with the standardization of assessment of PDX growth and antitumor activity¹³⁷. Having universally adopted standards for evaluating responses, and ways to make PDX models and data available publicly around the world, could lessen the time taken to identify novel therapies that translate to meaningful therapeutic responses for patients¹³⁸. Additional publicly available data and validated model repositories include CancerModels.org, the NCI Patient-Derived Models Repository, the Pediatric Preclinical In Vivo Testing Consortium, the Singapore Translational Cancer Consortium and EurOPDX.

Time, cost, and scalability

While PDX are arguably the most relevant in vivo preclinical models to discover and test new cancer therapies⁸, their high cost, low take-rates, and high turnaround time brings challenges to precision medicine research and FPO. In the long run, optimizing precision oncology approaches may have the potential to lower costs for patient care by choosing effective therapies from the outset, and avoiding overtreatment and unnecessary toxicities of ineffective drugs, but these types of detailed analyses still need to be conducted¹³⁹. In the meantime, efficiencies for PDX research are being improved. Establishment rates of PDX depends on multiple factors such as the tumor type, sampling site, quality and quantity of tumor tissue, the aggressiveness of the cancer, and the mouse strain used⁸. The time needed to develop and appropriately validate PDX models before drug testing reduces turnaround time and can be problematic, especially for FPO assays where patient treatment is concurrent with the trial. As some aggressive cancers evolve very rapidly, FPO with PDX may not be able to keep up with cancer progression and clinical treatment, and previously established PDXs representing earlier stages of cancer might not always be informative for the patient’s subsequent relapsed or refractory disease.

When more than a few drugs need to be tested, increased scalability is important, thus many groups are turning to PDXO or other ex vivo systems¹⁴⁰, which sometimes have lower turnaround times (Figure 3). The use of PDXOs also reduces cost of the drug testing portion of the study, while retaining the ability to validate findings in vivo in the matching PDX¹⁴¹. Although organoids and other ex vivo systems have benefits for FPO drug testing (as reviewed elsewhere^5,6) there are multiple situations in which the in vivo environment is critical, and novel assays, such as the mini-PDX now allow assessment of drug response within 7 days^72,142.

Figure 3. Scalability of drug testing in various patient derived models.

Emerging patient-derived xenograft (PDX) models, such as chicken egg chorioallantoic membrane (CAM) PDX, zebrafish PDX, short-term ex vivo or long-term stable PDX-derived organoids (PDXO), and 1 mouse, 1 tumor, 1 drug (1×1×1) models sometimes offer faster and more scalable models for functional precision oncology, as numerous drugs and drug combinations can be tested at the same time. While standard mouse PDX remain the most highly translatable models, they can take a longer time to establish and validate and are therefore limited to smaller scale drug testing

Lack of functional immunity in PDX models

The tumor microenvironment is an important contributor to tumor progression and therapeutic response. While PDX tumors in animal models preserve many tumor characteristics and are suited for many types of studies outlined here, a key caveat of PDX models is the absence of human tumor microenvironmental components including extracellular matrix and stromal cells, especially immune cells, from standard PDX models. This limitation precludes the use of these human tumor models to test various aspects of immunotherapies, which have shown high efficacy in combating certain tumor types, but only for a subset of patients¹⁴³. To address this issue, humanized PDX models have been developed, whereby human bone marrow, cord blood, or other immune cell populations are reconstituted in immunocompromised mice prior to or during PDX development^144,145. As an example, immunotherapies elicited tumor regression in ovarian PDX spheroid and gastric cancer PDX models in mice injected with autologous peripheral blood mononuclear cells¹⁴⁴. In another instance, the humanized mice were useful in testing the potential efficacy of immunotherapies against PDX representing platinum-based chemotherapy and PARP inhibitor resistant BRCA-associated pancreatic cancer with persistent homologous recombination deficiency. They found that the immunotherapy reduced the tumor growth in humanized mice (injected with CD34⁺ cells were isolated from umbilical cord blood)¹⁴⁶. While informative for assessing the efficacy of specific immunotherapies on a more personalized level, these studies require collection of blood and tumors from the same patient, and need to be performed in window of time that avoids graft versus host disease¹⁴⁷. An alternative method for autologous immune incorporation to PDX models is isolating infiltrating lymphocytes from the same tumor tissue sample that is used to establish the PDX, which has been tested in melanoma and ovarian cancer PDX models, enhancing the effect of immunotherapy on tumor control¹⁴⁴. Another alternative is to engineer human thymus organoids from induced pluripotent stem cells [G] that can be used to reconstruct a human functional T cell compartment in mice, capable to reject allogeneic tumors¹⁴⁸. Future studies incorporating components of the human immune system into PDX models in various types of PDX hosts will allow the expansion of our understanding of immune-tumor interactions, and how immune cells influence response to drug treatments, not necessarily restricted to today’s immunotherapies. Although human immune cells can be xenotransplanted into zebrafish¹⁴⁹, rats¹⁵⁰, and pigs¹⁵¹ to reconstitute some aspects of the human immune system, we are not yet aware of humanized PDX studies for cancer in these hosts.

Looking forward: PDX models and artificial intelligence

In addition to improvements made by addressing the limitations discussed above, we believe the future of PDX in precision oncology will truly be transformed by innovations in computational analysis and artificial intelligence (AI). AI is broadly defined as a technology that harnesses computation to process large amounts of data, recognize and learn patterns from representative data, solve problems, and make predictions¹⁵². This technology encompasses machine learning [G] and deep learning [G] via neural networks [G], which can work together to process vast datasets and uncover complex patterns at speeds and depth beyond human capability¹⁵³. The intersection of precision oncology, PDX models, and AI represents untapped potential for improving cancer care. Deep learning networks have the ability to characterize complex diseases, factoring in patient-centered data, clinical responses, molecular features, and PDX modeling of therapeutics to build robust, predictive computational models. AI-based algorithms can take advantage of this multimodal data, expediting the development of novel diagnostics and therapeutics, which can be rigorously tested in PDX models. AI can accelerate drug discovery by comprehensively analyzing the genetic, epigenetic or transcriptional, proteomic, and/or stroma and immune-related profiles of tumors (patient and PDX) and associating these data with drug efficacy. As an example, to identify vulnerabilities across a range of cancer types, the OncoTarget and OncoTreat¹⁵⁴ platforms were developed, which use integrated PDX gene expression and protein activity prediction using the Virtual Interference of Protein-activity by Enriched Regulon analysis (VIPER) algorithm¹⁵⁵ to predict patient-specific responses to >30 therapies. Both platforms leverage knowledge of master regulator [G] and ´tumor checkpoint proteinś to propose inhibitors that might interfere with tumor viability. OncoTarget focuses on individual master regulators and the anticipated direct effects from pharmacological inhibition, whereas OncoTreat employs a more global approach, leveraging known tumor transcriptional state signatures and predicting how inhibitors might disrupt the ´modulé of master regulators linked through each signature. In PDX models representing a range of different malignancies, both platforms predicted treatments that induced significant tumor response compared to vehicle treatment or randomly selected treatments. In addition, both platforms held strong prediction capabilities alone or combined, indicating that individual or clustered master regulators can be targeted to produce a similar antitumor effect. This approach provides a potential avenue for precision oncology drug prediction for multiple cancer types based on features of the tumor transcriptional state¹⁵⁴.

Because machine learning algorithms work best with larger and more diverse datasets, data obtained with large numbers of PDX and drug testing experiments, such as those being generated by the PDXNet, can be used to develop algorithms to accurately predict drug response without the need for FPO testing of tumor models for individual patients. This is especially timely with the 2021 FDA ruling, where showing therapeutic efficacy in animal models prior to clinical testing is no longer a requirement¹⁵⁶. Indeed, AI is already being used to develop prognostic tests applicable to breast cancer patients¹⁵⁷, to shorten interpretation time of results by radiologists in lung cancer¹⁵⁸, and for the selection of ideal drugs for lung cancer patients¹⁵⁹. AI can also be used to discover network-based biomarkers from patient tumors and PDX for monitoring treatment response. This utility has been demonstrated in a pan-cancer study, where integration of multimodal data predicted optimal therapeutic regimens with maximum efficacy and minimal toxicity in cell lines. These AI-driven findings suggest optimal combinations that can be further validated in PDX models, prior to being tested in patients¹⁶⁰. In ovarian cancer, immunofluorescence imaging paired with machine learning is being used to predict trends in treatment outcomes in patients⁷⁴. PDX models can be used as models in which to test or validate these types of hypotheses. In addition, PDX models may also facilitate advances in early cancer detection for future precision oncology applications. Training convolutional neural network algorithms on a wide range of images from patients and matched PDXs can potentially identify cancer at earlier stages and track how the tumor changes in response to treatment or immune infiltration over time, thereby allowing for early-stage cancer interventions that save lives¹⁶¹.

Despite its promise, the integration of AI into PDX-based FPO faces numerous challenges. Key considerations must be given to the collection and standardization of data across institutions and around the world. Model and data standardization is critical to ensure the collection of high-quality datasets, comprising results that can be reliably reproduced and interpreted and serve as foundational models for machine learning. Algorithms will need to be shared and rigorously tested across datasets and model systems to optimally advance data analysis, comprehensive predictive modeling, drug discovery and personalized treatment strategies during this new era of FPO. This will empower a paradigm shift in healthcare and create a new era of patient-centric care coupled with improvements in clinical outcomes. While several notable challenges exist, addressing these challenges is feasible with technological advancement, interdisciplinary collaboration, and commitment to enhance the translational relevance of PDX models in precision oncology.

Conclusions

Optimizing precision oncology holds tremendous promise for improving cancer care by utilizing individual tumor data comprising genetic vulnerabilities, epigenetic states, and immune status to predict drug responses. In the near future, we envision that, in addition to preclinical research involving PDX, precision oncology is conducted in real time using FPO assays to direct personalized therapy. In cases where the FPO assays are based on PDX, the PDX also serves as a renewable resource for future preclinical research for that tumor type, providing a virtuous circle to advance precision oncology strategies and identify new therapies.

Despite existing challenges, investing in enhancing FPO approaches can yield highly relevant and comprehensive preclinical models that rigorously test new targets and therapies. This will lead to better-informed clinical trials and more personalized treatments that ultimately improve patient outcomes. By fostering cross-disciplinary collaboration, adopting rigorous model validation standards and integrating AI-based tools, the future of FPO presents exciting possibilities for a new frontier of cancer care.