A Patient-Derived Xenograft Repository Capturing Clinical and Molecular Heterogeneity of Large B-cell Lymphoma Authors: Haopeng Yang, Kotaro Arita, Kevin Bowman, Dai Chihara, Jared Henderson, Griffin Rost, Estela Rojas, Sydney Parsons, Priya Lakra, Aneela Syeda Abedin, Sattva S. Neelapu, Paolo Strati, Loretta J. Nastoupil, Luis E. Fayad, Swaminathan P. Iyer, Maria Alma Rodriguez, Fredrick B. Hagemeister, Luis Malpica, Hun Lee, Laura K. Hilton, David W. Scott, R. Eric Davis, Christopher R. Flowers, Jason R. Westin, Giorgio Inghirami, Francisco Vega, Michael R. Green Published: 2026 ABSTRACT The development and characterization of large B-cell lymphoma patient-derived xenograft models that span the transcriptional and molecular LBCL landscape and capture important clinical scenarios like CAR T-cell resistance. Introduction Large B-cell lymphomas (LBCL) are an aggressive but clinically and molecularly heterogeneous group of malignancies. Although the majority of cases can be cured with first line chemo-immunotherapy regimens (1), relapsed or refractory cases (rrLBCLs) are difficult to manage, highlighting a need for model systems to study resistance mechanisms and test new therapeutics. To date, cell lines have been the predominant model system for the study of LBCL. However, the most used cell lines were developed in the pre-rituximab era, largely from effusions or leukemic-phase disease, and thus may not fully capture disease biology and/or resistance mechanisms to modern therapies. Although resistance to small molecules can be induced by chronic low-dose exposure to a drug (2), there is no clear path toward modeling resistance to cell therapy in vitro. Mechanistic and preclinical studies within the context of cell therapy resistance therefore currently rely on the development of models from patient biopsies following cell therapy failure for which resistance was developed in vivo. However, such model systems have not yet been described or made available for academic research, to the best of our knowledge. Patient-derived xenograft (PDX) models can be generated by implanting patient tumor biopsy material into immunodeficient mice, with a higher success rate for growth than in vitro culture. Some efforts have been made toward generating and characterizing LBCL PDX models (3, 4), demonstrating fidelity to the originating patient biopsy (4–6). However, these predate the CD19 chimeric antigen receptor T cell (CART19) era, the number of PDX models remains too small to capture the molecular diversity of LBCL, and not all are openly available to investigators outside of the originating center. We therefore undertook an initiative to generate a large repository of LBCL PDX models, including from relapses occurring following CART19, to provide an academic resource to advance preclinical and mechanistic studies of LBCL. Here, we describe our experience in this initiative, detail the first cohort of 48 models, and show evidence of their utility as models of CART19 resistance. Results Generation of PDX Models of High-Risk LBCL We generated PDX models by renal capsule implantation of core needle biopsies (entire core) of patients with LBCL in NOD.Cg-PrkdcscidIl2rgtm1Wjl/SzJ (NSG) mice, according to standard operating procedures from the Public Repository of Xenografts (3). No alternative sites of implantation or alternative murine strains were evaluated for efficiency. Mice were monitored for up to 6 months after implantation, and those with no detectable tumor by MRI were euthanized. No graft versus host disease was observed in any of the mice over this time. Tumor growth was regularly monitored by palpation, and mice that developed tumors or became sick were euthanized. The tumors (F1) were harvested and passaged fresh into the renal capsule of additional NSG mice. F2 tumors were extensively characterized, and fragments were cryopreserved for future use (Fig. 1A). Using patient biopsy pathology reports from synchronous cores that were clinically evaluated for diagnostic purposes, we determined that 133 of the implanted biopsies were LBCLs in origin, including 98 diffuse LBCL (DLBCL), 19 high-grade BCL (HGBCL), 6 primary mediastinal large BCL (PMBL), and 10 other LBCL subtypes. HGBCLs had the highest success rate for PDX generation (53%), with other histologies having similar take rates (30%–34%), resulting in 48 new LBCL PDXs from 45 patients, including three pairs of PDX biopsies from the same patient (Fig. 1B). Follicular lymphoma (FL) grade 1 to 3A biopsies (n = 35) generated no successful PDX models. Over a period of ∼6 months, 55 biopsies with a suspicion of lymphoma were implanted the same day of the biopsy procedure without cryopreservation to evaluate whether fresh implantation would improve success rates for model generation. All other biopsies were viably cryopreserved and thawed for implantation. We evaluated the effect of implant timing on PDX success rate and noted a slight decline with cryopreservation and additional decline if the biopsy was cryopreserved for more than 30 days (Fig. 1C). However, fresh implantation of biopsies meant that diagnostic hematopathology reports were not available, and these biopsies therefore included nonmalignant tissue and other lymphomas that made them poorly selective for LBCL (Fig. 1D). In contrast, viable cryopreservation allowed the consultation of hematopathology reports from synchronous cores to select biopsies for implantation without a major reduction in implantation efficiency if implantation occurred within 30 days (Fig. 1D). PDXs tended to be from biopsies taken at later lines of therapy (median = 3, range = 0–10), in line with a selection for higher-risk disease (Fig. 1E). Tissue microarrays (TMA) were constructed from F2 tumors of each model to facilitate IHC staining for key proteins of diagnostic interest, including CD20, BCL2, CD10, MUM1, MYC, BCL6, and CD19. Comparison with available clinical work-up of synchronous cores showed strong concordance for the majority of these select proteins when comparing dichotomous (positive/negative) classifications (Fig. 1F). However, heterogeneous patterns of expression were also observed across proteins and samples in tumors and their corresponding PDX (Fig. 1G). Histologies for the 48 new models are shown in Fig. 1H; clinical features are provided in Supplementary Table S1. LBCL PDX models can therefore be generated with high efficiency via renal capsule implantation into NSG mice, but it is recommended to use prior viable cryopreservation (<30 days) to target implantation toward subtypes of interest and optimize resource utilization. Generation of LBCL PDX models. A, A schematic of the workflow used for generating and characterizing PDX models. B and C, Success rates of implanted biopsies according to LBCL histology (B) and timing of implantation (C). D, Pie chart of the histologies of biopsies implanted without prior cryopreservation. E, Number of lines of prior therapy for successful PDX models. F, Agreement between IHC of PDX models and clinical workup of synchronous cores. Gain or loss indicates gain or loss of expression in the PDX model relative to the synchronous patient biopsy, as determined by an expert hematopathologist. G, Example of heterogenous MYC expression in a PDX model and synchronous biopsy. H, H&E staining of 48 newly generated PDX models. Dark boxes (outlined in thick black borders) indicate pairs of models from the same patient. Models are marked according to the histology assigned to the matching patient biopsy; #DLBCL (gray background), *HGBCL (blue background), and ^Other (green background). Full details of each model are in Supplementary Table S1. Mint-ChIP, multiplexed indexed native chromatin immunoprecipitation. [A, Created in BioRender. Yang, H. (2025) https://app.biorender.com/illustrations/6297dc307aa7149474a15bb7] PDX Models Span the Molecular Landscape of LBCL Molecular characteristics of PDXs were evaluated by whole-exome sequencing (WES), with germline control when available (25/48), low-pass whole-genome sequencing (lpWGS), and transcriptional subtyping with the DLBCL90 assay (Supplementary Table S2; ref. 7). For reference, we obtained and profiled LBCL PDX models from the Public Repository of Xenografts (PRoXe; ref. 3) and Weill Cornell Medical College (WCM) for a total of 80 models. Tumor material was not available for the implanted biopsies because the entire core was engrafted in mice. However, a subset of six patients had additional cores that were evaluated as part of a prior study (8), and three patients had residual cell pellets available from fine-needle aspirates, providing direct comparisons of WES data between the PDX models and the antecedent tumor for nine patients (Fig. 2). In all models, the variant allele frequencies (VAF) of somatic mutations were highly correlated (Pearson’s correlation P < 0.001 for all models) and mirrored the correlation observed for germline variants, suggesting that the observed variance was driven by the accuracy of VAF estimates from WES data (Fig. 2). There were no somatic mutations that were lost or gained in any of the models, including those in genes implicated as lymphoma drivers (Fig. 2; ref. 9), supporting the genetic stability of PDX models and fidelity to the antecedent tumor. Correlation between somatic mutation variant allele frequencies in PDX models and their antecedent tumor. Scatter plots show the variant allele frequencies (VAF) of germline variants (gray), somatic mutations in lymphoma driver genes (red), and other somatic mutations (blue) for nine models with available data. Pearson’s correlation R2 and P value are calculated from all somatic mutations. To further evaluate the observed frequencies of somatic mutations and molecular subtypes in our full PDX cohort (n = 80) and compare with those expected in a standard-of-care population of patients with rrLBCL, we compared with an independent cohort of rrLBCL tumors (n = 80; ref. 8). DLBCL90 analysis of the 48 newly generated PDXs identified four with PMBL molecular subtype, namely two PMBLs and two gray zone lymphomas, 36 activated B cell–like (ABC) subtype, 30 germinal center B cell–like (GCB) subtype, and 10 unclassified. Of the 30 GCB subtype, 15 were positive for dark-zone signature (DZsig; Fig. 3A). WES analysis identified recurrent mutations in genes with well-established roles in DLBCL (10–12), and GISTIC analysis identified significant gains and losses that have been previously described in DLBCL (Fig. 3A; Supplementary Fig. S1; refs. 11, 13). All LymphGen subtypes except for N1 were observed within the PDX models (Fig. 3B); however, the high genetic complexity led to ambiguous classifications of many models by LymphGen (Fig. 3A), a subtyping system that was generated using newly diagnosed DLBCL samples. Multiple LymphGen-informing mutations were enriched in PDXs compared with the comparator cohort (Fig. 3C), including EZB subtype genes such as CREBBP and KMT2D, although the frequencies of EZB classified samples were similar between cohorts (Fig. 3B). BN2-informing genes, NOTCH2 and SPEN, were also mutated at higher frequencies in PDXs, resulting in a higher frequency of BN2 classified PDXs compared with the comparator cohort (Fig. 3B and C). However, NOTCH2 mutations were not restricted to the canonical pattern of N-terminal truncating events that are seen in newly diagnosed DLBCLs of BN2 subtype (Fig. 3D; ref. 14), highlighting the importance of interpreting data according to mutation type rather than presence or absence. Mutations in TP53 were present in 42% of PDXs, and complex DNA copy-number profiles were common (Fig. 3E), though the A53 LymphGen subtype was called in a lower frequency of PDXs compared with the comparator cohort. Mutations in MYD88 and CD79B, and MCD classified tumors were also modestly underrepresented in PDXs, but this effort generated two new models with combined MYD88-L265P and CD79B-Y197 hotspot mutations that are characteristic of the MCD subtype (Supplementary Fig. S2). These LBCL PDX models therefore encompassed the molecular spectrum of rrLBCL but have biases in molecular features that may require future targeted efforts toward PDX generation to ensure adequate diversity. Molecular characteristics of PDX models. A, Oncoplot of DLBCL90 transcriptional subtype (NanoString), FISH, LymphGen classification scores (WES), mutations of individual genes (WES), and DNA copy-number gains and losses (lpWGS) across 80 F2-generation LBCL PDX models. B, Relative frequencies of LymphGen classes in PDX models and independent cohort of rrLBCLs (8), using whole-exome data. Samples are counted multiple times if ambiguously assigned to multiple classes, according to data in Supplementary Table S2. C, Relative frequencies of mutations in individual genes in PDX models and an independent cohort of rrLBCLs, using WES. D, Lollipop plot of NOTCH2 mutations from PDX models, using WES. E, Example DNA copy-number profiles from lpWGS. Models of CART19 Resistance Among the new PDXs, 23 were generated from biopsies obtained after CART19 failure, including 19 with CART19 as the immediate prior line of therapy (Fig. 4A; Supplementary Table S1). In addition, 11 PDXs from 10 patients were generated from biopsies obtained prior to receiving CART19, of which four patients progressed following CART19, and six patients were in complete response at their last follow-up (Fig. 4A; Supplementary Table S1). These models included three pairs of PDXs from the same patient. LPMI011 was generated from a biopsy taken following a partial response to CART19 and LPMI037 from the same patient after a second CART19 administration with progressive disease as the best response. LPMI495 and LPMI797 were generated from biopsies 13 months apart with five intervening lines of therapy, including CART19. LPMI550 and LPMI612 were generated from biopsies taken 3 months apart with two intervening lines of therapy, with the patient subsequently achieving a complete response to CART19 and remaining in remission at the last follow-up. CART19 progression in PDX models and CD19 mutation/loss. A, Swimmer plot showing lines of therapy, centered on the day of the biopsy that generated the PDX model. B, Examples of CD19 IHC in CD19-mutant (above) and CD19-negative (below) cases. C, Volcano plot of differentially expressed genes between CD19 IHC-negative and IHC-positive PDX models showing selective downregulation of CD19, using RNA-seq. D, Box plots showing CD19, CD22, and CD79B expression according to CD19 mutation and expression status. Center line represents the median, boxes span quartiles 25th to 75th percentile, and whiskers extend to minimum and maximum. E, Example tracks of chromatin accessibility over the CD19 gene locus showing cases from (B), using ATAC-seq. FPKM, fragments per kilobase per million; TPM, transcripts per million reads. CD19 antigen escape has been described as a mechanism of resistance to CART19 (15). WES analysis identified four PDXs carrying CD19 mutations, all generated from post-CART19 progression biopsies. Two possessed monoallelic truncating mutations that reduced CD19 protein expression, and two had point mutations within the loop recognized by FMC63, including mutation of arginine 163 that has been previously shown to abrogate CAR binding (Fig. 4B; ref. 16). CD19 DAB staining also revealed loss of CD19 protein expression in six PDXs, of which three were CART19-naïve and three were from after CART19 progression (Fig. 4B). Differential analysis of RNA sequencing (RNA-seq) of CD19-negative (n = 6) compared with CD19-positive PDX models (n = 74) showed significant loss of CD19 transcript without other broad accompanying signatures or loss of known CD19 regulators such as PAX5 (Fig. 4C; Supplementary Table S3; ref. 17), suggestive of selective silencing of CD19 expression. This was supported by Assay for Transposase-Accessible Chromatin using sequencing (ATAC-seq) data that showed a loss of chromatin accessibility at the CD19 locus in CD19-negative cases but not in CD19-positive cases, indicative of epigenetic silencing of CD19 expression (Fig. 4D). Notably, comparison of CD19 IHC from patient biopsies with the corresponding PDX model identified two instances in which CD19 was negative in the patient biopsy and positive in the PDX model, supporting the notion that CD19 expression is epigenetically regulated and may be subject to restoration. Furthermore, we did not observe loss of CD19 expression in PDX models when the patient biopsy was positive, suggesting that this is not a PDX-specific phenomenon. Importantly, CD19-mutant and protein-negative PDX models maintained variable levels of expression of other CAR targets such as CD22 and CD79B (Fig. 4E), establishing these as potential systems in which to test the efficacy of alternative CAR designs. Tumor Cell–Intrinsic Resistance to CART19 In Vivo The majority of CART19 progressions retain CD19 expression, highlighting tumor-intrinsic (18) or lymphoma microenvironment (LME) (8) factors in CART19 resistance. PDXs in immunodeficient mice do not maintain their human LME and are therefore not a tractable system for evaluating LME-driven CART19 resistance that we and others have described previously (8, 19). However, we aimed to investigate whether post-CART19 progression PDXs possessed tumor cell–intrinsic resistance and could thereby serve as models to test novel cell therapy designs or other therapeutics that would be positioned after CART19 in a clinical setting. We generated T-cell receptor knockout (TCR-KO) CART19 cells from three healthy donors to avoid the confounding effect of alloreactivity (Fig. 5A; Supplementary Fig. S3) and confirmed their comparable activity to TCR-intact CART19 cells in vitro (Fig. 5B). Tumors from seven PDX models, one CART19-naïve and six after CART19 progression, were treated with 2 × 106 CAR-positive TCR-KO CART19 cells (4 mice per donor, 12 per PDX model). Half of the mice were euthanized 14 days after CART19 infusion, and the tumors were evaluated for T-cell infiltration by flow cytometry and IHC, with the remaining mice monitored by weekly MRI for 4 weeks or until moribund (Fig. 5C). The CART19-naïve model (LPMI020) showed an increase in tumor size at 1 week after treatment, but with a notable influx of T cells, and responded thereafter with a reduction in tumor size at week 2 and clearance by week 4 (Fig. 5D and E). The remaining six models derived from post-CART19 progression biopsies did not clear the tumor following CART19 treatment but showed different patterns of T-cell infiltration and response. Three models showed T-cell infiltration within the body of the tumor and a modest reduction in tumor size (LPMI331) or slowing of tumor growth (LPMI462 and LPMI204). One model (LPMI408) displayed extensive T-cell infiltration in the periphery of the tumor, but with a failure to infiltrate the core, and had rapid progression. The remaining two models (LPMI406 and LPMI431) showed very few detectable T cells within the tumor and had rapid progression. These LBCL PDXs therefore capture important clinical scenarios such as post-CART19 progression and retain CART19 resistance in vivo, providing experimental systems in which to test therapeutics in this area of high clinical need. Tumor cell–intrinsic CART19 resistance. A, Example images of flow cytometry plots showing CAR and TCR positivity following CRISPR KO and CAR infection, before and after purification by magnetic beads. B, Incucyte analysis of cytotoxicity against the HT cell line comparing TCR-intact and TCR-KO CART19 cells from healthy donors. C, Schematic for the approach taken to compare CART19 sensitivity of PDX models. D, H&E staining and IHC for human CD4 and CD8 in exemplar tumors for each PDX model harvested at day 14 following CART19 treatment. Scale bar, 300 μM. E, Tumor volume measured by MRI, relative to the volume on the day of CART19 administration (day 0). Error bars represent the standard error of the mean (SEM). [C, Created in BioRender. Yang, H. (2025) https://app.biorender.com/illustrations/693c1ab12cfd6a03bbe27e4c] Molecular Features of Post-CART19 Models PDX models in NSG mice show variable levels of murine macrophage infiltration by IHC for the marker F4/80 (Supplementary Fig. S4) as well as nonhematopoietic cells that may play a role in immune infiltration and activity. The degree to which these murine components interact with human lymphoma cells and/or transplanted human T cells requires further study to deconvolve the direct (via the tumor cell) versus indirect (tumor cell influencing the murine microenvironment) nature of the CAR T-cell resistance patterns that we observed. To partially address this, we further probed the direct features of the PDX tumors by evaluating genetic and transcriptional features that are enriched within post-CART19 progression models (n = 23) compared with CART19-naïve models (n = 25). Analysis of somatic mutations identified modest enrichment of PIM1 (3/25 CART19-naïve; 9/23 post-CART; Fisher P = 0.046) and HLA-B (0/25 CART19-naïve; 4/23 post-CART19; Fisher P = 0.046) mutations in post-CART19 PDX models compared with CART19-naïve models (Fig. 6A). Genetic deletion of the HLA locus has been previously implicated in CART19 resistance (20); however, these excisional HLA locus deletions were not detected by our lpWGS approach. Assessment of DNA copy-number gains and losses identified a significant overrepresentation of 8p24.23 DNA copy-number gains (2/25 CART19-naïve; 10/23 post-CART19; Fisher P = 0.015) targeting the MYC enhancer and a significant under-representation of 13q32.3 (8/25 CART-naïve; 1/23 post-CART19; Fisher P = 0.007) DNA copy-number gains targeting the miR-17-92 locus in post-CART19 progression PDX models (Supplementary Fig. S5). The association between higher frequencies of PIM1 mutations and MYC enhancer amplifications and lower frequencies of miR-17-92 amplifications in the post-CART19 setting have not been previously observed (17, 21), so these associations require experimental validation to establish causative relationships. Genetic and gene expression characteristics of post-CART19 progression PDX models. A, Relative mutation frequencies for lymphoma driver genes shown in Fig. 3 comparing post-CART19 progression models (n = 23) with CART19-naïve models (n = 25), using whole exome sequencing. *Fisher P < 0.05. B, Volcano plot of significantly differentially expressed genes between post-CART19 progression models and CART19-naïve models, using DEseq2 analysis of RNA-seq data. Red, increased expression in post-CART19 models. Blue, decreased expression in post-CART19 models. Dark blue, gene with decreased expression in post-CART19 model and in pathway shown in (C). C, Heatmap of hypergeometric enrichment P values for genes with increased (Up genes) or decreased (Down genes) expression in post-CART19 progression models. D, GSEA plots for two exemplar pathways shown in (C). KEGG, Kyoto Encyclopedia of Genes and Genomes. Differential gene expression analysis of CART19-naïve and post-CART19 PDX models identified 82 genes with significantly lower expression and 200 genes with significantly higher expression (log2 fold change>0.58, adjusted P < 0.05) in post-CART19 models (Fig. 6B; Supplementary Table S4). Hypergeometric gene set enrichment analysis (GSEA) of downregulated and upregulated genes using curated gene sets from the molecular signature database (22) showed no significant pathway enrichment within upregulated genes. However, downregulated genes were significantly (hypergeometric enrichment adjusted P < 0.05) enriched for multiple gene sets associated with cytokine and chemokine signaling and leukocyte migration (Fig. 6C). These were corroborated by GSEA showing significant loss of expression (negative enrichment) for the same pathways in post-CART19 models compared with CART19-naïve models (Fig. 6D; Supplementary Table S4). These observations therefore highlight reduced expression of cytokine, chemokine, and leukocyte migration pathways in post-CART19 settings that may underlie their resistance to CART19. However, these observations require independent experimental validation to show causation. Lowering Barriers to Use Renal capsule implantation and imaging with MRI require specialized skillsets and equipment that may limit broad accessibility of the models. To address this, we tested subcutaneous implantation in 32 models using tissue fragments from freshly harvested F2 tumors, of which 29 successfully engrafted. Intravenous injections were not tested. Although subcutaneous engraftment could be achieved from cryopreserved tumor fragments, the engraftment and subsequent rate of growth was less predictable across cohorts of mice than that observed with fresh tumor fragments. Viral infection with a luciferase and near-infrared fluorescence protein (iRFP720) containing lentiviral vector was attempted on 17 of the PDX models with successful subcutaneous growth using optimized spleen focus-forming virus (SFFV) pseudotyping (Fig. 7A). Although viral infection efficiencies varied between models (median, 7.15%; range <1% to 48.4%; Fig. 7B; Supplementary Fig. S6), all attempted models could be successfully infected as single-cell suspensions and reimplanted subcutaneously in Matrigel followed by fluorescence-activated cell sorting of infected cells from the tumor to achieve sufficient numbers of high-purity cells for further propagation, banking, and use for luminescence imaging–based tumor measurements (Fig. 7C). Importantly, although this process required passaging of PDX models through three generations, WES analysis comparing the luciferized F5 tumor with the initially characterized F2 PDX model did not identify any dropout or acquisition of somatic mutations, suggesting that the models are genetically stable through serial passages (Supplementary Fig. S7). Together, this shows feasibility of subcutaneous implantation and viral manipulation of PDX models for use in future studies, with development of additional luciferized models ongoing. Subcutaneous implantation and luciferization of PDX models. A, Schema for subcutaneous implantation, viral infection, purification, and propagation of PDX models for luciferization. B, Example infection efficiencies from five PDX models, with lentivirally infected cells using the construct shown in (A) marked by iRFP. The gate indicates the percentage of lentivirally infected cells, demonstrating the infection efficiency. C, Example luminescence imaging of tumors following initial infection and following purification by fluorescence-activated cell sorting. Timing of the imaging shown is not consistent from mouse to mouse and depends on the growth rate of each tumor. Data are not representative of growth kinetics; they are demonstrative of the detectability of luciferase signal at the time of imaging. [A, Created in BioRender. Yang, H. (2026) https://app.biorender.com/illustrations/697ceb1c09b65479923d70da] Discussion Since the 2010s, there have been approvals of multiple new therapies for LBCL, which have introduced new areas of clinical need. Prior efforts to improve patient outcomes in frontline LBCLs have been centered on the ABC transcriptional subtype. However, the approval of polatuzumab vedotin in first-line chemo-immunotherapy combinations for DLBCL has improved the outcome of patients with the ABC subtype (23), highlighting “dark-zone” or “molecular high-grade” subtypes of LBCL as an emerging clinical need in the first line (24, 25). The description of genetic subtypes of DLBCL (10–12) has also driven interest in tailoring therapy to genetically defined subsets of patients (26). However, there remains a paucity of mechanistically rational therapeutic options for molecular subsets of LBCLs. The introduction of CART19 has transformed therapy for rrLBCL (27, 28), but patients with post-CART19 progression have dismal outcomes (29). Furthermore, resistance mechanisms to CART19 are largely nongenetic (8, 17) and difficult to replicate using cell line models, highlighting the complexity of the underlying biology. New drug development for emerging areas of clinical need will therefore require the proactive development of model systems that capture the molecular complexity of LBCL, together with tumor-intrinsic resistance mechanisms that are acquired within patients under the selective pressure of therapy. In this study, we show that this can be accomplished through prospective generation of PDX models from high-risk patients. Molecular characterization of our LBCL PDX series shows that it encompasses molecular subsets of disease such as cell-of-origin (COO) and dark-zone transcriptional subtypes as well as genetic subtypes. We note that PDX models have over-representation of mutations in genes such as TP53, KMT2D, CREBBP, NOTCH2, and SPEN, among others. Biases in genetics and other molecular characteristics are to be expected when generating PDX models but must be accounted for when subsequently selecting models for mechanistic or preclinical studies. Furthermore, we show that underrepresented subsets such as the MCD genetic subtype can still be generated, suggesting that targeted efforts toward PDX generation could be utilized to fill in gaps of underrepresented characteristics. We show that models generated from post-CART19 progression biopsies retain tumor-intrinsic characteristics associated with CART19 resistance that can be reproduced in vivo. These included different patterns of T-cell infiltration that have been noted in rrLBCL and found to be associated with CART19 outcomes (8). A major limitation of PDX models generated in NSG mice is the lack of human immune components and the unknown degree of crosstalk with murine components that may participate in T cell exclusion or contribute to exhaustion. However, gene expression characteristics of the PDX models showed mechanisms of CD19 loss via mutation or transcriptional silencing, together with downregulation of cytokine response and/or leukocyte migration gene expression pathways, suggesting that resistance mechanisms may in part be encoded within the tumor cell. Furthermore, there are currently no publicly available models of post-CART19 progression that have been shown to retain CART19 resistance in vivo. Thus, these models provide important preclinical systems in which to perform mechanistic studies and/or test new cellular, biologic, or small-molecule therapies in a post-CART19 setting. PDXs have multiple important limitations that need to be considered. First, their use is resource- and time-intensive, with in vivo experiments commonly taking up to 6 months each. This can be mitigated by the generation of luciferized models that can be implanted subcutaneously, alleviating the need for complex surgeries and imaging, but vascularization and other anatomic differences between sites must be carefully considered. Second, our PDX models are developed in immunodeficient mice that lack the immune contexture that is central to LBCL biology. Future advancements in humanization of immune deficient mice might facilitate the reproduction of immune-intact microenvironments, but these will be hampered by the need for HLA matching and rigorous validation of cell–cell communication patterns compared with primary tumors. Further work is therefore required for development of in vivo systems that will be applicable for mechanistic immunology and the development of novel immune therapies. In conclusion, X-LYMPH provides a publicly available repository of molecularly annotated LBCL PDX models that include important clinical scenarios such as post-CART19 progression. These models provide a unique resource that is available to the academic community for exploration of disease biology and rigorous preclinical testing of new therapies for high-risk patients. Methods Resource Availability All models and anonymized data are available to not-for-profit academic institutions through a materials transfer agreement. To request models or standard operating procedures, please email lymphomaPDXrequest@mdanderson.org. Sample Collection All work was approved by the Internal Review Board of MD Anderson (MDA). All patients in this study provided written informed consent (MDA Lymphoma Tissue Bank Protocol 2005-0656), and work was carried out in alignment with the Declaration of Helsinki. Core-needle biopsies were collected during routine diagnostic procedures by the MDA Lymphoma Tissue Bank and either viably cryopreserved or used fresh. A total of 133 biopsies were used for this study. Renal Capsule Implantation Core-needle biopsies or previously banked PDX seeds (1.5 mm2) were implanted into the subrenal capsule of NSG mice. For comparison, we additionally acquired and propagated available models from the PRoXe (n = 11; ref. 3), WCM (n = 18), and three Richter transformation models previously established at MDA. Of the WCM PDX models, 11 were previously established at WCM, and 7 were newly established at MDA from WCM specimens. Validation of PDX Models Tumors from F1 and F2 generations were formalin-fixed, paraffin-embedded (FFPE) and then evaluated by hematoxylin and eosin (H&E) staining by an expert hematopathologist (F. Vega) to confirm LBCL histomorphology. All models were evaluated for EBV status by Epstein–Barr virus-encoded RNA (EBER) in situ hybridization (ISH), and EBV-positive models (n = 1) were discarded. F2 specimens from all models were assembled onto TMAs and stained for PAX5 to confirm B-cell identity as well as other clinically relevant immunostains (CD10, CD19, CD20, BCL2, BCL6, MUM1, and MYC). TMAs were also evaluated for rearrangements of the MYC, BCL2, and BCL6 loci by fluorescent ISH (FISH). The results are summarized in Supplementary Table S1. Where available, IHC and FISH results were compared with those obtained from clinical workup of synchronous biopsies. TMA Generation Freshly harvested PDX tumors were fixed in 10% formalin for 24 to 48 hours, followed by an additional 24-hour incubation in 70% ethanol prior to embedding in paraffin. FFPE tumor blocks were used to generate TMAs. Representative tumor regions were identified and marked on H&E-stained slides by a pathologist. Using a tissue microarrayer, cylindrical cores of 1.0 mm diameter were extracted from donor blocks and arrayed into recipient paraffin blocks, with normal human tonsil tissue included as controls. Each case was represented by three cores to capture intratumoral heterogeneity. TMA blocks were sectioned at 4 µm thickness for subsequent immunohistochemistry (IHC), ISH, and FISH analyses. IHC IHC was performed using the Leica BOND RX autostainer and BOND Polymer Refine Detection Kit (Leica Biosciences, DS9800). Antibody conditions are listed as follows: BCL2 (Bcl-2/100/D5, Leica Biosystems, RTU), BCL6 (LN22, Leica Biosystems, RTU), CD10 (56C6, Leica Biosystems, RTU), CD19 (BT51E, Leica Biosystems, RTU), CD20 (L26, Leica Biosystems, RTU), CD30 (BER-H2, Cell Marque, RTU), CD38 (SPC32, Leica Biosystems, 1:75), CD79a (JCB117, Dako, 1:50), c-MYC (Y69, Roche-Ventana, RTU), HLA-DR (EPR3692, Abcam, 1:4000), Ki-67 (MIB-1, Biocare Medical, RTU), MUM-1 (MUM1p, Dako, 1:35), and PAX-5 (BSAP, Dako, RTU). FFPE TMAs (4 µm thickness; 3 cores of 1 mm diameter per case) and whole tissue slides (Ki-67 only) were used. All incubation steps were performed at room temperature with wash cycles between each step unless otherwise noted. Slides underwent baking and dewaxing, followed by heat-induced epitope retrieval using Epitope Retrieval solutions 1 (ER1) and 2 (ER2) for 20 minutes at 100°C. Endogenous peroxidase activity was blocked with hydrogen peroxide for 5 minutes, followed by incubation with the primary antibody for 60 minutes. Slides were then treated with post-primary and polymer reagents for 15 minutes each. Antibody–antigen complexes were visualized using DAB Refine for 10 minutes and counterstained with hematoxylin for 5 minutes. Slides were dehydrated through graded ethanol solutions, cleared in xylene, and coverslipped with Cytoseal XYL (Electron Microscopy Sciences, cat. #18009). All slides were evaluated by standard light microscopy. HLA-DR positivity was defined as >30% of neoplastic cells staining positive. COO classification was determined using the Hans algorithm (CD10, BCL6, and MUM1) with a positivity cutoff of >30%. Double-expressor status was defined as >40% of neoplastic cells positive for c-MYC and >50% positive for BCL2. ISH ISH for EBER was performed on whole tissue sections using a fluorescein-labeled peptide nucleic acid probe (Ventana) with the Ventana peptide nucleic acid ISH Detection Kit for FFPE tissues (Roche, cat. #08318883001). FISH FISH analysis for MYC, BCL2 and/or IGH/BCL2, and BCL6 was performed on TMA FFPE tissue sections using break-apart probes for MYC (Abbott, cat. #05J9101), BCL6 (Abbott, cat. #01N2320), and BCL2 (Abbott, cat. #07J7501) and/or IGH/BCL2 dual-fusion probes (Abbott, cat. #05J7101) following the manufacturers’ instructions. For each probe set, a total of 200 interphase nuclei were analyzed. Abnormal signal patterns were defined using laboratory-established cutoffs for FFPE tissues: 1R1GF1 (8.1%), 1G1F (4.6%), 1R1F (5.1%), 1G2F (5%), and 1R2F (4.2%). lpWGS WES and lpWGS were performed on F2 tumors and matched germline DNA when available (MDA models only, n = 25) to characterize mutations and DNA copy-number alterations (CNA), respectively, and for LymphGen classification (10), as previously described (8). PDX tumors were homogenized, and genomic DNA was extracted using the AllPrep DNA/RNA Kit (Qiagen, 80204) according to the manufacturer’s instructions. DNA quantity was measured using the Qubit High Sensitivity dsDNA Kit (Life Technologies, Q32854), and DNA integrity was assessed on a TapeStation 4200 (Agilent, G2991BA). A total of 500 ng of genomic DNA was used for library preparation using the KAPA HyperPlus Kit (Roche, KK8512) with enzymatic fragmentation and 6 to 8 cycles of PCR amplification. Libraries were purified with AMPure XP beads (Beckman Coulter, cat. #A63881), validated for size distribution on a TapeStation 4200, quantified by Qubit, multiplexed, and sequenced on an Illumina NovaSeq 6000 platform to achieve a target genome-wide coverage of 1×, using a 100-10-10-100 bp read configuration. Raw sequencing reads were trimmed using Trimmomatic, and both pre- and post-trimming quality metrics were assessed with FastQC. Trimmed reads were aligned to the hg38 human reference genome using the Burrows–Wheeler Aligner (BWA). Following deduplication, BAM files were processed with XenoFilter to remove contaminating mouse reads and subsequently realigned around insertions/deletions with base quality score recalibration (BQSR) using the Genome Analysis Toolkit (GATK). The processed Binary Alignment Map (BAM) files were used to generate copy-number segmentation data with CopyWriteR using 100-kb bins, and tumor fractions were estimated with IchorCNA. CNAs identified by IchorCNA were further analyzed using the LymphGen classifier and GISTIC2 to assess recurrent genomic alterations. For visualization, data were represented as log2 copy-number changes and segmented using CopyWriteR, and peaks of significant DNA copy-number gains and losses were identified using GISTIC2.0. WES WES was performed using the same genomic DNA libraries generated for low-pass whole-genome sequencing to maintain consistency across assays. For exome capture, equal amounts of each library were multiplexed at a 6 to 8 sample ratio per capture and processed according to the manufacturer’s protocols (IDT or NimbleGen). Captured libraries were subsequently amplified, quantified, and sequenced to achieve a target coverage of 50X for tumor samples, enabling reliable detection of coding mutations, small insertions/deletions, and other functional variants. For the corresponding peripheral blood mononuclear cell (PBMC) samples, the same workflow was used, except sequencing depth was adjusted to a target of 30× coverage. WES data processed using an in-house computational pipeline. Raw sequencing reads were trimmed with Trimmomatic, and both pre- and post-trimming quality metrics were evaluated using FastQC. Trimmed reads were aligned to the hg38 human reference genome using the BWA. Following deduplication, BAM files were processed with XenoFilter to remove contaminating mouse reads and then realigned around insertions and deletions with GATK BQSR. Coverage metrics were calculated using Picard GATK-CollectHsMetrics, variants were called using GATK-Mutect2 and functionally annotated with GATK-Funcotator, and high-confidence variants were selected using GATK-SelectVariants. When matched PBMC samples were available, variants were called with GATK-Mutect2 using the PBMC as a germline control. RNA-seq PDX tumors were homogenized, and total RNA was extracted using the AllPrep DNA/RNA Kit (Qiagen, 80204) and quantified using the Qubit High Sensitivity RNA Kit (Life Technologies, Q32852). One microgram of RNA was used to generate sequencing libraries using the KAPA HyperPrep RNA Kit with RiboErase (Roche, KK8560), using eight cycles of PCR amplification according to the manufacturer’s instructions. Libraries were assessed for fragment size distribution on a TapeStation 4200 (Agilent, G2991BA), quantified using the Qubit High Sensitivity dsDNA Kit (Life Technologies, Q32854), multiplexed, and sequenced on an Illumina NovaSeq 6000 using a 100-10-10-100 bp read configuration. The DLBCL90 NanoString assay was conducted to determine PMBL signature, COO, and DZsig status, as previously described (7). RNA-seq data were processed using an in-house computational pipeline. Raw reads were trimmed with Trimmomatic, and both pre- and post-trimming quality metrics were assessed using FastQC. Trimmed reads were aligned to the human reference genome (hg38) using STAR in 2-pass mode. To remove contaminating mouse reads, trimmed reads were filtered using XenoFilteR. Only samples with at least 20 million uniquely mapped reads were retained for downstream analyses. Gene-level quantification was performed with HTSeq using the following parameters: -m union -f bam -i gene_id -s reverse -t exon –additional-attr = gene_name. The resulting DESeqDataSet was filtered to retain genes with a minimum of 10 total reads across all samples. Transcripts per million reads values were calculated for expression normalization. For GSEAs, normalized counts were evaluated using curated BCL gene sets and the SignatureDB database (https://lymphochip.nih.gov/signaturedb/). ATAC-seq Viably cryopreserved PDX tumor cells were thawed and processed using the Dead Cell Removal Kit to eliminate non-viable cells. Fifty thousand live tumor cells were washed in ATAC-seq resuspension buffer, resuspended in 50 μL of ATAC-seq resuspension buffer containing 0.1% NP-40, 0.1% Tween-20, and 0.01% digitonin, and incubated on ice for 3 minutes. Following lysis, nuclei were washed with cold ATAC-seq resuspension buffer containing 0.1% Tween-20 (without NP-40 or digitonin) and then resuspended in 50 μL of transposition mixture and incubated at 37°C for 30 minutes in a thermomixer at 1,000 rpm. Fragmented DNA was purified using the Zymo DNA Clean and Concentrator-5 Kit (Zymo, D4024). ATAC-seq libraries were prepared using the KAPA HyperPrep Kit (Roche, KK8502) using 10 cycles of PCR amplification. Libraries were assessed on a TapeStation 4200 (Agilent, G2991BA), quantified using the Qubit High Sensitivity dsDNA Kit (Life Technologies, Q32854), multiplexed, and sequenced on an Illumina NovaSeq 6000 using a 100-10-10-100 bp read configuration. Technical replicates for each ATAC-seq sample were merged prior to quality control assessment with FastQC. Adapter sequences were removed from paired-end FASTQ files through Trimmomatic trimming, followed by alignment to the hg19 human reference genome using the BWA. Post-alignment processing included coordinate-based sorting and filtering to exclude reads mapping to mitochondrial DNA and blacklisted genomic regions via Samtools. PCR duplicates were marked and removed using Picard Tools’ MarkDuplicates command. Deduplicated BAM files were then processed with XenoFilteR to remove contaminating mouse reads. The processed BAM files, having completed alignment, sorting, and filtering, served as input for subsequent analyses. For track visualization purposes, BAM files were transformed into BigWig format through deepTools’ bamCoverage function, incorporating Tn5 transposase insertion site correction. Peak calling was performed on Tn5-adjusted insertion sites using MACS2 with the following parameters: -g hs (genome size), -q 0.05 (q-value threshold), -f BAMPE (paired-end BAM format), and -B (bedGraph output). Genomic tracks were visualized using pyGenomeTracks version 3.9. Viral Transduction and PDX Luciferization The luciferization construct (Addgene # 104587, a gift from Antonius Plagge) was modified by replacing its original EF-1α promoter with a B cell–specific promoter, SFFV. Lentivirus was produced in 293T cells by co-transfecting modified luciferization construct with packaging vectors psPAX2 (Addgene #12260, a gift from Didier Trono) and pMD2.G (Addgene #12259, a gift from Didier Trono) at a 12:10:5 ratio in OptiMEM (Thermo Fisher Scientific, 31985070) with Lipofectamine 2000 transfection reagent (Thermo Fisher Scientific, 11668019). Virus particles were collected from the supernatant 48 hours after transfection, filtered through a 0.45-μm filter, and mixed with polybrene infection reagent (Sigma-Aldrich TR-1003-G) prior to the 2-hour spinfection with PDX cell suspensions. Transduced cells were subsequently injected subcutaneously into NSG mice for in vivo expansion. Once luciferized tumors became palpable, tumors were harvested and sorted for iRFP720 expression as a selection marker, followed by an additional round of subcutaneous injection to expand cells of high transduction purity. Generation of TCR-KO CD19 CAR T Cells PBMCs obtained from three healthy donors (55 F, 53 F, and 40 M, respectively; Gulf Coast Regional Blood Center) were isolated using Ficoll-Paque (Cytiva) and cryopreserved in FBS with 10% DMSO until use. Thawed PBMCs were resuspended in CTS OpTmizer T-cell expansion SFM supplemented with CTS Immune Cell SR and GlutaMAX (all from Gibco; hereafter referred to as CAR T cell medium) at 1 × 106 cells/mL and were stimulated with 25 µL/mL ImmunoCult Human CD3/CD28/CD2 T-cell activator (STEMCELL) and 200 IU/mL recombinant human interleukin 2 (rhIL2; GenScript) for 3 days (day 0). On day 3, activated T cells were resuspended in Buffer T (Thermo Fisher Scientific) at 4 × 106 cells/100 µL, and ribonucleoproteins that consisted of 200 pmol Streptococcus pyogenes Cas9 and 500 pmol sgRNA targeting TRAC (both from Synthego) were electroporated into cells using 100 µL Neon reaction tips (Invitrogen) with 1,400 V, 10 ms, 3 pulses. Cells were subsequently plated in OptiMEM (Gibco) containing 25% v/v CD19 CAR lentiviral supernatant with 10 µg/mL Vectofusin-1 (Miltenyi Biotec) at 1 × 106 cells/mL and spinoculated at 400 × g for 2 hours. On day 4, cells were resuspended in CAR T-cell medium supplemented with 200 IU/mL rhIL2. On day 6, TCR-positive cells were depleted using a biotin-conjugated anti-human TCRα/ß antibody (clone IP26, BioLegend) and MACS systems (Miltenyi Biotec). Then, the retrieved TCR-KO cells were expanded in G-Rex 6M plates (Wilson Wolf) by day 10 and cryopreserved until use. CAR positivity and TCR-KO rate were evaluated by flow cytometry using VioBright B515 anti-CD19 CAR FMC 63 idiotype antibody (clone REA1297, Miltenyi Biotec) and Alexa Fluor 647 anti-TCRα/ß antibody (clone IP26, BioLegend) together with ViaDye Red (Cytek). Coculture assay with HT cells (ATCC) and CAR T cells at 1:1 E:T ratio was conducted using Incucyte SX5 (Sartorius) according to the manufacturer’s instructions with minor modifications. For lentiviral supernatant production, CD19-28z CAR plasmid and third-generation packaging mix (abm) were cotransfected into HEK293T cells (ATCC) using Lipofectamine 3000 (Invitrogen) according to the manufacturer’s protocol with minor modifications. The viral supernatant was harvested 2 days after transfection, filtered through a 0.45-µm polyvinylidene difluoride membrane (Millipore), and cryopreserved at −80°C until use. In Vivo Assessment of CART19 Response All experiments evaluating CART19 responses in vivo were conducted using the secondary implantation model. Cryopreserved PDX tumor fragments were implanted in NSG mice to generate seed tumors. After tumors became palpable and mice showed signs of illness, fresh tumor fragments were implanted into experimental mice. Weekly MRIs on Bruker systems were performed at the MDA small animal imaging facility. Once the median tumor volume reached 50 mm3, 2 × 106 CART19 cells, thawed and rested overnight in CAR T-cell medium with rhIL2, were administered via the tail vein as a one-time treatment. Histologic Analysis of T-cell Infiltration All tumor samples were sectioned along with the long axis of kidney, and one part of them was fixed in neutral formalin. Thereafter, H&E staining and IHC for CD4 (clone RBT-CD4, Bio SB) and CD8 (polyclonal, NBP2-29475, Novus Biologicals) were conducted by the MDA Department of Veterinary Medicine and Surgery core facility. Scanned images were evaluated in Aperio ImageScope (Leica Biosystems). Animal Studies We confirm that all mouse procedures to generate this PDX repository were approved by the MDA Institutional Animal Care and Use Committee. Supplementary Material Clinical features of 48 cases LBCL PDX from MD Anderson. Molecular characteristics of 48 cases LBCL PDX from MD Anderson. Differentially expressed genes between CD19-positive and CD19-negative PDX. Differentially expressed genes between post-CART19 and CART19-naïve models. Supplementary Figure S1: GISTIC analysis of DNA copy number from 80 cases of LBCL PDX. Graphs show the statistical significance of DNA copy number loss (left, blue) or gain (right, red) and the location of significant peaks (annotated cytogenetic location) determined by GISTIC2. Supplementary Figure S2: Lollipop plot of MYD88 and CD79B mutations from PDX models. Plots show the frequency of protein coding mutations in MYD88 (above) and CD79B (below), including hotspots at L252 and Y196, respectively. The cell of origin (purple, ABC; yellow, GCB) and LymphGen subtype for cases with MYD88 and/or CD79B mutations are summarized on the right. Supplementary Figure S3: Gating hierarchy for Figure 5A. Cells shown in Figure 5A have been gated on intact single cells by forward scatter, side scatter and autofluorescence, and viable cells using viability dye. Supplementary Figure S4: Immunohistochemical staining of the murine macrophage marker F4/80. IHC corresponding to Figure 5D showing zoomed out images of H&E, CD4 IHC and CD8 IHC, together with IHC staining for the myeloid marker F4/80. Supplementary Figure S5: Comparison of copy number alteration frequencies between post-CART19 and CART19-naïve models. Statistical comparison of the frequency of DNA copy number alterations between post-CAR T and CAR-naïve PDX models. Lesions correspond to GISTIC peaks shown in Supplementary Figure S1. *Fisher exact P < 0.05. Supplementary Figure S6: Flow cytometry gating hierarchy for Figure 7B. Cells shown in figure 7B are gated on viable lymphocytes and singlets by forward and side scatter. Supplementary Figure S7: Comparison of variant allele frequencies between F2 and F5 tumors. Scatter plots show the variant allele frequencies of germline variants (grey), somatic mutations in lymphoma driver genes (red) and other somatic mutations (blue) for 14 models obtained from the F2 generation and F5 generation. Pearson’s correlation R2 and P value are calculated from all somatic mutations.