A biobank of pediatric patient-derived-xenograft models in cancer precision medicine trial MAPPYACTS for relapsed and refractory tumors

Patient characteristics and PDX establishment

Between February 2016 and July 2020, 787 pediatric, adolescent and young adult patients with recurrent or refractory malignancies were enrolled in the MAPPYACTS trial;² 756 (96%) patients and their parents consented to the optional ancillary study of preclinical model development (Fig. 1a). 744 patients had a cancer tissue sampling procedure performed; 12 patients were screening failures. Median age at time of inclusion was 11.9 years (range, 0.5–38.5), 59% were male. All patients had received anticancer treatment before study procedure, aligned with the inclusion criteria of MAPPYACTS. All had prior systemic chemotherapy regimens, except two patients who had radiation therapy only; 31 PDX were derived from tumor lesions that had been irradiated. Fifty patients had a second and five patients had a third sample acquisition, in most cases because of insufficient tumor sample or unsuccessful sequencing; six patients had two successful sequencing analyses at different time points of their disease.

**Fig. 1: Workflow for PDX development in MAPPYACTS and established models.**

Figure 1, Supplementary Fig. S1 and Supplementary Table S1 depict the workflow and patient flowchart, samples processed as well as all PDX models established, respectively. From 799 procedures in 744 patients, 505 cancer samples (63%) were received in 10 research laboratories for PDX development. To date, 302 samples were transplanted directly into mice, 147 engrafted successfully (49%) in the first passage (P0) and 131 models were considered as stable/established (i.e., reached the 2nd passage in vivo (P2) in 123 PDX and eight leukemia considered established at first tumor take) (Fig. 1a–c). A total of 223 patient tumor samples have been soft-frozen in laboratories, including from 209 patients for whom no model could be transplanted or established (Supplementary Table S1).

Primary tumor take rates and tumor model establishment rates varied for the malignancies between 0% and 82% (Fig.1b and Supplementary Table S1). Figure 1c displays the 131 established PDX models per histology and the laboratory that developed it.

Characteristics of the established PDX models

One hundred thirty-one PDX obtained from patients with 0.5–30.8 years old were considered as stable and established models, 113 were solid tumors (76 sarcoma, 25 other non-central nervous system (CNS) tumors, 12 CNS tumors), 3 lymphomas and 15 leukemias presenting with more than 50% leukemic blasts. Supplementary Data 1 details the characteristics of all 131 PDX models and their originating patient tumor with origin, prior therapies, molecular alterations retained in the clinical molecular tumor board that were considered as actionable, of interest or disease specific, IMI2 ITCC-P4 identifiers and other model names. Established PDX models underwent further characterization for histology, WES, RNA-Seq, HLA typing and growth behavior. To date, 90 of the 131 solid tumor MAPPYACTS PDX models were shared with the European Innovative Medicines Initiative 2 (IMI2) ITCC-P4 (Innovative Therapies for Children with Cancer-Pediatric Preclinical Proof-of-concept Platform) initiative (Project Number: 116064; IMI2/INT/2015-03842 v.2020; www.itccp4.eu/), where they will be used for drug screening and exploration.

Growth characteristics of established PDX

The in vivo characteristics at P2 passage for 70 PDX models were grouped by tumor type (Fig. 2 and Supplementary Fig. S2). Tumor take of the first passage (P0) ranged between 12 and 285 days (median, 41 days), while growth accelerated in general for all models during P1 and P2 to a median of 21 days (range, 5 and 104).

Morphologic primary tumor type is preserved in established PDX

At each passage, hematoxillin-eosin-safranin (HES) staining was performed on PDX tissue sections for morphological analysis of concordance with the primary tumor. Fifty-one pairs of primary tumors versus PDX were evaluated centrally in one laboratory by one pathologist with expert experience in pediatric tumors (JY Scoazec). In a first step, representative sections of PDX tumors were examined without knowledge of the initial patient diagnosis and a tentative diagnosis based on morphology alone was proposed. In a second step, representative sections of both primary tumors and their corresponding PDX were compared to classify their respective histological features as similar (morphology fully preserved, including some distinctive features of the primary), comparable (morphology slightly divergent but still suggestive, loss of distinctive features of the primary) or different. In 39 cases, the PDX-based diagnosis was concordant with the initial diagnosis; histological features were similar in 29 cases and comparable in 10. In 12 cases, PDX-based diagnosis was discordant; in all cases, the initial diagnosis was confirmed after review of histological, immunohistochemical and/or molecular data. Three causes of discrepancies were identified: (1) loss of distinctive histological features resulting in a morphologically undifferentiated tumor (n = 5), even if additional studies confirmed the preservation of diagnostic immunohistochemical markers such as MyoD1 in rhabdomyosarcoma; (2) expansion of a blastic component without preservation of differentiated features (n = 3); (3) true diagnostic errors, mainly in cases in which morphology alone was not sufficient to reach a final diagnosis (n = 4).

Representative examples of the primary tumor and at stable passage (P2) are presented in Fig. 3 and detailed results of the comparative study are given in Supplementary Table S2. In these 51 PDX tumors, necrosis was usually absent (n = 48/51) or focal (n = 3/51), which confirms the quality and sustainability of their growth. Mouse stromal areas ranged from 1% to maximum 61% (mean 13.8%) of the PDX sample at P2 as defined by a TBP murine and human specific RT-qPCR assay in 38 PDX; osteosarcoma PDX at P2 had significantly more (p = 0.046) murine stroma (mean 24.2 ± 18.9% standard deviation) as compared to the other PDX entities (10.2 ± 8.6%).

**Fig. 3: Comparative histological study between primary and PDX tumors.**

Molecular characteristics in PDX and primary tumors

To further evaluate the faithfulness of these models, we compared WES and RNA-Seq of the primary tumors with their derived PDX focusing here on 39 PDX models from Gustave Roussy. The genome/transcriptome aligned on human (tumor clones) and mouse (tumor microenvironment) sequences were analyzed separately using Xenome. Recurrent alterations, found in both the patient tumor sample and the corresponding PDX, and considered by the clinical molecular tumor board as actionable or of interest in the disease were identified (Fig. 4a). PDX models exhibited high consistency of the genetic alterations found in the patient samples. Main discordance in molecular phenotyping was observed in the WES of GR-RMS-1 where the PIK3CA and TP53 pathogenic variants and the MYCN amplification were not detected, whereas the PAX3::FOXO1 gene fusion was present in the RNA-Seq analysis suggesting a technical concern with the WES analysis. The GR-BALL-1 PDX with distinct pattern had been engrafted with a minor clone (see leukemia chapter). The Burkitt lymphoma GR-BL-1 shares the TP53 mutation with the primary tumor but not the MYC variant.

**Fig. 4: Pangenomic profile and comparison of WES and RNA-Seq between matched primary patient tumors and PDX model.**

Taking advantage of the PDX model, a human tumor growing in a mouse microenvironment, we deconvoluted bulk RNA-Seq of matched patient/PDX samples to isolate the transcriptomic program of tumor cells from the microenvironment in both patient and PDX tumors (see Methods). Biased by the sampling and the tumor purity, the quantification of the alteration conservation with the Jaccard distance between matched patient and PDX suggests that osteosarcoma and rhabdomyosarcoma are the most divergent, in accordance with their unstable genetic (Supplementary Fig. S3a and b). Transcriptional landscape of tumor cells, summarized by five Tumor Principal Components (TPCs; Supplementary Fig. S3c), is highlighted after functional enrichment analysis of the tumor cell origins for each pathology (Supplementary Fig. S3d). Clustering of patient and PDX samples in this tumor latent space, is emphasized by the UMAP of the Fig. 4b, suggesting for 97% of the patients a high conservation of tumor cell phenotypic traits from primary tumors to PDX. The single exceptions are non-Hodgkin lymphoma or non-rhabdomyosarcoma soft tissue sarcoma, both classified by this methodology near osteosarcoma. This result is probably explained by the small number of patients for these pathologies, a challenge for any attempt of phenotypic characterization with bulk RNA-Seq data. Deconvoluting the microenvironment present in both patient and PDX models, summarized by the Microenvironment Principal Components (MPCs, Supplementary Fig. S3e, S3f and S3g), did not demonstrate any clear-cut conservation of the microenvironment (Supplementary Fig. S3e) in mice despite an interesting trend in osteosarcoma to cluster together suggesting a partial reconstitution of osteosarcoma microenvironment in PDX. This clustering of the osteosarcoma microenvironment fraction in mice and patients is driven by the MPC1 (Supplementary Fig. S3f), and is identified as the regulation of the immune response by functional enrichment analysis (Supplementary Fig. S3h). Most contributing genes to the MPC1 such as IL10RA, CD48, SLA, suggest that MPC1 describes the antigen presenting cell (APC) role in the immunosuppressive prone context of osteosarcoma relapse (MPC1 wordcloud in Supplementary Fig. S3i).

HLA class I typing and allele and supertype coverage of the established PDX models

In the emerging era of personalized cancer immunotherapy, the characterization of HLA alleles is a critical aspect for preclinical studies using PDX models such as adoptive transfer of neoantigen-specific T cells. Recently, we characterized the HLA class I and II genotypes of patients from the MOSCATO-01¹⁹ and MAPPYACTS trials². Thus we first compared HLA class I (HLA-A, -B, -C) genotypes for the 34 PDX models of solid tumors developed at Gustave Roussy with available WES and RNA-Seq data with their equivalent patient normal and primary tumor samples, using a combination of dedicated HLA typing algorithms (see Methods). Because of their consistency, we inferred the HLA genotypes for all solid tumor PDX models developed in MAPPYACTS from the corresponding patient sequencing data. Overall, 27, 47 and 28 HLA-A, -B, and -C alleles, respectively, were detected among the 110 PDX models established from patients with solid tumors and available normal WES data (Fig. 5, Supplementary Fig. S4 and Supplementary Data 2). When considering PDX models with available RNA-Seq in comparison to normal WES, all alleles from the three classical HLA class I loci were detected in Ewing sarcoma, non-rhabdomyosarcoma soft tissue sarcoma and nephroblastoma PDXs, and all except single HLA-B and -C alleles were detected in alveolar rhabdomyosarcoma and high-grade glioma PDXs. In contrast, only 2 out of 6 HLA-A alleles and none of the HLA-B and-C alleles were detected from RNA-Seq of neuroblastoma PDXs, consistent with low HLA-I transcriptional levels in this tumor type^20,21 (Supplementary Fig. S4 and Supplementary Data 2).

**Fig. 5: HLA class I alleles detected in PDX models from pediatric solid cancers.**

Furthermore, we looked at the genetic ancestry fractions of all patients included in the trials using the EthSEQ pipeline and compared the distribution of HLA allele frequencies in the PDX models with those observed in our European (EUR) patients study population. The cumulative HLA allele frequency coverage per tumor type or subtype in the corresponding PDX models ranged from 0.40 to 0.96 for HLA-A, from 0.27 to 0.80 for HLA-B, and from 0.23 to 0.97 for HLA-C (Supplementary Fig. S4). Notably, the HLA allele frequency coverage was >0.6 for both HLA-A and -C genes and ranged from 0.4 to 0.80 for HLA-B in the PDX panels representing the main pediatric solid tumor types and subtypes, with the exception of other rhabdomyosarcoma, carcinoma, hepatoblastoma, pleuropulmonary blastoma, ependymoma and atypical teratoid rhabdoid tumor, with limited numbers of models (n = 1–3).

It has been shown that the peptide-binding specificities of multiple HLA class I alleles largely overlap and can be grouped as supertype-specific binding motifs, corresponding to 5 HLA-A (A1, A2, A3, A24, A26), 7 HLA-B (B7, B8, B27, B39, B44, B58, B62) and two HLA-C (C1, C2) supertypes^22,23. A total of 97 out of 102 (95.1%) HLA-I alleles detected in the PDX models could be assigned to known supertypes, with only three alleles from osteosarcoma (A*24:07, A*43:01 and B*50:02) and two from hepatoblastoma (B*13:01 and B*44:27) unclassified. HLA-I supertypes were largely represented among the PDX models (Fig. 5, Supplementary Fig. S5 and Supplementary Table S3): all 5 HLA-A supertypes in sarcoma, neuroblastoma, high-grade glioma PDXs, and 5–7 HLA-B supertypes in sarcoma and neuroblastoma. Both HLA-C1 and -C2 supertypes were represented in the PDXs corresponding to all specific tumor types and subtypes.

Metabolic signatures in PDX maintain consistency with primary tumors

Metabolic reprogramming has long been identified as a hallmark in cancer²⁴. This dynamic reprogramming is driven by tumor intrinsic and extrinsic factors, such as microenvironment and nutrient availability. We investigated the conservation of metabolic networks by evaluating the metabolic capacity of the human fraction of the PDX and primary tumors to carry defined metabolic tasks (Fig. 6a). We inferred metabolic function (=performing metabolic tasks) by quantitatively evaluating the expression of essential enzymes to carry a set of core metabolic tasks. Comparison of metabolic network content showed similar metabolic function of PDX to their corresponding cancer types (Fig. 6b). Highly significant similarity (p = 5.48e−10) was found comparing matched patient primary tumor-PDX to randomly selected pairs (Fig. 6b). After confirming maintenance of global metabolic reprogramming phenotype, we investigated whether specific metabolic tasks are conserved. Figure 6c shows an example, the synthesis of dopamine from tyrosine that is present in the two neuroblastoma patient-PDX pairs. Capacity to fulfill this metabolic task can be expected in those tumors originating from adrenal gland. Interestingly, one carcinoma pair also showed the capacity to make dopamine. We expanded this analysis to all PDXs, including the ones without matched primary tumor, and found higher task score only in neuroblastoma and few other cancers, including HGG that can be expected to be derived from dopaminergic neurons (Fig. 6d). This confirms the ability of the PDX model system to reproduce expected metabolic functions and describe unexpected findings. We next evaluated the capacity of tumors to synthesize L-kynurenine from tryptophan, an oncometabolite known to exert immunosuppressive effects²⁵. We found that L-kynurenine can be produced by many cancer types, most strikingly in sarcomas, including osteosarcoma. Its synthesis was confirmed by numerous corresponding PDXs, which can be used in follow-up work to investigate its immunosuppression role and potential therapeutic implications (Fig. 6e).

**Fig. 6: Metabolic functional signatures of 42 PDX models depicted on the background of the whole MAPPYACTS patient population.**

Selected tumor types of PDX models

A description of osteosarcoma PDX developed at Gustave Roussy, both as subcutaneous and as orthotopic (paratibial) models, and their comparison with the originated patient sample at relapse and the primary tumor at diagnosis are presented in a separate manuscript²⁶. We further wished to provide additional context on models from selected pediatric tumor types for which multiple PDX models were established or for some tumor types that have so far rarely been reported.

Acute Leukemia PDX

Leukemia is the most common type of cancer in children and, despite significant advances in treatment, it remains the second cause of cancer-related deaths. In this study, we transplanted 27 ‘relapsed’ acute leukemia samples into sublethal irradiated NSG mice and established 15 models of childhood acute leukemias (7 B-ALL, 6 T-ALL, 2 AML-M1 and -M4/5; 56% successful engraftment rate overall with special difficulty for AML (20% versus 78% and 75% for B- and T-ALL respectively). Phenotypic analyses of the B-ALL models are depicted in Supplementary Fig. S6 and show that the major phenotypic clone, based on CD34/CD38 expression, was amplified in NSG recipients established from GR-BALL-2, GR-BALL-3 and GR-BALL-7 samples. Two B-ALL models were engrafted with minor clones: GR-BALL-1 and GR-BALL-8; the GR-BALL-1 profile indeed looks distinct from the originating primary leukemia profile (Fig. 4a). Notably, we also established a B-ALL cell line from the GR-BALL-3 sample, validated at the phenotypic and genomic levels (Supplementary Fig. S6b and c), in which expression of the TCF3::HLF fusion transcript associated with aggressive leukemia has been confirmed (Supplementary Fig. S6d). Together, although a relatively low number of leukemia samples where included in this study, however we developed models of aggressive leukemia.

Rhabdomyosarcoma (RMS)

RMS is the most common type of soft-tissue sarcoma of childhood with a poor prognosis in patients with metastatic or recurrent disease²⁷. The most frequent subtypes are embryonal (eRMS) and alveolar RMS (aRMS) in ~70% and 30% of childhood rhabdomyosarcoma, respectively. aRMS are characterized by FOXO1 gene fusions resulting from the stable reciprocal translocation of chromosomes 1 or 2 and 13 encoding two paired box transcription factors PAX7 and PAX3, respectively, while RMS that are FOXO1 fusion-negative harbor mutations in key signaling pathways²⁸.

Here we established 25 RMS PDX models, 12 FOXO1 fusion-positive aRMS, 9 eRMS and one each epithelioid, sclerosing, spindle cell and NOS. All had previously been treated with chemotherapy; 14 were from initially well responding tumors that relapsed, 11 were from primary refractory diseases. Ten models were derived from primary recurrent/refractory tumor sites, 15 from metastases. Ten of 12 established PDX exhibited morphological features concordant with their derived patient tumor (Supplementary Table S2). All aRMS PDX maintained the PAX3::FOXO1 fusion; aggressive features like TP53 mutations ± loss of heterozygosity, FGFR4 mutations ± amplification, PIK3CA or NRAS mutations, and cell cycle alterations like CDK4 amplification or CDK2NA/B loss were present in the rhabdomyosarcoma PDX. The sclerosing RMS PDX exhibits NSD3::FGFR1 and NSD3 (WHSC1L1)::MYOCD fusions and the spindle cell RMS the characteristic VGLL::NCOA2 fusion.

Ewing sarcoma

Ewing sarcoma is the second most frequent bone tumor of childhood and adolescence after osteosarcoma. Genetically it is characterized by a balanced chromosomal translocation involving EWSR1 or FEV and a gene of the ETS transcription factor family, with the most common fusion being EWSR1::FLI1. Other recurrent genetic alterations are infrequent except for STAG2 and TP53 mutations and CDKN2A deletions²⁹. Combination of these three alterations with the translocation appear to favor the oncogenic transformation of Ewing sarcoma cell of origin³⁰. Patients with STAG2 mutations alone or associated with TP53 mutations have a lower overall survival as compared to those with a wild type status²⁹. Mechanistically, STAG2 loss of function favors invasiveness and metastasis of Ewing sarcoma cells by reducing cis-mediated EWSR1-FLI1 activity³¹. Fifteen Ewing sarcoma PDX were established, 13 displayed the EWSR1::FLI1 translocation, one each FUS::ERG and EWSR1::FEV. PDX models display TP53 mutation/LOH, STAG2 mutation or deletion or CDKN2A deletion; all exhibit the typical small round cell morphology.

Nephroblastoma

Nephroblastoma or Wilms tumor is the most frequent renal malignancy in childhood and occurs mainly in young children³². Nephroblastoma is considered a failure of embryonic kidney development, a complex process in which different transcription factors, proto-oncogenes, and various types of growth factors are effective³². A recent initiative develops organoids from primary nephroblastoma within the UMBRELLA registry³³, however there is a scarcity of preclinical models overall and particularly at recurrence. In our hands, nephroblastoma had one of the highest rates for PDX establishment (73%) where 8/11 transplanted samples engrafted in P0 and were stable. Established PDX exhibited a variable morphological pattern; in the four cases examined histologically, two retained the distinctive epithelial structures typical of nephroblastoma whereas two exhibited only undifferentiated, “blastic” tumor cell morphology.

Pleuropulmonary blastoma (PPB)

PPB is a very rare, highly aggressive and malignant tumor, with poor prognosis originating from either the lungs or pleura. It occurs mainly in children of <5 or 6-years, but may rarely occur in adults³⁴. From four patients with PPB, three samples were transplanted, two engrafted and were stable. Both established models are from 4-year-old patients with DICER1 germline mutations with tumors displaying an additional somatic DICER1 pathogenic variant; both also showed NOTCH3 pathogenic mutations, one of them combined with a KRAS and TP53 hot spot mutation, or a HRAS mutation with BCOR and KAT6A mutations, respectively.

High-grade glioma (HGG)

Gliomas are the most common primary tumors of the brain and account for almost half of all CNS tumors in children and adolescents. HGG are aggressive, malignant lesions, accounting for 8–12% of all pediatric CNS tumors and of highly heterogenous nature distinct from adult HGG. Molecular characteristics mainly distinguish tumors harboring histone mutations, H3.K27M in midline and H3.G34 in hemispheric, tumors harboring the BRAF V600E mutation, and H3-/IDH wild type tumors³⁵. The prognosis for patients with pediatric HGG remains poor with limited treatment options.

Three of the six established HGG are pleomorphic xantho-astrocytoma, all with anaplastic features. Two of them exhibited the characteristic BRAF p.Val600Glu mutation and had homozygotic focal deletions of CDKN2A/2B. The third exhibited a TP53 p.Arg337Cys hot spot mutation as well as a pathogenic variant in TSC2, ATRX and PTPN20A. In contrast, all three other HGG exhibited alterations in the PI3K pathway. The relapsed primary hemispheric giant cell glioblastoma had a PIK3R1 pathogenic variant, associated with pathogenic variants of NF1, ATRX, and a heterozygotic TP53 mutation. The second relapsed temporal glioblastoma displayed a PIK3CA mutation, a TSC1 VUS, focal homozygotic deletion of NF1 and CDKN2A/B, as well as mTOR germline VUS associated with an isodisomy. The pontine glioma exhibited a PIK3CA hot spot mutation and pathogenic variants in CDKN2C and ACVR1. However, despite a loss of trimethylation of lysine 27 of histone H3 in immunohistochemistry, the classically associated H3.1 mutation was not detected. These three could be excellent models to further explore the role of targeting the PI3K/AKT/mTOR pathway. Specific inhibitors at all three levels are in clinical development and the optimal agent to be taken forward has not yet been defined, as we recently discussed in the paper exploring a dual mTORC1/2 inhibitor in AcSé-ESMART³⁶. The BRAF mutated anaplastic xanto-astrocytoma will contribute to the understanding of resistance mechanisms to BRAF ± MEK targeting that is not yet elucidated in pediatric glioma^37,38.