Between 2010 and 2019, thanks to a close collaboration with the Pathology and Urology departments of Strasbourg Hospital and the urological Clinic of Nantes, we collected 240 PCa samples from 237 patients. Human prostate cancer tissues were retrieved directly after surgery (radical prostatectomy, palliative TURP, surgical resection of node metastasis, pelvectomy) or biopsy ( Supplementary Table 1 ). The use of patients’ tissues complied with a protocol approved by both the “Comité de Protection des Personnes Est IV” and the “Comité de Protection des Personnes Sud-Ouest et Outre-mer” (Approval numbers: DC-2010-1193 and DC-2019-3565, respectively). Patients enrolled in the study provided written informed consent allowing the use of discarded surgical samples for research purposes. In addition, relevant clinical information was recorded from the patients’ data including age, PSA levels, treatments, and treatment responses when available. Tumor regions within surgical specimens were identified by uro-pathologists. Once collected, prostate tissues were transported in an appropriate solution (Custodiol). Prostate tissues were implanted within 24 hr from collection.

Four to five-week-old immuno-deficient mice including athymic Swiss-Nude (Crl : NU(Ico)-Foxn1nu), NMRI-Nude (Rj : NMRI-Foxn1^nu/nu) or Shrn (NOD.Cg-Prkdc^scidHr^hr/NCrHsd) male mice were purchased from Charles River Laboratories (L’Abresle, France), Envigo (Gannat, France) or Janvier Labs (Saint-Berthevin, France). Animals were handled under specific pathogen-free conditions. Their care and housing complied with guidelines set out in French animal welfare regulations referred as the European Community Council Directive 2010/63/UE. All tumorgraft studies were reviewed by CEE-35 and CEE-122 (ethical committees for the protection of animals used for scientific purposes) and approved by the French Ministry for National Education, Higher Education and Research under the numbers APAFIS#2949-2015113017594629v5 and APAFIS#14811-2018042316405732v6. The animal facility was maintained under standardized conditions: artificial 12h light-dark cycles between 7:00 a.m. and 7 p.m., ambient temperature of 22 ± 2°C and relative humidity maintained at 55 ± 10%.

The establishment of PDX models was carried out as previously described (21, 22). Briefly, different adult male immunodeficient mice strains were used for tissue implantation: Swiss-nude mice (G266, C901 and C1022) and Shrn mice (PCU-012 and PCU-018) for primo-implantations. At the time of grafting, mice were intact (G266, C901 and C1022) or had androgen supplementation (PCU-012 and PCU-018) ( Supplementary Table 2 ). Grafts were implanted into the interscapular fat pad and monitored weekly for tumor growth for up to 9 months post-implantation for initial growth. When xenografted tumors reached ~1500 mm³, they were sequentially passaged into new mice under the same conditions and using the same protocol as the original implants. A PDX model was defined as established when stable growth over at least three passages and regrowth after a freeze-thaw cycle could be observed. At each mouse-to-mouse passage, representative samples were cryopreserved, snap-frozen in liquid nitrogen and/or FFPE processed.

FFPE patient tumors and PDX blocks were sectioned and slides immunostained with the following antibodies: cytokeratin cocktail AE1AE3 (M3515, Agilent), androgen receptor (M3562, Agilent), hCD45 (IR751, Agilent), PSA (M0750, Agilent), ERG (AC-0105, Clinisciences), NKX3.1 (AC-0314, Clinisciences), PTEN (ab228466, Abcam), Ki67 (M7240, Agilent), P53 (GA616, Agilent), synaptophysin (IR660, Agilent) and chromogranin A (M0869, Agilent). After heat antigen retrieval as specified by provider, experiments were performed using Dako Omnis Instrument, EnVision FLEX, High pH kit for revelation (GV800, Agilent) for hCD45, cytokeratin cocktail AE1AE3, PSA, P53, synaptophysin, ERG and chromogranin A, or Leica Bond III Instrument for androgen receptor, Ki67 and NKX3.1 and on Autostainer 480S instrument for PTEN. The conditions are described in the Supplementary Table 3 .

Frozen samples were processed on ice. For DNA and RNA isolation, patient and PDX tumor fragments were processed using Qiagen Allprep DNA/RNA mini kit according to manufacturer’s instructions. Briefly, DNA and RNA were simultaneously extracted and purified. Nucleic acid yield and quality was assessed by NanoDrop spectophotometer ND8000 and RNA quality was further evaluated using Agilent 2100 Bioanalyzer.

The identity of each PDX was periodically authenticated by profiling STRs. Patient tumors and corresponding PDX DNA samples were subjected to STR using PowerPlex^® 16 HS System (Promega, ref DC2101) that amplifies 16 STR loci and the amelogenin gender-determining marker, according to manufacturer’s instructions. PDXs passed authenticity when >80% match in alleles was obtained. PCR products were separated by capillary electrophoresis on ABI prism 3500 and results were analyzed using GeneMapper software (v5).

DNA extracted from tumorgraft tissues with adequate quality was subjected to WES by IntegraGen (France, Evry).

Libraries were prepared with NEBNext^® Ultra™ II Directional RNA Library Prep Kit for Illumina protocol according to supplier recommendations.

Briefly, the key stages of this protocol are successively: the purification of PolyA containing mRNA molecules using poly-T oligo-attached magnetic beads from 100ng total RNA (with the Magnetic mRNA Isolation Kit from NEB); a fragmentation using divalent cations under elevated temperature to obtain approximately 300bp pieces; double strand cDNA synthesis, and finally Illumina adapter ligation and cDNA library amplification by PCR for sequencing. Sequencing was then carried out on Paired-end 100b reads of Illumina NovaSeq. Image analysis and base calling is performed using Illumina Real Time Analysis with default parameters.

Quality of reads was assessed for each sample using FastQC (V.0.11.4; http://www.bioinformatics.babraham.ac.uk/projects/fastqc/).

RNA-SeQC provided key measures of data quality. These metrics were shown within Reporting and included yield, alignment, and duplication rates, rRNA content, regions of alignment (exon, intron and intragenic). Alignment was performed by STAR (https://github.com/alexdobin/STAR).

The duplicate reads (e.g., paired-end reads in which the insert DNA molecules have identical start and end locations in the Human genome) were removed using Sambamba tools (https://github.com/biod/sambamba).

Variant calling for the identification of SNVs (Single Nucleotide Variations) and small insertions/deletions (up to 20bp) was performed via the Broad Institute’s GATK Haplotype Caller GVCF tool.

Ensembl’ VEP (Variant Effect Predictor, Release) program was used to process variants for further annotation. This tool annotates variants, determines the effect on relevant transcripts and proteins, and predicts the functional consequences of variants. This included considering data available in gnomAD, the 1000 Genomes Project and the Kaviar databases. Moreover, an in-house database enabled to filter out sequencing artefacts.

Five bioinformatics algorithms for pathogenicity were used to predict the functional, molecular, and phenotypic consequences of coding and non-coding SNPs. This included DANN, FATHMM, MutationTaster, SIFT and Polyphen. The clinical and pathological significance was also added from the ClinVar database. Other information reported included quality score, homozygote/heterozygote status, count of variant allele reads, and presence of the variant in the COSMIC and OncoKB databases.

RegulomeDB was used to annotate SNPs in known and predicted regulatory elements in the intergenic regions.

To detect fusion-genes candidates in RNA-seq data, FusionCatcher (start with fastq files) and STAR-Fusion (start with alignment files) were used to achieve higher detection efficiency (30, 31). These were run using default configurations. In silico validation of a list of fusion transcript predictions was then performed using FusionInspector, a component of the Trinity Cancer Transcriptome Analysis Toolkit (CTAT) (32). FusionInspector assisted in fusion transcript discovery by performing a supervised analysis of fusion predictions, attempting to recover and re-score evidence for such predictions. Fusion-gene candidates were annotated according to several databases of known fusion genes found in healthy samples (known false positives) including the 1000 Genome Project, ChimerDB2, GTEx and cancer databases such as COSMIC, 18Cancers.

Counting of reads per gene was performed using STAR with –quantMode GeneCounts option. Next, raw count was normalized using the Transcripts Per Kilobase Million (TPM) method.

Analyses were performed with Galiléo™, a cloud-based app for dynamic exploration of RNA-Seq expression data. https://integragen.com/fr/bioinformatique/galileo

For preclinical in vivo drug testing, tumor fragments were implanted into the interscapular fat pad of NMRI nude immunodeficient mice strain as described above. When tumors reached a volume comprised between 65 and 270 mm³, mice were randomly assigned to the vehicle or treatment groups (n=6-8 per group). Mice were then treated with 20 mg/kg docetaxel (1 dose every 3 weeks by intraperitoneal route, MedChemExpress, HY-B0011), 12 mg/kg leuprorelin (1 dose/week by subcutaneous injection, MedChemExpress HY-13665), 60 mg/kg enzalutamide (5 doses/week by oral gavage, MedChemExpress HY-70002), 200 mg/kg abiraterone (daily by oral gavage, MedChemExpress HY-70013) and 75 mg/kg olaparib (daily by oral gavage, MedChemExpress HY-10162). Olaparib was also administered in combination with abiraterone and enzalutamide at the same doses and schedule. Drugs were prepared in different solution; in 20% of a mix (50% Tween80 + 50% Ethanol), 80% NaCl 0.9% for docetaxel, in NaCl 0.9% for leuprorelin, in 10% DMSO + 40% PEG300 + 50% (0,5% Tween80) in NaCl 0.9% for enzalutamide, and in 10% DMSO + 90% (10% HP-β-CD) in PBS 1x for abiraterone and olaparib. Mice were treated until one animal reached the maximum ethical limit tumor volume of 1500 mm³ and/or a body weight loss > 20% for 3 consecutive measurements compared to the first day of treatment, in this case the entire group was removed from the study. Tumor volume was measured twice a week with a caliper and calculated as: TV (mm³) = [length (mm) x width (mm) ²]*π/6, where the length and width are the longest and shortest diameters of the tumor, respectively. For evaluation of therapeutic response, tumor growth was calculated as ΔT/ΔC in percentage where ΔT and ΔC are tumor volume changes relative to initial mean tumor volume for treated group (T) and control group (C), respectively, at a specific day. Response to treatment was also classified using the modified Response Evaluation Criteria in Solid Tumors (mRECIST) using the percentage of tumor volume change at the last day of treatment compared with the tumor volume at day 0 and classified as follows: complete response, BestResponse < −95% and BestAvgResponse < −40%; partial response, BestResponse < −50% and BestAvgResponse < −20%; stable disease, BestResponse < 35% and BestAvgResponse < 30%; progressive disease, not otherwise categorized (33). Health status and body weight for all mice were recorded twice weekly to control any adverse effects.

Comparisons between PDX treatment responses were presented as mean ± SEM. ^ns, P>0.05; *, P<0.05; **, P<0.01; ***, P<0.001; ****, P<0.0001 comparing treated to control groups using a two-way ANOVA followed by a Dunnet’s multiple comparisons post-test.

The aim of the present work was to generate PCa PDX models for preclinical applications. To this end, we worked in close collaboration with the Urology department of Strasbourg Hospital and the urological Clinic of Nantes to collect specimens from patients who had undergone surgery or, in a few cases, biopsy. Between 2010 and 2019, we processed 240 PCa samples from 237 patients. Collected samples were obtained from therapeutic or diagnostic procedures: 205 prostatectomies (85%), 31 transurethral resections of the prostate (TURP) (13%), 1 pelvectomy and 1 directly after biopsy ( Supplementary Table 1 ). All men with a presumed diagnosis of PCa were eligible. The median patient age was 64 years (range 43 - 89 years). All the implanted tumors were obtained from localized or regionally advanced diseases, namely 238 primary tumors and 2 samples that were derived from regional lymph node metastasis ( Supplementary Table 1 ). PDX models were developed from patients across the disease progression, i.e., treatment-naïve to castrate resistant disease.

Viable tumor tissues were xenografted subcutaneously into various strains of immune-deficient mice (for details see materials and methods). For each graft, 2 to 5 mice were used depending on material availability. Eighteen primary implants gave rise to a first tumor growth in mouse: 67% originated from prostatectomies (11 RP and 1 pelvectomy out of 18), 28% from specimens harvested following TURP (5 out of 18) and 5% from lymph node metastasis (1 out of 18). We obtained a low take rate probably due to the variability of the amount of viable tissue submitted for PDX development, mouse strain used and engraftment site. It took between 2.5 and 7.5 months to observe the first growth in mouse irrespectively of tumor stage. Nine tumors that grew did not survive after passage 1. The other half (9 tumors) were successfully propagated beyond three passages. Unfortunately, 3 models did not grow after freezing/thawing despite attempts in several strains of mice and 1 model was contaminated by lymphoma.

In total, we established 5 sequentially transplantable PDXs as working models. PCU-012 and PCU-018 PDX models were derived from primary local tumor samples from treatment-naïve patients. C901 PDX model was obtained from TURP. At the time of surgery, the corresponding patient was responsive to androgen-deprivation therapy (ADT). Cancer progressed and the patient underwent an additional palliative TURP eight months later. TURP chips were obtained and successfully implanted to generate the C1022 PDX model, which displays a neuroendocrine phenotype. PCa samples for the establishment of G266 were obtained at pelvectomy and correlated with a recurrence. Supplementary Table 4 summarizes the whole clinical annotated data from patients from whom PDXs were derived.

As described above, the PDX panel presented tumors with diverse clinical and genomic characteristics across the course of PCa disease, allowing to evaluate drug efficacy on either treatment-naïve tumors, hormone-sensitive tumors, or castrate-resistant tumors. We investigated the sensitivity to docetaxel, a standard of care chemotherapy, and enzalutamide, a new generation androgen pathway inhibitor in our established PDX cohort ( Figure 4A , Supplementary Figure 7 ). The response to each treatment reflected the heterogeneous clinical and genomic characteristics of the tumors. As expected, both G266 and C1022 CRPC PDX models were resistant to enzalutamide while the hormone-sensitive C901 PDX model was sensitive and exhibited a partial response. C901 and C1022 PDX models, derived from the same patient, were sensitive to docetaxel with a clear response for C901 and a partial response for C1022, while G266 was resistant ( Figure 4A ). Treatment responses of these two models reflected the hormone-resistance acquisition of the patient over time. Interestingly, PCU-018 from a localized and treatment-naïve tumor, was resistant to docetaxel and enzalutamide ( Figure 4A ).

In a clinical situation, the Response Evaluation Criteria in Solid Tumor (RECIST) are prevalently used to assess treatment responsiveness. Correspondingly, modified RECIST (mRECIST) was suggested to evaluate the treatment response in PDX model (33, 65). The mRECIST is estimated using the percentage of tumor volume change at the last day of treatment compared with tumor volume at day 0. We considered “responders” PDXs that showed a complete response (CR), partial response (PR) or stable disease (SD) and “non-responders” those with a progressive disease (PD) status. C901 was responsive to docetaxel treatment with a negative tumor volume change indicating shrinkage of the tumor after treatment, even though a tiny ball remained palpable; whereas using the modified RECIST classification, C901 was classed as a stable disease (SD) and not a responsive tumor ( Figure 4B ). The tumor response associated with C1022 was classed as a progressive disease (PD).

We then focused on C901 and C1022 PDX models which are interesting as they can provide a relevant preclinical tool to identify resistance mechanisms and to develop new therapeutic strategies. Both models were treated with androgen deprivation therapies (ADT), leuprorelin and enzalutamide, and with the chemotherapeutic docetaxel for at least 21 days. The hormone-sensitive model, C901, displayed relatively slow growth compared to the neuroendocrine castrate-resistant model C1022 as it took 60 days to reach tumor ethical size in the C901 vehicle group compared to 21 days for C1022 ( Figure 4C ). C901 tumor growth was significantly delayed when treated with enzalutamide while C1022 was a non-responder, correlating with their respective characteristics. Both models were responsive to docetaxel treatment, but an acquired resistance appeared for C1022. None of these two PDX responded to leuprorelin treatment which was used to treat the patient before and after the first surgery, the tumor of whom became resistant at the time of the second surgery, eight months later. Collectively, these analyses shed light on the occurrence of resistance in the same patient and the possibility of studying this through corresponding PDXs. Furthermore, the characteristics of PDX C1022 were consistent with the differentiation of neuroendocrine tumors, which were characterized by a highly proliferative profile, lack of responsiveness to hormonal therapies, and poor prognosis (66, 67).

PARP inhibitors have recently been approved for patients presenting defects in homologous recombination repair (68, 69). Since the neuroendocrine castrate resistant C1022 model was unresponsive to all tested drugs and had a homologous deletion of BRCA2, we assessed the antitumor activity of the PARP inhibitor, olaparib, on this model. Given that previous works have demonstrated the improved efficacy of olaparib compared to androgenic pathway inhibitors (70) and given that clinical trials are now focusing on combining PARP inhibitors with other treatments such as androgenic pathway inhibitors (71), we decided to evaluate the efficacy of olaparib also in combination with either enzalutamide or abiraterone. As previously observed ( Figure 4C ), this rapidly growing PDX reached maximum ethical tumor volume between days 21 and 22. The response to androgenic pathway inhibitors was also consistent with our previous results; C1022 PDX was unresponsive to enzalutamide and presented no response to the other inhibitor abiraterone ( Figure 4D ). Tumor growth was drastically reduced upon treatment with olaparib. In some mice, no tumors were even detectable at the end of the study. Of note, no adverse effects were noticed in any of the treated groups (data not shown).

Taken together, these data were consistent with the patient’s tumor responses and paved the way to new therapeutic strategies for patients with CRPC-NE.

In accordance with the patient’s consent, no sequencing of the patient’s tumor or fragment of normal tissue were performed. This introduced a background noise in the calculation of the tumor mutational burden estimated between 100 to 200 somatic variants as it was performed based on a panel of normal and not patient’s tissue.