*4.1. Tumor Sample Preparation and Xenograft Surgery Procedure*

LAPC9 and BM18 xenografts were maintained subcutaneously in 6-week-old CB17 SCID male mice under anesthesia (Domitor® 0.5 mg/kg, Dormicum 5 mg/kg and Fentanyl 0.05 mg/kg). All animal experiments were approved by the Ethical Committee of Canton Bern (animal licenses BE55/16 and BE12/17). Castration was achieved by bilateral orchiectomy. For androgen replacement, testosterone propionate dissolved in castor oil (86541-5G, Sigma-Aldrich, Buchs, Switzerland) was administered by single subcutaneous injection (2 mg per dosage, 25-G needle).

#### *4.2. RNA Isolation from Tissue Samples*

Tissue RNA was extracted using the standard protocol of Qiazol (79306, Qiagen AG, Hombrechtikon, Switzerland) tissue lysis by TissueLyser (2 min, 20 Hz). Quality of RNA was assessed by Bioanalyzer 2100 (Agilent Technologies, Basel, Switzerland). RNA from formalin-fixed-paraffin embedded (FFPE) material was extracted using the Maxwell® 16 LEV RNA FFPE Purification Kit (AS1260, Promega AG, Dübendorf, Switzerland).

#### *4.3. RNA Sequencing*

RNA extracted from BM18, and LAPC9 whole PDX tumor extracts (300 ng) were subjected to RNA sequencing. Specimens were prepared for RNA sequencing using Tru-Seq RNA Library Preparation Kit v2 or riboZero, as previously described [51]. RNA integrity was verified using the Bioanalyzer 2100 (Agilent Technologies, Basel, Switzerland). Complementary cDNA was synthesized from total RNA using Superscript III reverse transcriptase (18080093, Thermo Fisher Scientific, Basel, Switzerland). Sequencing was then performed on GAII, Hi-Seq 2000 or Hi-Seq 2500. The sample preparation was performed according to the protocol "NEBNext Ultra II Directional RNA Library Prep Kit (NEB #E7760S/L, Illumina GmbH, Zürich, Switzerland). Briefly, mRNA was isolated from total RNA using the oligo-dT magnetic beads. After fragmentation of the mRNA, a cDNA synthesis was performed. This was used for ligation with the sequencing adapters and PCR amplification of the resulting product. The quality and yield after sample preparation was measured with the Fragment Analyzer. The size of the resulting products was consistent with the expected size distribution (a broad peak between 300–500 bp). Clustering and DNA sequencing using the NovaSeq6000 was performed according to manufacturer's protocols. A concentration of 1.1 nM of DNA was used. Image analysis, base calling and quality check was performed with the Illumina (Illumina GmbH, Zürich, Switzerland) data analysis pipeline RTA3.4.4 and Bcl2fastq v2.20.

Sequence reads were aligned using STAR two-pass to the human reference genome GRCh37 [52] and mouse reference genome GRCm38. Gene counts were quantified using the "GeneCounts" option. Per-gene counts-per-million (CPM) were computed and log2-transformed, adding a pseudo-count of 1 to avoid transforming 0. Genes with log2 CPM <1 in more than three samples were removed. Differential expression analysis was performed using the edgeR package [53]. Normalization was performed using the "TMM" (weighted trimmed mean) method, and differential expression was assessed using the quasi-likelihood F test. Genes with false discovery rate FDR <0.05 and >2-fold were considered significantly differentially expressed. RNA-Seq Expectation Maximization (RSEM) was used to obtain TPM (transcripts per million) counts.

Pathway analysis (over-representation analysis) was performed using clusterProfiler R package [54] for Gene Ontology biological processes and KEGG. For Venn Euler diagram analysis, expressed genes were identified using the zFPKM transformation [55]. For the comparison between the states of the BM18 and LAPC9 models, genes were considered expressed if a gene had zFPKM values > −3 [55] in all samples.
