Therapy-induced developmental reprogramming of prostate cancer cells and acquired therapy resistance

Treatment-induced neuroendocrine transdifferentiation (NEtD) complicates therapies for metastatic prostate cancer (PCa). Based on evidence that PCa cells can transdifferentiate to other neuroectodermally-derived cell lineages in vitro, we proposed that NEtD requires first an intermediary reprogramming to metastable cancer stem-like cells (CSCs) of a neural class and we demonstrate that several different AR+/PSA+ PCa cell lines were efficiently reprogrammed to, maintained and propagated as CSCs by growth in androgen-free neural/neural crest (N/NC) stem medium. Such reprogrammed cells lost features of prostate differentiation; gained features of N/NC stem cells and tumor-initiating potential; were resistant to androgen signaling inhibition; and acquired an invasive phenotype in vitro and in vivo. When placed back into serum-containing mediums, reprogrammed cells could be re-differentiated to N-/NC-derived cell lineages or return back to an AR+ prostate-like state. Once returned, the AR+ cells were resistant to androgen signaling inhibition. Acute androgen deprivation or anti-androgen treatment in serum-containing medium led to the transient appearance of a sub-population of cells with similar characteristics. Finally, a 132 gene signature derived from reprogrammed PCa cell lines distinguished tumors from PCa patients with adverse outcomes. This model may explain neural manifestations of PCa associated with lethal disease. The metastable nature of the reprogrammed stem-like PCa cells suggests that cycles of PCa cell reprogramming followed by re-differentiation may support disease progression and therapeutic resistance. The ability of a gene signature from reprogrammed PCa cells to identify tumors from patients with metastasis or PCa-specific mortality implies that developmental reprogramming is linked to aggressive tumor behaviors.


Antibodies and reagents
The antibodies and concentrations used for immunostaining, and Western blotting were: anti-BRN3A

Gene expression analysis
Processed signal was quantile normalized with Agilent GeneSpring 12.0. Gene clustering and Venn diagram were also generated using Agilent GeneSpring 12.0 to determine the 132 gene signature common to all three cell lines. Only genes with a fold-change > 2 and p-value < 0.05 were considered. This produced a list of 150 commonly altered genes, which contained 132 probe overlap with the Mayo Clinic II (MCII), The Cleveland Clinic Foundation (CCF) Memorial Sloan Kettering (MSKCC) cohort datasets. Cellular and Molecular functions enrichment in STM-reprogrammed PCa cells was assessed using Ingenuity Pathway Analysis (IPA ® , QIAGEN) by comparing the imported microarray data generated from our PCa cell lines with the Ingenuity ® knowledge base. A list of relevant networks, canonical pathways and algorithmically generated mechanistic networks based on their connectivity was obtained. A score (p-score = −log10[p-value]) according to the fit of the set of supplied genes and a list of biological functions stored in the Ingenuity Knowledge Base are generated. P-scores > 1.3 are significant. Only genes with a foldchange > 2 and p-value < 0.05 were considered.

Association of 132 gene signature with patients' cohort dataset Patient Cohorts
The Mayo Clinic I (MCI) cohort was designed as a case-cohort study as described previously [1,2]. In this study men are matched triples of metastatic progression (N = 213), biochemical recurrence after prostatectomy (N = 213) and patients with no evidence of disease (N = 213). The Mayo Clinic II (MCII) cohort is a casecohort study consisting of a cohort of 1010 men at highrisk for recurrence after receiving radical prostatectomy between 2001-2006 [2][3][4]. The Cleveland Clinic Foundation (CCF) patient cohort is a case-control study with a 1:3 ratio of metastatic cases (N = 53) versus nonmetastatic cases (N = 144) as described previously [2,5]. These patients were sampled from a total of 2,641 men that were conservatively treated after undergoing radical prostatectomy at Cleveland Clinic between 1987-2008. The Memorial Sloan Kettering (MSKCC) patient cohort has already been described [6].

Specimen selection and processing
After histopathologic review of formalin-fixed paraffin embedded (FFPE) tumor blocks from each case by two expert GU pathologists, the tumor block with the highest Gleason score, regardless of tumor volume was selected for specimen processing. Two x0.6 mm diameter biopsy punch tool cores were used to sample the primary Gleason grade of the index lesion and placed in a microfuge tube for RNA extraction. RNA extraction and microarray hybridization was performed using clinical-grade techniques in a Clinical Laboratory Improvement Amendments (CLIA)-certified laboratory facility (GenomeDx Biosciences, San Diego, CA, USA). CLIA certification was obtained through the Centers for Medicare and Medicaid Services through standard procedures, and laboratory facilities satisfied all criteria required for certification. Total RNA was extracted and purified as described previously [3]. Following microarray quality control using the Affymetrix Power Tools packages, probeset normalization and summarization was performed using the Single Channel Array Normalization (SCAN) algorithm, which normalizes each sample individually by modeling and removing probe-and array-specific background noise [7]. To calculate gene expression, we used Affymetrix Core level summaries for annotated genes. All RNA and micorarray processing was performed in a CLIA-certified laboratory.
Microarray data are available on the NCBI Gene Expression Omnibus as accession numbers GSE46691 (Mayo Clinic I), GSE62116 (Mayo Clinic II), GSE62667 (Cleveland Clinic) and GSE21032 (MSKCC).

Statistical analysis
Metastatic progression after prostatectomy was defined as a positive CT scan or bone scan. Secondary aims were to assess the prognostic value of the signature for biochemical recurrence and death from prostate cancer. A classifier to distinguish between metastatic vs nonmetatstatic cancers was constructed from the 132 gene signature, as a generalized linear model with elastic net regularization. Statistical analyses were performed in R, version 3.2.2 and all statistical tests were two-sided using a 5% significance level. To test the significance of the association with outcome, Wilcoxon rank sum were used. The classifier was constructed using the glmnet package (glmnet 2.0-2). A 10-fold cross-validation was used to tune the lambda value used in glmnet. The final classifier outputs a continuous variable score ranging between 0 and 1, with a higher score indicating a higher probability of metastasis. The classifier model was generated using the Mayo Clinic I cohort as training data. Samples in this cohort were assigned into training (n = 359) and validation (n = 186) as described in Erho 2013. The classifier model was trained on the training portion and the remainder was used for testing. Classifier scores were subsequently generated for every sample in the validation cohorts and performance was assessed separately in each cohort using area under receiver operating characteristic (AUC) curves (pROC_1.8). MCII Kaplan-Meier curve p-values were generated with a weighted Cox regression model (survival 2.38-3).