Abstract
Evidence suggests that altered adipose tissue homeostasis may be an important contributor to the development and/or progression of prostate cancer. In this study, we investigated the adipose transcriptional profiles of low- and high-risk disease to determine both prognostic potential and possible biological drivers of aggressive disease. RNA was extracted from periprostatic adipose tissue from patients categorised as having prostate cancer with either a low or high risk of progression based on tumour characteristics at prostatectomy and profiled by RNA sequencing. The expression of selected genes was then quantified by qRT-PCR in a cross-validation cohort. In the first phase, a total of 677 differentially transcribed genes were identified, from which a subset of 14 genes was shortlisted. In the second phase, a 3 gene (IGHA1, OLFM4, RERGL) signature was refined and evaluated using recursive feature selection and cross-validation, obtaining a promising discriminatory utility (area under curve 0.72) at predicting the presence of high-risk disease. Genes implicated in immune and/or inflammatory responses predominated. Periprostatic adipose tissue from patients with high-risk prostate cancer has a distinct transcriptional signature that may be useful for detecting its occult presence. Differential expression appears to be driven by a local immune/inflammatory reaction to more advanced tumours, than any specific adipose tissue-specific tumour-promoting mechanism. This signature is transferable into a clinically usable PCR-based assay, which in a cross-validation cohort shows diagnostic potential.
Introduction
Prostate cancer is the most widely diagnosed cancer among men in developed countries and the second most diagnosed cancer in men worldwide (Torre et al. 2015). The widespread use of serum PSA measurement as an ad hoc community-based screening test has meant that the majority of patients are diagnosed very early in the natural history of the disease, in the absence of any significant cancer-related symptoms (O’Brien et al. 2011). It is clear both from prospective cohort studies as well as the control arms of randomised intervention trials, that many of these cancers will not pose a threat to patient life or even wellbeing within their lifetime (Parker et al. 2006, Lu-Yao et al. 2009, Hamdy et al. 2016). Counterbalancing this is the recognition that numerically, many diagnosed cancers are biologically aggressive, and prostate cancer is the primary cause of death in over 160,000 men worldwide each year (Torre et al. 2015). However, current methods fail to reliably discriminate indolent tumours from those with metastatic potential (Corcoran et al. 2011, 2012). This uncertainty leads many clinicians and/or patients to choose to complete radical therapy, even when the chance of benefit is small or absent. In addition, although the majority of patients recover well, a minority will develop debilitating post-operative morbidity, which is both time consuming and costly to manage, and significantly impacts quality of life (Michaelson et al. 2008). At the same time, misidentification of biologically aggressive disease is equally troublesome, with the possibility of patients with potentially lethal tumours being inappropriately observed.
As the majority of tumours are neither palpable nor visible sonographically, a key limitation of the current diagnostic strategy is the reliance on prostate sampling, with transrectal ultrasound used to obtain needle biopsy cores from pre-specified areas within the gland (Gore et al. 2001). This inevitably leads to a ‘sampling error’, in which clinically significant tumours may be missed resulting in a false-negative biopsy, or tumour grade, consistently the most important predictor of disease natural history, under- or over-estimated (Corcoran et al. 2011). Indeed a recently reported large cohort study suggests that standard transrectal biopsy fails to diagnose 27–45% of clinically significant cancers (Ahmed et al. 2017).
Increasing evidence over the last decade indicates that altered adipose tissue homeostasis may be an important contributor to the development and/or progression of a number of solid organ tumours, including prostate cancer (Prieto-Hontoria et al. 2011). This was initially suggested by observations linking obesity to the development of aggressive prostate cancer (Snowdon et al. 1984, Andersson et al. 1997, Calle et al. 2003, Freedland et al. 2009, Moller et al. 2015), and findings that obesity-induced biochemical changes in periprostatic adipose tissue can directly affect tumour growth (Venkatasubramanian et al. 2014). As prostate tumours progress locally, they frequently invade the periprostatic adipose compartment, thereby gaining immediate access to adipose tissue and locally produced adipokines that have been associated with tumour cell growth, invasion and metastases (Cheng et al. 1999, Finley et al. 2009, van Roermund et al. 2011, Ribeiro et al. 2012a ). Similarly, tumour-derived factors may simultaneously influence the metabolic profile of periprostatic adipose tissue, perhaps culturing a niche within which the cancer can progress to a more advanced stage (Ribeiro et al. 2012b ). Although the evidence so far is the strongest in support of a role for periprostatic fat in promoting prostate cancer progression in obese patients, the association between changes in local adipose tissue metabolism and prostate cancer aggressiveness in non-obese patients is less clear.
Given the potential role of the local fat depot in prostate cancer progression, as well as the ongoing need for new strategies to improve risk stratification at the time of diagnosis, we investigated the possibility that alterations in periprostatic adipose tissue may be associated with disease risk. We performed RNA sequencing on periprostatic adipose tissue obtained at the time of prostatectomy, to determine if there existed a transcriptional signature that would allow differentiation between groups of patients at high or low risk of progression. We found significant alterations in expression affecting 677 genes and identified a distinct signature that we confirmed by qPCR in an expanded cohort. Analyses of the genes involved suggest that differential expression is due to reactive changes within the tissue rather than local adipose tissue acting as a distinct driver of tumour progression.
Materials and methods
Ethics statement
The collection and use of tissue for this study had Epworth Healthcare institutional review board approval and patients provided written informed consent (HREC approval number 34506).
Study cohort selection
Patients with localised prostate cancer from whom adipose tissue was obtained for research purposes at the time of prostatectomy were identified from a prospectively collected adipose tissue bank (Kerger et al. 2012). Prior to ligation of the dorsal venous complex and prostate pedicles, the anterior prostate was defatted and the specimen was removed immediately, placed in a sterile container and transferred on ice for long-term storage in the vapour phase of liquid nitrogen. Selected patients had sufficient quantities of periprostatic adipose tissue collected at the time of surgery, had no prior local or systemic prostate cancer therapies, could be categorised into low- or high-risk cohorts based on their pathological stage, prostatectomy Gleason score and tumour volume and were free from significant systemic medical conditions. CAPRA-S scores were calculated for all patients as described (Cooperberg et al. 2011). Patients were grouped into a discovery phase (n = 20) and then a second, cross-validation (n = 58) phase, in a balanced fashion between low and high- risk.
Gene expression screen
50–100 µg of adipose tissue were separated from fresh frozen samples stored at ~−160°C. RNA was isolated using the Qiagen RNeasy Lipid Tissue Mini Kit and eluted in 35 µL nuclease-free water. 0.5–1 µg of total RNA was used as the input for cDNA library synthesis using TruSeq RNA Sample Prep Kit v2 (Illumina), and libraries were constructed according to manufacturer’s instructions. Samples were sequenced on a HiSeq 2500 (Illumina) using 101 bp paired-end chemistry, aiming for 50 million mapped paired-end reads per sample.
Data pre-processing and differential expression analysis
A schematic summarising the overall workflow is provided in Supplementary Fig. 1 (see section on supplementary data given at the end of this article). The RNA sequencing (RNA-seq) quality for each sample was checked using the fastqc algorithm (Andrews 2010). Reads were trimmed for Illumina adapters and low-quality fragments using the trimmomatic algorithm, and short reads filtered out from the pools according to default settings (Bolger et al. 2014). The remaining reads were aligned to the reference genome hg19 using the STAR aligner with default settings (Dobin et al. 2013). The gene abundance for each sample was quantified in terms of nucleotide reads per gene (read-count) using featureCounts (Liao et al. 2014). Low abundant genes were filtered from the analysis if not present in at least 0.5 parts per million in two-thirds of the samples in each disease group (i.e., low- and high-risk). The gene read counts were normalised based on their log-medians for each sample. The normalised gene abundances were adjusted for unknown variation using RUVseq (using default settings) (Risso et al. 2014). The number of unwanted covariates (i.e., hidden batches) to account for was chosen through an iterative process. RUVseq was run for an increasing number of covariates (from 1 to 15); for each run, the covariate matrix calculated was added to the design matrix (i.e., low-/high-risk labels) and our samples tested for differential transcription using the edgeR package (Robinson et al. 2010, McCarthy et al. 2012). Based on the ranking of selected putative positive gene controls identified in previous studies (Prieto-Hontoria et al. 2011) (IL6, TNF, LEP, NFKB1, CD68) as well as the P-value distribution of each run, 14 covariates were chosen accordingly (Supplementary Figs 2 and 3). The differential transcription analysis was performed on the resulting data set, utilising the identified 14 covariate parameters in the linear model of edgeR. Pathway analyses were performed on the differentially transcribed genes (false discovery rate, FDR < 0.05) using two algorithms in parallel, SPIA and GSEA (Subramanian et al. 2005, Tarca et al. 2009). A potential transcriptional signature of 14 genes was selected for further analysis prioritised on low FDR, high fold-change and transcript abundance.
Classification using quantitative qRT-PCR
Selected genes were analysed in an extended cross-validation cohort of 58 patients (28 low-risk, 30 high-risk) through quantitative real-time (qRT-) PCR, using 1 µL of cDNA, 0.5–1 µL qRT-PCR primers (see below), 5 µL of TaqMan Fast Advanced Master Mix (Applied Biosystems) and made up to 10 µL volume per well with UltraPure distilled water (Gibco). Primers to 14 genes from the initial exploratory cohort including IGHA1 (Hs00733892_m1), SAA2 (Hs01667582_m1), MYH11 (Hs00224610_m1), RERGL (Hs00922947_m1), SOCS3 (Hs02330328_s1), PLA2G2A (Hs00179898_m1), SLC2A1 (Hs00892681_m1), COL6A6 (Hs01029204_m1), GPR34 (Hs00271105_s1), CLDN1 (Hs00221623_m1), PCDH10 (Hs00252974_s1), SELE (Hs00174057_m1), OLFM4 (Hs00197437_m1) and DES (Hs00157258_m1) were pre-designed and commercially available from Applied Biosystems. Samples were run on a 384-well plate using a Viia7 qRT-PCR machine (Applied Biosystems) under the following conditions: UNG incubation at 50°C for 2 min; polymerase activation at 95°C for 20 s; followed by 40 cycles of denature at 95°C for 1 s; anneal/extend at 60°C for 20 s. Expression levels of target genes were normalised to the geometric mean of GAPDH (Hs00266705_g1, Applied Biosystems), TBP (Hs00427621_m1, Applied Biosystems) and POLR2A (Hs00427621_m1, Applied Biosystems) using the formula 2−ΔC(T).
Machine learning algorithms including support vector machine (Bennett & Demiriz 1999) (with the following settings; SVM-Type: C-classification; SVM-Kernel: radial; cost: 1; gamma: 0.3), random forest (Svetnik et al. 2003) (with the following settings; Number of trees: 500; number of variables tried at each split: 1) and generalised linear model (Dobson & Barnett 2008) (with the following settings; degrees of freedom: 49; residuals: 46) were employed to classify patients as low or high risk both in the training and cross-validation phases. Genes were prioritised based on recursive feature selection using the rfe function from Caret R package (Kuhn 2008) on a randomly subsampled training portion of the data set (~70%). The classification performances of the resulting gene rank were cross-validated against the remaining data set (~30%) across 13 (gene N-1) iterations, first including only the two top genes, with increments of one gene per iteration. This gene feature selection/cross-validation procedure was iterated 30 times with different random, balanced subsampling of training and cross-validation fractions and the mean area under curve (AUC; expressing the prediction performances of a binary classifier) was calculated for each combination. The signature size N was chosen according to the performance trend observed using from 2 to 14 classifiers. The most recurrent N genes across cross-validation were then selected and a final round of cross-validation was performed using such genes. The patients in this cohort who were also present in the RNA-seq analysis were excluded from any cross-validation set, and limited to the training set for the qRT-PCR classifier. The Cibersort tool (Newman et al. 2015) was used to test for epithelial cell infiltration within the profiled periprostatic adipose tissue.
Analysis of TCGA data
Read counts and sample annotations of the TCGA prostate adenocarcinoma RNA-seq dataset were taken from the website portal.gdc.cancer.gov (Abeshouse et al. 2015). Data were filtered, normalised and tested for global differential expression accounting for unwanted variation as described earlier for our RNA-seq data set. The 14 genes analysed with qRT-PCR were employed to classify low-/high-grade patients in a cross-validation fashion analogously to the RNA-seq classification procedure described earlier. Furthermore, the potential infiltration of cancerous cell from the prostate into the periprostatic fat was tested with Cibersort (Newman et al. 2015) using an ad hoc signature based on LM22, with fibroblast and endothelial gene expression signatures were taken from ENCODE, BLUEPRINT and FANTOM5 data sets.
Data and computational algorithms
The raw data of sequence reads can be retrieved at ega-archive.org with the code EGAS00001002446. The informatics code used for the analyses in this work can be retrieved at github.com/stemangiola/Fat-classification-RNAseq-2017.
Results
Patient characteristics
For the initial screen, we selected 20 patients with high- or low-risk disease respectively, as determined by the prostatectomy Gleason score, pathological stage and total tumour volume (Supplementary Table 1). Patients in the low-risk group had median CAPRA-S score of 1 (range 0–4) with an estimated 91.0% 5-year progression-free survival, compared to a median of 7 (range 3–9) in the high-risk group and an estimated 5-year disease-free survival of only 26.9% (Cooperberg et al. 2011). Similarly, tumours in the high-risk patients were significantly larger than those with low-risk disease (mean 8.3 cc vs 0.8 cc, P = 0.003 Students t-test). The groups were however well matched for body mass index (BMI) (low-risk mean 28.0 vs high-risk mean 26.2, P = 0.56 Students t-test). Further patients with similar characteristics were later selected for validation studies (Table 1).
Clinical characteristics of study cohort.
| Low risk | High risk | ||
|---|---|---|---|
| N | 28 | 30 | |
| Age (years) | Median | 60 | 66 |
| Range | 44–74 | 49–80 | |
| PSA (ng/dL) | Median | 5.8 | 7.3 |
| Range | 0.7–41.0 | 2.7–81.0 | |
| <10 | 22 | 20 | |
| 10–20 | 5 | 5 | |
| >20 | 1 | 5 | |
| Pathological stage | pT2a | 8 | – |
| pT2c | 18 | 1 | |
| pT3a | 2 | 20 | |
| pT3b | – | 9 | |
| Gleason sum (ISUP Group) | 6 (1) | 20 | – |
| 3+4 (2) | 8 | 6 | |
| 4+3 (3) | – | 11 | |
| 8 (4) | – | 2 | |
| 9–10 (5) | – | 11 | |
| Tumour volume (cm3) | Mean | 0.63 | 6.9 |
| Range | 0.1–3.3 | 1.2–32.1 | |
| BMI (kg/m2) | Mean | 28.6 | 26.8 |
| s.d. | 4.2 | 3.2 | |
| CAPRA-S | Mean | 1.2 | 6.2 |
| s.d. | 1.2 | 2.2 | |
| Recurrence | No | 28 | 13 |
| Yes | 0 | 17 | |
| Follow-up (months) | Mean | 27 | 29 |
| Range | 2–59 | 3–50 |
Gene expression of adipose tissue in prostate cancer patients
Samples were sequenced to an average depth of 67 million reads. After data filtering and normalisation, the distribution of gene read counts followed an expected log-normal distribution. Preliminary clustering revealed two outlying samples (one each in the low and high-risk group), which were removed as previously recommended (Law et al. 2014, Liu et al. 2015). A multi-dimensional scaling (MDS) analysis (Ritchie et al. 2015) of gene read counts in the low- and high-risk cohorts demonstrated a noticeable separation for periprostatic adipose tissue (Figs 1 and 2A), which increased significantly after reduction of unwanted variation (RUV) that eliminated sample processing batch effects (Supplementary Fig. 4A). No significant clustering was noted based on BMI or statin use, indicating that the effect of tumour risk on transcription was greater than either of these two covariates (Supplementary Figs 5 and 6). On differential expression analysis a total of 677 differentially transcribed genes (FDR < 0.05) were identified between low- and high-risk disease in periprostatic tissue (of which the top 25 genes ranked by FDR P-values are listed in Table 2; full list provided as a supplementary excel file). Overall, the range of fold changes was low (from −3 to 2, Fig. 2B). Interestingly, the majority of the top ranked genes have roles in inflammation and immune response, including IGHA1, SAA1, SAA2, SELE, LYZ, CXCL2 and ITGAD as well SOCS3 (upregulated), which is known to be upregulated by increased levels of the inflammatory cytokines IL6 and IL10, suggesting that differences in immune activation are important to the differing severity of prostate cancers. When differentially transcribed genes were ranked according to their log fold-change (logFC), 5 of the top 20 genes overrepresented in high-risk disease encode various types of immunoglobulins. Further differences in gene expression were identified in genes involved in the transport of calcium or dependent upon calcium for their action; SCGN, GRIN2A, PLA2G2A, genes responsible for forming or maintaining the extracellular matrix; CLDN1, ITIH3, NPNT and genes encoding muscle proteins; MYH11, MUSTN1, DES, MYOZ1. Looking specifically at adipokines that have previously been implicated in driving obesity-related cancer progression, a significant increase in expression in high-risk disease was noted for IL6 (logFC 0.42, FDR = 0.04) and CCL2 (logFC 0.57, FDR = 6.4E-06), although the fold-change was quite small (Supplementary Table 2). No significant difference in expression was identified for TNF, LEP (leptin), ADIPOQ (adiponectin), IL10 or IL8.
Heatmap of the differentially expressed genes in periprostatic adipose tissue between patients with low- and high-risk prostate cancer.
Citation: Endocrine-Related Cancer 25, 5; 10.1530/ERC-18-0058
(A) Multi-dimensional-scaling (MDS) plot of RNA-seq samples after filtering, normalisation and removal of unwanted variation. (B) Smear plot of genes showing the relation between log fold changes (logFC) and gene abundance (count per million, CPM). The genes selected for qRT-PCR analysis are labelled in the plot.
Citation: Endocrine-Related Cancer 25, 5; 10.1530/ERC-18-0058
Top differentially transcribed genes; ordered based on increasing P-value.
| Gene symbol | logFC | logCPM | LR | P-value | FDR |
|---|---|---|---|---|---|
| IGHA1 | 1.7 | 4.6 | 138.9 | 4.5E-32 | 7.0E-28 |
| SAA2 | −1.1 | 6.3 | 111.9 | 3.6E-26 | 2.8E-22 |
| SYCE1 | −1.2 | 1.5 | 98.7 | 2.9E-23 | 1.5E-19 |
| SAA1 | −0.9 | 9.5 | 96.1 | 1.0E-22 | 4.0E-19 |
| CLDN1 | 1.4 | 1.5 | 91.9 | 8.7E-22 | 2.7E-18 |
| SCGN | 1.3 | 1.1 | 87.3 | 9.3E-21 | 2.4E-17 |
| MYH11 | 1.0 | 8.1 | 82.1 | 1.2E-19 | 2.7E-16 |
| SELE | 1.1 | 2.7 | 76.2 | 2.4E-18 | 4.7E-15 |
| MUSTN1 | 0.8 | 4.0 | 69.1 | 9.2E-17 | 1.5E-13 |
| LYZ | −0.8 | 4.7 | 68.9 | 1.0E-16 | 1.5E-13 |
| GRIN2A | 1.2 | 0.2 | 64.8 | 8.0E-16 | 1.1E-12 |
| DES | 1.1 | 4.4 | 64.5 | 9.4E-16 | 1.1E-12 |
| MYOZ1 | 1.1 | 0.5 | 64.4 | 9.8E-16 | 1.1E-12 |
| SHOX2 | 0.9 | 2.4 | 64.2 | 1.0E-15 | 1.2E-12 |
| ITIH3 | 1.2 | 1.8 | 63.0 | 2.0E-15 | 2.1E-12 |
| SUSD5 | 1.0 | 2.8 | 60.4 | 7.4E-15 | 7.1E-12 |
| KRT19 | 1.7 | 1.1 | 59.9 | 9.9E-15 | 8.9E-12 |
| CXCL2 | 0.8 | 3.9 | 59.8 | 1.0E-14 | 8.9E-12 |
| NPNT | 0.8 | 4.1 | 57.7 | 2.9E-14 | 2.3E-11 |
| SOCS3 | 0.8 | 6.8 | 57.2 | 3.8E-14 | 2.9E-11 |
| TBX5 | −1.3 | 2.1 | 56.7 | 4.9E-14 | 3.6E-11 |
| LOC284454 | 0.7 | 4.9 | 56.5 | 5.3E-14 | 3.8E-11 |
| ITGAD | −1.5 | 0.1 | 56.0 | 6.8E-14 | 4.6E-11 |
| PLA2G2A | 0.8 | 6.2 | 55.2 | 1.0E-13 | 6.9E-11 |
| RERGL | 0.8 | 4.7 | 54.6 | 1.4E-13 | 8.9E-11 |
Pathway analysis of the list of differentially transcribed genes using both SPIA and GSEA algorithms (Table 3) identified 18 differentially regulated pathways (FDR < 0.05), with an overlap of 7 pathways. The majority of differences observed in functional pathway regulation demonstrate cancer-related alterations to immune response and inflammation.
Pathways enriched in periprostatic adipose tissue derived from patients with low-risk compared to high-risk prostate cancer as determined by two independent algorithms, SPIA and GSEA.
| Name | ID | FDR | Status | Algorithm |
|---|---|---|---|---|
| Cytokine–cytokine receptor interaction | 4060 | 3.76–04 | Inhibited | SPIA |
| Malaria | 5144 | 2.3E-03 | Activated | SPIA |
| Graft vs host disease | 5332 | 3.8E-03 | Inhibited | SPIA |
| Circadian rhythm | 4710 | 3.8E-03 | Inhibited | SPIA |
| cAMP signalling pathway | 4024 | 8.9E-03 | Activated | SPIA |
| Autoimmune thyroid disease | 5320 | 1.9E-02 | Inhibited | SPIA |
| Allograft rejection | 5330 | 1.9E-02 | Inhibited | SPIA |
| Leishmaniasis | 5140 | 1.9E-02 | Activated | SPIA |
| Type I diabetes mellitus | 4940 | 2.1E-02 | Inhibited | SPIA |
| Intestinal immune network for IgA production | 4672 | 2.8E-02 | Activated | SPIA |
| Pathways in cancer | 5200 | 3.3E-02 | Activated | SPIA |
| Systemic lupus erythematosus | 5322 | 4.0E-02 | Activated | SPIA |
| Antigen processing and presentation | M16004 | ≈0 | NA | GSEA |
| Allograft rejection | M18615 | ≈0 | NA | GSEA |
| Type I diabetes mellitus | M12617 | ≈0 | NA | GSEA |
| Graft vs host disease | M13519 | 2.0E-04 | NA | GSEA |
| Autoimmune thyroid disease | M13103 | 6.0E-04 | NA | GSEA |
| Asthma | M13950 | 1.2E-03 | NA | GSEA |
| Leishmania infection | M3126 | 2.8E-03 | NA | GSEA |
| Intestinal immune network for IGA production | M615 | 8.2E-03 | NA | GSEA |
| Viral myocarditis | M12294 | 9.0E-03 | NA | GSEA |
| Metabolism of xenobiotics by cytochrome p450 | M16794 | 1.3E-02 | NA | GSEA |
| Systemic lupus erythematosus | M4741 | 2.1E-02 | NA | GSEA |
A gene signature to be refined and cross-validated with qRT-PCR was selected based on FDR and logFC, including IGHA1, SAA2, MYH11, DES, RERGL, SOCS3, PLA2G2A, SLC2A1, COL6A6, GPR34, CLDN1, PCDH10, SELE and OLFM4.
qRT-PCR refinement of the gene signature
qRT-PCR was used to interrogate the 14 selected genes across a larger cohort of 58 patients (28 low-risk, 30 high-risk). Of these genes, IGHA1, MYH11, RERGL, SOCS3, PLA2G2A, CLDN1 and OLFM4 were confirmed to be significantly differentially transcribed across the two groups (P-value <0.05, one-sided Student’s t-test) and GPR34, PCDH10 and SELE were found to have P-values <0.1 (Supplementary Fig. 7). Overall, qRT-PCR fold-change between low- and high-risk tumours (calculated on delta-TC-value) showed a positive linear correlation with the RNA-seq transcription fold-change (calculated on mean read counts), with a slope of 1.5 and a P-value (linear model; lm R package) (Wilkinson & Rogers 1973) of 0.06 and an R 2 of 0.26 (Fig. 3A). Of the four genes that did not validate, two (SLC2A1 and GPR34) had a logFC <0.7. The second phase feature selection using qRT-PCR gene abundance values led to the decision to set the signature size to 3 (Fig. 3B) as the performances of the three classifiers peaked at this value. The best performance expressed in mean AUC was 0.73 (s.d. = 0.14; Fig. 3B). Among the 3 gene signatures across iterations, the genes IGHA1, OLFM4 and RERGL occurred most often and further cross-validation using only these genes led to a mean AUC of 0.72 (s.d. = 0.14; Fig. 3C), with an out-of-bag estimate of error rate ranging from 31 to 39% across iterations.
(A) The correlation plot of the fold-change for the selected genes between qRT-PCR and RNA sequencing. (B) Area under curve (AUC), of the ROC curve representing the classification performances of three different classifiers on qRT-PCR for a range of gene signature sizes. (C) The performances of the classifier created with the selected gene signature (n = 3 genes) on the qRT-PCR cohort, represented as a ROC curve. For comparison, the dashed lines represent the performances for the same genes on our RNA-seq (Supplementary Fig. 4A) and the TCGA prostate adenocarcinoma data sets.
Citation: Endocrine-Related Cancer 25, 5; 10.1530/ERC-18-0058
Specificity of the gene signature to fat
To confirm that the signature was specific to fat and did not just represent sample contamination by invasive prostate cancer, a number of approaches were taken. Firstly, we examined global differential expression in the TCGA prostate cancer data set and found little overlap in differentially transcribed genes compared to the periprostatic fat cohort with just two of the genes in the 14-gene signature reaching FDR < 0.05 (Supplementary Table 3), neither of which was a component of the final 3-gene assay. Secondly, we applied the 14-gene signature developed for adipose tissue to the TCGA dataset, which showed an overall negligible ability to classify high- and low-risk cancer cancers (Fig. 3C and Supplementary Fig. 4B), with a mean AUC of 0.61 compared with 1.0 using the adipose tissue RNA-seq dataset, and 0.72 using the less complex qRT-PCR for our 3-gene signature (Fig. 3C), indicating that the signature is specific to adipose tissue. In addition, analysis of epithelial cell infiltration of the periprostatic fat using Cibersort indicated that epithelial cells were rare in the adipose tissue in both groups, and not significantly different between low- or high-risk prostate cancer patients (P = 0.84, Supplementary Fig. 8). A number of immune cell subtypes, however, were differed significantly between groups, including eosinophils (P = 0.009) and T memory effector cells (P = 9.8E-8). We also did not detect any expression of prostate epithelial cell-specific transcripts such as KLK3, KLK2 and PCA3 in any sample.
Discussion
Due to the sampling error associated with the standard prostate biopsy technique, there is always the inherent risk that potentially lethal tumours may be missed, even after multiple biopsies (Nafie et al. 2014). Although a number of advances have been made to reduce these errors, such as the use of transperineal mapping biopsies (Ahmed et al. 2017) as well as MRI-guided needle placement (Pokorny et al. 2014), patient misclassification remains a common clinical problem, and there are significant trade-offs in terms of increased costs and potential side effects. Contemporaneously with advances in biopsy technique, two transcriptional signatures have been developed, which are used clinically to predict the presence of more advanced pathological features or an increased risk of recurrence (Cooperberg et al. 2013, Klein et al. 2014), both of which can reduce misclassification error particularly in those patients diagnosed with low-risk disease. However, both these tests depend upon the presence of adequate tumour tissue in the diagnostic core to obtain usable RNA, which may be limited, particularly in patients with low volume disease, and the risk scores obtained can vary with the individual tissue core selected for input from any given cancer due to inherent genomic heterogeneity within primary prostate tumours (Wei et al. 2017). A potential alternative strategy involves the identification of a ‘field-change’ specifically within benign tissue that is associated with adverse pathological features or clinical outcome. Certainly, a number of studies have shown that gene expression in benign prostate tissue of cancer-bearing organs differ significantly from that of benign tissue obtained from cancer-free glands (Schlomm et al. 2009, Risk et al. 2010, Kosari et al. 2012), and the altered expression in 39 genes in tumour-adjacent normal-appearing tissue has recently been associated with the risk of recurrence post treatment (Magi-Galluzzi et al. 2016).
Given the epidemiological and experimental associations between periprostatic adipose tissue homeostasis and prostate cancer aggression, we believed that analysis of periprostatic fat could potentially yield a useful signal that could help detect the presence of high-risk tumours. We therefore performed a genome-wide analysis of gene expression in periprostatic adipose tissue obtained from patients with low- and high-risk disease, and despite the low dynamic range of differential expression, have identified a transcriptional signature that could distinguish between the two groups with high accuracy. The overall lack of significance in correlation between RNA-seq fold-change and qRT-PCR fold-change (Fig. 3A; P-value = 0.06) was expected, considering the increase of sample size and the change of technology for the cross-validation cohort. On the other hand, 7 genes in particular were shown to be consistent between the two cohorts, providing more confidence on the applicability of our signature into a clinical setting. We have then translated and refined this signature into a 3-gene qRT-PCR-based assay, which demonstrates promising discriminatory ability in an expanded cross-validation cohort, confirming the presence of a cancer-related ‘field effect’ in periprostatic adipose tissue that may be clinically exploitable. Indeed, the use of a fat-based expression assay has many advantages over a tumour tissue-based test, most importantly in that it does not rely on the actual detection of tumour in the diagnostic biopsy, and the anterior fat pad is amenable to biopsy, particularly with transperineal needle placement. There is also the potential that there is less heterogeneity in the signal across the fat depot, although this has not been tested.
The role of adipose tissue in promoting cancer initiation and development is best understood in relation to obesity, which has been the subject of intense research interest over the last two decades. It is now well recognised that obesity induces a low-grade chronic inflammatory state within adipose depots, with infiltration by cells of both the innate and specific immune system. This results in the elaboration of various cytokines such as IL6, MCP-1 and TNF-α, which given the prostate gland is surrounded in fatty tissue, may promote prostate cancer progression in a paracrine manner (Taylor et al. 2015). Certainly conditioned media derived from periprostatic adipose tissue from obese patients significantly increased the proliferation of PC-3 cells to a significantly greater extent than similar media from lean counterparts (Venkatasubramanian et al. 2014). However, multiple studies have demonstrated that similarly conditioned media derived from non-obese patients can promote the proliferation and invasion of human prostate tumour cells (Ribeiro et al. 2012a ; Sacca et al. 2012, Laurent et al. 2016). Indeed, prostate cancer has been demonstrated to induce many of these changes in the surrounding fat, establishing a reciprocal loop that promotes tumour progression (Ribeiro et al. 2012b , Laurent et al. 2016). Besides changes in cytokine expression, adipose tissue can secrete a number of fat-specific adipokines, including leptin and adiponectin, which can affect tumour cell growth and have previously been found to be altered in the anterior fat pad of prostate cancer patients (Zhang et al. 2016). Additionally, changes in the expression of proteins involved in adipocyte lipolysis and/or lipogenesis have been implicated in promoting prostate cancer growth, by increasing local supply of lipids required for intratumoural energy production (Guaita-Esteruelas et al. 2018).
Given these observations, a priori we anticipated that any distinguishing expression signature would largely comprise of these recognised changes, particularly given the differences in clinical aggressiveness between the two cohorts used for the screen. However we could only identify significant changes in two potentially relevant cytokines (IL6 and CLL2), and the differential expression was low with a logFC between cohorts being < 0.6. This is perhaps not surprising, given that the cohorts were well matched for BMI and suggests previously described changes may be obesity state specific, and not contributing significantly to the progression of high-grade high-stage disease in non-obese individuals. The optimal 3-gene signature that did emerge comprising IGHA1 (immunoglobulin heavy chain constant alpha 1), OLFM4 (olfactomedin 4) and RERGL (ras-related and oestrogen-regulated growth inhibitor-like protein). IGHA1 encodes the first part of the constant region the IgA1 isoform of the immunoglobulin IgA, a secretory antibody that is produced predominantly at mucosal surfaces where it binds to and prevents pathogen entry. The IgA1 isoform however is predominantly found in tissue and serum, and functions to opsonise foreign antigens to initiate phagocytosis and antibody-mediated cytotoxicity. Interestingly, prostate tumour-induced immune tolerance has previously been linked to the selective recruitment of IgA expressing B-cells, that suppress induction of a specific immune response through the expression of PD-L1 and IL10 (Shalapour et al. 2015). OLFM4 encodes an extracellular matrix protein, which is involved in cell adhesion and has anti-apoptotic effects. Counter-intuitively, intratumoural loss of expression of olfactomedin 4 has been linked to prostate cancer development and progression, reportedly through activation of the hedgehog signalling pathway (Li et al. 2015). In contrast, our data indicate that OLFM4 is overexpressed in periprostatic adipose tissue associated with high-risk disease, although the source of the transcript is unclear, as the expression level in normal fat is very low (Uhlen et al. 2015). One potential source is through tissue infiltration by a specific subset of neutrophils, which express the transcript and are defined by it (Clemmensen et al. 2012). The function of RERGL is unknown, but it shares significant sequence homology with the ras-superfamily member RERG, which encodes a tumour-suppressing GTPase and is predicted to share many of its functions (Liu & Zhang 2017). Loss of expression of RERGL has been identified in colorectal cancer, where it is associated with poorer overall survival (Liu & Zhang 2017). Although RERGL transcript is ubiquitously expressed, including in adipose tissue, preliminary immunohistochemical studies suggest that adipocytes do not elaborate the protein, although both fibroblasts and lymphocytes stain strongly (Uhlen et al. 2015). Certainly tissue decomposition analysis using Cibersort supports alteration in the immune cell composition of adipose tissue as a potential source of the discriminatory signal, with significant changes in the proportion of a number of immune cells types. However, given that immune cells make up a very small proportion of the tissue, and the performance of these types of algorithms in complex tissues is untested, further work is required to validate these finding.
Despite the use of a cross-validation rather than an independent validation cohort this pilot study shows a diagnostic potential for periprostatic tissue. Although the AUC of 0.72 is modest, it is considerable considering the size of the discovery cohort in this study, and it is of the same predictive range of other approved polygenic tests such as Mi-Prostate Score (AUC = 0.77), SelectMDx (AUC = 0.76) and ExoDx (AUC = 0.71) for high-risk disease (McKiernan et al. 2016, Tomlins et al. 2016, Van Neste et al. 2016). This suggests that an extended discovery cohort (n > 200) may result in a more robust signature with greater classification accuracy when translated into a PCR-based assay. Although we have not formally tested the adequacy of periprostatic tissue sampling, anecdotal experience (NMC) indicates that it may be biopsied inadvertently when obtaining anterior cores during transperineal prostate biopsy without adverse effects, and certainly yields sufficient RNA for qRT-PCR (up to 100 ng of RNA from a single biopsy core). Feasibility of this approach will however require formal testing in a prospective study.
In summary, we have identified a transcriptional signature in periprostatic fat that can distinguish patients with clinically localised prostate cancer at low or high risk of progression and have successfully translated it into a 3-gene qRT-PCR-based assay. The basis of this signature appears to be related more to a local immune and/or inflammatory reaction to the presence of high-risk tumour rather than a specific adipose tissue-based tumour-promoting mechanism as previously described, although the latter may be more obvious in obese and severely obese patients. Significant developmental work is required to assess utility in more marginal cases as well its specificity in the presence of benign prostatic conditions such as benign prostatic hyperplasia and prostatitis, before it can translated into a clinically usable test.
Supplementary data
This is linked to the online version of the paper at https://doi.org/10.1530/ERC-18-0058.
Declaration of interest
The authors declare that there is no conflict of interest that could be perceived as prejudicing the impartiality of the research reported.
Funding
The Australian Prostate Cancer Centre Epworth is supported by the Australian Government as represented by the Department of Health and Ageing. S Mangiola is supported by the David Mayor PhD Scholarship from the Prostate Cancer Research Foundation. N M Corcoran was supported by a David Bickart Clinician Research Fellowship from the Faculty of Medicine, Dentistry and Health Sciences at the University of Melbourne, as well a Movember – Distinguished Gentleman’s Ride Clinician Scientist Award through Prostate Cancer Foundation of Australia’s Research Programme.
Author contribution statement
S Mangiola, R Stuchbery, M J Clarkson, C M Hovens and N M Corcoran each contributed to the design of the study. R Stuchbery contributed to the generation of data for the study. S Mangiola, R Stuchbery, A Kowalczyk and G Macintyre each contributed to the analysis of data for the study. R Stuchbery, S Mangiola, C M Hovens and N M Corcoran each contributed to the design and writing of the manuscript.
Acknowledgements
The authors thank Adam Kowalczyk for his helpful advice on the implementation of the marker discovery methodology.
References
Abeshouse ACN, Costello AJ, Hovens CM & Zmdua E 2015 The molecular taxonomy of primary prostate cancer. Cell 163 1011–1025. (https://doi.org/10.1016/j.cell.2015.10.025)
Ahmed HU, El-Shater Bosaily A, Brown LC, Gabe R, Kaplan R, Parmar MK, Collaco-Moraes Y, Ward K, Hindley RG, Freeman A, et al. 2017 Diagnostic accuracy of multi-parametric MRI and TRUS biopsy in prostate cancer (PROMIS): a paired validating confirmatory study. Lancet 389 815–822. (https://doi.org/10.1016/S0140-6736(16)32401-1)
Andersson SO, Wolk A, Bergstrom R, Adami HO, Engholm G, Englund A & Nyren O 1997 Body size and prostate cancer: a 20-year follow-up study among 135006 Swedish construction workers. Journal of the National Cancer Institute 89 385–389. (https://doi.org/10.1093/jnci/89.5.385)
Andrews S 2010 FastQC: a quality control tool for high throughput sequence data. Cambridge, UK: Babraham Bioinformatics. (available at: http://www.bioinformatics.babraham.ac.uk/projects/fastqc/)
Bennett KP & Demiriz A 1999 Semi-supervised support vector machines. Advances in Neural Information Processing Systems 11 368–374.
Bolger AM, Lohse M & Usadel B 2014 Trimmomatic: a flexible trimmer for Illumina sequence data. Bioinformatics 30 2114–2120. (https://doi.org/10.1093/bioinformatics/btu170)
Calle EE, Rodriguez C, Walker-Thurmond K & Thun MJ 2003 Overweight, obesity, and mortality from cancer in a prospectively studied cohort of U.S. adults. New England Journal of Medicine 348 1625–1638. (https://doi.org/10.1056/NEJMoa021423)
Cantarutti A, Bonn SE, Adami HO, Gronberg H, Bellocco R & Balter K 2015 Body mass index and mortality in men with prostate cancer. Prostate 75 1129–1136. (https://doi.org/10.1002/pros.23001)
Cheng L, Darson MF, Bergstralh EJ, Slezak J, Myers RP & Bostwick DG 1999 Correlation of margin status and extraprostatic extension with progression of prostate carcinoma. Cancer 86 1775–1782. (https://doi.org/10.1002/(SICI)1097-0142(19991101)86:9<1775::AID-CNCR20>3.0.CO;2-L)
Clemmensen SN, Bohr CT, Rorvig S, Glenthoj A, Mora-Jensen H, Cramer EP, Jacobsen LC, Larsen MT, Cowland JB, Tanassi JT, et al. 2012 Olfactomedin 4 defines a subset of human neutrophils. Journal of Leukocyte Biology 91 495–500. (https://doi.org/10.1189/jlb.0811417)
Cooperberg MR, Hilton JF & Carroll PR 2011 The CAPRA-S score: a straightforward tool for improved prediction of outcomes after radical prostatectomy. Cancer 117 5039–5046. (https://doi.org/10.1002/cncr.26169)
Cooperberg MR, Simko JP, Cowan JE, Reid JE, Djalilvand A, Bhatnagar S, Gutin A, Lanchbury JS, Swanson GP, Stone S, et al. 2013 Validation of a cell-cycle progression gene panel to improve risk stratification in a contemporary prostatectomy cohort. Journal of Clinical Oncology 31 1428–1434. (https://doi.org/10.1200/JCO.2012.46.4396)
Corcoran NM, Hong MK, Casey RG, Hurtado-Coll A, Peters J, Harewood L, Goldenberg SL, Hovens CM, Costello AJ & Gleave ME 2011 Upgrade in Gleason score between prostate biopsies and pathology following radical prostatectomy significantly impacts upon the risk of biochemical recurrence. BJU International 108 E202–E210. (https://doi.org/10.1111/j.1464-410X.2011.10119.x)
Corcoran NM, Hovens CM, Hong MK, Pedersen J, Casey RG, Connolly S, Peters J, Harewood L, Gleave ME, Goldenberg SL, et al. 2012 Underestimation of Gleason score at prostate biopsy reflects sampling error in lower volume tumours. BJU International 109 660–664. (https://doi.org/10.1111/j.1464-410X.2011.10543.x)
Dobin A, Davis CA, Schlesinger F, Drenkow J, Zaleski C, Jha S, Batut P, Chaisson M & Gingeras TR 2013 STAR: ultrafast universal RNA-seq aligner. Bioinformatics 29 15–21. (https://doi.org/10.1093/bioinformatics/bts635)
Dobson AJ & Barnett AG 2008 An Introduction to Generalized Linear Models Third Edition. Boca Raton, FL, USA: Taylor & Francis.
Finley DS, Calvert VS, Inokuchi J, Lau A, Narula N, Petricoin EF, Zaldivar F, Santos R, Tyson DR & Ornstein DK 2009 Periprostatic adipose tissue as a modulator of prostate cancer aggressiveness. Journal of Urology 182 1621–1627. (https://doi.org/10.1016/j.juro.2009.06.015)
Freedland SJ, Banez LL, Sun LL, Fitzsimons NJ & Moul JW 2009 Obese men have higher-grade and larger tumors: an analysis of the duke prostate center database. Prostate Cancer and Prostatic Diseases 12 259–263. (https://doi.org/10.1038/pcan.2009.11)
Gore JL, Shariat SF, Miles BJ, Kadmon D, Jiang N, Wheeler TM & Slawin KM 2001 Optimal combinations of systematic sextant and laterally directed biopsies for the detection of prostate cancer. Journal of Urology 165 1554–1559. (https://doi.org/10.1016/S0022-5347(05)66347-1)
Guaita-Esteruelas S, Guma J, Masana L & Borras J 2018 The peritumoural adipose tissue microenvironment and cancer. The roles of fatty acid binding protein 4 and fatty acid binding protein 5. Molecular and Cellular Endocrinology 462 107–118.
Hamdy FC, Donovan JL, Lane JA, Mason M, Metcalfe C, Holding P, Davis M, Peters TJ, Turner EL, Martin RM, et al. 2016 10-year outcomes after monitoring, surgery, or radiotherapy for localized prostate cancer. New England Journal of Medicine 375 1415–1424. (https://doi.org/10.1056/NEJMoa1606220)
Kerger M, Hong MK, Pedersen J, Nottle T, Ryan A, Mills J, Peters JS, Moon D, Crowe H, Costello AJ, et al. 2012 Microscopic assessment of fresh prostate tumour specimens yields significantly increased rates of correctly annotated samples for downstream analysis. Pathology 44 204–208. (https://doi.org/10.1097/PAT.0b013e3283511c96)
Klein EA, Cooperberg MR, Magi-Galluzzi C, Simko JP, Falzarano SM, Maddala T, Chan JM, Li J, Cowan JE, Tsiatis AC, et al. 2014 A 17-gene assay to predict prostate cancer aggressiveness in the context of Gleason grade heterogeneity, tumor multifocality, and biopsy undersampling. European Urology 66 550–560. (https://doi.org/10.1016/j.eururo.2014.05.004)
Kosari F, Cheville JC, Ida CM, Karnes RJ, Leontovich AA, Sebo TJ, Erdogan S, Rodriguez E, Murphy SJ & Vasmatzis G 2012 Shared gene expression alterations in prostate cancer and histologically benign prostate from patients with prostate cancer. American Journal of Pathology 181 34–42. (https://doi.org/10.1016/j.ajpath.2012.03.043)
Kuhn M 2008 Building predictive models in R using the caret package. Journal of Statistical Software 28 1–26. (https://doi.org/10.18637/jss.v028.bib5)
Laurent V, Guerard A, Mazerolles C, Le Gonidec S, Toulet A, Nieto L, Zaidi F, Majed B, Garandeau D, Socrier Y, et al. 2016 Periprostatic adipocytes act as a driving force for prostate cancer progression in obesity. Nature Communications 7 10230. (https://doi.org/10.1038/ncomms10230)
Law CW, Chen Y, Shi W & Smyth GK 2014 voom: precision weights unlock linear model analysis tools for RNA-seq read counts. Genome Biology 15 R29. (https://doi.org/10.1186/gb-2014-15-2-r29)
Li H, Liu W, Chen W, Zhu J, Deng CX & Rodgers GP 2015 Olfactomedin 4 deficiency promotes prostate neoplastic progression and is associated with upregulation of the hedgehog-signaling pathway. Scientific Reports 5 16974. (https://doi.org/10.1038/srep16974)
Liao Y, Smyth GK & Shi W 2014 featureCounts: an efficient general purpose program for assigning sequence reads to genomic features. Bioinformatics 30 923–930. (https://doi.org/10.1093/bioinformatics/btt656)
Liu HY & Zhang CJ 2017 Identification of differentially expressed genes and their upstream regulators in colorectal cancer. Cancer Gene Therapy 24 244–250. (https://doi.org/10.1038/cgt.2017.8)
Liu R, Holik AZ, Su S, Jansz N, Chen K, Leong HS, Blewitt ME, Asselin-Labat ML, Smyth GK & Ritchie ME 2015 Why weight? Modelling sample and observational level variability improves power in RNA-seq analyses. Nucleic Acids Research 43 e97. (https://doi.org/10.1093/nar/gkv412)
Lu-Yao GL, Albertsen PC, Moore DF, Shih W, Lin Y, DiPaola RS, Barry MJ, Zietman A, O’Leary M, Walker-Corkery E, et al. 2009 Outcomes of localized prostate cancer following conservative management. JAMA 302 1202–1209. (https://doi.org/10.1001/jama.2009.1348)
Magi-Galluzzi C, Maddala T, Falzarano SM, Cherbavaz DB, Zhang N, Knezevic D, Febbo PG, Lee M, Lawrence HJ & Klein EA 2016 Gene expression in normal-appearing tissue adjacent to prostate cancers are predictive of clinical outcome: evidence for a biologically meaningful field effect. Oncotarget 7 33855–33865.
McCarthy DJ, Chen YS & Smyth GK 2012 Differential expression analysis of multifactor RNA-Seq experiments with respect to biological variation. Nucleic Acids Research 40 4288–4297. (https://doi.org/10.1093/nar/gks042)
McKiernan J, Donovan MJ, O’Neill V, Bentink S, Noerholm M, Belzer S, Skog J, Kattan MW, Partin A, Andriole G, et al. 2016 A novel urine exosome gene expression assay to predict high-grade prostate cancer at initial biopsy. JAMA Oncology 2 882–889. (https://doi.org/10.1001/jamaoncol.2016.0097)
Michaelson MD, Cotter SE, Gargollo PC, Zietman AL, Dahl DM & Smith MR 2008 Management of complications of prostate cancer treatment. CA: A Cancer Journal for Clinicians 58 196–213. (https://doi.org/10.3322/CA.2008.0002)
Moller H, Roswall N, Van Hemelrijck M, Larsen SB, Cuzick J, Holmberg L, Overvad K & Tjonneland A 2015 Prostate cancer incidence, clinical stage and survival in relation to obesity: a prospective cohort study in Denmark. International Journal of Cancer 136 1940–1947.
Nafie S, Pal RP, Dormer JP & Khan MA 2014 Transperineal template prostate biopsies in men with raised PSA despite two previous sets of negative TRUS-guided prostate biopsies. World Journal of Urology 32 971–975. (https://doi.org/10.1007/s00345-013-1225-x)
Newman AM, Liu CL, Green MR, Gentles AJ, Feng W, Xu Y, Hoang CD, Diehn M & Alizadeh AA 2015 Robust enumeration of cell subsets from tissue expression profiles. Nature Methods 12 453–457. (https://doi.org/10.1038/nmeth.3337)
O’Brien BA, Cohen RJ, Ryan A, Sengupta S & Mills J 2011 A new preoperative nomogram to predict minimal prostate cancer: accuracy and error rates compared to other tools to select patients for active surveillance. Journal of Urology 186 1811–1817.
Parker C, Muston D, Melia J, Moss S & Dearnaley D 2006 A model of the natural history of screen-detected prostate cancer, and the effect of radical treatment on overall survival. British Journal of Cancer 94 1361–1368. (https://doi.org/10.1038/sj.bjc.6603105)
Pokorny MR, de Rooij M, Duncan E, Schroder FH, Parkinson R, Barentsz JO & Thompson LC 2014 Prospective study of diagnostic accuracy comparing prostate cancer detection by transrectal ultrasound-guided biopsy versus magnetic resonance (MR) imaging with subsequent MR-guided biopsy in men without previous prostate biopsies. European Urology 66 22–29. (https://doi.org/10.1016/j.eururo.2014.03.002)
Prieto-Hontoria PL, Perez-Matute P, Fernandez-Galilea M, Bustos M, Martinez JA & Moreno-Aliaga MJ 2011 Role of obesity-associated dysfunctional adipose tissue in cancer: a molecular nutrition approach. Biochimica et Biophysica Acta 1807 664–678. (https://doi.org/10.1016/j.bbabio.2010.11.004)
Ribeiro R, Monteiro C, Cunha V, Oliveira MJ, Freitas M, Fraga A, Principe P, Lobato C, Lobo F, Morais A, et al. 2012a Human periprostatic adipose tissue promotes prostate cancer aggressiveness in vitro. Journal of Experimental and Clinical Cancer Research 31 32. (https://doi.org/10.1186/1756-9966-31-32)
Ribeiro RJ, Monteiro CP, Cunha VF, Azevedo AS, Oliveira MJ, Monteiro R, Fraga AM, Principe P, Lobato C, Lobo F, et al. 2012b Tumor cell-educated periprostatic adipose tissue acquires an aggressive cancer-promoting secretory profile. Cellular Physiology and Biochemistry 29 233–240. (https://doi.org/10.1159/000337604)
Risk MC, Knudsen BS, Coleman I, Dumpit RF, Kristal AR, LeMeur N, Gentleman RC, True LD, Nelson PS & Lin DW 2010 Differential gene expression in benign prostate epithelium of men with and without prostate cancer: evidence for a prostate cancer field effect. Clinical Cancer Research 16 5414–5423. (https://doi.org/10.1158/1078-0432.CCR-10-0272)
Risso D, Ngai J, Speed TP & Dudoit S 2014 Normalization of RNA-seq data using factor analysis of control genes or samples. Nature Biotechnology 32 896–902. (https://doi.org/10.1038/nbt.2931)
Ritchie ME, Phipson B, Wu D, Hu Y, Law CW, Shi W & Smyth GK 2015 limma powers differential expression analyses for RNA-sequencing and microarray studies. Nucleic Acids Research 43 e47. (https://doi.org/10.1093/nar/gkv007)
Robinson MD, McCarthy DJ & Smyth GK 2010 edgeR: a Bioconductor package for differential expression analysis of digital gene expression data. Bioinformatics 26 139–140. (https://doi.org/10.1093/bioinformatics/btp616)
Sacca PA, Creydt VP, Choi H, Mazza ON, Fletcher SJ, Vallone VB, Scorticati C, Chasseing NA & Calvo JC 2012 Human periprostatic adipose tissue: its influence on prostate cancer cells. Cellular Physiology and Biochemistry 30 113–122. (https://doi.org/10.1159/000339051)
Schlomm T, Hellwinkel OJ, Buness A, Ruschhaupt M, Lubke AM, Chun FK, Simon R, Budaus L, Erbersdobler A, Graefen M, et al. 2009 Molecular cancer phenotype in normal prostate tissue. European Urology 55 885–890. (https://doi.org/10.1016/j.eururo.2008.04.105)
Shalapour S, Font-Burgada J, Di Caro G, Zhong Z, Sanchez-Lopez E, Dhar D, Willimsky G, Ammirante M, Strasner A, Hansel DE, et al. 2015 Immunosuppressive plasma cells impede T-cell-dependent immunogenic chemotherapy. Nature 521 94–98. (https://doi.org/10.1038/nature14395)
Snowdon DA, Phillips RL & Choi W 1984 Diet, obesity, and risk of fatal prostate cancer. American Journal of Epidemiology 120 244–250. (https://doi.org/10.1093/oxfordjournals.aje.a113886)
Subramanian A, Tamayo P, Mootha VK, Mukherjee S, Ebert BL, Gillette MA, Paulovich A, Pomeroy SL, Golub TR, Lander ES, et al. 2005 Gene set enrichment analysis: a knowledge-based approach for interpreting genome-wide expression profiles. PNAS 102 15545–15550. (https://doi.org/10.1073/pnas.0506580102)
Svetnik V, Liaw A, Tong C, Culberson JC, Sheridan RP & Feuston BP 2003 Random forest: a classification and regression tool for compound classification and QSAR modeling. Journal of Chemical Information and Computer Sciences 43 1947–1958. (https://doi.org/10.1021/bib34160g)
Tarca AL, Draghici S, Khatri P, Hassan SS, Mittal P, Kim JS, Kim CJ, Kusanovic JP & Romero R 2009 A novel signaling pathway impact analysis. Bioinformatics 25 75–82. (https://doi.org/10.1093/bioinformatics/btn577)
Taylor RA, Lo J, Ascui N & Watt MJ 2015 Linking obesogenic dysregulation to prostate cancer progression. Endocrine Connections 4 R68–R80. (https://doi.org/10.1530/EC-15-0080)
Tomlins SA, Day JR, Lonigro RJ, Hovelson DH, Siddiqui J, Kunju LP, Dunn RL, Meyer S, Hodge P, Groskopf J, et al. 2016 Urine TMPRSS2:ERG plus PCA3 for individualized prostate cancer risk assessment. European Urology 70 45–53. (https://doi.org/10.1016/j.eururo.2015.04.039)
Torre LA, Bray F, Siegel RL, Ferlay J, Lortet-Tieulent J & Jemal A 2015 Global cancer statistics, 2012. CA: A Cancer Journal for Clinicians 65 87–108. (https://doi.org/10.3322/caac.21262)
Uhlen M, Fagerberg L, Hallstrom BM, Lindskog C, Oksvold P, Mardinoglu A, Sivertsson A, Kampf C, Sjostedt E, Asplund A, et al. 2015 Proteomics. Tissue-based map of the human proteome. Science 347 1260419. (https://doi.org/10.1126/science.1260419)
Van Neste L, Hendriks RJ, Dijkstra S, Trooskens G, Cornel EB, Jannink SA, de Jong H, Hessels D, Smit FP, Melchers WJ, et al. 2016 Detection of high-grade prostate cancer using a urinary molecular biomarker-based risk score. European Urology 70 740–748. (https://doi.org/10.1016/j.eururo.2016.04.012)
van Roermund JG, Hinnen KA, Tolman CJ, Bol GH, Witjes JA, Bosch JL, Kiemeney LA & van Vulpen M 2011 Periprostatic fat correlates with tumour aggressiveness in prostate cancer patients. BJU International 107 1775–1779. (https://doi.org/10.1111/j.1464-410X.2010.09811.x)
Venkatasubramanian PN, Brendler CB, Plunkett BA, Crawford SE, Fitchev PS, Morgan G, Cornwell ML, McGuire MS, Wyrwicz AM & Doll JA 2014 Periprostatic adipose tissue from obese prostate cancer patients promotes tumor and endothelial cell proliferation: a functional and MR imaging pilot study. Prostate 74 326–335. (https://doi.org/10.1002/pros.22756)
Wei L, Wang J, Lampert E, Schlanger S, DePriest AD, Hu Q, Gomez EC, Murakam M, Glenn ST, Conroy J, et al. 2017 Intratumoral and intertumoral genomic heterogeneity of multifocal localized prostate cancer impacts molecular classifications and genomic prognosticators. European Urology 71 183–192. (https://doi.org/10.1016/j.eururo.2016.07.008)
Wilkinson GN & Rogers CE 1973 Symbolic description of factorial models for analysis of variance. Royal Statistical Society Series C: Applied Statistics 22 392–399.
Zhang Q, Sun LJ, Yang ZG, Zhang GM & Huo RC 2016 Influence of adipocytokines in periprostatic adipose tissue on prostate cancer aggressiveness. Cytokine 85 148–156. (https://doi.org/10.1016/j.cyto.2016.06.019)
