Abstract
Purpose:
Improved early diagnosis and determination of aggressiveness of prostate cancer (PC) is important to select suitable treatment options and to decrease over-treatment. The conventional marker is total prostate specific antigen (PSA) levels in blood, but lacks specificity and ability to accurately discriminate indolent from aggressive disease.
Experimental design:
In this study, we sought to identify a serum biomarker signature associated with metastatic PC. We measured 157 analytes in 363 serum samples from healthy subjects, patients with non-metastatic PC and patients with metastatic PC, using a recombinant antibody microarray.
Results:
A signature consisting of 69 proteins differentiating metastatic PC patients from healthy controls was identified.
Conclusions and clinical relevance:
The clinical value of this biomarker signature requires validation in larger independent patient cohorts before providing a new prospect for detection of metastatic PC.
Keywords: affinity proteomics, antibody microarrays, biomarkers, cancer, prostate cancer
1. Introduction
Prostate cancer (PC) is the fifth leading cause of cancer death among men worldwide and the second most commonly diagnosed cancer, with an estimated number of 1.3 million new cases in 2018 [1]. PC is a heterogeneous disease, exhibiting both indolent slow-growing tumors and more aggressive forms with rapid progression. The majority of PC patients are diagnosed with localized low or intermediate risk disease with a 5-year survival rate of nearly 100%, but a significant minority present with advanced or metastatic disease where the 5-year relative survival is only 30% [2]. Early and accurate determination of the disease state, particularly those with micro-metastases is crucial for selecting suitable therapeutic strategies to improve survival outcomes [3].
In current clinical practice, diagnosis of PC is made by digital rectal examination and measurement of total prostate specific antigen (PSA) levels in blood to decide if further investigations are necessary, such as imaging in the form of multiparametric magnetic resonance imaging (mpMRI) followed by transrectal or transperineal targeted/systematic ultrasound guided biopsies [4]. Despite the utility of PSA in triggering further tests, this marker lacks sufficient specificity as moderately elevated PSA levels appear also in non-malignant conditions, such as benign prostate hyperplasia and prostatitis. Consequently, many men without cancer or clinically insignificant PC undergo unnecessary investigations including biopsies with potential side effects, including life-threatening sepsis [5].
Intensive efforts have been made to find alternative diagnostic methods to stratify patients and improve early detection of clinically significant PC, such as e.g. PSA velocity, PSA density, proPSA isoforms, complexed PSA, and percentage of free PSA [6]. In addition, combinatorial approaches, such as prostate health index (PHI) and the four kallikrein panel (4Kscore), have been evaluated and shown improved accuracy at time of diagnosis and improved ability to discriminate Gleason≥7 [7]. Similarly, the Stockholm-3 (STHLM3) model and the later updated version, which is comprised of clinical variables, blood biomarkers, a genomic SNP panel and an explicit variable for the HOXB13 SNP, has displayed an improved prediction of Gleason ≥7 [8]. Furthermore, practice is changing rapidly in a large number of countries particularly in Europe whereby a pre-biopsy mpMRI is now triggered by an elevation in serum PSA, and guides targeted biopsies of the prostate, thus reducing the over-detection of low-risk disease, and enhancing the diagnosis of intermediate and high-risk PC [9, 10]. However, this continues to cause controversy, as targeted biopsies alone of visible lesions on mpMRI can miss up to 15% of important PC lesions [11].
In this study, we have focused on identifying blood-based biomarkers associated with metastatic PC disease at time of diagnosis, as this group of patients requires immediate intervention. We have used a high-throughput, recombinant antibody microarray targeting 157 analytes, mainly involved in immune regulation and cancer [12, 13], to identify (i) serum proteins that discriminate patients with a metastatic prostate condition from healthy subjects; (ii) key pathways, and (iii) regulators, involved in metastatic PC disease.
2. Material and methods
2.1. Clinical samples
In this retrospective study, 380 cryopreserved serum samples from the Oxford Biobank, consisting of serum from prostate cancer (PC) patients and healthy controls (HC) were assayed, out of which 363 samples were finally analyzed (Table 1). Samples were collected at different clinics in England from 2002 to 2014, centrifuged and serum was stored at −80°C. All prostate cancer samples, except one (discarded from analysis), were collected at time of diagnosis. Healthy controls (HC) were divided in two groups, (i) HC-1, including men with non-elevated total PSA (≤1 ng/ml) derived from the large ProtectT trial [14], and (ii) HC-2, including men with elevated total PSA (>1 ng/ml) and at least two negative biopsies. All samples were transported on dry ice to CREATE Health Translational Cancer Center at Lund University, Sweden, and upon arrival immediately stored at −80°C. Trial approval was obtained from the UK East Midlands (formerly Trent) Multicentre Research Ethics Committee (01/4/025), and informed consent was obtained from all participants.
Table 1.
Clinical samples included in the data analysis.
Healthy controls (HC) | Prostate cancer (PC) | |||
---|---|---|---|---|
HC-1 | HC-2 | nmPC | mPC | |
Number of samples | 138 | 45 | 147 | 33 |
tPSA | 0.7 (0.5–0.8) | 8.1 (6.0–11.1) | 6.7 (5.2–9.6) | 270 (73.7–723.2) |
Age | 64 (62–68) | 68 (63–72) | 65 (62–71) | 72 (64–82) |
Digital rectal examination | NA | Normal | Normal or abnormal | Normal or abnormal |
Total Gleason | ||||
≤6 | 51 (34.9) | 1 (2.9) | ||
≥7 | 96 (65.3) | 21 (63.6) | ||
unknown | 0 (0) | 11 (32.4) | ||
T stage | ||||
T1/T2 | 108 (73.5) | 4 (12.1) | ||
T3/T4 | 15 (10.3) | 28 (82.4) | ||
unknown | 24 (16.4) | 1 (2.9) | ||
M stage | ||||
M0 | 31 (21.1) | 0 (0) | ||
M1 | 0 (0) | 33 (100)b) | ||
unknown | 116 (78.9) a) | 0 (0) |
All values are median (interquartile range) or frequency (%). tPSA, total PSA (ng/ml); T stage, tumor stage; M stage, metastasis stage; HC-1, healthy controls with total PSA ≤1 ng/ml; HC-2, healthy controls with total PSA >1 ng/ml; nmPC, non-metastatic prostate cancer; mPC, metastatic prostate cancer; NA, not applicable
cases with no indication or suspicion of metastatic disease at time of diagnosis
indicated cases diagnosed with M1b.
2.2. Biotinylation of serum samples
Sample IDs were recoded and all samples were randomized before biotinylation. Biotinylation steps were performed on ice or at 4°C in a Nunc™ 96-well polypropylene plate (Thermo Fisher Scientific, Waltham, USA). Briefly, crude serum samples were centrifuged at 16 000 × g, and 10 μl cleared serum was diluted in PBS (D-PBS-Sterile w/o Mg, Ca, GE Healthcare Life Sciences, Marlborough, USA) after which 2.56 mM Biotin (EZ-LinkTM NHS-PEG4-Biotin, Thermo Fisher Scientific) solution diluted in PBS (GE Healthcare Life Sciences) was added, to a final biotin concentration of 1.13 mM. The reaction was terminated after 2 hours by adding 0.5 M Tris-HCl pH 8.0 (Thermo Fisher Scientific) to a final concentration of 181 mM. For each biotinylation plate, three replicates of a reference serum sample (ERM-DA470k/IFCC) [15] were included as process control. Biotinylated samples were aliquoted and stored at −80°C.
2.3. Antibody production and purification
Human recombinant single-chain Fv (scFv) antibodies (n=371, Supplementary Table 1) directed against 158 proteins, mainly involved in immune regulation and cancer, were selected from in-house designed phage display scFv libraries [16, 17]. ScFvs were purified from E. coli using His MultiTrap FF 96 well plates (GE Healthcare Life Sciences), according to manufacturer’s protocol. Buffer exchange to PBS (GE Healthcare Life Sciences) was performed using Zeba™ 96-well desalt spin plates (Thermo Fisher Scientific), and thereafter 1% sodium azide (GBiosciences, Saint Louis, USA) was added to the purified scFv in PBS to a final concentration of 0.06%. The purity of the scFvs was evaluated by SDS-PAGE, using 8–16% Criterion™ TGX Stain-Free™ Protein Gel (BioRad, Hercules, USA). The concentration was measured using a SPECTROstar Omega microplate reader and analyzed with the included MARS software (BMG Labtech, Ortenberg, Germany).
The specificity, affinity, and on-chip functionality of the scFvs has been assured using stringent phage display selection protocols [16, 17], multiple clones (one to nine) per target, and a molecular design adapted for microarray application [18]. Moreover, the specificity of several of the antibodies has previously been validated using well-characterized human samples and orthogonal methods including ELISA, mass spectrometry, Meso Scale Discovery assay, cytometric bead assay and spiking and blocking experiments [13, 17, 19].
2.4. Generation of antibody microarrays
Produced scFvs (concentration range 0.025 – 0.3 mg/ml) were printed (~330 pl/drop) on black polymer MaxiSorp slides (NUNC, Roskilde, Denmark), using a non-contact printer (SciFlexarrayer S11, Scienion, Berlin, Germany). Each scFv was printed in triplicate. PBS and Biotin-LC-BSA (Thermo Fisher Scientific) were included as negative and positive controls, respectively.
2.5. Antibody microarray assay
Slides, consisting of multiple identical subarrays allowing up to seven different samples to be assayed on one single slide, were processed 6–9 days post microarray generation. A maximum of 15 slides were assayed at the same time. All samples were randomized over the different slides and assay days. In addition, on each array slide a quality control sample, denoted QCpool, consisting of pooled human serum from five healthy individuals, biotinylated at one occasion, was analyzed to allow evaluation of the inter-assay variability, meaning array-to array variations.
All incubation steps were performed at room temperature using a 3D rotator (PS-M3D, Progen Scientific, London, England) with reciprocal rotation (360°) for 9 sec and vibration (5°) for 5 sec. Briefly, slides were mounted in NEXTERION® IC-16 (Schott, Jena, Germany) multi-well incubation chambers, and blocked with MT-PBS (1% (w/v) non-fat dry milk (AppliChem GmbH, Darmstadt, Germany), 1% (v/v) Tween 20 (Merck KGaA, Darmstadt, Germany) in PBS) for 1 h. Arrays were washed four times with T-PBS (0.05% (v/v) Tween 20 in PBS) prior to incubation with diluted (1:50 in MT-PBS) biotinylated serum samples for 2 h. Next, slides were washed four times with T-PBS and then incubated with 1 μg/ml Alexa Fluor 647-Streptavidin (Thermo Fisher Scientific) for 1 h in darkness. Finally, slides were again washed four times with T-PBS, dismounted, quickly immersed in dH2O, and dried with compressed N2 before scanning at 635 nm using the InnoScan 710AL with Mapix software (Innopsys, Carbonne, France).
2.6. Data acquisition and pre-processing
The IMMray Evaluation Software (IES, Immunovia AB, Lund, Sweden) was used to align and quantify spot signal intensities. In total, 14 samples were not quantified due to poor quality images resulting from slide surface defects and/or uneven/high background. Additionally, one sample was removed due to other sampling time than at time of diagnosis. Using IES, local background was subtracted from each spot signal. Each data-point was represented by a mean spot intensity signal calculated based on three spot replicates per antibody. If the replicate coefficient of variation (CV) exceeded 15% from the mean value, the mean spot intensity was based on the two remaining replicates, which displayed most similar signals. At this stage, 7560 out of totally 135415 datapoints had a CV over 15%. Finally, all mean signals were trimmed, meaning that 5% of the lower and upper extreme values were discarded.
Quantified arrays were evaluated considering spot saturation, mean signal intensity, signal-to-noise ratio, and degree of missing points. Two samples (belonging to the HC-1 subgroup) were identified as outliers due to high degree of saturation over a large number of spots and were subsequently excluded. No antibodies were below the signal-to-noise ratio cutoff, defined as mean PBS signal ± 2 SD. In total, eight antibody clones targeting BIRC2, BTK, MCP-1, MUC1, ORP-3, PSA, PTN13 and TM-peptide had more than 5% missing values due to removal of spots with poor quality, and were consequently discarded from the final dataset, which consisted of 363 antibodies targeting 157 different proteins. Importantly, all eight targets except the TM-peptide, were still possible to measure using remaining clone(s) (one or several) targeting the same protein (Supplementary Table 1). Remaining missing values (corresponding to 0.06% of all data-points) were replaced by bagged tree imputation [20]. Finally, acquired mean signal intensities were log2 transformed, visualized using 3D principal component analysis (PCA) (Qlucore AB, Lund, Sweden), and normalized using the Bayes algorithm ComBat [21] to adjust for batch (print-to-print) variation.
To assess the inter-assay variability, a total number of 173 QCpool samples were assayed. Of those, 37 samples were not quantified due to poor quality images resulting from slide surface defects and/or uneven/high background. The remaining 136 QCpool samples were quantified and used for calculation of the inter-assay CV. Briefly, CV values were first calculated per scFv for all slides within an assay day, then a mean CV value over 363 scFvs was calculated per assay day. The inter-assay CV was finally calculated as the average CV over 13 assay days. Importantly, the inter-assay CV was assessed for both raw data, and processed data (bagged tree imputed [20], log2 transformed and ComBat [21] normalized for print-to-print variation).
Raw and processed data is available through Figshare (https://doi.org/10.6084/m9.figshare.12370187).
2.7. Data analysis
Sample and analyte distribution were assessed using PCA, and the performance of each analyte was evaluated using Wilcoxon rank sum test. Analytes with p-value ≤0.05 were considered as significantly different, either elevated or reduced in levels. Also, the Benjamini-Hochberg method was used for false discovery rate control. Moreover, for each analysis, selected sample groups were randomly divided into training and test set (70/30%), where the ratio of case versus control samples within the data sets was retained, and a backward elimination (BE) algorithm was applied using the training set, as previously described [22].
Briefly, using BE, proteins with the lowest discriminating contribution were eliminated, and analyte combinations with highest predictive power were identified. The signature length was selected based on the starting point where the lowest classification error occurred, and was used to build a support vector machine (SVM) classification model in the training set, which was then frozen and tested on the corresponding test set. The receiver operating characteristic (ROC) was used as error metric for all BE procedures and the area under curve (AUC) was used as a measure of the accuracy of performance of the signature in the test set. To minimize over-interpretation and to ensure robustness, a minimum of five independent BE procedures was performed, leading to a minimum of five signature lengths. Furthermore, each analyte in the different signatures was scored and ranked using a robust ranking algorithm (RRA) [23] based on the reverse order of elimination, where number one being the last analyte to be eliminated. To identify the signature with highest predictive classification accuracy, a consensus signature was generated by, (i) combining the results from the different signatures, (ii) ranking the analytes based on a median RRA score, and (iii) eliminating analytes with an occurrence <60%.
Finally, to identify over-represented analytes and pathways, enrichment analysis (i.e. Enrichment analysis in Pathway Maps) was performed using MetaCore (Thomson Reuters, New York, NY, USA). A list consisting of all analytes targeted in the antibody microarray (Supplementary Table 1) was set as background and used as reference for statistics in the enrichment analysis.
3. Results
3.1. Assessment and handling of technical variation
A sample pool, QCpool, was used to analyze technical variation in the antibody microarray. Principle component analysis (PCA) of non-normalized (raw) log2 transformed QCpool data revealed print-to-print variations (Supplementary Figure 1A). This variation was reduced using the Bayes algorithm ComBat[21], and confirmed by a PCA analysis of normalized data (Supplementary Figure 1B). In addition, assessment of the inter-assay variation revealed a mean CV of 16 % for non-normalized data, but only 2 % after ComBat normalization (log2 transformed data), indicating further that the variation observed initially in the raw data could be handled by data processing.
Moreover, signals from binders targeting unique epitopes of the same analyte were compared with each other, and a median correlation coefficient of 0.8463 (including all possible comparisons, n=336; interquartile range, 0.7049–0.9063) was observed. The variation is most probably due to differences in epitope specificity, as indicated by antibody sequence variances between clones targeting the same analyte.
3.2. Analysis of discriminatory protein levels
In total, 363 clinical samples (Table 1) and 363 scFv, corresponding to 157 analytes were analyzed using Wilcoxon rank sum test. We identified 60 proteins that were elevated (p≤0.05) in samples from metastatic PC (mPC) patients compared to non-metastatic PC (nmPC) samples (Table 2). In addition, we identified 16 proteins to be elevated in mPC compared to samples from all healthy controls (HC). Fifteen of those 16 were also identified (by identical antibody clone) in the 60 analytes above, indicating an mPC-associated profile. These proteins were AMOT, Apo-A1, C5, CD40, CFAB, CYTC, DLG1, IL-6, IL-8, IL-16, JAK3, RANTES, STAP1, TNF-alpha and VEGFA, and their levels were all significantly higher in mPC compared to both nmPC and HC (data not shown). When comparing nmPC with HC, two proteins, C1q and UBC9, had significantly lower levels in nmPC (Table 2). Interestingly, UBC9 was also significantly lower in nmPC compared to mPC.
Table 2.
Significantly (p≤0.05) differentially expressed analytes as derived from Wilcoxon rank sum test. Raw p-values were adjusted using the Benjamini-Hochberg method. For analytes targeted by multiple clones, the individual antibody clone suffix is shown within brackets.
mPC vs. nmPC | mPC vs. HC | nmPC vs. HC | ||||||
---|---|---|---|---|---|---|---|---|
Antibody | p-value | BH adj.a) p-value | Antibody | p-value | BH adj. a) p-value | Antibody | p-value | BH adj. a) p-value |
AMOT (2) | 2.86E-02 | 6.75E-02 | AMOT (2) | 4.50E-02 | 6.75E-02 | C1q | 4.99E-02 | 1.07E-01 |
ANM5 (2) | 2.62E-02 | 7.85E-02 | Apo-A1 (1) | 2.07E-02 | 6.21E-02 | UBC9 (3) | 3.28E-02 | 4.92E-02 |
Apo-A1 (1) | 4.55E-02 | 6.82E-02 | C5 (2) | 1.20E-03 | 1.81E-03 | |||
Apo-A4 (3) | 3.14E-02 | 9.43E-02 | CD40 (1) | 4.51E-02 | 6.77E-02 | |||
ATP-5B (1) | 1.38E-02 | 4.14E-02 | CFAB (3) | 1.52E-02 | 2.27E-02 | |||
C1 Inh (3) | 4.73E-02 | 1.42E-01 | CYTC (1) | 1.78E-02 | 2.67E-02 | |||
C3 (1) | 3.69E-02 | 1.11E-01 | DLG1 (2) | 4.58E-02 | 6.87E-02 | |||
C3 (5) | 4.77E-02 | 1.43E-01 | IL-16 (3) | 3.18E-02 | 4.77E-02 | |||
C4 (2) | 2.85E-02 | 8.23E-02 | IL-6 (2) | 4.11E-02 | 6.16E-02 | |||
C4 (3) | 8.53E-03 | 2.56E-02 | IL-6 (3) | 3.95E-02 | 5.92E-02 | |||
C4 (4) | 3.57E-02 | 1.07E-01 | IL-8 (1) | 1.59E-02 | 2.67E-02 | |||
C5 (2) | 2.38E-04 | 7.14E-04 | IL-8 (3) | 1.77E-02 | 3.72E-02 | |||
CCL11 (1) | 2.48E-02 | 7.45E-02 | JAK3 | 4.82E-02 | 7.22E-02 | |||
CCL11 (2) | 1.67E-02 | 5.00E-02 | MCP-1 (8) | 4.09E-02 | 1.23E-01 | |||
CD40 (1) | 1.71E-02 | 5.13E-02 | RANTES (1) | 3.93E-02 | 7.13E-02 | |||
CFAB (1) | 3.07E-02 | 9.21E-02 | STAP1 (1) | 3.36E-02 | 5.05E-02 | |||
CFAB (3) | 8.48E-03 | 2.27E-02 | STAP1 (2) | 4.90E-02 | 7.35E-02 | |||
CYTC (1) | 7.20E-03 | 2.16E-02 | TNF-alpha (2) | 4.82E-02 | 7.22E-02 | |||
CYTC (4) | 3.98E-02 | 1.04E-01 | VEGFA (3) | 2.35E-02 | 3.53E-02 | |||
DLG1 (2) | 2.80E-02 | 6.87E-02 | ||||||
DLG2 (1) | 4.82E-02 | 1.44E-01 | ||||||
DLG2 (2) | 3.72E-02 | 1.12E-01 | ||||||
DLG4 (1) | 4.65E-02 | 1.40E-01 | ||||||
DPOLM (2) | 3.66E-02 | 1.10E-01 | ||||||
FAS (2) | 4.63E-02 | 1.39E-01 | ||||||
FER (2) | 4.88E-02 | 1.46E-01 | ||||||
GNAI3 (2) | 3.20E-02 | 9.61E-02 | ||||||
HCD2 (1) | 3.56E-02 | 1.01E-01 | ||||||
HCD2 (4) | 4.47E-02 | 1.34E-01 | ||||||
HER2 (2) | 3.16E-02 | 9.47E-02 | ||||||
HsMAD2 (2) | 3.29E-02 | 9.87E-02 | ||||||
IFN-gamma (2) | 4.51E-02 | 1.35E-01 | ||||||
IL-12 (1) | 2.96E-02 | 7.78E-02 | ||||||
IL-16 (3) | 2.36E-02 | 4.77E-02 | ||||||
IL-18 (1) | 4.06E-02 | 1.15E-01 | ||||||
IL-1a (1) | 4.67E-02 | 1.40E-01 | ||||||
IL-1a (3) | 2.06E-02 | 6.17E-02 | ||||||
IL-2 (1) | 3.54E-02 | 8.09E-02 | ||||||
IL-4 (1) | 1.10E-02 | 3.31E-02 | ||||||
IL-6 (2) | 1.78E-02 | 5.34E-02 | ||||||
IL-6 (3) | 2.60E-02 | 5.92E-02 | ||||||
IL-6 (5) | 3.06E-02 | 9.17E-02 | ||||||
IL-8 (1) | 1.78E-02 | 2.67E-02 | ||||||
IL-8 (3) | 2.48E-02 | 3.72E-02 | ||||||
IL-9 (3) | 4.73E-02 | 9.22E-02 | ||||||
ITCH (1) | 3.61E-02 | 1.08E-01 | ||||||
JAK3 | 4.55E-02 | 7.22E-02 | ||||||
KCC2B (1) | 3.25E-02 | 9.74E-02 | ||||||
KCC2B (2) | 4.22E-02 | 1.09E-01 | ||||||
MAPK1 (2) | 2.48E-02 | 7.45E-02 | ||||||
MAPK8 (1) | 4.06E-02 | 1.14E-01 | ||||||
MARK2 (1) | 3.20E-02 | 9.61E-02 | ||||||
MCP-1 (6)/CCL2 (6) | 1.25E-02 | 3.75E-02 | ||||||
MUC1 (5) | 2.93E-02 | 8.79E-02 | ||||||
MYOM2 (2) | 4.49E-02 | 1.17E-01 | ||||||
NOS1 (2) | 2.63E-02 | 7.89E-02 | ||||||
OSTP (1) | 3.69E-02 | 1.11E-01 | ||||||
P85A (3) | 3.69E-02 | 1.11E-01 | ||||||
PGAM5 (1) | 4.84E-02 | 1.45E-01 | ||||||
PRDM8 (1) | 3.23E-02 | 9.69E-02 | ||||||
PRDM8 (2) | 3.19E-02 | 9.56E-02 | ||||||
PTK6 | 3.86E-02 | 1.16E-01 | ||||||
PTN1 (1) | 2.65E-02 | 7.96E-02 | ||||||
RANTES (1) | 4.75E-02 | 7.13E-02 | ||||||
RANTES (3) | 4.43E-02 | 9.25E-02 | ||||||
R-PTP-kappa (2) | 4.84E-02 | 1.45E-01 | ||||||
R-PTP-O (1) | 2.12E-02 | 6.35E-02 | ||||||
STAP1 (1) | 1.53E-02 | 4.59E-02 | ||||||
STAP1 (2) | 1.76E-02 | 5.29E-02 | ||||||
STAT1 (2) | 3.83E-02 | 1.15E-01 | ||||||
TGF-beta-1 (1) | 4.75E-02 | 1.43E-01 | ||||||
TNF-alpha (2) | 2.57E-02 | 7.22E-02 | ||||||
UBC9 (3) | 3.16E-03 | 9.47E-03 | ||||||
UBE2C (2) | 4.37E-02 | 1.31E-01 | ||||||
UCHL5 | 3.61E-02 | 1.08E-01 | ||||||
VEGFA (3) | 1.90E-02 | 3.53E-02 |
mPC, metastatic prostate cancer; nmPC, non-metastatic prostate cancer; HC, healthy controls
BH adj. = Benjamini-Hochberg adjusted
Using the recombinant antibody microarray, we did not observe any significant difference in PSA values between the different sample groups (HC, nmPC, mPC). However, these data demonstrated a lack of correlation to clinically measured PSA levels (Table 1), possibly due to dissimilarities in epitope specificities between the antibodies used in the two different assays.
3.3. Identification of signatures for classification
To define a biomarker signature that classified mPC from healthy controls with non-elevated PSA (HC-1) we utilized BE-SVM, using five different training and test sets to test the model stability. This approach gave corresponding area under the curve (AUC) values in the test sets ranging from 0.68 to 0.81 (mean 0.74) (Figure 1). This variation depends on the limited number of samples included in the training sets, but still the data supports the fact that a discriminatory power exists in the different test sets classifying mPC from HC-1. Subsequently, the results were combined into a consensus signature consisting of 69 analytes targeted by 75 antibodies (Table 3). As expected, analytes from the consensus signature differed from those derived from the Wilcoxon test (Table 2), since different algorithms and sample groups (mPC vs. nmPC/HC in Wilcoxon test, and mPC vs. HC-1 in BE-SVM) were used. However, a total number of 27 analytes, including ATP-5B, C1 Inh, C4, C5, CCL11, CD40, CYTC, DLG2, DLG4, DPOLM, FAS, GNAI3, HCD2, HER2, IL-4, IL-16, IL-18, ITCH, MAPK1, MAPK8, MCP-1, NOS1, OSTP, PRDM8, PTN1, R-PTP-O and UBC9 were observed in both analyses, and were significantly upregulated in mPC compared to nmPC or HC, or both (Figure 2).
Figure 1.
Backward elimination and support vector machine (BE-SVM) based classification of metastatic prostate cancer (mPC) versus healthy controls with total PSA ≤1 ng/ml (HC-1). Signature lengths generated from five different BE iterations, and corresponding AUC values from SVM model evaluation on the test sets.
Table 3.
Consensus signature of analytes classifying samples as mPC or HC-1. The analytes in the consensus signature are based on the occurrence in five backward elimination iterations, sorted by highest combined score. For analytes targeted by multiple antibody clones, the individual antibody clone suffix is shown within brackets.
mPC vs. HC-1 | Occurrence (%) in total no. of signatures | Median RRAa) score |
---|---|---|
CYTC (1) | 100 | 4 |
IgM (2) | 100 | 8 |
C5 (2) | 100 | 9 |
MCP-1 (9) | 100 | 36 |
PTPRD (2) | 100 | 45 |
MAPKK 2 (1)/ MEK2 (1) | 100 | 56 |
GLP-1R | 80 | 10 |
Lewis y | 80 | 15 |
HER2 (1)/ ERBB2 (1) | 80 | 19.5 |
MATK (3) | 80 | 20.5 |
GNAI3 (2) | 80 | 26 |
IL-16 (3) | 80 | 29 |
OSTP (3) | 80 | 29 |
HCD2 (2) | 80 | 32.5 |
PRD14 (1) | 80 | 39.5 |
ITCH (1) | 80 | 41 |
HCD2 (3) | 80 | 42 |
KSYK (1) | 80 | 42.5 |
OSTP (2) | 80 | 44.5 |
KKCC1 (1) | 80 | 48 |
C4 (2) | 80 | 49 |
OTUB2 (2) | 80 | 49 |
PRD14 (2) | 80 | 59.5 |
ATP-5B (2) | 80 | 64.5 |
IL-18 (2) | 80 | 66.5 |
SHC1 (2)/ Shc (2) | 80 | 77.5 |
UBC9 (1) | 80 | 82 |
GLP-1 | 60 | 8 |
C1 Inh (4) | 60 | 10 |
SOX11 | 60 | 10 |
PROP | 60 | 12 |
HLA-DR/DP | 60 | 15 |
NOS1 (1) | 60 | 16 |
CCL11 (2) | 60 | 18 |
GM-CSF (4) | 60 | 20 |
C1 Inh (2) | 60 | 22 |
KGP2 (1) | 60 | 24 |
MAPK1 (3)/ Erk1/2 (3) | 60 | 26 |
CKIe (2) | 60 | 27 |
PTN1 (3) | 60 | 30 |
DLG4 (1) | 60 | 30 |
GAK (3) | 60 | 31 |
R-PTP-T (2) | 60 | 32 |
MAPK8 (3) | 60 | 34 |
EGFR | 60 | 36 |
MCP-3 (1) | 60 | 39 |
OTUB1 (1) | 60 | 39 |
IL-4 (2) | 60 | 41 |
APLF (2) | 60 | 43 |
KCC4 (1) | 60 | 44 |
PTPR2 (2) | 60 | 45 |
PTPRD (1) | 60 | 46 |
KRAS | 60 | 48 |
R-PTP-O (1) | 60 | 51 |
TNR14 (1) | 60 | 54 |
FAS (4) | 60 | 55 |
TENS4 | 60 | 55 |
MAPKK 6 (2) | 60 | 55 |
PAR-6B (1) | 60 | 58 |
CD40 (3) | 60 | 58 |
IL-10 (1) | 60 | 61 |
HsHec1 (2) | 60 | 62 |
DLG2 (1) | 60 | 62 |
IL-1ra (3) | 60 | 69 |
ARHGC (1) | 60 | 69 |
OTU6B (2) | 60 | 74 |
PARP-1 | 60 | 75 |
BIRC2 (1) | 60 | 77 |
PRDM8 (2) | 60 | 78 |
PKB gamma (1)/ AKT (1) | 60 | 79 |
MCP-4 (1) | 60 | 83 |
TNF-beta (4) | 60 | 84 |
DPOLM (1) | 60 | 86 |
MATK (1) | 60 | 98 |
TNR3 (3) | 60 | 99 |
mPC, metastatic prostate cancer; HC-1, healthy controls with total PSA ≤1 ng/ml
RRA = robust ranking algorithm
Figure 2.
Significantly (p≤0.05, Wilcoxon rank sum test) elevated analytes in metastatic prostate cancer (mPC) compared to non-metastatic prostate cancer (nmPC) and/or healthy controls (HC) overlapping (either by identical or non-identical antibody clone(s)) with the consensus signature.
3.4. Pathway analysis
To elucidate which pathways that are represented by the proteins found in the consensus signature discriminating mPC from HC-1 (Table 3), enrichment analysis using MetaCore (Thomson Reuters) was performed.
Among the significantly (p≤0.05) enriched pathway maps two PC associated pathways were defined, including (i) EGFR signaling in PC, and (ii) ligand-independent activation of androgen receptor in PC (Supplementary Figure 2). Important top regulators in the identified pathway maps involve EGFR and ERBB2 (HER2), which together with downstream targets, promote cell proliferation and inhibition of apoptosis, supporting PC cell survival and growth (Supplementary Figure 2).
Enrichment analysis was also performed for the analytes showing significant differential expression between mPC and nmPC (Table 2), and demonstrated significance considering regulation of angiogenesis in PC (Supplementary Figure 3). Importantly, several proteins including among others TGF-beta-1, TNF-alpha, IL-1a and CCL2 (MCP-1) induce expression of VEGFA and IL-8, which are important regulators of angiogenesis and tumor neovascularization in PC (Supplementary Figure 3).
Moreover, among the significantly enriched pathway maps, also immune response through the three complement pathways (classical, lectin and alternative pathway) was enhanced (data not shown), as levels of complement associated proteins including C1 Inh, C3, C4, C5 and CFAB were higher in mPC compared to nmPC (Table 2).
4. Discussion
Identification of metastases in PC patients has a major impact on treatment selection and overall survival outcome [24]. In spite of novel emerging imaging techniques showing improved ability to detect clinically significant PC [10] and metastatic PC (mPC) [25], there are still limitations considering the diagnostic accuracy including the dependence on an experienced radiologist interpretation [26]. As of today, there is a lack of alternative non-invasive methods that could discriminate mPC with sufficient sensitivity and specificity at time of diagnosis.
In the present study we used an antibody-based microarray to identify a consensus biomarker signature that discriminate mPC from healthy controls with non-elevated PSA (HC-1). Due to the limited number of samples from metastatic patients the outcome of the analysis and the biomarker signature indicated model instability. To circumvent and improve this and to finally identify a signature with clinical utility, several development steps are needed. First, patient cohorts with similar diagnosis collected from different hospitals need to be tested to estimate the individual contribution of each biomarker to the performance of the signature. Secondly, this will allow us to reduce and refine the signature to minimize potential false positive analytes. Finally, the resulting biomarker signature should be evaluated with blinded samples in a multicenter program to be able to deliver clinical utility.
Nevertheless, using part of our data as a training set, we employed backward elimination and a supervised learning model (SVM) for reduction and classification purpose, and identified a consensus biomarker signature consisting of 69 analytes, which demonstrated the possibility to discriminate mPC from HC-1. In parallel, we applied a non-parametric statistical test (Wilcoxon rank sum test) to the entire set of mPC and non-metastatic PC (nmPC) samples, to identify significantly altered analytes. This analysis revealed 60 analytes as elevated in mPC compared to nmPC, where 27 of the 60 analytes overlapped with the consensus signature. The two different bioinformatic strategies gave, as expected, somewhat different outcome due to variations between the control groups (HC-1 in BE-SVM vs. nmPC in Wilcoxon rank sum test), as well as in the utilized algorithms (BE-SVM vs. Wilcoxon rank sum test). We believe that this comparative approach is crucial to circumvent any biases and limitations related to the use of one single method, in particular in absence of an independent validation cohort. Moreover, to further understand the biological context of the obtained results we performed enrichment analysis, which allowed us to identify enriched pathways in our datasets. This could also elucidate which of the identified proteins in the consensus biomarker signature and among the 60 analytes elevated in mPC vs. nmPC, were enriched in PC associated pathways and are key findings with respect to mPC. Based on enrichment analysis in pathways, as a holistic approach to combine the results derived from BE-SVM and the Wilcoxon rank sum test, we identified EGFR, ERBB2 (HER2), IL-8 and VEGFA as potential important regulators of mPC. Noteworthy, despite the fact that only HER2 being a common denominator between the two bioinformatic approaches (BE-SVM and Wilcoxon rank sum test), we believe that the biological relevance of EGFR, IL-8 and VEGFA may be equally important in the context of mPC, which is also supported by previous findings reported in the literature, as discussed below.
Both EGFR and ERBB2 are members of the ERBB family of signaling receptors. Recent finding shows that ERBB2 and EGFR overexpression on a cellular level support metastatic progression of PC to bone [27]. In addition, higher expression of EGFR has been observed in exosomes derived from PC serum compared to healthy subjects [28], and serum levels of ERBB2 have been correlated with presence of metastatic prostate disease [29]. In line with these observations, we found upregulation of ERBB2 in mPC compared to nmPC, and presence of ERBB2 and EGFR in the consensus biomarker signature discriminating mPC from HC-1. Of note, high serum EBBR2 and EGFR levels have also been associated with advanced stage of breast cancer [30] and malignant pleural mesothelioma [31], respectively, potentially indicating a common role in disease progression, not necessarily limited to PC.
We also observed IL-8 to be significantly upregulated in mPC compared to both nmPC and HC. In consistency with our results, a significant elevation of IL-8 has been observed in men with PC bone metastases, compared to men with localized PC disease. [32] In addition, recent findings demonstrate that IL-8 is associated with PC aggressiveness and androgen receptor loss in primary and metastatic PC disease.[33] Importantly, expression of IL-8 has also been shown to correlate with PC tumorigenicity, metastasis and angiogenesis. [34] Likewise, also VEGFA plays a critical role in promoting angiogenesis in PC, leading to enhanced formation of blood vessels and further support of tumor growth [35], which likely contributes to tumor spread and development of metastasis. In this context, stimulation of VEGFA from prostate tumor cells has been demonstrated to be specifically mediated through binding of the cytokine CCL2 to its receptor on PC cells. [35] Importantly, also elevated plasma levels of VEGFA have been reported in metastatic PC compared with localized PC or healthy controls. [36] In agreement with these findings, we observed upregulation of VEGFA and CCL2 in mPC compared to both nmPC and HC, further emphasizing their involvement in metastatic PC disease.
Together, our findings suggest that EGFR, ERBB2, IL-8 and VEGFA are potential important regulators involved in metastatic PC disease. Also, additional analytes, including among others different cytokines (IL-1a, IL-2, IL-4, IL-6, IL-9, IL-12, IL-16, IL-18, RANTES), complement proteins (CFAB, C3, C4, C5) and apolipoproteins (Apo-A1, Apo-A4) were observed to be elevated in mPC compared to nmPC. Some of them have been related to PC aggressiveness (IL-4) [37], progressiveness (IL-6, IL-16) [38], invasion (RANTES) [39], metastases (IL-8) [40], and transition of androgen-dependent to androgen-independent phenotype (IL-6) [41], while others have not, thus opens up for further research of their association to PC.
In summary, we have identified the first biomarker signature that discriminates mPC at diagnosis from HC-1. Furthermore, we also identified both low and high abundant proteins as elevated in serum derived from mPC compared to nmPC and/or HC. These findings indicate that our antibody-based microarray platform can be used to retrieve information on the systemic response of prostate cancer. To translate the identified biomarker signature into clinical utility additional studies are needed, as described. Furthermore, cohort sizes and inclusion of additional control samples from asymptomatic patients should be carefully designed to obtain clinical actionable information with an acceptable statistical power. Nevertheless, our findings presented herein represents a first step in paving the way towards a clinically useful biomarker signature associated with metastatic prostate cancer.
Supplementary Material
Clinical Relevance.
Prostate cancer (PC) is the second most commonly diagnosed cancer worldwide, and represents a heterogeneous disease, displaying both an indolent and aggressive behavior. Prognosis of PC is dependent on early and specific diagnosis of the disease. Prostate specific antigen (PSA) is a well-established diagnostic biomarker used in PC, but has limited specificity and ability to discriminate indolent from aggressive disease. Consequently, there is an urgent need for additional biomarkers that can add value to current diagnostic practices. In this study, we have shown that a number of low and high abundant proteins are elevated in serum derived from metastatic PC patients, compared to non-metastatic PC patients and/or healthy control patients. Importantly, we have identified a candidate serum biomarker signature that discriminates metastatic PC from healthy controls. Once validated in larger well-characterized clinical cohorts, these results might potentially provide a novel opportunity for the detection of metastatic PC, which requires immediate intervention.
5. Acknowledgement
This study was supported by funding from Vinnova (UDI 2016–00462) and CREATE Health Cancer Center at Lund University. Hans Lilja was supported in part by the National Institutes of Health/National Cancer Institute (NIH/NCI) with a Cancer Center Support Grant to Memorial Sloan Kettering Cancer Center (MSKCC) [P30 CA008748], a SPORE grant in Prostate Cancer [P50-CA92629], the Sidney Kimmel Center for Prostate and Urologic Cancers, David H. Koch through the Prostate Cancer Foundation, the Swedish Cancer Society (CAN 2017/559), and the Swedish Research Council (VR-MH project no. 2016–02974). The ProtecT trial is funded by the UK National Institute for Health Research Health Technology Assessment Program (projects 96/20/06, 96/20/99, http://www.nets.nihr.ac.uk/projects/hta/962099) with the University of Oxford as sponsor. Department of Health disclaimer: The views and opinions expressed therein are those of the authors and do not necessarily reflect those of the Department of Health. Freddie Hamdy is supported in part by the National Institute for Health Research Oxford Biomedical Research Centre (Oxford NIHR BRC).
Abbreviations:
- BE
backward elimination
- HC
healthy controls
- HC-1
healthy controls with total PSA ≤1 ng/ml
- HC-2
healthy controls with total PSA >1 ng/ml
- IES
IMMray Evaluation Software
- mPC
metastatic prostate cancer
- mpMRI
multiparametric magnetic resonance imaging
- nmPC
non-metastatic prostate cancer
- PC
prostate cancer
- PSA
prostate specific antigen
- RRA
robust ranking algorithm
- scFv
single-chain fragment variable
- SVM
support vector machine
Footnotes
Conflicts of Interest/Disclosures
Hans Lilja holds a patent on assays to measure intact PSA and is named on a patent for a statistical method to detect prostate cancer that has been commercialized as 4Kscore test by OPKO Health. Hans Lilja receives royalties from sales of the test and has stock in OPKO Health.
7 References
- [1].Bray F, Ferlay J, Soerjomataram I, Siegel RL, Torre LA, Jemal A, CA: a cancer journal for clinicians 2018, 68, 394; [DOI] [PubMed] [Google Scholar]; Filippou P, Ferguson JE 3rd, Nielsen ME, Seminars in interventional radiology 2016, 33, 182. [DOI] [PMC free article] [PubMed] [Google Scholar]
- [2].Institute NC, National Cancer Institute, 2018; [Google Scholar]; Siegel RL, Miller KD, Jemal A, CA: a cancer journal for clinicians 2015, 65, 5. [DOI] [PubMed] [Google Scholar]
- [3].Johansson JE, Andren O, Andersson SO, Dickman PW, Holmberg L, Magnuson A, Adami HO, Jama 2004, 291, 2713. [DOI] [PubMed] [Google Scholar]
- [4].Heidenreich A, Bastian PJ, Bellmunt J, Bolla M, Joniau S, van der Kwast T, Mason M, Matveev V, Wiegel T, Zattoni F, Mottet N, European urology 2014, 65, 124. [DOI] [PubMed] [Google Scholar]
- [5].Loeb S, Vellekoop A, Ahmed HU, Catto J, Emberton M, Nam R, Rosario DJ, Scattoni V, Lotan Y, European urology 2013, 64, 876. [DOI] [PubMed] [Google Scholar]
- [6].Brawer MK, Aramburu EA, Chen GL, Preston SD, Ellis WJ, The Journal of urology 1993, 150, 369; [DOI] [PubMed] [Google Scholar]; Lein M, Semjonow A, Graefen M, Kwiatkowski M, Abramjuk C, Stephan C, Haese A, Chun F, Schnorr D, Loening SA, Jung K, The Journal of urology 2005, 174, 2150; [DOI] [PubMed] [Google Scholar]; Catalona WJ, Partin AW, Slawin KM, Brawer MK, Flanigan RC, Patel A, Richie JP, deKernion JB, Walsh PC, Scardino PT, Lange PH, Subong EN, Parson RE, Gasior GH, Loveland KG, Southwick PC, Jama 1998, 279, 1542; [DOI] [PubMed] [Google Scholar]; Steuber T, O’Brien MF, Lilja H, European urology 2008, 54, 31. [DOI] [PubMed] [Google Scholar]
- [7].Braun K, Sjoberg DD, Vickers AJ, Lilja H, Bjartell AS, European urology 2016, 69, 505; [DOI] [PMC free article] [PubMed] [Google Scholar]; Morote J, Celma A, Planas J, Placer J, Ferrer R, de Torres I, Pacciuci R, Olivan M, Actas urologicas espanolas 2016, 40, 378. [DOI] [PubMed] [Google Scholar]
- [8].Gronberg H, Adolfsson J, Aly M, Nordstrom T, Wiklund P, Brandberg Y, Thompson J, Wiklund F, Lindberg J, Clements M, Egevad L, Eklund M, The Lancet. Oncology 2015, 16, 1667; [DOI] [PubMed] [Google Scholar]; Strom P, Nordstrom T, Aly M, Egevad L, Gronberg H, Eklund M, European urology 2018, 74, 204. [DOI] [PubMed] [Google Scholar]
- [9].Ahmed HU, El-Shater Bosaily A, Brown LC, Gabe R, Kaplan R, Parmar MK, Collaco-Moraes Y, Ward K, Hindley RG, Freeman A, Kirkham AP, Oldroyd R, Parker C, Emberton M, Lancet (London, England) 2017, 389, 815. [DOI] [PubMed] [Google Scholar]
- [10].Kasivisvanathan V, Rannikko AS, Borghi M, Panebianco V, Mynderse LA, Vaarala MH, Briganti A, Budäus L, Hellawell G, Hindley RG, Roobol MJ, Eggener S, Ghei M, Villers A, Bladou F, Villeirs GM, Virdi J, Boxler S, Robert G, Singh PB, Venderink W, Hadaschik BA, Ruffion A, Hu JC, Margolis D, Crouzet S, Klotz L, Taneja SS, Pinto P, Gill I, Allen C, Giganti F, Freeman A, Morris S, Punwani S, Williams NR, Brew-Graves C, Deeks J, Takwoingi Y, Emberton M, Moore CM, New England Journal of Medicine 2018, 378, 1767. [DOI] [PMC free article] [PubMed] [Google Scholar]
- [11].Elwenspoek MMC, Sheppard AL, McInnes MDF, Merriel SWD, Rowe EWJ, Bryant RJ, Donovan JL, Whiting P, JAMA network open 2019, 2, e198427; [DOI] [PMC free article] [PubMed] [Google Scholar]; Bryant RJ, Hobbs CP, Eyre KS, Davies LC, Sullivan ME, Shields W, Sooriakumaran P, Verrill CL, Gleeson FV, MacPherson RE, Hamdy FC, Brewster SF, The Journal of urology 2019, 201, 510; [DOI] [PubMed] [Google Scholar]; Bryant RJ, Yang B, Philippou Y, Lam K, Obiakor M, Ayers J, Chiocchia V, Gleeson F, MacPherson R, Verrill C, Sooriakumaran P, Hamdy FC, Brewster SF, BJU international 2018, 122, 794. [DOI] [PubMed] [Google Scholar]
- [12].Carlsson A, Wingren C, Ingvarsson J, Ellmark P, Baldertorp B, Ferno M, Olsson H, Borrebaeck CA, European journal of cancer (Oxford, England : 1990) 2008, 44, 472; [DOI] [PubMed] [Google Scholar]; Pauly F, Smedby KE, Jerkeman M, Hjalgrim H, Ohlsson M, Rosenquist R, Borrebaeck CA, Wingren C, Leukemia research 2014, 38, 682; [DOI] [PubMed] [Google Scholar]; Borrebaeck CA, Sturfelt G, Wingren C, Methods in molecular biology (Clifton, N.J.) 2014, 1134, 67; [DOI] [PubMed] [Google Scholar]; Gerdtsson AS, Wingren C, Persson H, Delfani P, Nordstrom M, Ren H, Wen X, Ringdahl U, Borrebaeck CA, Hao J, Molecular oncology 2016, 10, 1305. [DOI] [PMC free article] [PubMed] [Google Scholar]
- [13].Ingvarsson J, Larsson A, Sjoholm AG, Truedsson L, Jansson B, Borrebaeck CA, Wingren C, Journal of proteome research 2007, 6, 3527. [DOI] [PubMed] [Google Scholar]
- [14].Hamdy FC, Donovan JL, Lane JA, Mason M, Metcalfe C, Holding P, Davis M, Peters TJ, Turner EL, Martin RM, Oxley J, Robinson M, Staffurth J, Walsh E, Bollina P, Catto J, Doble A, Doherty A, Gillatt D, Kockelbergh R, Kynaston H, Paul A, Powell P, Prescott S, Rosario DJ, Rowe E, Neal DE, The New England journal of medicine 2016, 375, 1415.27626136 [Google Scholar]
- [15].Zegers I, Keller T, Schreiber W, Sheldon J, Albertini R, Blirup-Jensen S, Johnson M, Trapmann S, Emons H, Merlini G, Schimmel H, Clinical chemistry 2010, 56, 1880. [DOI] [PubMed] [Google Scholar]
- [16].Soderlind E, Strandberg L, Jirholt P, Kobayashi N, Alexeiva V, Aberg AM, Nilsson A, Jansson B, Ohlin M, Wingren C, Danielsson L, Carlsson R, Borrebaeck CA, Nature biotechnology 2000, 18, 852. [DOI] [PubMed] [Google Scholar]
- [17].Sall A, Walle M, Wingren C, Muller S, Nyman T, Vala A, Ohlin M, Borrebaeck CAK, Persson H, Protein engineering, design & selection : PEDS 2016, 29, 427. [DOI] [PubMed] [Google Scholar]
- [18].Borrebaeck CA, Wingren C, Methods in molecular biology (Clifton, N.J.) 2011, 785, 247. [DOI] [PubMed] [Google Scholar]
- [19].Pauly F, Dexlin-Mellby L, Ek S, Ohlin M, Olsson N, Jirstrom K, Dictor M, Schoenmakers S, Borrebaeck CA, Wingren C, Journal of proteome research 2013, 12, 5943; [DOI] [PubMed] [Google Scholar]; Kristensson M, Olsson K, Carlson J, Wullt B, Sturfelt G, Borrebaeck CA, Wingren C, Proteomics. Clinical applications 2012, 6, 291; [DOI] [PubMed] [Google Scholar]; Gustavsson E, Ek S, Steen J, Kristensson M, Algenas C, Uhlen M, Wingren C, Ottosson J, Hober S, Borrebaeck CA, New biotechnology 2011, 28, 302; [DOI] [PubMed] [Google Scholar]; Carlsson A, Wuttge DM, Ingvarsson J, Bengtsson AA, Sturfelt G, Borrebaeck CA, Wingren C, Molecular & cellular proteomics : MCP 2011, 10, M110.005033; [DOI] [PMC free article] [PubMed] [Google Scholar]; Dexlin-Mellby L, Sandstrom A, Centlow M, Nygren S, Hansson SR, Borrebaeck CA, Wingren C, Proteomics. Clinical applications 2010, 4, 794; [DOI] [PubMed] [Google Scholar]; Ingvarsson J, Wingren C, Carlsson A, Ellmark P, Wahren B, Engstrom G, Harmenberg U, Krogh M, Peterson C, Borrebaeck CA, Proteomics 2008, 8, 2211; [DOI] [PubMed] [Google Scholar]; Wingren C, Ingvarsson J, Dexlin L, Szul D, Borrebaeck CA, Proteomics 2007, 7, 3055. [DOI] [PubMed] [Google Scholar]
- [20].Kuhn M, Journal of Statistical Software, Foundation for Open Access Statistics 2008, 28, 1. [Google Scholar]
- [21].Johnson WE, Li C, Rabinovic A, Biostatistics (Oxford, England) 2007, 8, 118. [DOI] [PubMed] [Google Scholar]
- [22].Carlsson A, Wingren C, Kristensson M, Rose C, Ferno M, Olsson H, Jernstrom H, Ek S, Gustavsson E, Ingvar C, Ohlsson M, Peterson C, Borrebaeck CA, Proceedings of the National Academy of Sciences of the United States of America 2011, 108, 14252. [DOI] [PMC free article] [PubMed] [Google Scholar]
- [23].Kolde R, Laur S, Adler P, Vilo J, Bioinformatics (Oxford, England) 2012, 28, 573. [DOI] [PMC free article] [PubMed] [Google Scholar]
- [24].Damodaran S, Kyriakopoulos CE, Jarrard DF, The Urologic clinics of North America 2017, 44, 611. [DOI] [PMC free article] [PubMed] [Google Scholar]
- [25].Smith CP, Laucis A, Harmon S, Mena E, Lindenberg L, Choyke PL, Turkbey B, Current oncology reports 2019, 21, 31. [DOI] [PubMed] [Google Scholar]
- [26].Sonn GA, Fan RE, Ghanouni P, Wang NN, Brooks JD, Loening AM, Daniel BL, To’o KJ, Thong AE, Leppert JT, European urology focus 2019, 5, 592. [DOI] [PubMed] [Google Scholar]
- [27].Day KC, Lorenzatti Hiles G, Kozminsky M, Dawsey SJ, Paul A, Broses LJ, Shah R, Kunja LP, Hall C, Palanisamy N, Daignault-Newton S, El-Sawy L, Wilson SJ, Chou A, Ignatoski KW, Keller E, Thomas D, Nagrath S, Morgan T, Day ML, Cancer research 2017, 77, 74. [DOI] [PMC free article] [PubMed] [Google Scholar]
- [28].Kharmate G, Hosseini-Beheshti E, Caradec J, Chin MY, Tomlinson Guns ES, PloS one 2016, 11, e0154967. [DOI] [PMC free article] [PubMed] [Google Scholar]
- [29].Osman I, Mikhail M, Shuch B, Clute M, Cheli CD, Ghani F, Thiel RP, Taneja SS, The Journal of urology 2005, 174, 2174. [DOI] [PubMed] [Google Scholar]
- [30].Lee MH, Jung S-Y, Kang SH, Song EJ, Park IH, Kong S-Y, Kwon YM, Lee KS, Kang H-S, Lee ES, PloS one 2016, 11, e0163370. [DOI] [PMC free article] [PubMed] [Google Scholar]
- [31].Gaafar R, Bahnassy A, Abdelsalam I, Kamel MM, Helal A, Abdel-Hamid A, Eldin NA, Mokhtar N, Lung cancer (Amsterdam, Netherlands) 2010, 70, 43. [DOI] [PubMed] [Google Scholar]
- [32].Lehrer S, Diamond EJ, Mamkine B, Stone NN, Stock RG, Technology in cancer research & treatment 2004, 3, 411. [DOI] [PubMed] [Google Scholar]
- [33].Maynard JP, Ertunc O, Kulac I, Baena Del Valle JA, De Marzo AM, Sfanos KS, Molecular Cancer Research 2019, molcanres.0595.2019. [DOI] [PubMed] [Google Scholar]
- [34].Kim SJ, Uehara H, Karashima T, McCarty M, Shih N, Fidler IJ, Neoplasia 2001, 3, 33; [DOI] [PMC free article] [PubMed] [Google Scholar]; Aalinkeel R, Nair MPN, Sufrin G, Mahajan SD, Chadha KC, Chawda RP, Schwartz SA, Cancer research 2004, 64, 5311. [DOI] [PubMed] [Google Scholar]
- [35].Li X, Loberg R, Liao J, Ying C, Snyder LA, Pienta KJ, McCauley LK, Cancer research 2009, 69, 1685. [DOI] [PMC free article] [PubMed] [Google Scholar]
- [36].Duque JL, Loughlin KR, Adam RM, Kantoff PW, Zurakowski D, Freeman MR, Urology 1999, 54, 523. [DOI] [PubMed] [Google Scholar]
- [37].Goldstein R, Hanley C, Morris J, Cahill D, Chandra A, Harper P, Chowdhury S, Maher J, Burbridge S, Cancers 2011, 3, 4281. [DOI] [PMC free article] [PubMed] [Google Scholar]
- [38].Comperat E, Roupret M, Drouin SJ, Camparo P, Bitker MO, Houlgatte A, Cancel-Tassin G, Cussenot O, The Prostate 2010, 70, 1622; [DOI] [PubMed] [Google Scholar]; Nakashima J, Tachibana M, Horiguchi Y, Oya M, Ohigashi T, Asakura H, Murai M, Clinical cancer research : an official journal of the American Association for Cancer Research 2000, 6, 2702; [PubMed] [Google Scholar]; Rojas A, Liu G, Coleman I, Nelson PS, Zhang M, Dash R, Fisher PB, Plymate SR, Wu JD, Oncogene 2011, 30, 2345. [DOI] [PMC free article] [PubMed] [Google Scholar]
- [39].Vaday GG, Peehl DM, Kadam PA, Lawrence DM, The Prostate 2006, 66, 124. [DOI] [PubMed] [Google Scholar]
- [40].Inoue K, Slaton JW, Eve BY, Kim SJ, Perrotte P, Balbay MD, Yano S, Bar-Eli M, Radinsky R, Pettaway CA, Dinney CPN, Clinical Cancer Research 2000, 6, 2104; [PubMed] [Google Scholar]; Kim SJ, Uehara H, Karashima T, McCarty M, Shih N, Fidler IJ, Neoplasia 2001, 3, 33. [DOI] [PMC free article] [PubMed] [Google Scholar]
- [41].Lee SO, Lou W, Hou M, de Miguel F, Gerber L, Gao AC, Clinical cancer research : an official journal of the American Association for Cancer Research 2003, 9, 370. [PubMed] [Google Scholar]
Associated Data
This section collects any data citations, data availability statements, or supplementary materials included in this article.