Usefulness of the top-scoring pairs (TSP) of genes for prediction of prostate cancer progression

Abstract

Prediction of cancer progression after radical prostatectomy (RP) is one of the most challenging problems in the management of prostate cancer. Gene-expression profiling is widely used to identify genes associated with such progression. Usually candidate genes are identified according to a gene-by-gene comparison of expression. Recent reports suggested that relative expression of a gene pair more efficiently predicts cancer progression than single-gene analysis does. The top-scoring pair (TSP) algorithm classifies phenotypes according to the relative expression of a pair of genes. We applied the TSP approach to predict which patients would experience systemic tumor progression after RP. Relative expression of TPD52L2/SQLE and CEACAM1/BRCA1 gene pairs identified those patients, with more than 99% specificity but relatively low sensitivity (~10%). These two gene pairs were validated in three independent datasets. Additionally, combining two pairs of genes improved sensitivity without compromising specificity. Functional annotation of the TSP genes demonstrated that they cluster by a limited number of biologic functions and pathways, suggesting that relatively lower expression of genes from specific pathways can predict cancer progression. In conclusion, comparative analysis of the expression of two genes may be a simple and effective classifier for prediction of prostate cancer progression. The TSP approach can be used to identify patients whose prostate cancer will progress after they undergo radical prostatectomy. Two gene pairs can predict which men would experience progression to the metastatic form of the disease. However, because our analysis was based on a relatively small number of genes, a larger study will be needed to identify the best predictors of disease outcome overall.

Introduction

Screening for prostate cancer according to serum prostate-specific antigen (PSA) level has improved early detection of the disease, resulting in increased identification of patients with localized disease that is still curable with surgery and radiotherapy. However, 20–30% of treated patients experience a relapse.^1,2 Thus, from the clinical perspective, it is important to be able to predict which patients will experience a relapse.^3–6

Understanding the biology of prostate cancer progression is essential for enabling the development of prognostic markers and effective therapeutic targets. A number of studies have been conducted to characterize the dynamics of gene expression in prostate cancer progression by using DNA microarrays.^5,7–12 In some studies, tumor-expression signatures associated with clinical parameters and outcome have been identified.^13–16

Normally, researchers use one gene at a time to find any association between gene expression and phenotype. The results of recent studies, however, suggest that assessing the expression of more than one gene (i.e., co-expression analysis) yields a better prediction of tumor progression than the analysis of individual genes does.^5,6 Motivated by these findings, we analyzed whether co-expression patterns can be useful in predicting prostate cancer progression by using the paired-gene approach of the top-scoring pairs (TSP) algorithm as described by Geman et al. in 2004.¹⁷ However, we modified the algorithm by controlling the specificity and sensitivity of the TSPs. This may be particularly important when the results of classification influence the selection of treatment modality or follow-up procedures that can be associated with serious side effects. For our purpose, we took this approach to analyze gene-expression data from a recently reported large case-control microarray study.⁵

Materials and methods

Datasets

To overcome the disadvantageous pattern of microarray data (small sample size and large number of variables), we searched the Gene Expression Omnibus (GEO) database and selected four studies of prostate cancer progression that each had at least 10 samples per phenotype. We used one as the discovery dataset and the other three as validation datasets (Table 1). Three pairs of phenotypes were compared in the four datasets: (1) systemic progression (SYS; development of distant metastasis after radical prostatectomy [RP]) vs no evidence of disease (NED) (the GSE10645 dataset),⁵ (2) primary prostate tumors vs distant metastases (the GSE6752 and GSE6919 datasets),^7,11 and (3) hormone-sensitive vs hormone-refractory prostate cancer (the GSE6811 dataset).¹⁰

Table 1.

Gene-expression datasets used in our analysis

GEO ID	PMID	Phenotypes	Number of samples	Platform	Number of probes
GSE10645	18846227	NED vs SYS	213 vs 213	GPL5858 and GPL5873	1028
GSE6752	17430594	Primary prostate tumor vs prostate tumor metastases	10 vs 21	CodeLink UniSet Human 20K I Bioarray (GPL2891)	23 572
GSE6811	17545589	HSPCs vs HRPCs	10 vs 25	YN Human 36K (sets 1–8) (GPL4747)	36 864
GSE6919	15254046	Primary prostate tumor vs prostate tumor metastases	196 vs 75	Affymetrix Human Genome U95 Version 2, U95C, and U95B Array (GPL8300, GPL93, and GPL92)	12 625 12 646 12 620

Abbreviations: NED, no evidence of disease; SYS, systemic prostate cancer progression; HSPC, hormone-sensitive prostate cancers; HRPC, hormone-refractory prostate cancers.

For our discovery set, we used gene-expression data from the study by Nakagawa et al.⁵ (i.e., dataset 1 as described above). Those authors analyzed the association between gene expression and outcome after initial therapy, comparing data from a group of 213 patients who had NED during the 7 years after undergoing retropubic RP (RRP) with those from a group of 213 patients who experienced SYS during the 5 years after the initial rise in PSA level. The authors analyzed the expression of 1028 genes for which previous evidence had implicated their involvement in prostate cancer progression.

For our validation studies, we did not analyze exactly the same phenotypes that we analyzed in the discovery dataset. As counterparts for the NED data, we used primary tumor data from the GSE6752 and GSE6919 datasets and hormone-sensitive prostate cancer data from the GSE6811 dataset. Our counterparts for the SYS phenotype were data on distant metastases from GSE6752 and GSE6919 and hormone-refractory prostate cancer from GSE6811.

Top-scoring pairs algorithm

The original TSP classifier algorithm was described by Geman et al.¹⁷ Here we give a simplified description of the method. Given one expression matrix X = [x_gn] of G genes and N samples, x_gn represents the gth gene-expression value from the nth sample. Assume that every sample can be labeled as either 1 or 2. For example, the N₁ samples from 1 to N₁ (N₁ < N) are labeled class 1 (e.g., NED), and N₂ samples from (N₁ + 1) to N are labeled class 2 (e.g., SYS), in which N₁ + N₂ = N. The focus is on detecting “marker gene pairs” (i, j) for which there is a significant difference in the probability of observing X_i < X_j from class 1 to class 2. The conditional probabilities of observing X_i < X_j in each class are defined as follows:

$$p_{ij}(1)=\text{Pr}(X_i<X_j|c=1)=\frac{1}{N_1}\sum _{n=1}^{N_1}I(x_{\mathit{\text{in}}}<x_{jn})p_{ij}(2)=\text{Pr}(X_i<X_j|c=2)=\frac{1}{N_2}\sum _{n=N_1+1}^{N}I(x_{in}<x_{jn})$$

in which I(x_in < x_jn) is the indicator function, defined as

$$I(x_{\mathit{\text{in}}}<x_{jn})=\{\begin{array}{cc}\hfill 1,\hfill & \hfill x_{\mathit{\text{in}}}<x_{jn}\hfill \\ \hfill 0,\hfill & \hfill x_{\mathit{\text{in}}}\ge x_{jn}\hfill \end{array}n=1,2,\dots ,N$$

The typical TSP method is based on maximizing the scores of (i,j) Δ_ij = |p_ij(1) − p_ij(2)|. This approach provides superior performance to that of support vector machines and other sophisticated methods for classifying cancer samples.^18,19 Although maximizing the delta value identifies the best classifier, that classifier may be associated with relatively low sensitivity and specificity, which is a concern when the classifier is used in selecting a treatment. Because of this, we included an assessment of specificity and sensitivity in our analysis, keeping at least 99% specificity and trying to maximize the sensitivity.

The statistical significance of classifiers can be assessed by randomly permuting the class labels maintaining the sample sizes N₁ and N₂. From this permutation analysis, we estimated P values associated with a given conditional probability. This P value can be interpreted as the probability that the gene pair is not informative for the classification.

Functional annotation

For functional annotation, we applied Ingenuity Pathways Analysis (Ingenuity Systems, Inc, Redwood City, CA, USA; http://www.ingenuity.com/), which evaluates the distribution of the top-ranked genes by both pathway and gene ontology categories, testing the null hypothesis that the genes are uniformly distributed. P values characterize the statistical evidence for the clustering of the genes by pathway or functional categories: the lower the P value, the stronger the statistical evidence that the top-ranking genes belong to a specific pathway or functional category.

Results

Top-scoring pairs

Table 2 lists the TSPs we identified as useful for predicting which patients will develop SYS and which will have NED after they undergo RRP. These TSPs were selected on the basis of a delta value of at least 0.1 and on the condition that the specificity was ≥ 99%. We identified 22 TSPs for predicting SYS but only two for predicting NED.

Table 2.

Top-scoring pairs of genes for prediction of systemic progression (SYS) and no evidence for disease (NED). The last column is the estimated P value for pair genes by 10⁶ permutation testing

Predict	Gene1	Gene2	P(X1<X2| NED)	P(X1<X2| SYS)	Delta = |P1−P2|	Sensitivity	Specificity	RR	P value
NED	F5	ESR2	0.108	0.000	0.108	0.108	1.000	0.000	<10E-6
NED	COL1A1	CDKN1B	0.149	0.010	0.139	0.149	0.990	0.119	<10E-6
SYS	PAGE4	ANXA2	0.000	0.130	0.130	0.130	1.000	2.115	<10E-6
SYS	HLF	BLM	0.000	0.105	0.105	0.105	1.000	2.081	<10E-6
SYS	PASK	SQLE	0.000	0.105	0.105	0.105	1.000	2.055	10E-6
SYS	MYBPC1	TMEM65	0.005	0.175	0.170	0.175	0.995	2.121	<10E-6
SYS	SNTB2	SQLE	0.005	0.165	0.160	0.165	0.995	2.046	<10E-6
SYS	JUN	MKI67	0.005	0.155	0.150	0.155	0.995	2.037	10E-6
SYS	PAGE4	TAF2	0.005	0.140	0.135	0.140	0.995	2.037	5 × 10E-6
SYS	PTN	IMMT	0.005	0.135	0.130	0.135	0.995	2.027	<10E-6
SYS	EFS	WDR67	0.005	0.130	0.125	0.130	0.995	2.018	3 × 10E-6
SYS	EDG7	LGR4	0.005	0.125	0.120	0.125	0.995	2.018	<10E-6
SYS	TCF7L2	BIRC5	0.005	0.120	0.115	0.120	0.995	2.096	<10E-6
SYS	PTEN	DVL3	0.005	0.120	0.115	0.120	0.995	2.009	<10E-6
SYS	GDF15	MKI67	0.005	0.120	0.115	0.120	0.995	2.009	<10E-6
SYS	SLC45A3	ENY2	0.005	0.115	0.110	0.115	0.995	2.009	2 × 10E-6
SYS	MADH2	NOTCH2	0.005	0.115	0.110	0.115	0.995	2.009	<10E-6
SYS	PXN	TK1	0.005	0.115	0.110	0.115	0.995	2.089	<10E-6
SYS	CEACAM1	BRCA1	0.005	0.110	0.105	0.110	0.995	1.999	<10E-6
SYS	PAGE4	LGR4	0.005	0.110	0.105	0.110	0.995	1.999	10E-6
SYS	TPD52L2	SQLE	0.005	0.110	0.105	0.110	0.995	1.999	<10E-6
SYS	PAGE4	ST3GAL1	0.005	0.110	0.105	0.110	0.995	1.999	<10E-6
SYS	TMEM71	TPX2	0.005	0.110	0.105	0.110	0.995	1.999	10E-6

Abbreviation: RR, relative risk.

We validated the top 22 TSPs by using the three independent validation datasets (Table 3). To estimate the overall consistency between the discovery and validation sets, we looked at the correlation between proportions of positive decisions in the two sets. The discovery and validation sets tended to have a similar proportion of positive classifying decisions (i.e., X_i < X_j): the correlation coefficient was 0.44 (N = 104, P = 0.000003). We identified two pairs of genes—TPD52L2/SQLE and CEACAM1/BRCA1—that were significant in all three validation datasets. Analysis of the interactions between genes by using Pathways Studio (Ariadne, Inc, Rockville, MD, USA; http://www.ariadnegenomics.com/products/pathway-studio/) identified no interactions between the genes, which suggests that TPD52L2, SQLE, CEACAM1, and BRCA1 act independently. Figure 1 shows the classification plots for comparing the expression values of the gene pairs TPD52L2 and SQLE in the discovery and validation sets. Similar patterns can also be plotted with gene pair of CEACAM1 and BRCA1.

Table 3.

In silico validation of the 24 top-scoring pairs listed in Table 2; the two top validated pairs of genes are shown in bold

Predicted phenotype	Dataset 1	Gene1	Gene2	P(X1<X2|NED)	P(X1<X2|SYS)
SYS	GSE10645	TPD52L2	SQLE	0.005128	0.11
Meta	GSE6752	TPD52L2	SQLE	0	0.095238
HR	GSE6811	TPD52L2	SQLE	0	0.52
Metastasis	GSE6919	TPD52L2	SQLE	0.015385	0.24
SYS	GSE10645	CEACAM1	BRCA1	0.005128	0.11
Meta	GSE6752	CEACAM1	BRCA1	0	0.238095
HR	GSE6811	CEACAM1	BRCA1	0.2	0.72
Metastasis	GSE6919	CEACAM1	BRCA1	0.061538	0.64
NED	GSE10645	COL1A1	CDKN1B	0.148718	0.01
HS	GSE6811	COL1A1	CDKN1B	1	0.64
Primary	GSE6919	COL1A1	CDKN1B	0.230769	0.08
SYS	GSE10645	EDG7	LGR4	0.005128	0.125
Meta	GSE6752	EDG7	LGR4	0	0.238095
HR	GSE6811	EDG7	LGR4	0.6	0.68
SYS	GSE10645	EFS	WDR67	0.005128	0.13
Meta	GSE6752	EFS	WDR67	0	0.047619
HR	GSE6811	EFS	WDR67	0.5	0.84
Metastasis	GSE6919	EFS	WDR67	0.030769	0.12
NED	GSE10645	F5	ESR2	0.107692	0
HR	GSE6811	F5	ESR2	0.1	0.36
Primary	GSE6919	F5	ESR2	0.753846	0.68
SYS	GSE10645	GDF15	MKI67	0.005128	0.12
Meta	GSE6752	GDF15	MKI67	0	0.142857
HR	GSE6811	GDF15	MKI67	0.4	0.68
SYS	GSE10645	HLF	BLM	0	0.105
Meta	GSE6752	HLF	BLM	0.8	1
HS	GSE6811	HLF	BLM	0.9	0.84
Metastasis	GSE6919	HLF	BLM	0.507692	0.72
SYS	GSE10645	JUN	MKI67	0.005128	0.155
HR	GSE6811	JUN	MKI67	0.8	0.96
Metastasis	GSE6919	JUN	MKI67	0.015385	0.04
SYS	GSE10645	PAGE4	ANXA2	0	0.13
HS	GSE6811	PAGE4	ANXA2	0.1	0.04
Metastasis	GSE6919	PAGE4	ANXA2	0.169231	0.92
SYS	GSE10645	PAGE4	LGR4	0.005128	0.11
Meta	GSE6752	PAGE4	LGR4	0.2	0.857143
HR	GSE6811	PAGE4	LGR4	0.6	0.68
SYS	GSE10645	PAGE4	ST3GAL1	0.005128	0.11
Meta	GSE6752	PAGE4	ST3GAL1	0.2	1
HR	GSE6811	PAGE4	ST3GAL1	0.1	0.32
Metastasis	GSE6919	PAGE4	ST3GAL1	0.153846	0.96
SYS	GSE10645	PAGE4	TAF2	0.005128	0.14
Meta	GSE6752	PAGE4	TAF2	0.1	1
HR	GSE6811	PAGE4	TAF2	0.4	0.68
Metastasis	GSE6919	PAGE4	TAF2	0.353846	1
SYS	GSE10645	PASK	SQLE	0	0.105
HR	GSE6811	PASK	SQLE	0	0.84
Metastasis	GSE6919	PASK	SQLE	0.723077	0.92
SYS	GSE10645	PTEN	DVL3	0.005128	0.12
HR	GSE6811	PTEN	DVL3	0.4	0.48
Metastasis	GSE6919	PTEN	DVL3	0.984615	1
SYS	GSE10645	PTN	IMMT	0.005128	0.135
HR	GSE6811	PTN	IMMT	0.7	0.88
Metastasis	GSE6919	PTN	IMMT	0.076923	0.64
SYS	GSE10645	PXN	TK1	0.005128	0.115
Meta	GSE6752	PXN	TK1	0.9	0.952381
HR	GSE6811	PXN	TK1	0.6	0.8
Metastasis	GSE6919	PXN	TK1	0.261538	0.68
SYS	GSE10645	SLC45A3	ENY2	0.005128	0.115
HR	GSE6811	SLC45A3	ENY2	0.3	0.52
SYS	GSE10645	SNTB2	SQLE	0.005128	0.165
Meta	GSE6752	SNTB2	SQLE	0	0.952381
HR	GSE6811	SNTB2	SQLE	0.2	0.72
Metastasis	GSE6919	SNTB2	SQLE	0.266317	0.42
SYS	GSE10645	TCF7L2	BIRC5	0.005128	0.12
Meta	GSE6752	TCF7L2	BIRC5	0.8	0.952381
HR	GSE6811	TCF7L2	BIRC5	0.4	0.76
Metastasis	GSE6919	TCF7L2	BIRC5	0.230769	0.44
SYS	GSE10645	TMEM71	TPX2	0.005128	0.11
HR	GSE6811	TMEM71	TPX2	0.4	0.72

Figure 1.

Scatterplots of the expression of TPD52L2 and SQLE genes in the discovery set (GSE10645) and the three validation sets (GSE6752, GSE6811, and GSE6919). Each data point represents a patient: red dots are patients with systemic cancer progression (SYS) for the discovery dataset and corresponding counterparts for the validation datasets. The slanted lines are decision rules: above the decision line, the expression of TPD52L1 is lower than that of SQLE. NED, no evidence of disease; HRPC, hormone-refractory prostate cancer; HSPC, hormone-sensitive prostate cancer.

Double top-scoring pairs

Although the specificity for predicting SYS was > 99% for the two TSPs we identified, the sensitivity was relatively low: we could make a confident prediction for only ~10% of patients. Therefore, to increase the sensitivity, we combined two TSPs, i.e., we used double TSPs (DTSPs). The decision rule for DTSP was the condition that X_i < X_j, in which X_i is the expression level of gene i and X_j is the expression level of gene j, which was met in at least one pair of genes. Figure 2 shows the scatterplots for the combination of the two pairs, TPD52L2/SQLE and CEACAM/BRCA1. The shaded quadrants in the figure represent the decision region. Combining these two pairs of genes increased the sensitivity to 21.5% without compromising the specificity, which was 99.5%.

Figure 2.

Combination of TPD52L2/SQLE and CEACAM/BRCA1 pairs for predicting prostate tumor progression. The x axis is the log-transformed expression ratio log(CEACAM/BRCA1), and the y axis is that of log(TPD52L2/SQLE). The top right quadrant is the area where the conditions CEACAM > BRCA1 and TPD52L2 > SQLE were met; the top left quadrant is where the conditions CEACAM < BRCA1 and TPD52L2 > SQLE were met; the bottom right is where the conditions TPD52L2 > SQLE and CEACAM < BRCA1 were met; and the bottom left is where the conditions TPD52L2 < SQLE and CEACAM < BRCA1 were met. According to the decision criteria of the double top-scoring pairs analysis, the three shaded quadrants comprise the decision region for predicting systemic prostate cancer progression (SYS). NED, no evidence of disease.

The statistical significance of the DTSPs was estimated by using permutation testing. Figure 3 illustrates the approach we took and shows the distribution of the conditional SYS probability for at least one (X_i < X_j) of the two gene pairs. The estimated probability value of 0.215 was not found among 10 000 permutations, suggesting that the P value for the DTSPs TPD52L2/SQLE and CEACAM/BRCA1 for predicting SYS was less than 0.0001.

Figure 3.

Graph illustrates the permutation distribution of P (X1 < X2|SYS) for the gene pairs TPD52L2/SQLE and CEACAM/BRCA1. The arrow indicates that the estimated P value (x axis), 0.215, was not found among the 10 000 permutations (y axis).

Primary tumor vs distant metastasis

An obvious advantage of the study by Nakagawa et al.⁵ is its large sample size, > 200 individuals in each group. A disadvantage, though, is the limited number of genes analyzed—a total of 1028. Although those investigators selected genes that have been implicated in prostate cancer progression, it is probable that some good candidate genes were missed. The study by Chandran et al.⁷ was done with a smaller sample: they analyzed the gene-expression profiles of only 24 androgen-refractory metastatic samples and 64 primary tumors, but they assessed genome-wide gene expression by using Affymetrix HGU95av2, HGU95b, and HGU95c arrays (Affymetrix, Inc, Santa Clara, CA, USA). We reanalyzed the data from Chandran et al. by using all 23 572 probes from the GSE6919 dataset. The resulting TSPs for distant metastasis and primary tumors are shown in Supplementary Table S1. Three gene pairs—GRB2/ADD1, IDH3G/LARP1, and HNRNPUL1/LARP1 were perfect classifiers, with 100% sensitivity and 100% specificity. The classification plot for IDH3G/LARP1, which is representative of the three plots, is shown in Figure 4.

Figure 4.

Classification plot for discrimination between primary tumors and distant metastasis, using data from Chandran et al.⁵ This exemplary pair of genes, IDH3G/LARP1, has 100% sensitivity and 100% specificity. The x axis represents the log-expression value of IDH3G, and the y axis is that of LARP1.

Functional annotation

Functional annotation of the genes listed in Supplementary Table S1 was conducted using Ingenuity Pathways Analysis. We subdivided those genes into four groups by their relative expression: (1) high expression in the primary tumor, (2) low expression in the primary tumor, (3) high expression in the distant metastasis, and (4) low expression in the distant metastasis. Figure 5 shows the significant canonical pathways identified for each group of genes with top five P values. Among them, the most significant pathways include molecular mechanisms of cancer (group 2), insulin receptor signaling (group 2), integrin signaling (group 2), and regulation of actin-based motility by Rho (group 4).

Figure 5.

Ingenuity-defined canonical pathways for four groups of genes: (1) relatively low expression in the primary tumor (N = 463), (2) relatively high expression in the primary tumor (N = 341), (3) relatively low expression in the distant metastasis (N = 162), and (4) relatively high expression in the distant metastasis (N = 598). The numbers in parentheses in each quadrant are −log(P).

To find out whether the classifying pairs of genes form specific classifying pairs of pathways, we took the genes from each of the 17 different pathways listed in Figure 5 and identified the genes with which they made pairs. Then we looked for functional enrichment of those pairs of genes. We found that the relatively higher expression of genes associated with arginine and proline metabolism than that of genes associated with insulin receptor signaling predicts for primary tumor (P = 1.6E–7). We also found that the relatively lower expression of ILK-signaling genes than that of xenobiotic metabolism–signaling genes predicts for distant metastases. The genes in the arginine and proline metabolism, insulin receptor signaling, ILK signaling, and xenobiotic-metabolism pathways are listed in Supplementary Table S2.

Discussion

By applying the modified TSP approach to analyze the reported data from the largest retrospective study to date that has evaluated for an association between gene expression in primary tumor and disease progression after RRP, we identified two pairs of genes whose relative expression predicts such tumor progression. We validated their usefulness by using three independent datasets. The overall results of our analyses suggest that the TSP-based approach can be used to identify predictors of prostate cancer progression.

One major limitation associated with the TSP approach, however, is that a large number of tests can lead to a large number of false-positive results. That problem is expected to be most severe when the sample size is small and the number of probes is large, as they are in a typical microarray design. For example, if the number of samples in each of two groups is 10 and the number of probes is ~10 000, the expected number of perfect classifiers that can appear by chance is $47.6=\frac{C_{10000}^2}{2^{20}}$. To overcome this problem, we used the reported data from the largest published study conducted to date on prostate cancer progression.⁵ In that study, the SYS and NED groups both contained 213 individuals. This, in combination with a relatively small number of probes, contributed to the low P values estimated by our permutation testing (Table 2). For validation, we used data from studies that had much smaller sample sizes, but because we validated only the top 22 gene pairs, false-positive results were not a serious concern.

The two gene pairs we identified for use in predicting which patients will develop distant metastases of prostate cancer after undergoing RRP were TPD52L2/SQLE and CEACAM1/BRCA1. All four of those genes have some connection to prostate cancer. TPD52 reportedly regulates the migration of prostate tumor cells.²⁰ SQLE is a gene from the 8q24 region that has been strongly associated with prostate cancer risk in several genome-wide association studies.^21–23 Further, CEACAM1 plays an important role in the change from endothelial cells to an angiogenic phenotype during prostate cancer development,²⁴ and BRCA1 is regulated by the androgen receptor and often mutates in prostate tumors.²⁵

In the three studies from which we took our data for validation, the phenotypes were not exactly the same as those in the discovery set: Chandran et al.⁷ and Yu et al.¹¹ both compared gene expression in distant metastases with that in primary tumors, and Tamura et al.¹⁰ used hormone-refractory vs hormone-sensitive prostate cancer. Despite the obvious differences in those analyzed phenotypes, we believe that it was reasonable to use them for validation because hormone-refractory tumors and distant metastases are likely to have a common gene-expression signature.

Whether the genes from the TSPs belong to specific pathways or have specific biologic functions is an interesting question. It was problematic to use genes from the study of Nakagawa et al.⁵ for functional annotation because those genes were selected on the basis of published evidence of their involvement in prostate cancer progression. Thus, we used the GSE6911 dataset and separately annotated four groups, as described in Results. We found that lower expression of genes from arginine and proline metabolism relative to the expression of genes from the insulin receptor signaling pathway may used to classify primary tumors vs distant metastases. Arginine is one of the most versatile amino acids in animal cells, serving as a precursor for the synthesis of proteins and nitric oxide (NO). NO has an important function in cancer, including prostate cancer, in that increased NO generation contributes to angiogenesis by up-regulating vascular endothelial growth factor.²⁶ In addition, a high concentration of NO is associated with chronic inflammation and tumorigenesis. In prostate cancer, NO is involved in inhibiting androgen receptor activity and is associated with disease progression.^{27, 28} Our analyses also showed that the arginine and proline metabolism pathway may be coupled with the insulin receptor signaling pathway to classify primary prostate tumors. Insulin receptor signaling also plays an important role in prostate tumorigenesis. These two signaling pathways seem to be independent: analysis of the relationship between the genes from those pathways using all possible types of interactions identified no cross talk between them (data not shown). The pathways paired in predicting for distant metastasis were the xenobiotic metabolism and integrin-linked kinase (ILK) signaling pathways. Some experimental evidence exists that both of those pathways may also be associated with prostate cancer.^{29, 30}

In summary, the TSP approach can be used to identify patients whose prostate cancer will progress after they undergo radical prostatectomy. Two gene pairs can predict which men would experience progression to the metastatic form of the disease. However, because our analysis was based on a relatively small number of genes, a larger study will be needed to identify the best predictors of disease outcome overall.