Figures
Abstract
Molecular analysis of patient tissue samples is essential to characterize the in vivo variability in human cancers which are not accessible in cell-lines or animal models. This applies particularly to studies of tumor metabolism. The challenge is, however, the complex mixture of various tissue types within each sample, such as benign epithelium, stroma and cancer tissue, which can introduce systematic biases when cancers are compared to normal samples. In this study we apply a simple strategy to remove such biases using sample selections where the average content of stroma tissue is balanced between the sample groups. The strategy is applied to a prostate cancer patient cohort where data from MR spectroscopy and gene expression have been collected from and integrated on the exact same tissue samples. We reveal in vivo changes in cancer-relevant metabolic pathways which are otherwise hidden in the data due to tissue confounding. In particular, lowered levels of putrescine are connected to increased expression of SRM, reduced levels of citrate are attributed to upregulation of genes promoting fatty acid synthesis, and increased succinate levels coincide with reduced expression of SUCLA2 and SDHD. In addition, the strategy also highlights important metabolic differences between the stroma, epithelium and prostate cancer. These results show that important in vivo metabolic features of cancer can be revealed from patient data only if the heterogeneous tissue composition is properly accounted for in the analysis.
Citation: Tessem M-B, Bertilsson H, Angelsen A, Bathen TF, Drabløs F, Rye MB (2016) A Balanced Tissue Composition Reveals New Metabolic and Gene Expression Markers in Prostate Cancer. PLoS ONE 11(4): e0153727. https://doi.org/10.1371/journal.pone.0153727
Editor: Craig N. Robson, Northern Institute for Cancer Research, UNITED KINGDOM
Received: January 26, 2016; Accepted: April 1, 2016; Published: April 21, 2016
Copyright: © 2016 Tessem et al. This is an open access article distributed under the terms of the Creative Commons Attribution License, which permits unrestricted use, distribution, and reproduction in any medium, provided the original author and source are credited.
Data Availability: Microarray data used in this study were obtained from Array Express, accession E-MTAB-1041 and Gene Expression Omnibus, accession GSE8218. Additional relevant data are within the paper and its Supporting Information files.
Funding: This work was funded by research grants to MBR and MBT from the Liaison Committee between the Central Norway Regional Health Authority (RHA) and the Norwegian University of Science and Technology (NTNU). The funders had no role in study design, data collection and analysis, decision to publish, or preparation of the manuscript.
Competing interests: The authors of this manuscript have the following competing interests: Tone Bathen is an Academic Editor for PLOS ONE. This does not alter the authors' adherence to all the PLOS ONE policies on sharing data and materials.
Introduction
Cancer is a heterogeneous disease where the understanding of underlying molecular mechanisms characteristic for each individual type could be important to provide efficient personalized and targeted therapy or choice of treatment. Human tissue samples generate a molecular snapshot of tumor states which is essential for characterization of in vivo cancer heterogeneity that is lost in animal models and cell-lines [1]. This applies to studies of tumor metabolism in particular [2]. A common but complicating factor when analyzing human tissue samples is the heterogeneous mixture of various cells and tissue types which can confound results from molecular analysis. This risk of confounding applies not only to prostate cancer (PCa) tissue, but to most cancers tissue types.
A simplistic model divides the human prostate tissue into three different components; benign epithelium, stroma and cancer tissue. Based on this model, a differential molecular analysis between cancer and normal samples should ideally emphasize differences between the benign epithelium and the PCa tissue. However, in such differential studies, normal samples are composed of two tissue types (benign epithelium and stroma) while PCa samples consist of three tissue types (benign epithelium, stroma and PCa), all in various proportions. This introduces a systematic sample bias of increased stroma content in the normal samples which confounds the molecular differences between epithelium and cancer. This problem is generally acknowledged [3,4]. However, thorough assessment of the composition of these three tissue types is not routinely accounted for during sample harvesting, preparation and analysis of PCa tissue samples. Additionally, the tumor environment introduces changes to the surrounding stroma tissue, termed stromogenic cancer [5,6], which represent a fourth tissue component in prostate cancer samples, but was not considered in this study.
Previously presented strategies to handle sample tissue heterogeneity have generally used computational models to adjust each sample for the influence of various tissue types in differential analysis [7,8]. Such computational strategies can either require prior information on tissue composition [9–11], pre-defined gene signatures [12,13], or be purely data driven [14,15]. Computational models usually handle tissue heterogeneity by making adjustment to the original expression values prior to differential analysis. This increase the risk of introducing a model bias, especially when the signal from each tissue component is not homogeneous. This is the case for tissue samples from many cancers, including PCa [16–19].
In this study we apply an alternative and simpler approach to account for the confounding effect of stroma when comparing PCa samples with normal sample histology for differential analysis. The strategy is to construct datasets of cancer and normal samples where the average content of stroma is balanced between the two datasets. The intention is to attenuate the differential signal due to various amounts of stroma in the analysis, and emphasize the true molecular differences between cancer and normal tissue. Unlike most computational methods used for tissue heterogeneity adjustments, the strategy does not perform corrections to the original molecular expression values. However, pathological characterization of tissue composition (benign epithelium, stroma and PCa) for all samples is required.
In this study, fresh frozen prostate tissue slices were collected after radical prostatectomy and each tissue core sample was histopathologically evaluated for tissue composition prior to analysis [20]. The tissue was then analyzed by HR-MAS (high resolution magic angle spinning) MRS (magnetic resonance spectroscopy) allowing 23 quantified metabolites, followed by gene expression measurements by microarray. HR-MAS is a non-destructive method, which permits gene expression analysis to be performed on the exact same tissue sample [21]. This integration of gene and metabolite data offers a unique opportunity to investigate central molecular pathways in PCa.
This study illustrates that variations in stromal tissue composition can be a systematic confounding factor in molecular analysis when PCa and normal tissue samples are compared. This bias can be removed by balancing the stromal composition between the sample sets. Only when such tissue biases are accounted for, integrated differential changes in genes and metabolites in central metabolic pathways can be simultaneously revealed.
Methods
Sample collection, microarray and HR-MAS analysis
Samples were obtained using a standardized harvesting procedure previously described by Bertilsson et al. [20]. This procedure includes cutting and handling of frozen tissue slices removed from the prostate gland during prostatectomy. Tissue sample cores were carefully selected from the slices based on clinical histopathology. Additionally, histopathological assessment of the amount of cancer, stroma and benign epithelium was carried out on each sample prior to the HR-MAS and microarray analyses (S1 File). Microarray and HR-MAS protocols and analysis are thoroughly described previously in Bertilsson et al. and Giskeødegård et al., respectively [21,22]. The complete dataset (95 PCa samples and 34 normal samples) contained microarray expression measurements for 16,381 genes, and HR MAS metabolite concentrations for 23 different metabolites measured on the exact same tissue samples. The use of human tissue material was approved by the Regional Committee for Medical and Health Research Ethics approval no 4-2007-1890. Consent forms were obtained from all patients. Other ethical aspects regarding the specific samples used in this study have been described in previous publications [20–22]. The microarray gene experiment is published in Array Express with accession E-MTAB-1041. Metabolite concentrations measured on each samples is given in S2 File.
Controlling the influence of stroma tissue by sorting the original complete data into, balanced, unbalanced and unstratified datasets
The balanced, unbalanced and unstratified datasets were created by the following procedure: Cancer and normal samples were sorted independently according to their histopathologically determined content of stroma. The balanced dataset was created by splitting the sorted cancer and normal samples in two halves, and then selecting the 47 PCa samples with the highest stroma content (top half) and the 17 normal samples with the lowest stroma content (bottom half) from the sorted sample lists. This procedure minimizes the difference in average stroma content between PCa and normal samples for the balanced dataset (Fig 1A). Likewise, the unbalanced dataset was created by selecting the 48 PCa samples with the lowest stroma content and the 17 normal samples with the highest stroma content. In the unbalanced dataset, the difference in average stroma content between the cancer and normal samples is maximized (Fig 1A). The separation of samples into balanced and unbalanced datasets is given in S1 File. To create an unstratified dataset with the same statistical power as the balanced and unbalanced datasets, 50 random selections of 47 cancer and 17 normal samples were drawn from the complete dataset.
(A) By canceling out the same types of tissues in PCa and normal samples, the balanced dataset compares in average 51% cancer to 43% benign epithelium and 8% stroma. Likewise, the unbalanced dataset compares 75% cancer to 58% stroma and 17% benign epithelium, while the complete and unstratified datasets compares 63% cancer to 33% stroma and 30% benign epithelium. The balanced, unbalanced and unstratified dataset have the same statistical power. Our strategy predicts that differentially expressed genes identified in the balanced dataset set should reflect changes in PCa rather than differences in tissue composition. (B) Histogram of tissue percentage distributions for cancer, stroma and benign epithelium in the balanced and unbalanced datasets. (C) Percentage of DEGs shared between the balanced and unbalanced datasets compared to the average percentage of DEGs shared between the 50 random subsets in the unstratified dataset. Percentages of shared genes are calculated using increasing numbers of the most significant DEGs in each dataset. The random numbers were calculated as the average number of genes shared over all 1225 possible comparisons between the 50 random subsets. For the 200 most significant DEGs, 20% were shared between the balanced and unbalanced datasets, and 39% were shared for the 2000 most significant DEGs. The corresponding percentages for the confounded dataset were 65% and 73% respectively. (D) Microarray probes with a p-value changing in various orders of magnitude between the balanced and unbalanced datasets greatly exceeds p-value changes expected by chance in the unstratified dataset. Of all probes, 2830 showed a p-value fold change of more than four orders of magnitude, and 1460 changed their p-values by more than six orders of magnitude between the balanced and unbalanced datasets compared to 42 and 1 probes for the unstratified dataset respectively.
Differentially expressed genes and metabolites
Raw expression values were filtered, log2 transformed and quantile normalized using the limma R package [23] as described in Bertilsson et al. [21]. All operations were performed prior to sample separations into balanced, unbalanced and unstratified datasets. Differentially expressed genes were identified for the complete, unstratified, balanced and unbalanced datasets as described in Bertilsson et al. [21]. P-values were corrected for multiple testing using Benjamini-Hochberg false discovery rate [24]. For the unstratified dataset, the average p-value over the 50 datasets was used. P-values for differentially expressed metabolites were calculated by the t.test function in R. The t.test function initially gives an assessment of whether the two sample groups display equal variance, which is used to decide the optimal significance test. Using the initial variance equality assessment, we used the t.test function with equal variance if the p-value of the initial test was below 0.05, and the test for unequal variance otherwise.
Validation data
To validate the strategy of tissue balancing, an independent dataset [11] were downloaded from Gene Expression Omnibus (accession GEO ID, GSE8218). This data set consisted of microarray gene expression from 65 PCa samples and 71 normal samples together with histopathologic evaluation of four different tissue components (prostate cancer, stroma, epithelium from BPH and atrophic gland). To compare tissue composition in the validation set with the main dataset (the dataset used in the current study), percentages of epithelium from BPH and atrophic gland were combined into a single component corresponding to benign epithelium in the main dataset. Dataset balancing and identification of differentially expressed genes were performed in the same way as for the main data. To compare the similarity of differentially expressed genes highlighted in the balanced and unbalanced datasets between the validation and main data, we developed a Significance Ratio (SR) score calculated for each gene:
(1)
for the balanced dataset and
(2)
for the unbalanced dataset, where pB and pS are the p-value in the balanced and unbalanced datasets, respectively. The score is introduced to compare the relative p-values for each gene, and is calculated independently for the validation and the main data. Genes are ranked in ascending order based on the SR score for the balanced and unbalanced datasets in both the validation and main data. A high number of shared genes among the top-ranked genes in the main and validation data indicate that the compared lists have similar relative alterations in p-values between the balanced and unbalanced dataset.
Results and Discussion
Balancing the content of stroma tissue when human PCa tissue samples are compared to samples from normal prostate tissue
Samples from the original complete dataset are presented as three differently arranged subsets, balanced, stroma and unstratified, based on the difference in average stroma content between PCa and histopathologically normal samples (Fig 1A, Methods). The balanced dataset was designed to have an average stroma content as equal as possible between PCa and normal samples (37% vs 45% respectively). This was to emphasize the molecular transformations between cancer and normal tissue, without the confounding effect due to different amounts of stroma. In contrast, the unbalanced dataset was created with a maximized difference in average stroma content (14% vs 72% for PCa and normal samples respectively) to emphasize the molecular variations observed when the amount of stroma is not accounted for. A unstratified dataset was created with the same tissue properties as the complete dataset, but with a reduced number of samples to permit direct comparison of p-values with the balanced and unbalanced datasets. The unstratified dataset had an intermediate difference in average stroma content (26% vs 59% for PCa and normal samples respectively), representing the tissue composition in samples from a typical patient cohort. There was no significant difference in average Gleason score between the PCa samples in the balanced, unbalanced and unstratified datasets (7.17, 7.27 and 7.22 respectively). To simplify interpretation of the results, we assume that molecular signals from equal amounts of the same tissue type in PCa and normal samples cancel each other out in a differential analysis. This analysis does not take into account the changes between healthy stroma and stromogenic cancer. However, it will serve as a good approximation to study confounding signals from stroma due to different average tissue compositions in the cancer and normal samples. If we remove the contributions from equal amounts of tissue, analysis of the complete and unstratified datasets compare the molecular signal from 63% cancer tissue in the PCa samples to 33% stroma and 30% benign epithelium in the normal samples, which represents an almost complete confounding between stroma and benign epithelium. In this setting it is impossible to conclude whether the observed differentially expressed genes (DEGs) and metabolites are due to changes between PCa and normal tissue, or due to different amounts of stroma in PCa and normal sample groups. In contrast, the observed DEGs and metabolites are directly attributable to molecular changes between cancer and normal tissue for the balanced dataset, and cancer and stroma tissue in the unbalanced dataset. The differential signals in these two dataset comparisons thus highlight genes and metabolites, as well as the relationships between them, which are otherwise hidden in the complete and unstratified datasets.
Two criteria were used to divide samples between the balanced and unbalanced datasets: i) the average percentage of stroma should be as equal as possible between cancer and normal samples, and ii) each dataset should include as many samples as possible to ensure maximum statistical power. To improve the analysis even further, a third criterion can be introduced, which ensures that individual samples in each dataset are homogenous with respect to the percentage of stroma. This criterion should, in theory, reduce the expression value variance within each group to improve differential statistics. Due to the limited number of available samples with similar tissue composition, it was concluded that a selection also including the third criterion would compromise the statistical power of the analysis. Nevertheless, the first two criteria proved to be sufficient to highlight important differences between the balanced and unbalanced datasets.
An important feature of the balanced and the unbalanced datasets is that they indirectly also assess the difference between benign epithelium and stroma tissue. Genes and metabolites displaying differential change only in the unbalanced dataset, and not in the balanced dataset, are not markers for PCa transformation. These will rather be markers of stroma that changes in the complete and unstratified datasets only due to differences in the amount of stroma tissue. Further, genes and metabolites which display similar changes in both the balanced and the unbalanced dataset are markers for PCa which is unaffected by the presence of stroma. Finally, genes and metabolites displaying differential changes only in the balanced data set, but not in the unbalanced dataset, represent markers for PCa which are confounded by the presence of stroma. These assessments are important when separating observed changes directly related to PCa transformation from changes resulting only from biased amounts of stroma tissue. Such interpretations will be performed in the subsequent analysis.
Balanced and unbalanced datasets display different patterns of gene expression at the global level
Considerable differences in gene expression patterns were observed when comparing DEGs between the balanced and unbalanced datasets. A far lower percentage of DEGs were shared between the balanced and the unbalanced datasets when compared to the percentage of DEGs expected to occur by chance (Fig 1B), where the expected percentage was estimated as the average number of shared genes between the 50 random subsets in the unstratified dataset. The fact that the balanced and unbalanced datasets capture significant biological differences which are unlikely to be observed by chance was confirmed by the high number of gene probes with a p-value changing by several orders of magnitude between the balanced and the unbalanced datasets compared to the unstratified dataset (Fig 1C). All three datasets displayed expected up and downregulation of seven well validated genes in PCa (Figure C in S3 File). To conclude, the observed differences between these datasets are attributable to differences in the amount of stroma tissue. The applied strategy of tissue balancing is able to highlight these differences.
The balanced dataset highlights metabolites related to PCa change
In the balanced dataset 17 of the 23 metabolites measured by HR-MAS were found to display significant changes in concentration levels between PCa and normal tissue (Fig 2). This is a considerably higher number compared to the 8 significant metabolites identified in the unstratified and 13 identified in the complete dataset. This indicates a high degree of confounding from the stroma tissue in these datasets, which specifically applies to the nine metabolites putrescine, spermine, glycine, glutamine, alanine, valine, succinate, isoleucine and citrate that were significant in the balanced but not the unstratified dataset. These nine metabolites are thus changes which are characteristic for PCa but are difficult to detect due to confounding from unbalanced amounts of stroma tissue. The identification of citrate serves as a proof-of-principle for the applied analysis strategy, due to the absence of citrate in stroma and reduced levels observed in PCa tissue [22,25]. Additionally, changes in the metabolites succinate, valine and glycine were not detectable in the complete dataset, despite the higher statistical power compared to the balanced set. No metabolites were specific for the unbalanced dataset, possibly with an exception of taurine with borderline significance (p = 0.053). Other typical metabolite biomarkers for PCa like glucose, lactate and the three choline-containing metabolites [22,26] were observed in all datasets, thus showing stability as biomarkers regardless of tissue composition.
Positive -log10(p-value) indicate increased concentration and negative -log10(p-value) indicate reduced concentration in PCa. The unstratified, balanced and unbalanced datasets all have the same statistical power.
Integration of metabolite and gene expression data pinpoints changes in genes involved in PCa metabolism
We predict that performing differential expression analysis in a dataset balanced for the content of stroma tissue between PCa and normal samples highlights genes and metabolites directly relevant for PCa changes. Moreover, the simultaneous gene and metabolite data measured on the exact same samples facilitate the integration of these data on specific metabolic pathways. In this study we focused on pathways of the TCA cycle, polyamine synthesis and fatty acid synthesis, and on DEGs related to the metabolites putrescine, citrate and succinate, as these pathways were specifically highlighted in the balanced dataset. Differential expression is presented in terms of p-values. Corresponding changes with respect to gene rank and fold change are presented in Figures A and B in S3 File and in S4 File.
Reduced putrescine levels relates to increased expression of SRM in the polyamine pathway.
Increased production of polyamines is important for cancer progression by promoting cell growth, degrading surrounding tissue and diminishing anti-tumor immune functions [27]. In the polyamine pathway (Fig 3A), ODC1 converts ornithine to putrescine, which is further converted to spermidine by SRM. Spermidine is then converted to spermine by SMS, all in a forward reaction. SRM and SMS are assisted by AMD1 in the conversion process, while SAT1 and SMOX are responsible for the further downstream fate of spermidine and spermine [28]. Previous studies have reported that genes in the polyamine pathway are upregulated in cancer [29,30]. A concordant upregulation of polyamine genes will increase the flux of all polyamines, but not necessarily change the relative levels of individual metabolites. In the balanced dataset, we observe a significant level reduction for the metabolite putrescine. This reduction coincides with a specific upregulation of the SRM gene (Fig 4A). Mechanistically, upregulation of SRM will consume putrescine for the production of spermidine, and without putrescine replacement from ornitihine by upregulation of ODC1, the concentration of putrescine should be reduced. This is exactly what is observed from the MR metabolic measurements. Upregulation of SRM without concordant upregulation of other genes in the polyamine pathway is observed only in the balanced dataset, while a general upregulation of polyamine genes is a feature observed mostly in the unbalanced dataset. Applying the indirect relationship between the balanced and unbalanced dataset, this general upregulation are most likely caused by differences between benign epithelium and stroma, and the upregulation of SRM and concordant putrescine decline is due to PCa transformation. Interestingly, SRM has recently been suggested as a drug target for the polyamine pathway in a model of B-cell lymphomas [31]. In addition, a significant decrease in spermine levels (p = 0.047) is also observed, which fits with a significant upregulation of SAT1 and SMOX, without significant upregulation of SMS. However, this relation is more subtle. A significant difference in spermine concentrations was previously reported to separate aggressive (Gleason grade ≥7) from more indolent types of PCa tissue (Gleason grade = 6) [22]. However, in this comparison the confounding of stroma is less pronounced since only cancer samples are compared.
(A) Polyamine pathway. (B) TCA cycle, fatty acid synthesis and glutamine consumption. The genes up- and downregulated in the balanced dataset are highlighted in red and blue respectively. Gene names: ODC1: ornithine decarboxylase 1, SRM: spermidine synthase, SMS: spermine synthase, AMD1: adenosylmethionine decarboxylase 1, SAT1: spermidine/spermine N1-acetyltransferase 1, SMOX: spermine oxidase, ACO1/2: aconitase 1/2, CS: citrate synthase, ACLY: ATP citrate lyase, ACACA/B: acetyl-CoA carboxylase alpha/beta, FASN: fatty acid synthase, SUCLA2: succinate-CoA ligase ADP-forming beta subunit, SUCLG1/2: succinate-CoA ligase beta subunit, SDHA/B/C/D: succinate dehydrogenase complex subunit A/B/C/D.
(A) Top: SRM and putrescine in the polyamine pathway. Middle: SUCLA2 and SDHD and succinate in the TCA cycle. Bottom: Citrate in fatty acid synthesis. Positive -log10(p-value) indicate upregulation and negative -log10(p-value) indicate downregulation. The metabolites displayed on the left side are a subset of the metabolites shown in Fig 2. (B) Validation of significant genes in PCa compared to normal samples with respect to the polyamine pathway, succinate in the TCA cycle and fatty acid synthesis in an independent dataset. (C) Average tissue compositions in the balanced, unbalanced and complete/unstratified datasets from the validation study. Histogram of tissue distributions are shown in Figure D in S3 File. (D) Number of shared genes among the N first genes sorted by the significance ratio score when various datasets are compared. Genes with opposite changes (3 in the balanced and 116 in the unbalanced datasets) were removed from analysis. The number shared genes are much higher when balanced and unbalanced datasets are compared between themselves, than when balanced are compared with unbalanced datasets. This indicates that the validation and control studies display a highly concordant ranking of genes for both the balanced and unbalanced datasets.
Reduced citrate levels relates to increased fatty acid synthesis.
Significantly higher levels of citrate in normal samples were only detectable in the balanced dataset. In the complete dataset, where the statistical power is doubled, citrate levels did not reach statistical significance between cancer and normal samples. Citrate is normally converted to isocitrate by ACO1 and ACO2 in the TCA cycle (Fig 3B), which is the preferred mode of energy production in most cells. The normal prostate deviates from this mode, because the conversion from citrate to isocitrate is partly blocked, leading to accumulation of citrate. Accumulated citrate can be secreted, or converted to acetyl-CoA and utilized for fatty acid synthesis (Fig 3B). Secretion leads to increased detectable levels of citrate, while higher TCA cycle activity and/or fatty acid synthesis reduces citrate levels. Our data supports a model where high level of citrate is a hallmark of normal prostate epithelial cells, and that citrate levels are reduced in both stroma and cancer cells.
The most pronounced gene differences related to citrate metabolism and fatty acid synthesis is found in the unbalanced dataset (Fig 4A). We therefore detect more gene differences between cancer and stroma than between cancer and benign epithelium. This result is in accordance with general TCA cycle activity in the healthy stroma [32], while TCA cycle function is altered in both benign epithelium and PCa. Concordantly, we observe a highly significant downregulation of ACO1, upregulation of fatty acid synthesis genes ACLY, ACACA and FASN, and downregulation of fatty acid oxidation gene ACACB in the unbalanced dataset where PCa are compared to the stroma. In the balanced dataset, changes in gene expression are more subtle. We observe significant downregulation of ACO1, ACO2 and ACACB, and upregulation of fatty acid synthesis gene ACACA, but not ACLY and FASN. Altogether, this indicates that increased fatty acid synthesis is a hallmark feature for both the benign epithelium and PCa.
In the transformation to PCa (the balanced dataset) fatty acid synthesis can be further activated by increased expression of ACLY, FASN and/or ACACA (the rate limiting enzyme of fatty acid synthesis), leading to increased consumption of accumulated citrate. Consumption of citrate for fatty acid synthesis, rather than restoration of normal TCA cycle function may be the main reason for citrate depletion, since downregulation of both ACO1 and ACO2 in the balanced dataset corroborates decreased forward TCA cycle. This conclusion is in agreement with other studies highlighting increased fatty acid synthesis as a common feature in PCa [33]. However, compared to these studies, ACACA rather than FASN [33–35] is here highlighted as the main source of increased fatty acid synthesis. Inhibition of ACACA in PCa cell-lines induced growth arrest and apoptosis [36,37]. ACACA has also been shown to be a target for androgen receptor AR [38], and may be a drug target in PCa that directly targets in vivo changes in fatty acid synthesis.
Increased succinate levels relates to changed expression of SUCLA2 and SDHD.
Increased succinate levels were only observed in the balanced dataset. In the forward TCA cycle, the genes SUCLG1, SUCLG2 and SUCLA2 participate in the reversible conversion of succinyl-CoA to succinate, which is further converted to fumarate by the four succinate dehydrogenases SDHA/B/C/D (Fig 3B). Downregulation of these genes is most pronounced in the unbalanced dataset, especially for SUCLA2, SUCLG1, SDHA and SDHD. The cancer tissue is therefore more different from stroma than from benign epithelium, and this study indicates that stroma cells harbor canonical TCA cycle metabolism whereas benign epithelium and PCa do not. In benign epithelium and PCa these genes are downregulated, due to impaired TCA cycle and possibly increased fatty acid synthesis (Fig 4A). In the balanced dataset, downregulation of SUCLA2 was particularly pronounced. Downregulation of SUCLA2 without concordant downregulation of the related enzymes may disturb the metabolic flux balance, leading to the observed increased levels of succinate in PCa. However, since the conversion of succinyl-CoA to succinate is a reversible reaction, the actual effect of SUCLA2 expression will depend on additional factors such as the equilibrium state between succinate and succinyl-CoA. A decreased, but less pronounced, expression of SDHD is also observed in the balanced dataset. In contrast to SUCLA2, conversion of succinate to fumarate by SDHD is irreversible, and downregulation of this enzyme will lead to accumulation of succinate in PCa. Earlier studies have suggested a tumor promoting function of increased succinate levels through activation of hypoxic regulatory programs by the transcription factor HIF1A, but did not mention PCa explicitly [39]. It is thus concluded that this mechanism, as well as other possible functions of increased succinate levels has yet to be explored for PCa.
Validation in an independent cohort
Currently there exist only one additional dataset for PCa where the tissue composition of cancer, stroma and benign epithelium are characterized in both PCa and normal samples [11]. This enables a direct transfer of the balancing strategy applied in this study. Gene expression using microarrays were measured on 65 PCa and 71 normal tissue samples in this dataset, however no measurements of metabolism was collected and only validation of gene expression was possible. To our knowledge, there are no other study which simultaneously measure gene expression and metabolism in PCa. In both studies, histopathology showed the same bias of increased stroma tissue in the normal samples (Fig 4C). Using the 9723 genes with measurements in both datasets, concordant variations in differentially expressed genes (ranked according to the SR-score) were identified in the balanced and the unbalanced datasets (Fig 4D, Methods).
At the pathway level, we observed similar patterns for both the polyamine and fatty acid synthesis in the balanced and unbalanced datasets of both studies (Fig 4B). At the individual gene level, the specific cancer-related increase in expression of SRM in the polyamine pathway, as well as a general activation of all pathway genes in epithelium compared to stroma, was confirmed. In the fatty acid synthesis pathway, the validation analysis highlights FASN and ACLY as significantly upregulated, but not ACACA (p = 0.13). This supports the notion of increased fatty acid synthesis in PCa, but also indicates that this activity can be caused by the activation of various key genes in this pathway. Both datasets strongly confirmed increased fatty acid synthesis and downregulation of ACO1 as a feature mostly characteristic of PCa when compared to stroma, while the transformation from epithelium to PCa involves more subtle changes. Succinate and the TCA-cycle showed less similarity and lower p-values for the involved genes. The expression of succinate genes also showed a comparable expression pattern in the validation data, though SUCLA2 and SDHD were only borderline significant (p = 0.054 and p = 0.066 respectively).
Generally, the number of differentially expressed genes for the balanced dataset were fewer in the validation data compared to the main data (1011 versus 2800 respectively), and the p-values were less significant. This is likely attributed to the higher average content of normal stroma in the validation data (74% and 46% in validation and main data, respectively), which attenuates the differential signal. Overall, the validation datasets generally confirm the findings in the main data, and thus supports the viability of the proposed strategy to analyze tissue sample collection with large variations in tissue compositions.
An advantage of the applied strategy of removing stroma confounding the reduced amount of samples needed to obtain similar results of differentially expressed genes. To illustrate this, we performed differential expression analysis on the complete dataset using the histopathological analysis as contrasts in a factorial design (S4 File). When PCa were compared to benign epithelium, p-values in this analysis were highly comparable to the p-values in the balanced analysis. Since the influence of confounding tissue is under control in the balanced dataset, comparable p-values could be obtained by using only half the number of samples.
Conclusion
A strategy to balance the original dataset based on pathology-reviewed tissue composition has been applied. Separation of changes between PCa and normal tissue, from changes between PCa and stroma, makes it possible to identify relevant DEGs and metabolites characteristic for each tissue component. A unique integration of gene expression and metabolite data on the exact same tissue sample highlights new information on important cancer pathways which can be possible future drug targets. The presented strategy is transferable to all heterogeneous cancer tissue types and similar results may be expected. The results from balanced and unbalanced datasets also show that adding more samples to a cohort study will not improve differential expression analysis unless the sample groups have a balanced composition of confounding tissue.
Supporting Information
S1 File. Sample metadata.
Sample histopathology and assignment to balanced and unbalanced datasets.
https://doi.org/10.1371/journal.pone.0153727.s001
(XLSX)
S2 File. Metabolite concentrations.
Metabolite concentrations measured by HR-MAS for each sample.
https://doi.org/10.1371/journal.pone.0153727.s002
(XLSX)
S3 File. Supplementary Figures A-D.
Supplementary figures referenced in the manuscript, and figure captions.
https://doi.org/10.1371/journal.pone.0153727.s003
(PDF)
S4 File. Extended gene analysis table.
Table of gene p-values, fold changes and gene ranks for the genes analyzed.
https://doi.org/10.1371/journal.pone.0153727.s004
(XLSX)
Acknowledgments
This work was funded by the Liaison Committee between the Central Norway Regional Health Authority (RHA) and the Norwegian University of Science and Technology (NTNU). The authors would also like to acknowledge Maria Karoline Andersen for assisting in the layout of Fig 3.
Author Contributions
Conceived and designed the experiments: HB MBT MBR TFB AA. Performed the experiments: HB MBT. Analyzed the data: MBR MBT FD. Contributed reagents/materials/analysis tools: MBT HB. Wrote the paper: MBT MBR HB TFB FD AA.
References
- 1. Valkenburg KC, Williams BO (2011) Mouse models of prostate cancer. Prostate Cancer 2011: 895238. pmid:22111002
- 2. Keshari KR, Sriram R, Van Criekinge M, Wilson DM, Wang ZJ, et al. (2013) Metabolic reprogramming and validation of hyperpolarized 13C lactate as a prostate cancer biomarker using a human prostate tissue slice culture bioreactor. Prostate 73: 1171–1181. pmid:23532911
- 3. Liotta L, Petricoin E (2000) Molecular profiling of human cancer. Nat Rev Genet 1: 48–56. pmid:11262874
- 4. Tomlins SA, Mehra R, Rhodes DR, Cao XH, Wang L, et al. (2007) Integrative molecular concept modeling of prostate cancer progression. Nature Genetics 39: 41–51. pmid:17173048
- 5. Barron DA, Rowley DR (2012) The reactive stroma microenvironment and prostate cancer progression. Endocr Relat Cancer 19: R187–204. pmid:22930558
- 6. Jia Z, Wang Y, Sawyers A, Yao H, Rahmatpanah F, et al. (2011) Diagnosis of prostate cancer using differentially expressed genes in stroma. Cancer Res 71: 2476–2487. pmid:21459804
- 7. Zhao Y, Simon R (2010) Gene expression deconvolution in clinical samples. Genome Med 2: 93. pmid:21211069
- 8. Venet D, Pecasse F, Maenhaut C, Bersini H (2001) Separation of samples into their constituents using gene expression data. Bioinformatics 17 Suppl 1: S279–287. pmid:11473019
- 9. Stuart RO, Wachsman W, Berry CC, Wang-Rodriguez J, Wasserman L, et al. (2004) In silico dissection of cell-type-associated patterns of gene expression in prostate cancer. Proc Natl Acad Sci U S A 101: 615–620. pmid:14722351
- 10. Shen-Orr SS, Tibshirani R, Khatri P, Bodian DL, Staedtler F, et al. (2010) Cell type-specific gene expression differences in complex tissues. Nat Methods 7: 287–289. pmid:20208531
- 11. Wang Y, Xia XQ, Jia Z, Sawyers A, Yao H, et al. (2010) In silico estimates of tissue components in surgical samples based on expression profiling data. Cancer Res 70: 6448–6455. pmid:20663908
- 12. Zhong Y, Wan YW, Pang K, Chow LM, Liu Z (2013) Digital sorting of complex tissues for cell type-specific gene expression profiles. Bmc Bioinformatics 14: 89. pmid:23497278
- 13. Wang M, Master SR, Chodosh LA (2006) Computational expression deconvolution in a complex mammalian organ. Bmc Bioinformatics 7: 328. pmid:16817968
- 14. Lahdesmaki H, Shmulevich L, Dunmire V, Yli-Harja O, Zhang W (2005) In silico microdissection of microarray data from heterogeneous cell populations. Bmc Bioinformatics 6: 54. pmid:15766384
- 15. Repsilber D, Kern S, Telaar A, Walzl G, Black GF, et al. (2010) Biomarker discovery in heterogeneous tissue samples -taking the in-silico deconfounding approach. Bmc Bioinformatics 11: 27. pmid:20070912
- 16. Lapointe J, Li C, Higgins JP, van de Rijn M, Bair E, et al. (2004) Gene expression profiling identifies clinically relevant subtypes of prostate cancer. Proc Natl Acad Sci U S A 101: 811–816. pmid:14711987
- 17. Markert EK, Mizuno H, Vazquez A, Levine AJ (2011) Molecular classification of prostate cancer using curated expression signatures. Proc Natl Acad Sci U S A 108: 21276–21281. pmid:22123976
- 18. Taylor BS, Schultz N, Hieronymus H, Gopalan A, Xiao Y, et al. (2010) Integrative genomic profiling of human prostate cancer. Cancer Cell 18: 11–22. pmid:20579941
- 19. Schoenborn JR, Nelson P, Fang M (2013) Genomic profiling defines subtypes of prostate cancer with the potential for therapeutic stratification. Clinical Cancer Research 19: 4058–4066. pmid:23704282
- 20. Bertilsson H, Angelsen A, Viset T, Skogseth H, Tessem MB, et al. (2011) A New Method to Provide a Fresh Frozen Prostate Slice Suitable for Gene Expression Study and MR Spectroscopy. Prostate 71: 461–469. pmid:20860008
- 21. Bertilsson H, Tessem MB, Flatberg A, Viset T, Gribbestad I, et al. (2012) Changes in Gene Transcription Underlying the Aberrant Citrate and Choline Metabolism in Human Prostate Cancer Samples. Clinical Cancer Research 18: 3261–3269. pmid:22510345
- 22. Giskeodegard GF, Bertilsson H, Selnaes KM, Wright AJ, Bathen TF, et al. (2013) Spermine and citrate as metabolic biomarkers for assessing prostate cancer aggressiveness. PLoS One 8: e62375. pmid:23626811
- 23.
Smyth GK (2005) Limma: linear models for microarray data. In: Gentleman R, Carey V, Dudoit S, Irizarry R, Huber W, editors. Bioinformatics and Computational Biology Solutions Using R and Bioconductor. New York: Springer. pp. 397–420.
- 24. Benjamini Y, Hochberg Y (1995) Controlling the False Discovery Rate—a Practical and Powerful Approach to Multiple Testing. Journal of the Royal Statistical Society Series B-Methodological 57: 289–300.
- 25. Swanson MG, Vigneron DB, Tabatabai ZL, Males RG, Schmitt L, et al. (2003) Proton HR-MAS spectroscopy and quantitative pathologic analysis of MRI/3D-MRSI-targeted postsurgical prostate tissues. Magn Reson Med 50: 944–954. pmid:14587005
- 26. Bathen TF, Sitter B, Sjobakk TE, Tessem MB, Gribbestad IS (2010) Magnetic resonance metabolomics of intact tissue: a biotechnological tool in cancer diagnostics and treatment evaluation. Cancer Res 70: 6692–6696. pmid:20699363
- 27. Soda K (2011) The mechanisms by which polyamines accelerate tumor spread. Journal of Experimental & Clinical Cancer Research 30.
- 28. Gerner EW, Meyskens FL Jr. (2004) Polyamines and cancer: old molecules, new understanding. Nat Rev Cancer 4: 781–792. pmid:15510159
- 29. Bettuzzi S, Davalli P, Astancolle S, Carani C, Madeo B, et al. (2000) Tumor progression is accompanied by significant changes in the levels of expression of polyamine metabolism regulatory genes and clusterin (sulfated glycoprotein 2) in human prostate cancer specimens. Cancer Research 60: 1472–1472.
- 30. Rhodes DR, Barrette TR, Rubin MA, Ghosh D, Chinnaiyan AM (2002) Meta-analysis of microarrays: Interstudy validation of gene expression profiles reveals pathway dysregulation in prostate cancer. Cancer Research 62: 4427–4433. pmid:12154050
- 31. Forshell TP, Rimpi S, Nilsson JA (2010) Chemoprevention of B-Cell Lymphomas by Inhibition of the Myc Target Spermidine Synthase. Cancer Prevention Research 3: 140–147. pmid:20103729
- 32. Ghesquiere B, Wong BW, Kuchnio A, Carmeliet P (2014) Metabolism of stromal and immune cells in health and disease. Nature 511: 167–176. pmid:25008522
- 33. Wu X, Daniels G, Lee P, Monaco ME (2014) Lipid metabolism in prostate cancer. Am J Clin Exp Urol 2: 111–120. pmid:25374912
- 34. Migita T, Ruiz S, Fornari A, Fiorentino M, Priolo C, et al. (2009) Fatty acid synthase: a metabolic enzyme and candidate oncogene in prostate cancer. J Natl Cancer Inst 101: 519–532. pmid:19318631
- 35. Baron A, Migita T, Tang D, Loda M (2004) Fatty acid synthase: a metabolic oncogene in prostate cancer? J Cell Biochem 91: 47–53. pmid:14689581
- 36. Beckers A, Organe S, Timmermans L, Scheys K, Peeters A, et al. (2007) Chemical inhibition of acetyl-CoA carboxylase induces growth arrest and cytotoxicity selectively in cancer cells. Cancer Res 67: 8180–8187. pmid:17804731
- 37. Brusselmans K, De Schrijver E, Verhoeven G, Swinnen JV (2005) RNA interference-mediated silencing of the acetyl-CoA-carboxylase-alpha gene induces growth inhibition and apoptosis of prostate cancer cells. Cancer Res 65: 6719–6725. pmid:16061653
- 38. Massie CE, Lynch A, Ramos-Montoya A, Boren J, Stark R, et al. (2011) The androgen receptor fuels prostate cancer by regulating central metabolism and biosynthesis. Embo Journal 30: 2719–2733. pmid:21602788
- 39. King A, Selak MA, Gottlieb E (2006) Succinate dehydrogenase and fumarate hydratase: linking mitochondrial dysfunction and cancer. Oncogene 25: 4675–4682. pmid:16892081