Abstract
Despite improvements in the management of liver cancer, the survival rate for patients with hepatocellular carcinoma (HCC) remains dismal. The survival benefit of systemic chemotherapy for the treatment of liver cancer is only marginal. Although the reasons for treatment failure are multifactorial, intrinsic resistance to chemotherapy plays a primary role. Here, we analyzed the expression of 377 multidrug resistance (MDR)-associated genes in two independent cohorts of patients with advanced HCC, with the aim of finding ways to improve survival in this poor-prognosis cancer. Taqman-based quantitative polymerase chain reaction revealed a 45-gene signature that predicts overall survival (OS) in patients with HCC. Using the Connectivity Map Tool, we were able to identify drugs that converted the gene expression profiles of HCC cell lines from ones matching patients with poor OS to profiles associated with good OS. We found three compounds that convert the gene expression profiles of three HCC cell lines to gene expression profiles associated with good OS. These compounds increase histone acetylation, which correlates with the synergistic sensitization of those MDR tumor cells to conventional chemotherapeutic agents, including cisplatin, sorafenib, and 5-fluorouracil. Our results indicate that it is possible to modulate gene expression profiles in HCC cell lines to those associated with better outcome. This approach also increases sensitization of HCC cells toward conventional chemotherapeutic agents. This work suggests new treatment strategies for a disease for which few therapeutic options exist.
Introduction
Liver cancer is the third most common cancer in the world, causing approximately 745,000 deaths per year (Ferlay et al., 2015). Hepatocellular carcinoma (HCC) is by far the most prevalent type, accounting for approximately 80%–85% of primary liver cancer cases (Singal and El-Serag, 2015), whereas cholangiocarcinoma (Ghouri et al., 2015) and fibrolamellar carcinoma (Lim et al., 2014; Cornella et al., 2015) occur at a frequency of only ∼14% and ∼1%, respectively. The epidemiology of HCC is well known, and in the vast majority of cases, it arises as a consequence of underlying liver disease, usually a viral hepatitis (Singal and El-Serag, 2015). In the case of hepatitis B, integration of the viral DNA into the hepatocyte genome results in loss of chromosomal stability, deregulation of tumor-suppressor genes, and activation of proto-oncogenes, eventually leading to the development of HCC (Su et al., 2014).
Patients with early stage tumors undergo either surgical resection or liver transplantation if their HCC meets the so-called Milan criteria (Mazzaferro et al., 1996; Waller et al., 2015). When surgery is not a suitable option, local ablation, including radiofrequency ablation and percutaneous ethanol injection, are standard treatment. Transcatheter arterial chemoembolization is recommended for patients with intermediate stage HCC (EASL-EORTC, 2012; Villanueva et al., 2013). With the exception of sorafenib, a multi-tyrosine kinase inhibitor for which a survival benefit of 3 months was demonstrated, no effective systemic therapy exists for patients with advanced HCC (Llovet et al., 2008; Llovet et al., 2015). Although sorafenib is now established as the first line of therapy for advanced HCC, it was shown to be a substrate of ABCB1 and ABCG2, two major ABC transporters involved in multidrug resistance (MDR) and expressed in hepatocytes and hepatomas (Lagas et al., 2010; Tang et al., 2013). Comprehensive molecular profiling contributed to substantial improvement in our knowledge of the biology of liver cancer and provides a road map to facilitate the development of targeted therapies (Andersen and Thorgeirsson, 2012; Pinyol et al., 2014; Bruix et al., 2015; Simon et al., 2015). Besides proof-of-concept trial testing signaling pathway inhibitors or biomarker-based trial enrichment for defining cancer subpopulations, there is still a need for unspecific drugs that target all patients (Llovet and Hernandez-Gea, 2014).
Although the reasons for treatment failure are multifactorial, intrinsic resistance to chemotherapy plays a primary role. Here, the expression of MDR-associated genes was analyzed in two independent cohorts of patients with advanced HCC. We hypothesized that the Connectivity Map Tool might reveal compounds that reverse the gene expression profile of cancers from patients with poor prognosis to that of cancers from patients who respond well to treatment (Lamb et al., 2006; Lamb, 2007; Zhang and Gant, 2008). The ultimate aim is to find a new strategy to sensitize intrinsically MDR cancer.
Materials and Methods
Tumor Samples.
Anonymized clinical samples of 38 HCCs and 13 normal liver tissues were provided by Lee et al. (2004). Most of the patients had a hepatitis B virus background, but a few had hepatitis C virus and alcoholic backgrounds. All the samples originated from untreated primary resected tumors. Importantly, even though most of the patients in this cohort were infected with HBV (which is also the common background for Chinese HCC patients), the cohort also includes hepatitis B virus–, hepatitis C virus– and alcoholic-related HCC cases, simulating a “true” clinical situation, as patients with HCC are not normally a homogeneous group. Thirty-eight HCC samples were randomly selected to be reanalyzed among samples that were previously classified into two groups based on overall survival (Lee et al., 2004). Lee et al. (2006) demonstrated that although patients may originate from different ethnic groups, the cohort could still be considered homogeneous at the molecular level. A total of 17 normal liver samples were analyzed. Total RNA for N1-N4 was purchased commercially, N1 from Ambion (catalog no. AM7960; Austin, TX), N2 from Stratagene (catalog no. 540017; Santa Clara, CA), N3 from Clontech (catalog no. 636531; Mountain View, CA), N4 from Biochain Institute (catalog no. R1234149-50; Newark, CA). RNA from N5-N17 was provided by the Laboratory of Experimental Carcinogenesis, National Cancer Institute (Bethesda, MD).
TaqMan Low-Density Arrays.
Expression levels of 377 MDR-associated genes were measured in the previously mentioned samples using a custom-made Taqman Low-Density Array (TLDA; Applied Biosystems, Foster City, CA), as previously reported (Calcagno et al., 2010).
Normalization and Filtering.
The median expression of each sample was subtracted from all gene-expression data for that sample. Two of the normal (N1, N2) and seven of the HCC samples (A1, A5, A8, A9, A11, B16, B18) were analyzed in duplicate. For these nine samples, the Pearson correlation between the duplicates was greater than 99%. These duplicates were averaged together. One of the genes (18S) was present as multiple probes. The expression data from the multiple probes for that gene were averaged together. Genes that were expressed in 10 or fewer samples were removed from the analysis.
Comparison of HCC and Normal Liver Samples.
Genes expressed differentially in HCC and normal samples were detected using the t statistic. The obtained P values were adjusted for multiple comparisons using the Benjamini-Hochberg method (Benjamini and Hochberg, 1995).
Comparison of Samples from Groups A and B.
Unsupervised clustering (average linkage algorithm with 1-Pearson correlation as the distance) was done on the normalized, filtered data. Genes differentially expressed in groups A and B were detected using the t-statistic, with the P values adjusted for multiple comparisons using the Benjamini-Hochberg method.
Validation of the 45 MDR-Linked Genes as a Prognostic Signature for Poor Overall Survival.
The clinical data of this cohort and the analysis of the gene expression profile were published by Andersen et al. (2010). Analysis of survival data were performed by Kaplan-Meier using Mantal-Cox (log-rank) statistics (GraphPad.Prism v5; GraphPad Software, Inc., La Jolla, CA).
Connectivity Map Analysis.
The upregulated and downregulated genes found in 38 HCC samples that had an adjusted P < 0.05 in the TLDA data were used as input to the Connectivity Map online web tool (http://www.broadinstitute.org/science/projects/connectivity-map/connectivity-map).
Validation of the Compounds Highlighted by Using the Connectivity Map.
All 20 cell lines were grown to 70%–80% confluence. Twenty-four hours before RNA extraction using an RNeasy Micro kit (Qiagen, Valencia, CA), all cell lines were grown in Dulbecco’s modified Eagle’s medium/F12 (Invitrogen, Carlsbad, CA). The RNA was prepared and profiled as previously mentioned in the section entitled Validation of the 45 MDR-linked Genes as a Prognostic Signature for Poor Overall Survival. Integration was performed by z-transforming each data set separately, and the hierarchical clustering was performed using Cluster 3.0 and TreeView1.6, software developed by Michael Eisen, Stanford University, Stanford, CA.
Cell Lines.
HUH7, PLC, and HEP3B, which clustered with samples from patients with poor overall survival, and HUH1, SNU182, and FOCUS clustering with samples from patients with better overall survival were maintained in RPMI-1640 (Life Technologies, Invitrogen, Grand Island, NY), supplemented with 10% fetal bovine serum, 100 units of penicillin/streptomycin/ml at 37°C in 5% CO2 humidified air.
Cytotoxicity Assay to Determine Synergism of Added Drugs.
To assess the synergistic effects of sorafenib, 6TG, 8-Aza, doxorubicin, apigenin, cisplatin, and 5-FU in combination, Huh7, Hep3B, and PLC cells were each treated with a matrix of two different drugs with serial 1:2 dilutions from 100–0.001526 µM. Five thousand cells were seeded per well 16 hours before the addition of the drug combinations. An assay using the reagent (3-(4,5-dimethylthiazol-2-yl)-2,5-diphenyl tetrazolium bromide) (MTT) was performed according to the manufacturer's (Trevigen, Gaithersburg, MD) protocol to measure the proliferation of cells at 72 hours after the addition of drug. To evaluate sensitization effects, 20 µM of one drug was added for 24 hours and removed. The second drug was added in a serial 1:2 dilution from 100–0.3906 µM, and proliferation was measured at 72 hours using the MTT assay. Three technical replicates were performed for the sensitization experiments. All additions of cells to 96-well plates, drug dilutions/additions, and the MTT assay were carried out using a Hamilton Star Liquid Handler (Reno, Nevada). Synergy calculations were done using custom scripts on R (version 2.15.2).
Western Blot Immunoassay.
The following antibodies were used for a Western blot immunoassay: rabbit anti-acetyl histone H3 (1:1,000, catalog no. 06-599; Millipore, Billerica, MA), rabbit anti-GAPDH (1:5000; catalog no. 14C10; Cell Signaling, Danvers, MA). Horseradish peroxidase–linked secondary antibodies (1:10,000) were from DakoCytomation (Carpinteria, CA). Bands were visualized by chemiluminescence using X-ray film.
Silencing the Histone Acetyltranserase GCN5.
A pGIPZ shRNA construct for stable knockdown of human GCN5 (KAT2A gene) was obtained from Open Biosystems (Huntsville, AL)-Thermo Scientific (catalog no. RHS4430-99293180; Somerset, NJ). A nonsilencing pGIPZ shRNA (catalog no. RHS4346) was used as a control. Lentiviral particles were made via Lipofectamine 2000 (Invitrogen)–mediated triple transfection of 293T cells with pGIPZ shRNA plasmids along with the lentiviral envelope plasmid (pMD2.G, Addgene plasmid 12259) and the lentiviral packaging plasmid (psPAX2, Addgene plasmid 12260). Liver cancer cells were transduced with either nonsilencing or GCN5-specific shRNA containing lentiviral particles in the presence of 8 μg/ml polybrene and stable cells were selected using 3 μg/ml puromycin for 1 week and pooled before determining knockdown efficiency. Knockdown efficiency was determined via Western blot analysis using a GCN5-specific antibody (no. 3305) from Cell Signaling Technology.
Results
Genes Differentially Expressed in HCC Compared with Normal Liver Samples.
We conducted a study on 17 normal liver samples and 38 HCC samples (Lee et al., 2004, 2006) to compare the expression profile of 377 MDR-linked genes in normal and HCC samples. These genes, selected from the literature published over the last 30 years, were reported to have a role in MDR, based primarily on in vitro studies (Calcagno et al., 2010; Gillet et al., 2011). Our analysis revealed 103 genes that are differentially expressed in HCC compared with normal liver samples. Eighty-two genes have a false discovery rate (FDR) < 0.05 and P value < 0.01, and 21 additional genes fulfill the less stringent criteria of FDR < 0.05 and P value < 0.05 (Table 1). More precisely, 32 genes were found to be downregulated in HCC compared with normal samples, whereas 71 were found to be upregulated (Table 1). Notably, eight ABC transporters were overexpressed in HCC. Many of these drug efflux transporters are important mediators of MDR (Gillet et al., 2007). Within this group of eight genes, the involvement of ABCC1, ABCC4, ABCC5, and ABCC10 in MDR has been well characterized (Gillet et al., 2007). Our analysis also highlighted a cluster of genes involved in cell-cycle regulation, including CDKN2A, CDK2 and 4, CCD42, and CCNE1. Moreover, the cell-cycle checkpoints CHEK 1 and 2 and RAD1 were found to be overexpressed, as well as a large cluster of DNA repair genes, including TOP2A, overexpressed 67-fold.
Four main groups of downregulated genes were uncovered, including several ABC transporters, CYP450s, metallothioneins, and solute carriers. None of the downregulated ABC transporters were involved in MDR except ABCC11, whereas several downregulated SLCs have been previously identified as drug transporters. They include SLC21A8/SLCO1B3, SLC22A1/OCT1, SLC28A1/CNT1, and the copper transporter SLC31A1/CTR1, which also transports cisplatin, oxaliplatin, and carboplatin (Huang, 2007; Kuo et al., 2007).
Identification of an MDR-Linked Gene Signature in a Previously Established Group of Poor Overall Survival Patients.
We characterized the differential MDR-linked gene expression of the 38 HCC samples, which were previously classified into two groups based on overall survival (Lee et al., 2004). Group A consisted of 20 samples taken from Chinese patients with a poor overall survival; group B contained eleven samples taken from Chinese patients and seven samples from Belgian patients. All group B patients demonstrated better overall survival.
Unsupervised clustering of all the genes expressed in more than 10 of the 38 samples yields distinct clusters for the HCC group A subtype (poor overall survival), the HCC group B subtype (good overall survival), and the normal samples, with few exceptions (Fig. 1A). Supervised class comparisons highlighted 45 genes that are differentially expressed in groups A and B (FDR < 0.05), of which 29 were found to be upregulated in patients with poor overall survival compared with patients with good overall survival, whereas 16 were downregulated (Table 2). Perhaps the most striking finding is the upregulation of 12 genes related to DNA repair, including cell-cycle checkpoints CHEK1 and ATM; the regulators BRCA1 and 2; the double-strand break repair genes MRE11A, TOP2, RAD51, XRCC1, 2, and 5; and the single-strand DNA repair gene TOP1. Three gene families were found to be downregulated. They include five ABC transporter genes (ABCA6, B4, B11, C9, and G8), four cytochrome P450s (CYP2A6, 2C8, 2C9, and 2C19) and three SLCs (SLC10A1, 22A1, and 28A1), which could be attributed to the concomitant downregulation of the nuclear receptor genes NR1I2 and NR1I3 (di Masi et al., 2009).
Validation of the 45 MDR-Linked Genes as a Prognostic Signature for Poor Overall Survival.
We assessed the predictive power of the 45-gene signature identified as differentially expressed in A versus B subtypes on an independent cohort of 53 HCCs obtained from white and Chinese patients (Andersen et al., 2010). The gene expression profiling was performed using Illumina bead chips (Andersen et al., 2010). Figure 1B shows that the 45-gene signature effectively predicts overall survival of patients with HCC (P < 0.02), validating the clinical relevance of the gene signature.
Identification of Compounds that Sensitize Chemoresistant HCC.
The next step in this study was to pinpoint drugs that might efficiently alter the poor prognosis gene signature in HCC. For this, we used the Connectivity Map tool published by Lamb and colleagues (Lamb et al., 2006), designed to reveal connections among drugs, genes, and pathologic states. The Connectivity Map algorithm (Lamb et al., 2006) compares the direction of gene expression change from one disease state to another with the change due to a drug treatment. Drugs that cause an expression change similar to the change between poor-prognosis and better-prognosis tumors may be able to change the outcome in those with HCC, possibly by causing the resistant cells to become more drug-sensitive or by changing the physiology of the tumor.
We were interested in drugs that cause a change in gene expression that matches the gene expression change from group A to group B. From the upregulated and downregulated genes obtained by TLDA that had an adjusted P value < 0.05, we found four drugs with high positive concordance, low P value (P < 0.001), and a low specificity score (specificity score < 0.05). These drugs were 8-azaguanine, 6-thioguanosine, apigenin, and 0175029-0000, a pyrimidine derivative (2-[4-(2-diethylaminoethyloxy)anilino]-8-phenyl-pyrido[2,3-day]pyrimidin-7-one) (Supplementary Table S1).
To confirm our findings regarding these drugs, we performed an integrative clustering using the 45-gene signature in 20 HCC cell lines and the 53 HCC clinical samples of our validation set (Fig. 2A). Two HCC cell lines randomly selected, HUH7 and PLC, which clustered with samples from patients with poor overall survival, were treated for 72 hours with a subcytotoxic dose of each of the drugs individually, except for compound 0175029-0000, which is unavailable. The data indicate that the treatment caused a change in the gene expression profile of the cell lines from that of poor overall survival to that of better overall survival (Fig. 2B).
6-TG, 8-AZG, and Apigenin Mediate Increased Acetylation of Histone Protein.
We next hypothesized that the mechanism underlying the ability of these three compounds to change gene expression patterns might be associated with increased acetylation of histone protein. This was confirmed using an antibody directed against acetylated histone H3 in three HCC cell lines (HUH7, PLC, and HEP3B) treated for 6 hours with 10 μM of any one of these drugs (Fig. 3A). Optimization of the treatment (24 hours with 20 μM) dramatically increased the effect observed, with a 3- to 6-fold increase in acetylated histone H3 and as much as a 13.5-fold increase when treated with depsipeptide as a positive control (Fig. 3B).
To explore further the hypothesis that changes in gene expression patterns were due to increased histone H3 acetylation, we knocked down the expression of the major histone H3 acetyltransferase GCN5 (Fig. 4, C–F). As hypothesized, drug treatment-induced histone acetylation decreased in the HUH7 and PLC cell lines when GCN5 was knocked down and to some extent in the HEP3B cell line, when treated with apigenin.
Drug Combinations Show Synergistic Cytotoxicity.
For each pair of drugs tested, we used combinations of varying concentrations to determine whether the combination of these drugs resulted in increased cytotoxicity and whether this effect was additive or synergistic. The combination of three drug pairs had a significant synergistic effect on all three HCC cell lines: 6-TG/apigenin, doxorubicin/apigenin, and sorafenib/apigenin, and to some extent, 6-TG/5-FU (Fig. 4), indicating that the changes we observed in gene expression patterns were also associated with increases in treatment efficacy over and above the toxicity of the drugs themselves.
Discussion
Using a TaqMan-based quantitative reverse transcription-polymerase chain reaction array, we studied the expression profile of 377 MDR-linked genes and found a signature of 103 genes differentially expressed in normal liver cells and HCC. The MDR genotype consists of the upregulation of several members of the ABCC family (known as MRPs), of genes involved in cell proliferation through regulation of the G1/S cell-cycle transition, and of DNA repair genes. We also identified downregulation of several solute carriers involved in platinum drug uptake, potentially resulting in a dramatic decrease in the cellular entry of this drug (Huang, 2007; Kuo et al., 2007). The intrinsic expression of several additional ABC transporters known to efflux standard chemotherapeutics, including ABCB1, ABCB4, ABCB11, and ABCG2, leaves limited treatment modalities to clinicians when coupled with the MDR-linked gene signature of HCC. Many of these ABC transporters have been shown to transport doxorubicin (Szakacs et al., 2006). It should be noted that this poor-prognosis MDR gene signature probably reflects a biologic state of the HCC rather than being the sole cause of the poor prognosis, since the patients who were the source of the analyzed HCC samples were not treated with chemotherapeutic agents; however, the presence of these drug-resistance mechanisms in poor-prognosis HCC makes it difficult to design chemotherapy that might be effective against these cancers. On the other hand, this 103-gene signature not only confirms the expression of known markers of HCC such as TOP2A (Wong et al., 2009), which is a target for topoisomerase inhibitors, but also highlights new markers including the solute carriers SLC2A5/GLUT5, SLC16A3/MCT, SLC7A11, and the melphalan transporter SLC7A5/LAT1 (del Amo et al., 2008). It is possible that these uptake transporters might facilitate cellular entry of certain yet unidentified drug species, therefore facilitating therapy.
The HCC samples analyzed comprised two groups defined by overall survival rate. Since previous studies indicated the superiority of TLDAs over high-density microarrays (confirmed in this work), we used this technique to investigate the differences in gene expression profiles for patients with good and poor overall survival (Gillet and Gottesman, 2011). Interestingly, our analysis revealed a novel 45-gene signature that was shown to predict overall survival. In addition to well-established markers, we demonstrated 13-fold and 5-fold overexpression of SLC29A2 and SLC16A3/MCT, respectively, in the poor overall survival group. SLC29A2, a nucleoside uptake transporter, mediates the transport of gemcitabine, cladribine, and zidovudine (Baldwin et al., 2004; Huang and Sadee, 2006). Huang and colleagues correlated SLC gene expression profiles in the NCI-60 cancer cell line panel with the potencies of 119 standard anticancer drugs and identified new transporter substrates (Huang et al., 2004). More recently, a similar approach was carried out by Okabe et al. (2008) to investigate the SLC22 and SLCO gene families using quantitative reverse transcription-polymerase chain reaction array rather than oligonucleotide array. That work revealed the role of SLC22A4 in cellular sensitivity to doxorubicin and mitoxantrone. Studies of this type may help reveal additional compounds that could be used to treat SLC-overexpressing cancer cells because they are substrates for upregulated uptake transporters.
The ultimate goal of highlighting MDR-linked gene signatures is to be able to sensitize drug-resistant cancers. One strategy is the development of targeted therapy using newly identified biomarkers associated with drug resistance, as indicated here. Another approach highlighted in this work is to use a resource published earlier by the Golub group, the Connectivity Map, which allows the identification of compounds that might sensitize MDR cancers with a specific gene signature (Lamb et al., 2006; Lamb, 2007; Zhang and Gant, 2008). We found three compounds, a flavone (apigenin) and two nucleoside analogs (8-AZG and 6-TG), that convert the gene expression profiles of three HCC cell lines classified with poor overall survival patients into gene expression profiles associated with good overall survival. Application of these compounds, which we have demonstrated in this work, are histone deacetylase inhibitors in the cell lines studied and resulted in the synergistic sensitization of those MDR tumor cells to conventional chemotherapeutic agents, including cisplatin, sorafenib, and 5-fluorouracil. Of note, we made this observation using cell lines with three different TP53 mutational profiles: TP53 null (HEP3B), mutated (HUH7), and WT (PLC). Thorgeirsson and colleagues showed that HCC patients with p53 mutations had a shorter overall survival time compared with patients with wild-type p53 (Woo et al., 2011). Here we show that the efficiency of the proposed strategy is not dependent on TP53 status.
This study provides in vitro–based evidence that, if validated in vivo, suggests that the management of cancers intrinsically resistant to standard chemotherapeutic treatments would improve by changing transcriptional profiles, including genes whose expression is known to be associated with drug resistance. We can also envision applying such a strategy to any cancer highly resistant to standard chemotherapy with the goal of reversing its gene expression profile to that of cancers from patients who respond well to treatment.
Acknowledgments
The authors thank George Leiman for editorial assistance and Allison C. Meade in the Protein Expression Laboratory (SAIC-Frederick, Inc.) for assistance in liquid handling.
Authorship Contributions
Participated in research design: Gillet, Gottesman, Ambudkar, Bagni, Madigan, Thorgeirsson.
Conducted experiments: Gillet, Andersen, Madigan, Bagni, Powell, Burgan, Wu, Calcagno.
Performed data analysis: Gillet, Varma, Andersen, Wu.
Wrote or contributed to the writing of the manuscript: Gillet, Gottesman, Ambudkar, Andersen.
Footnotes
- Received August 14, 2015.
- Accepted December 11, 2015.
↵1 Current affiliation: Laboratory of Molecular Cancer Biology, Molecular Physiology Research Unit—URPhyM, Namur Research Institute for Life Sciences (NARILIS), Faculty of Medicine, University of Namur, Belgium.
↵2 Current affiliation: Biotech Research and Innovation Centre, University of Copenhagen, Ole Maaløes Vej 5, DK-2200 Copenhagen N, Denmark.
This research was funded by the Intramural Research Program of the National Institutes of Health (NIH). The project was funded in part with federal funds from the National Cancer Institute NIH [Contract HHSN2612008000001E]. The content of this publication does not necessarily reflect the views or policies of the Department of Health and Human Services, nor does mention of trade names, commercial products, or organizations imply endorsement by the U.S. Government.
↵This article has supplemental material available at (molpharm.aspetjournals.org).
Abbreviations
- HCC
- hepatocellular carcinoma
- MDR
- multidrug resistance
- OS
- overall survival
- TLDA
- TaqMan low-density array
- U.S. Government work not protected by U.S. copyright
References
MolPharm articles become freely available 12 months after publication, and remain freely available for 5 years.Non-open access articles that fall outside this five year window are available only to institutional subscribers and current ASPET members, or through the article purchase feature at the bottom of the page.
|