Molecular Cancer Research
HOME HELP FEEDBACK SUBSCRIPTIONS ARCHIVE SEARCH TABLE OF CONTENTS
Molecular Cancer Research 5, 443-454, May 1, 2007. doi: 10.1158/1541-7786.MCR-06-0337
© 2007 American Association for Cancer Research
This Article
Right arrow Abstract Freely available
Right arrow Full Text (PDF)
Right arrow Supplementary Data
Right arrow Alert me when this article is cited
Right arrow Alert me if a correction is posted
Services
Right arrow Similar articles in this journal
Right arrow Similar articles in PubMed
Right arrow Alert me to new issues of the journal
Right arrow Download to citation manager
Right arrow reprints & permissions
Citing Articles
Right arrow Citing Articles via HighWire
Right arrow Citing Articles via Google Scholar
Google Scholar
Right arrow Articles by Schwartz, D. R.
Right arrow Articles by Sloane, B. F.
Right arrow Search for Related Content
PubMed
Right arrow PubMed Citation
Right arrow Articles by Schwartz, D. R.
Right arrow Articles by Sloane, B. F.

Cancer Genes and Genomics

Hu/Mu ProtIn Oligonucleotide Microarray: Dual-Species Array for Profiling Protease and Protease Inhibitor Gene Expression in Tumors and Their Microenvironment

Donald R. Schwartz1, Kamiar Moin1,2, Bin Yao5, Lynn M. Matrisian7, Lisa M. Coussens8, Thomas H. Bugge9, Barbara Fingleton7, Heath B. Acuff7, Mark Sinnamon7, Hind Nassar3, Adrian E. Platts4,6, Stephen A. Krawetz1,4,5,6, Bruce E. Linebaugh2 and Bonnie F. Sloane1,2

1 Barbara Ann Karmanos Cancer Institute; Departments of 2 Pharmacology, 3 Pathology, and 4 Obstetrics and Gynecology; 5 Applied Genomics Technology Core; and 6 Center for Molecular Medicine and Genetics and the Institute for Scientific Computing, Wayne State University, Detroit, Michigan; 7 Department of Cancer Biology, Vanderbilt University School of Medicine, Nashville, Tennessee; 8 Department of Pathology, Cancer Research Institute and Comprehensive Cancer Center, University of California, San Francisco, California; and 9 Proteases and Tissue Remodeling Unit, Oral and Pharyngeal Cancer Branch, National Institute of Dental and Craniofacial Research, NIH, Bethesda, Maryland

Requests for reprints: Donald R. Schwartz, Biodiscovery, LLC, 5692 Plymouth Road, Ann Harbor, MI 48105. Phone: 734-998-0751; Fax: 734-998-0750. E-mail: dschwartz{at}biodiscovery-llc.com


   Abstract
Top
 Notes
 Abstract
Introduction
 Results
 Discussion
 Materials and Methods
 Acknowledgements
 References
 
Proteolysis is a critical regulatory mechanism for a wide variety of physiologic and pathologic processes. To assist in the identification of proteases, their endogenous inhibitors, and proteins that interact with proteases or proteolytic pathways in biological tissues, a dual-species oligonucleotide microarray has been developed in conjunction with Affymetrix. The Hu/Mu ProtIn microarray contains 516 and 456 probe sets that survey human and mouse genes of interest (proteases, protease inhibitors, or interactors), respectively. To investigate the performance of the array, gene expression profiles were analyzed in pure mouse and human samples (reference RNA; normal and tumor cell lines/tissues) and orthotopically implanted xenografts of human A549 lung and MDA-MB-231 breast carcinomas. Relative gene expression and "present-call" P values were determined for each probe set using dChip and MAS5 software, respectively. Despite the high level of sequence identity of mouse and human protease/inhibitor orthologues and the theoretical potential for cross-hybridization of some of the probes, >95% of the "present calls" (P < 0.01) resulted from same-species hybridizations (e.g., human transcripts to human probe sets). To further assess the performance of the microarray, differential gene expression and false discovery rate analyses were carried out on human or mouse sample groups, and data processing methods to optimize performance of the mouse and human probe sets were identified. The Hu/Mu ProtIn microarray is a valuable discovery tool for the identification of components of human and murine proteolytic pathways in health and disease and has particular utility in the determination of cellular origins of proteases and protease inhibitors in xenograft models of human cancer. (Mol Cancer Res 2007;5(5):443–54)


   Introduction
Top
 Notes
 Abstract
 Introduction
Results
 Discussion
 Materials and Methods
 Acknowledgements
 References
 
Proteolysis is a critical regulatory mechanism for a wide variety of physiologic and pathologic processes including inflammation, angiogenesis, tumor initiation, invasion, and metastasis (reviewed in refs. 1-5). Based on their common evolutionary origin, there are 47 clans of proteases comprised of more than 175 families that share similar tertiary structures or catalytic site residues (6). Further information on the diversity, complexity, and relationships of known proteases can be found on the MEROPS database online (6).10 Protease activity is controlled by endogenous protein inhibitors (reviewed in refs. 7-9). Like proteases, these inhibitors are numerous and diverse and have been organized into 33 clans of 55 families (6). The biological functions of proteases are also affected by their interactions with proteins that may modulate their behavior or location, such as activators and receptors. Identifying the proteases, protease inhibitors, and protease interactors expressed in healthy or diseased tissues and non-diseased host cells in response to the diseased tissue is essential to understanding tissue homeostasis and pathologic processes. Xenografts of human cancer cells implanted orthotopically in mice represent a tractable model to assess functional interactions of protease-mediated pathways in vivo.

Oligonucleotide microarrays have greatly facilitated gene expression analysis in diseases like cancer (for review, see ref. 10). Typically, oligonucleotide microarrays contain probes for interrogating gene expression from only one species. Consequently, simultaneous determination of both host (mouse) and graft (human) gene expression in murine orthotopic xenograft models of human cancer cannot be accomplished with one microarray. Although this could be accomplished by profiling with two microarrays, one specific for human and one for mouse gene expression, this can be cost- and time-prohibitive. Furthermore, such dual-species analyses may lead to misinterpretation, as probes on single-species arrays are not designed to prevent interrogation of transcripts from other species but rather to interrogate transcripts of only one species. Thus, cross-hybridization might be interpreted as a change in expression.

We report here the design and performance evaluation of an oligonucleotide microarray, the Hu/Mu ProtIn microarray, for measurement of protease, protease inhibitor, and protease interactor gene expression. Uniquely, the Hu/Mu ProtIn microarray has the capacity to differentiate murine from human gene expression. Here, we show the specificity and sensitivity of the probe sets on our microarray based upon gene expression analyses in pure mouse or human samples or orthotopic xenograft models of human breast and lung cancer. Thus, we show the utility of comparative analyses on one platform.


   Results
Top
 Notes
 Abstract
 Introduction
 Results
Discussion
 Materials and Methods
 Acknowledgements
 References
 
Hu/Mu ProtIn Microarray
This microarray contains custom-designed probe sets, each of which surveys one gene with a collection of 16 unique oligonucleotide (25-mer) probes that match the same gene transcript. There are 516 and 456 custom probe sets on the microarray that survey, respectively, 430 human and 390 mouse genes of interest (proteases, protease inhibitors, and genes that interact with proteases). Some genes are surveyed by more than one probe set. Annotation for all of the mouse and human genes surveyed on the Hu/Mu ProtIn microarray is available as Supplementary Table S1. The Hu/Mu ProtIn microarray also contains a number of control probe sets designed by Affymetrix. If not otherwise stated, probes and probe sets refer to the custom-designed probes and probe sets that survey genes of interest (e.g., proteases) and not control genes. The Hu/Mu ProtIn microarray is available commercially through Affymetrix by contacting the customer service department and requesting Protease520066F.

Theoretical Cross-Hybridization Analysis
To determine the theoretical potential for cross-hybridization of the custom probes, we did a Blast(n) analysis of each probe sequence against the human and mouse repeat-masked genome in the Ensembl database online.11 The threshold for potential cross-hybridization was set at ≥16 continuous matched base pairs within the 25-mer oligonucleotide probe. A summary of the probes on the Hu/Mu ProtIn microarray that might cross-hybridize is shown in Table 1 . A file listing all of the potentially cross-hybridizing probes is available as Supplementary Table S2. Each custom probe set on the microarray contains 16 unique perfect match (PM) probes. Within a probe set, individual probe sequences may overlap but are never identical. Consequently, there are 8,256 unique human and 7,296 unique murine probes on the Hu/Mu ProtIn microarray that comprise 516 human and 456 mouse probe sets, respectively. Ninety-eight percent of the probe sets contain at least one probe (of 16) that may theoretically cross-hybridize. If potentially cross-hybridizing probes are masked (i.e., probe information is eliminated before any data processing or analysis), 952 or 98% of the probe sets retain at least six probes (Table 1), a number sufficient for reliable data analysis, suggesting that each probe set, on average, contains 4 of 16 probes that may exhibit cross-hybridization. Probe masks were generated to remove probes with a potential for cross-hybridizing between species from the dChip analysis. To test the potential for cross-hybridization, several different masks were developed with varying degrees of stringency. These ranged from probe masks for the condition where ≥15 bp in a probe matched continuously in both human and mouse genomes up to the condition where ≥21 bp matched. In accord with expectations, when mouse fat pad and carcinoma tissues were compared, specificity was serially increased by reducing the continuous match criteria for masking, whereas sensitivity was affected if a criterion of <17 bp was used. Masking those probes with >21 consecutive base pairs in common marginally decreased the number of genes reported from the human probe sets and somewhat improved the sensitivity of mouse probe sets (differentially expressed genes Mouse/Human at P < 0.01, 181:63). Masking more rigorously above 14 bp significantly reduced the differential gene expression (DGE) reported by the human probe sets, but masking mouse probes at this level was found to impair sensitivity of mouse probes to differential expression (differentially expressed genes Mouse/Human at P < 0.01, 142:32).


View this table:
[in this window]
[in a new window]

  		Table 1. Theoretical Cross-hybridization Summary

 
Hu/Mu ProtIn Probe Set Response to Mouse or Human Universal Reference RNA
We have assessed the specificity and selectivity of the custom probe sets to determine the extent, if any, of cross-hybridization of one species' transcripts to probe sets designed for the other species (i.e., non–species-specific transcript hybridization) using pure human or mouse Universal Reference total RNA. Universal Reference RNA are expected to contain a broad representation of transcripts because these reference RNA are derived from a mixture of many tissues. The number of probe sets giving P values smaller than 0.05, 0.01, and 0.001, as determined by the default detection algorithm in MAS5 software (Affymetrix), is shown for mouse (Fig. 1A ) and human (Fig. 1B) Universal Reference RNA. As expected, hundreds of transcripts were detected by mouse (Fig. 1A, black columns) and human (Fig. 1B, gray columns) probe sets when the Hu/Mu ProtIn microarray was hybridized to mouse or human Universal Reference RNA, respectively. In contrast, <3% of the total present calls (at P < 0.05) represent probe sets responding to other-species reference RNA (Fig. 1A, gray columns or Fig. 1B, black columns). This is preliminary evidence that the probe design was sufficiently stringent to prevent detection of most transcripts from the other species.


Figure 1
View larger version (22K):
[in this window]
[in a new window]

  		FIGURE 1. Species-identical probe response to universal reference total RNA. Three independent replicates of XpressRef (SuperArray Bioscience) universal reference total RNA from mouse (A and C) or human (B and D) tissues were profiled with the Hu/Mu ProtIn microarray. Only probe sets were evaluated for all subsequent figures. The presence of a transcript was determined by the detection algorithm in MAS5 software (Affymetrix), which assigns P as a measure of the detection call confidence. The number of probe sets giving present call P values smaller than 0.05, 0.01, and 0.001, for all three replicates (A and B). In the presence of only mouse transcripts, there were extremely few present detection calls by human probe sets (A, gray columns), and vice versa (B, black columns), despite the high sequence similarity among many of the protease and protease inhibitor genes and their potential for cross-hybridization (see Table 1). Relative gene expression was calculated with dChip software. Distribution of the log 2 transformed relative gene expression (C and D). In the presence of mouse (C) or human (D) transcripts, the relative gene expression distribution measured by species-identical probes (C, black columns and D, gray columns) has a tail to the right of a normal-shaped distribution. This distribution likely results from a low level of gene expression and nonspecific hybridization (background) to probes both in the absence (C, gray columns and D, black columns) or the presence (left-hand bell-shaped distribution in C, black columns and D, gray columns) of species-identical transcripts. Frequency represents the number of probe sets that satisfy the selection criterion.

 
The distribution of relative gene expression measured by mouse (black columns) or human (gray columns) probe sets in response to either mouse or human Universal Reference RNA is shown in Fig. 1C and D, respectively. To normalize and quantify the raw expression data, we chose a popular data processing method called dChip (11). Within dChip, we chose the option to use only the PM probe information. This approach was chosen based on literature disputing the ability of mismatch (MM) probes to accurately assess cross-hybridization to PM probes (12-14). In the absence of species-identical RNA (gray columns in Fig. 1C and black columns in Fig. 1D), the distribution of relative gene expression seems normal (Gaussian) and is centered (mean ± SD: Fig. 1C, 5.6 ± 1.2; Fig. 1D, 5.4 ± 1.0) in the area of poor or very low signal. This likely represents background hybridization but may include some cross-species hybridization. In contrast, the presence of species-identical RNA (Fig. 1C, black columns and Fig. 1D, gray columns) results in a shift of the relative gene expression distribution to the right (mean ± SD: Fig. 1C, 7.7 ± 2.0; Fig. 1D, 6.9 ± 1.8), with the right-hand tail containing highly expressed genes. The gene expression distributions derived from mouse and human probe sets in the presence of either mouse or human RNA were clearly different and suggested that irrelevant probe sets (e.g., mouse probe sets in the presence of human RNA) did not exhibit significant cross-species hybridization. This was investigated further using DGE analysis coupled with false discovery analysis, as described below.

Hu/Mu ProtIn Probe Set Response to RNA Derived from Mouse or Human Mammary Carcinomas
To further validate the high fidelity of species-identical transcript hybridization to species appropriate probe sets on the Hu/Mu ProtIn microarray, we challenged the microarray with RNA derived from human or mouse mammary carcinomas (Fig. 2 ). These samples would be expected to have transcript signatures distinct from Universal Reference RNA. A detection call was calculated (MAS5 software) for the presence of a transcript for each probe set. The number of present calls giving P values smaller than 0.05, 0.01, or 0.001, in response to RNA derived from mouse and human mammary carcinomas, is shown in Fig. 2A and B, respectively. Many transcripts were detected by mouse (Fig. 2A, black columns) and human (Fig. 2B, gray columns) probe sets in response to mouse or human mammary carcinoma, respectively. In contrast, <2% of the present calls (P < 0.01) represented probe sets responding to other-species mammary carcinoma-derived RNA (Fig. 2A, gray columns and Fig. 2B, black columns). These data confirm that the probe design was sufficiently stringent to prevent detection of most transcripts from the other species (cf. Fig. 1A and B).


Figure 2
View larger version (22K):
[in this window]
[in a new window]

  		FIGURE 2. Species-identical probe response to transcripts derived from mouse or human mammary carcinoma biopsies. Total RNA derived from spontaneous mammary carcinomas from MMTV-PyMT+(FVB/n) (n = 8) and MMTV-PyMT+/uPARAP–/–(FVB/n) (n = 2) transgenic mice (A and C) or human breast ductal carcinoma biopsies (n = 18) from women with stage II or III disease (B and D) were profiled. Number of probe sets giving present call P values smaller than 0.05, 0.01, and 0.001 for ≥80% of the samples (A and B). Transcripts derived from mammary carcinoma were rarely detected by non–species-identical probes (A, gray columns and B, black columns). The relative gene expression distributions of species-identical (C, black columns and D, gray columns) and non–species-identical (C, gray columns and D, black columns) probe sets were shaped differently. Frequency represents the number of probe sets that satisfy the selection criterion.

 
The distribution of mouse and human mammary carcinoma transcript expression illustrates that gene expression distributions determined by relevant probe sets [mouse probe sets in the presence of mouse mammary carcinoma RNA (Fig. 2C, black columns) and human probe sets in the presence of human mammary carcinoma RNA (Fig. 2D, gray columns)] were clearly different compared with gene expression distributions derived from irrelevant probe sets (e.g., mouse probe sets in the presence of human RNA). Furthermore, these results were similar to those for universal reference RNA (Fig. 1C and D), providing additional evidence that cross-species hybridization, if present, was minimal.

The detection algorithm analyses shown above (Fig. 1A and B and Fig. 2A and B) suggest that most mouse or human probe sets on the Hu/Mu ProtIn microarray do not hybridize to transcripts from the opposite species sufficiently to trigger a present call even at a liberal P (0.05). Thus, the specificity of the probes seems to be sufficient to allow interrogation of a sample in which there is a mixture of transcripts from both species, permitting simultaneous analysis of human and mouse genes.

Expression Analysis of Orthotopic Xenografts of Human Carcinoma Cell Lines
To show the utility of the Hu/Mu ProtIn microarray in detecting and differentiating expression of genes in human xenograft tumors from those in the mouse tumor microenvironment, we profiled A549 human lung carcinoma cells and MDA-MB-231 human breast carcinoma cells grown in vitro and as orthotopic xenografts in mice. In addition, we profiled normal murine lung and mammary fat pads (i.e., the orthotopic sites for implantation of the lung and breast carcinoma cells, respectively). The number of probe sets giving P values smaller than 0.05, 0.01, and 0.001 for normal orthotopic sites, cultured human cells, and xenografts are shown in Fig. 3 . As was evident in the experiments reported above (Figs. 1 and 2), species-identical transcripts were detected in RNA derived from mouse lung (Fig. 3A, black columns), mouse mammary fat pads (Fig. 3B, black columns), cultured A549 cells (Fig. 3C, gray columns), and cultured MDA-MB-231 (Fig. 3D, gray columns) cells. In contrast, <2% of the present calls (at P < 0.01) represented probe sets detecting other-species transcripts (Fig. 3A and B, gray columns and Fig. 3C and D, black columns). These data support the evidence presented in Figs. 1 and 2 that probes on the Hu/Mu ProtIn microarray, in response to biologically diverse RNA, retained their intended species-identical fidelity despite interspecies homology of proteases, protease inhibitors, and protease interactors.


Figure 3
View larger version (20K):
[in this window]
[in a new window]

  		FIGURE 3. Frequency of transcript detection by mouse and human probe sets in samples derived from normal mouse lung and mammary fat pads, cultured human cell lines, and orthotopic xenografts. Total RNA derived from normal mouse lung (n = 3; A) and mammary fat pads (n = 4; B), cultures of human A549 lung adenocarcinoma cells (n = 2; C) and MDA-MB-231 breast carcinoma cells (n = 3; D), and orthotopically implanted xenografts of A549 cells (n = 3; E) and MDA-MB-231 cells (n = 3; F) was profiled. Number of probe sets giving present call P values smaller than 0.05, 0.01, and 0.001 for all replicates in each group. Non–species-identical probes rarely detect transcripts in mouse lung (A, gray columns), mouse mammary fat pads (B, gray columns), A549 cells (C, black columns), and MDA-MB-231 cells (D, black columns). Importantly, transcripts from A549 (E) and MDA-MB-231 (F) orthotopic xenografts were detected by mouse probes (black columns) and human probes (gray columns), suggesting that both mouse (host) and human (tumor) transcripts present in the xenograft can be detected simultaneously by the Hu/Mu ProtIn microarray. Frequency represents the number of probe sets that satisfy the selection criterion.

 
The presence of both mouse and human transcripts was clearly detected in A549 (Fig. 3E) and MDA-MB-231 (Fig. 3F) orthotopic xenografts, even at a highly stringent P (<0.001). At the P values tested, >35% of the present detection calls represented probe sets interrogating mouse transcripts in RNA derived from the human xenografts. The mouse transcripts represent gene expression in host (murine) cells that have either infiltrated the human tumor xenograft (e.g., cells composing the vasculature, fibroblasts, or immune cells) or were adjacent to the human tumor xenograft and removed at necropsy. The relative contribution of mouse RNA to the total RNA present in the xenograft sample is not known and will vary from sample to sample. The data do suggest that our microarray can detect mouse transcripts in the presence of human RNA and will be a useful tool for detecting host gene expression in xenograft models of human disease. The relative gene expression distributions for these samples are available as Supplementary Fig. S1.

Present Call Summary
A summary of the detection call data, organized by sample type, is shown in Table 2 . Greater than 98% of the present calls (P < 0.01) within each group were from probe sets and sample RNA of the same species (bold font). Probe sets rarely detected the presence of RNA from another species (≤2%, P < 0.01) in single-species samples. In the xenografts, by contrast, 38% to 54% of the present calls were from mouse probe sets, indicating the ability to detect mouse transcripts in a sample containing mouse and human RNA. A global summary of present call analyses of combined species-identical or other-species RNA experiments is shown in Table 3 . The analyses labeled as "same species" consider all probe set and RNA combinations from the same species (e.g., human RNA) and only probe sets designed to survey human transcripts. In contrast, the analyses labeled as "other species" consider combinations of either human RNA and "mouse" probe sets or mouse RNA and "human" probe sets. In both types of analyses, there are over 23,500 data points. The number and percentage of present detection calls are shown at the indicated P thresholds. Whereas numerous same-species transcripts were present even at stringent P values (<0.001), only 4% of the other-species combinations resulted in a present call at P < 0.05. Interestingly, only two mouse (Mm.1485_s_at and Mm.2287_s_at) and human (Hs.118797_s_at and Hs.178761_s_at) probe sets were present (P < 0.01) in all samples tested (human or mouse) despite their diverse biological origins. These probe sets survey ubiquitin conjugating enzymes and proteasome components. In contrast, 129 mouse and 122 human probe sets detected the presence of transcripts (P < 0.01) in all of the same-species samples tested; the genes detected by these probe sets were detected in <1% of the other-species comparisons (data not shown). Together, these detection call data emphasize the species specificity of the probe set design and suggest that the Hu/Mu ProtIn microarray is a useful tool for DGE analyses of either single- or dual-species samples.


View this table:
[in this window]
[in a new window]

  		Table 2. "Present Call" Summary

 

View this table:
[in this window]
[in a new window]

  		Table 3. Global Present Call Summary

 
DGE and False Discovery Rate Analyses
To further assess the performance of the microarray, we used DGE and false discovery rate (FDR) analyses. This was accomplished by comparing normal mammary tissue with mammary carcinoma for both mouse and human. We expected many differentially expressed genes in species-relevant situations (e.g., mouse probe sets and mouse samples). In contrast, for species-irrelevant situations (e.g., human probe sets and mouse samples), we expected few differentially expressed genes, or no more than can be anticipated by chance alone (based on FDR), assuming little or no cross-species hybridization.

The results of the DGE and FDR analyses are presented in Table 4 . The relative gene expression values for the data analyzed in Table 4 were generated using the dChip (PM only) algorithm (see Materials and Methods). As expected, many differentially expressed genes (Table 4, bold font) were observed in relevant species analyses (e.g., mouse probe sets interrogating the mouse sample comparison). In relevant species analyses, importantly, <6% (human) or 2% (mouse) of the differentially expressed genes (when combined with a fold change threshold) were estimated to be falsely discovered. When mouse probe sets were used to assess the human sample comparison, at all P and fold change combinations, only two or fewer differentially expressed genes were discovered. The corresponding FDR were low, suggesting that only a few mouse probe sets exhibited sufficient hybridization with human transcripts to elicit significant DGE (although this does not exclude the caveat that some human transcripts may cross-hybridize to mouse probes sets but are not differentially expressed). These data indicate that DGE analysis of dual-species samples by mouse probe sets is feasible.


View this table:
[in this window]
[in a new window]

  		Table 4. DGE and False Discovery Analyses

 
Surprisingly, many differentially expressed genes were discovered by human probe sets in the mouse sample comparison and, as indicated by a low FDR (<5%), were not due to chance variation in the data but rather to some mouse transcripts that cross-hybridize with the human probes. This result differs dramatically from the MAS5 present detection call data presented in Figs. 1, 2, and 3 and summarized in Tables 2 and 3, which show that human probe sets infrequently (<3% of the present detection calls at P < 0.01) detected the presence of mouse transcripts. This discrepancy may be due to just a few human probes (within a probe set) that hybridized to mouse transcripts sufficiently to reveal differential expression but do not trigger a significant present call. Consistent with this interpretation, masking of sequences above 16 bp in length was found to marginally improve both the sensitivity of mouse probe sets and the specificity of human probe sets (Table 5 ). Taken together, these data suggest that the choice of data processing and normalization algorithms may be critically important when processing human probe set data derived from samples that contain both mouse and human RNA.


View this table:
[in this window]
[in a new window]

  		Table 5. False Discovery Analysis of Mouse Samples That Were Processed by Different Methods

 
There are several popular data processing methods that normalize and quantify raw expression values for microarray data before any formal analysis [e.g., dChip (11), RMA and GCRMA (15), MAS5 (16), PDNN (13), and TM (17-19)]. We selected three of these methods {dChip, either with [dChip (PM + MM)] or without [dChip (PM only)] mismatch probe information; MAS5 (the default Affymetrix algorithm); and TM (using both 20% and 25% trimming)} to process the raw mouse sample data before DGE and FDR analyses. Our goal was to determine if there may be a more accurate and reliable data processing method than dChip (PM only) for processing murine samples that would also eliminate or substantially reduce the influence of the human probes that are apparently cross-hybridizing to mouse transcripts. This would be important when processing xenograft samples containing mouse and human RNA and for preventing overestimation of human gene expression due to mouse transcript cross-hybridization. In Fig. 4 , for each data processing method, the fraction of qualifying mouse (A) and human (B) probe sets, at a range of FDR values, are illustrated, as derived from comparing mouse mammary fat pads to mouse spontaneous mammary tumors. Of the processing methods tested, dChip (PM only) seems to provide enhanced sensitivity when analyzing probe sets and transcripts from the same species (Fig. 4A), as a higher proportion of probe sets were selected at low FDR thresholds. For example, at an FDR threshold of 0.05, dChip (PM only) noticeably out-performed all other methods that were tested in that ~50% of the mouse probe sets satisfied this selection threshold compared with ≤40% by the other methods. In contrast, when analyzing mouse samples with human probe sets (Fig. 4B), processing methods other than dChip (PM only) resulted in fewer qualifying human probe sets even at stringent FDR values. For example, when the data were processed with MAS5 or TM25, ≤2% of the human probe sets satisfied the selection criterion at a liberal FDR threshold of 0.1 (i.e., 10% of the differentially expressed genes will be falsely discovered). These results can be readily observed in tabular format (Table 5). dChip (PM only) seemed to be the more sensitive method for data processing when same-species analyses (e.g., mouse probes and RNA) were investigated. On the other hand, when non–identical-species analyses were done (i.e., human probes with mouse RNA), several data processing methods (e.g., MAS5 and TM25), other than dChip (PM only), provided more protection against cross-species hybridization, presumably due to reducing the influence of human probes that give outlying data by employing mismatch probe information. A limited improvement in dChip (PM only) results was suggested by analyses in which probes with a potential for cross-species hybridization were masked.


Figure 4
View larger version (14K):
[in this window]
[in a new window]

  		FIGURE 4. Sensitivity of five data processing methods. Mouse mammary fat pads (n = 4) and mammary carcinomas (n = 10) were profiled, and the raw data were processed by five previously published algorithms (see Materials and Methods for details). The resulting relative gene expression data from fat pads and carcinomas were compared using the t statistic to assess differential expression. FDR were estimated by 1,000 random permutations (see Materials and Methods). A. Data from mouse probe sets. B. Data from human probe sets.

 

   Discussion
Top
 Notes
 Abstract
 Introduction
 Results
 Discussion
Materials and Methods
 Acknowledgements
 References
 
The Hu/Mu ProtIn microarray is a unique discovery tool with an obvious and focused function to survey gene expression of proteases, their endogenous inhibitors, and selected proteins that interact with proteases. The Hu/Mu ProtIn microarray was designed to survey human and mouse gene expression simultaneously (or independently if desired). Consequently, the Hu/Mu ProtIn microarray may be used to detect host gene expression in mouse models of human disease, such as host protease expression within the tumor microenvironment of mouse xenograft models of human cancer.

There is no doubt that many proteases are dysregulated in human cancers (e.g., see ref. 20). Furthermore, proteases act in networks and pathways that involve multiple proteases, their inhibitors, and proteins that interact with them. To understand these complexities, it is necessary to identify which proteases and endogenous inhibitors are functional regulators within tumor microenvironments. Therefore, it is critically important to have capabilities enabling discovery not only of which proteases and endogenous inhibitors are present in the tumor microenvironment but also whether they derive from the tumor or the host cells infiltrating and regulating the tumor. This information will greatly facilitate targeted drug design and strategies against proteolytic pathways. The Hu/Mu ProtIn microarray is the first profiling discovery tool to possess the ability to not only assess the gene expression of proteases, protease inhibitors, and protease interactors in the tumor microenvironment but to also discriminate the species origin of these transcripts in xenograft models. Thus, the Hu/Mu ProtIn microarray will be an important discovery tool for use with orthotopic xenograft models of human cancers.

The Hu/Mu ProtIn microarray has several advantages over the alternative of using multiple arrays (one for mouse and one for human gene expression) to assess both host (mouse) and tumor (human) gene expression in xenograft models. First, the probe design for the Hu/Mu ProtIn microarray included consideration of probe sequence homologies for both mouse and human transcripts, whereas arrays designed to survey one species logically only focus on the appropriate species. Consequently, probes on single-species arrays are likely to cross-hybridize to transcripts from the other species in dual-species samples. This is especially true for highly conserved families of genes such as proteases and protease inhibitors. Probe sequences for the Hu/Mu ProtIn microarray were chosen based on having as little theoretical cross-species transcript hybridization as feasible. Despite our efforts, due to the high sequence conservation between mouse and human proteases and protease inhibitors, some of the probes on our array have the potential to hybridize with transcripts of the other species (Table 1). The data presented here, however, empirically show that the limited amount of cross-species hybridization can be controlled by a judicious selection of data processing methods (see Tables 2 and 5). Second, the Hu/Mu ProtIn microarray conserves time, effort, and expense. Only one chip is needed per xenograft sample compared with two for an alternative method. Third, by processing one array per xenograft sample, the chance for introducing human error is less than when processing two chips.

To our knowledge, the Hu/Mu ProtIn microarray is the first commercially available (Affymetrix) oligonucleotide array that contains, on a single platform, probes for both mouse and human genes. Overall et al. have created an in-house spotted array, called the CLIP-CHIP for human proteases and inhibitors (21), and have recently reported a similar array for mouse proteases and inhibitors (22). The Hu/Mu ProtIn microarray includes probes for proteases and inhibitors as well as genes whose protein products interact with proteases or inhibitors. Among the interactors are receptors involved in trafficking and cell surface localization of proteases, transcription factors that enhance expression of a number of proteases, and protease-associated proteins involved in endocytosis of matrix substrates. Thus, the Hu/Mu ProtIn microarray is suitable for assessing not only expression of proteases and inhibitors but also for assessing expression of proteins that affect proteolytic pathways. The Hu/Mu ProtIn microarray is the first array that is capable of simultaneous (on the same platform) survey of both mouse (host) and human (tumor) gene expression in orthotopic xenograft models.

To control for variation in the proportion of human and mouse RNA across xenograft replicates, we advocate normalizing mouse and human probe sets separately. With this normalization method, FDR did not exceed 1% for differentially expressed genes [at least 30% of the probe sets qualify at P < 0.01 (t test) with ≥2-fold change in expression] in xenografts compared with either normal orthotopic tissues (mouse expression) or three-dimensional matrix overlay cultures of the grafted human cells (data not shown). With this data normalization method and employing the data processing algorithms presented here, we have used the Hu/Mu ProtIn microarray to discover host protease and protease inhibitor gene expression in xenografts of human lung carcinoma cells (23) and human breast carcinoma cells.12 We were able to identify increases in expression of a number of human and mouse proteases and protease inhibitors in the human lung carcinoma xenografts. One of the mouse proteases, matrix metalloproteinase-12, was localized to cells that had infiltrated into the tumor; subsequent studies in matrix metalloproteinase-12–null mice revealed a protective role for stromal matrix metalloproteinase-12 in growth of lung tumors (24). Thus, the Hu/MuProtIn microarray chip has already shown its value in identifying tumor-host interactions that affect tumor development.

In the present study, we began the task of masking individual probes that may potentially cross-hybridize (see Table 5), noting the performance of the microarray following individual probe level masking based on various thresholds of potential cross-species hybridization. There remains a considerable opportunity for refining this approach and hence to bioinformatically improve both chip sensitivity and specificity.

In the present study, we have introduced a unique dual-species (human and mouse) oligonucleotide array, the HuMuProtIn chip. We have shown that the chip has excellent fidelity with minor cross-species hybridization that was controlled by employing different data processing methods depending on the type of data analysis. Therefore, we believe that the Hu/Mu ProtIn microarray is an ideal tool for gene profiling, particularly profiling of xenografts, and will allow comparative studies to be done on one platform to select mouse models that will most closely recapitulate human diseases.


   Materials and Methods
Top
 Notes
 Abstract
 Introduction
 Results
 Discussion
 Materials and Methods
Acknowledgements
 References
 
Array Design
All known human and mouse protease, protease inhibitor, and selected protease interactor mRNA sequences were harvested from the National Center for Biotechnology Information database and submitted to Affymetrix for probe design. Probe design was based on Blast(n) analysis of the human and mouse genomes to minimize cross-species hybridization. This eliminated some genes from further consideration because of the inability to design probe sets that could differentiate orthologues. Due to the space limitations of the Affymetrix custom microarray platform, not all known human and mouse proteases and protease inhibitors could be included. Genes on the microarray were selected by participants in the Protease Consortium (25) and a Department of Defense Breast Cancer Center of Excellence grant13 on the basis of known or putative biological significance to cancer. Each gene is interrogated by a probe set that contains 16 probe pairs of the typical PM and MM design.14

RNA Sources
Universal reference human (GA-004) and mouse (GA-005) total RNA was purchased from SuperArray Bioscience. Three independent aliquots of each reference RNA were processed and hybridized separately.

Ten spontaneous murine mammary carcinomas were obtained from MMTV-PyMT+(FVB/n) transgenic mice (n = 4; L. Matrisian's laboratory) or MMTV-PyMT+(FVB/n) transgenic mice (n = 4; T. Bugge's laboratory) and MMTV-PyMT+/uPARAP–/–(FVB/n) transgenic mice (n = 2; T. Bugge's laboratory) by macroscopic dissection. As controls, normal mouse mammary fat pads (n = 4; female Rag-1 tm1Mom retired breeders at >6 weeks after lactation) were used. Samples harvested from individual mice were processed and profiled independently.

Human breast carcinoma biopsies were obtained from discarded tissue at the time of surgical removal of the tumor, following informed consent. The experimental protocol was approved by the Wayne State University Institutional Review Board. Only biopsies from patients with ductal histology and either stage II or III disease were profiled (n = 18). Normal breast tissue biopsies (n = 13) were obtained from discarded uninvolved tissue from cancer patients or from women undergoing reduction mammoplasty.

MDA-MB-231 human breast carcinoma cell line was obtained from the American Type Culture Collection. In vitro cultures of MDA-MB-231 (n = 3) were profiled. Methods for the culture of these cells are available online.15 Orthotopic xenografts of MDA-MB-231 cells were harvested by gross surgical dissection 5 to 6 weeks following implantation into abdominal mammary fat pads of Rag-1tm1Mom female mice (The Jackson Laboratory). The mice were irradiated with 600 rad 2 to 4 h before implantation. Mice were injected bilaterally with 1 x 106 cells in 50 µL of PBS containing 5 mg/mL Matrigel (Invitrogen) into the abdominal fat pads by s.c. injection at the base of the nipple. At 5 to 6 weeks of growth, the average diameter of the xenograft was <1 cm, and the encapsulated tumors had not invaded into surrounding tissues. At this stage, it was easier to remove the surrounding host mammary fat pads, which are relatively acellular. Three independent xenografts were evaluated. All animals were maintained in accordance with the guidelines of the Committee for Protection of Animal Subjects at Wayne State University.

Sample preparation of cultured A549 human lung adenocarcinoma cells (n = 2), A549 orthotopic xenografts (n = 3), and normal mouse lung tissue (n = 3; male Rag-2 null/C57Bl/6) were reported elsewhere (24).

RNA Isolation
All tissue samples were immediately placed in RNALater (Ambion) and stored for at least 24 h at 4°C before removing the preservative and freezing the sample at –80°C. Tissue (50-100 mg) was transferred to a QIAshredder column tissue homogenizer (Qiagen) and briefly centrifuged (1,000 x g) at room temperature to remove residual RNALater. The tissue was removed from the QIAshredder column and transferred to a nuclease-free microcentrifuge tube containing 1 mL Trizol Reagent (Invitrogen) on ice. Isolation of total RNA from Trizol proceeded as recommended by the manufacturer. Following isolation, RNA was further purified with an RNeasy Mini-Kit (Qiagen), which included DNA removal using an Rnase-Free Dnase set (Qiagen).

Cultured cells were handled as follows. The medium was discarded, and the monolayers were washed with Dulbecco's PBS without calcium or magnesium (Invitrogen) thrice. The residual wash was aspirated, and the cells were removed directly into 1 mL Trizol Reagent per 35-cm2 dish using a plastic cell scraper. Cell homogenates were collected and allowed to incubate at room temperature for 10 min. The homogenates were stored at –80°C until RNA isolation, which followed the manufacturer's recommendations. Following isolation, RNA was further purified with an RNeasy Mini-Kit (Qiagen), which included DNA removal using an RNase-Free DNase (Qiagen).

Concentration and yield of RNA samples were determined by using a NanoDrop ND-1000 Spectrophotometer (NanoDrop Technologies). Samples with total RNA yields of <5 µg were not analyzed further. RNA integrity was determined by analysis on an Agilent 2100 Bioanalyzer (Agilent Technologies) following the manufacturer's recommendations. Samples with rRNA ratios <1 were not profiled.

RNA Processing, cRNA Production, and cRNA Labeling
The processing of RNA for hybridization, including cRNA synthesis and labeling, was by standard protocols described in the Affymetrix GeneChip Expression Analysis Technical Manual (version 2, section 2, chapters 1-3, 2004).

Microarray Processing Methods
Hybridization of the labeled and fragmented cRNA to the Hu/Mu ProtIn microarray and subsequent staining, washing, and scanning of the arrays strictly adhered to standard protocols described in the Affymetrix GeneChip Expression Analysis Technical Manual (version 2, section 2, chapters 1-3, 2004).

Microarray Data Processing Methods
The following data processing methods were used: dChip (11), Microarray Suite version 5 (MAS5, Affymetrix; ref. 16), and TM (17, 19). Unless stated otherwise, relative gene expression values were generated with dChip using the option to ignore the mismatch (MM) probe information [dChip (PM only)]. This method was chosen based on the controversy in the literature indicating that a mismatch probe may not be able to accurately assess cross-hybridization that may be occurring with its perfect match probe (12-14). To compare arrays, the raw data from each array were normalized using the invariant set normalization algorithm within the dChip software. Gene expression values were calculated using the Model-Based Expression Index. For some experiments, normalization and Model-Based Expression Index was done with PM and MM probe information (dChip PM + MM). In some experiments, gene expression values using a trimmed-mean algorithm (TM) followed by quantile normalization to a standard array were calculated as previously reported (18). TM was done by removing either 20% (TM20) or 25% (TM25) of the PM-MM probe pair values within each probe set before calculating the gene expression value. All gene expression values were log 2 transformed before further analysis. Gene expression values generated with TM were transformed by log 2 (Max [x + 40,0] + 40).

To generate a P for the likelihood of the presence of a transcript, the default detection algorithm in Affymetrix MAS5 software was used, which employs the one-sided Wilcoxon's signed rank test. To calculate the frequency of a present detection call, 100% of the universal reference RNA, xenograft, orthotopic site, and cultured cell replicates or ≥80% of the mouse mammary carcinoma or human breast carcinoma replicates were required to have a present detection call at a particular P (>0.05, >0.01, and >0.001).

Computational Methods
To determine the intraspecies and interspecies theoretical cross-hybridization for all probes on the Hu/Mu ProtIn microarray, each probe sequence was blasted against the human and mouse repeat-masked genome in the Ensembl database online.11 If the blast returned a run of ≥16 continuous (of 25) bases identical to another gene, the probe was considered to have the potential to cross-hybridize to an inappropriate transcript under the standard Affymetrix hybridization conditions. Probe-level mask files were generated from this data and used to determine the effect of cross-hybridizing probes on chip sensitivity and specificity.

FDR analysis estimates, by permutation testing, the probability of selecting false-positive genes (i.e., equally expressed genes in the rejection population; ref. 23). FDRs were determined as previously reported (26). Briefly, log 2 transformed normalized gene expression data were used. Mouse spontaneous mammary carcinoma samples (n = 10) were compared with normal mouse mammary fat pad samples (n = 4), or human breast carcinoma (n = 18) were compared with normal human breast tissue (n = 13). Sample labels were randomly permuted 1,000 times. Following each permutation, the number of differentially expressed genes at the Student's t test P thresholds of <0.05, <0.01, and <0.001 with or without a mean fold change ratio (at either 1.5 or 2) were recorded. If the real sample groups were re-created by any random permutation of the sample labels, that permutation was excluded from the FDR analysis. FDR was estimated as the average number of qualifying probe sets from the permutated data normalized by the number of qualifying probe sets in the actual data.


   Acknowledgements
Top
 Notes
 Abstract
 Introduction
 Results
 Discussion
 Materials and Methods
 Acknowledgements
References
 
We thank the Applied Genomics Technology Core at Wayne State University, especially Dan Lott, for sample and microarray processing; Dr. Sorin Draghici (Bioinformatics Core, Karmanos Cancer Institute) for harvesting probe sequences from National Center for Biotechnology Information; and Dr. Kerby Shedden and Rork Kuick for numerous insightful discussions and access to the False Discovery Rate software (26).


   Notes
Top
 Notes
Abstract
 Introduction
 Results
 Discussion
 Materials and Methods
 Acknowledgements
 References
 
Grant support: Department of Defense Breast Cancer Center of Excellence grant DAMD17-02-1-0693 (B.F. Sloane), National Institute of Dental and Craniofacial Research Intramural Program (T.H. Bugge), and NIH grants P50 CA90949 (L.M. Matrisian) and P50 CA58207 (L.M. Coussens). Applied Genomics Technology Core and the Bioinformatics Core are supported in part by grant P30 CA22453 to the Karmanos Cancer Institute.

The costs of publication of this article were defrayed in part by the payment of page charges. This article must therefore be hereby marked advertisement in accordance with 18 U.S.C. Section 1734 solely to indicate this fact.

Note: Supplementary data for this article are available at Molecular Cancer Research Online (http://mcr.aacrjournals.org/).

D.R. Schwartz and K. Moin contributed equally to this work.

The microarray data are deposited at Gene Expression Omnibus-National Center for Biotechnology Information under accession no. GSE6413.

10 http://merops.sanger.ac.uk/ Back

11 http://www.ensembl.org/Multi/blastview Back

12 Article in preparation. Back

13 http://bccoe.med.wayne.edu/bccoe/index.jsp Back

14 For more information, visit http://www.affymetrix.com/index.affx Back

15 http://www.atcc.org Back

Received 10/ 6/06; revised 1/24/07; accepted 2/22/07.


   References
Top
 Notes
 Abstract
 Introduction
 Results
 Discussion
 Materials and Methods
 Acknowledgements
 References
 

  1. DeClerck YA, Mercurio AM, Stack MS, et al. Proteases, extracellular matrix, and cancer: a workshop of the path B study section. Am J Pathol 2004;164:1131–9.[Abstract/Free Full Text]
  2. Lynch CC, Matrisian LM. Matrix metalloproteinases in tumor-host cell communication. Differentiation 2002;70:561–73.[CrossRef][Medline]
  3. van Kempen LC, de Visser KE, Coussens LM. Inflammation, proteases and cancer. Eur J Cancer 2006;42:728–34.[Medline]
  4. Coussens LM, Fingleton B, Matrisian LM. Matrix metalloproteinase inhibitors and cancer: trials and tribulations. Science 2002;295:2387–92.[Abstract/Free Full Text]
  5. Podgorski I, Sloane BF. Cathepsin B and its role(s) in cancer progression. Biochem Soc Symp 2003;(70):263–76.[Medline]
  6. Rawlings ND, Morton FR, Barrett AJ. MEROPS: the peptidase database. Nucleic Acids Res 2006;34:D270–2.[Abstract/Free Full Text]
  7. Abrahamson M, Alvarez-Fernandez M, Nathanson CM. Cystatins. Biochem Soc Symp 2003;(70):179–99.[Medline]
  8. van Gent D, Sharp P, Morgan K, Kalsheker N. Serpins: structure, function and molecular evolution. Int J Biochem Cell Biol 2003;35:1536–47.[CrossRef][Medline]
  9. Murphy G, Knauper V, Lee MH, et al. Role of TIMPs (tissue inhibitors of metalloproteinases) in pericellular proteolysis: the specificity is in the detail. Biochem Soc Symp 2003;(70):65–80.[Medline]
  10. Bucca G, Carruba G, Saetta A, Muti P, Castagnetta L, Smith CP. Gene expression profiling of human cancers. Ann N Y Acad Sci 2004;1028:28–37.[Abstract/Free Full Text]
  11. Li C, Wong WH. Model-based analysis of oligonucleotide arrays: expression index computation and outlier detection. Proc Natl Acad Sci U S A 2001;98:31–6.[Abstract/Free Full Text]
  12. Irizarry RA, Bolstad BM, Collin F, Cope LM, Hobbs B, Speed TP. Summaries of Affymetrix GeneChip probe level data. Nucleic Acids Res 2003;31:e15.[Abstract/Free Full Text]
  13. Zhang L, Miles MF, Aldape KD. A model of molecular interactions on short oligonucleotide microarrays. Nat Biotechnol 2003;21:818–21.[CrossRef][Medline]
  14. Millenaar FF, Okyere J, May ST, van Zanten M, Voesenek LA, Peeters AJ. How to decide? Different methods of calculating gene expression from short oligonucleotide array data will give different results. BMC Bioinformatics 2006;7:137.[CrossRef][Medline]
  15. Wu Z, Irizarry R, Gentleman R, Murillo F, Spencer F. A model-based background adjustment for oligonucleotide expression arrays. Baltimore (MD): John Hopkins University, Department of Biostatistics; 2003.
  16. Hubbell E, Liu WM, Mei R. Robust estimators for expression analysis. Bioinformatics 2002;18:1585–92.[Abstract/Free Full Text]
  17. Giordano TJ, Shedden KA, Schwartz DR, et al. Organ-specific molecular classification of primary lung, colon, and ovarian adenocarcinomas using gene expression profiles. Am J Pathol 2001;159:1231–8.[Abstract/Free Full Text]
  18. Schwartz DR, Kardia SL, Shedden KA, et al. Gene expression in ovarian cancer reflects both morphology and biological behavior, distinguishing clear cell from other poor-prognosis ovarian carcinomas. Cancer Res 2002;62:4722–9.[Abstract/Free Full Text]
  19. Rickman DS, Bobek MP, Misek DE, et al. Distinctive molecular profiles of high-grade and low-grade gliomas based on oligonucleotide microarray analysis. Cancer Res 2001;61:6885–91.[Abstract/Free Full Text]
  20. Sloane BF, Yan S, Podgorski I, et al. Cathepsin B and tumor proteolysis: contribution of the tumor microenvironment. Semin Cancer Biol 2005;15:149–57.[CrossRef][Medline]
  21. Overall CM, Tam EM, Kappelhoff R, et al. Protease degradomics: mass spectrometry discovery of protease substrates and the CLIP-CHIP, a dedicated DNA microarray of all human proteases and inhibitors. Biol Chem 2004;385:493–504.[CrossRef][Medline]
  22. Overall CM, Dean RA. Degradomics: Systems biology of the protease web. Pleiotropic roles of MMPs in cancer. Cancer Metastasis Rev 2006;25:69–75.[Medline]
  23. Benjamini Y, Hochberg Y. Controlling the false discovery rate: a practical and powerful approach to multiple testing. J R Statist Soc Ser B 1995;57:289–300.
  24. Acuff HB, Sinnamon M, Fingleton B, et al. Analysis of host- and tumor-derived proteinases using a custom dual species microarray reveals a protective role for stromal matrix metalloproteinase-12 in non-small cell lung cancer. Cancer Res 2006;66:7968–75.[Abstract/Free Full Text]
  25. Giranda VL, Matrisian LM. The protease consortium: an alliance to advance the understanding of proteolytic enzymes as therapeutic targets for cancer. Mol Carcinog 1999;26:139–42.[Medline]
  26. Shedden K, Chen W, Kuick R, et al. Comparison of seven methods for producing Affymetrix expression scores based on False Discovery Rates in disease profiling data. BMC Bioinformatics 2005;6:26.[CrossRef][Medline]



This article has been cited by other articles:


Home page
 CarcinogenesisHome page
M. J. Sinnamon, K. J. Carter, L. P. Sims, B. LaFleur, B. Fingleton, and L. M. Matrisian
A protective role of mast cells in intestinal tumorigenesis
Carcinogenesis, April 1, 2008; 29(4): 880 - 886.
[Abstract] [Full Text] [PDF]

This Article
Right arrow Abstract Freely available
Right arrow Full Text (PDF)
Right arrow Supplementary Data
Right arrow Alert me when this article is cited
Right arrow Alert me if a correction is posted
Services
Right arrow Similar articles in this journal
Right arrow Similar articles in PubMed
Right arrow Alert me to new issues of the journal
Right arrow Download to citation manager
Right arrow reprints & permissions
Citing Articles
Right arrow Citing Articles via HighWire
Right arrow Citing Articles via Google Scholar
Google Scholar
Right arrow Articles by Schwartz, D. R.
Right arrow Articles by Sloane, B. F.
Right arrow Search for Related Content
PubMed
Right arrow PubMed Citation
Right arrow Articles by Schwartz, D. R.
Right arrow Articles by Sloane, B. F.

HOME HELP FEEDBACK SUBSCRIPTIONS ARCHIVE SEARCH TABLE OF CONTENTS
Cancer Research Clinical Cancer Research
Cancer Epidemiology Biomarkers & Prevention Molecular Cancer Therapeutics
Molecular Cancer Research Cancer Prevention Research
Cancer Prevention Journals Portal Cancer Reviews Online
Annual Meeting Education Book Cell Growth & Differentiation