| HOME | HELP | FEEDBACK | SUBSCRIPTIONS | ARCHIVE | SEARCH | TABLE OF CONTENTS |
Regular Articles |
1 Ludwig Institute for Cancer Research and2 Hospital do Câncer A.C. Camargo;3 BioInfo and Instituto de Matemática e Estatística da Universidade de São Paulo; and4 SENAC College of Computer Science and Technology, São Paulo, Brazil
 |
ABSTRACT |
|---|
 Top ABSTRACT INTRODUCTIONMATERIALS AND METHODSRESULTSREFERENCES |
|---|
|
INTRODUCTION |
|---|
|
TopABSTRACTINTRODUCTIONMATERIALS AND METHODSRESULTSREFERENCES |
|---|
Gastric adenocarcinoma represents >95% of all gastric tumors and, following Lauren’s classification (3) , it can be divided into intestinal and diffuse type, according to tumor histology. These two histological types have a distinct pathology, epidemiology, and etiology. The intestinal type is more frequent and represents the dominant histological type in areas where stomach cancer is epidemic, suggesting an environmental etiology. The pathogenesis of intestinal type adenocarcinoma has been connected to precursor changes such as chronic active gastritis, multifocal atrophic gastritis, intestinal metaplasia, and dysplasia as proposed by Correa and Chen (4) , and with the presence of Helicobacter pylori infection (5) . In contrast, the diffuse type has not been related to precursor lesions and has a higher association with familial occurrence. It is widely accepted that genetic alterations in the CDH1 (e-cadherin) gene play an important role in the oncogenesis of the diffuse-type gastric cancer (6) .
The relationship between intestinal metaplasia and intestinal-type gastric adenocarcinomas has not been fully established. Some individuals with intestinal metaplasia will never develop gastric cancer. The molecular events related to progression from metaplasia to adenocarcinomas remains unknown. Certain molecular alterations have been associated with intestinal metaplasia. For example, mutation in p53 was detected in intestinal metaplasia adjacent to gastric tumors (7) . Overexpression of cyclooxygenase-2 was detected in intestinal metaplasia-associated gastritis (8) , and also detected in gastric cancer and in other tumors like colon cancer (9) . Kang et al. (10) detected promoter hypermethylation in the DNA mismatch repair gene (hMLH1) in intestinal metaplasia, and this epigenetic alteration is related to microsatellite instability. This author also describes hypermethylation of other genes, including p16, DAP-kinase, THBS1, and TIMP-3. Recently, Boussioutas et al. (11) described the expression profile of 124 gastric mucosa representing gastritis, intestinal metaplasia, and adenocarcinomas, and identified a series of genes that are typically expressed in the intestinal type and could be related to tumor progression. Together, these findings demonstrate that some of the molecular events associated with intestinal metaplasia can also be detected in cancer sample and, hence, comparing the molecular alterations between nonmalignant and malignant lesions could lead to the identification of genes involved in gastric carcinogenesis. Moreover, the identification of expression signatures that correlate with adenocarcinomas could be used for follow-up of patients with intestinal metaplasia and assessment of the risk of the premalignant stages becoming malignant.
In this work, we used a cDNA array with 376 genes, including those 141 genes described previously by our group (12) plus other genes known to be generally altered in human cancers. We determined the expression profile in 99 tissue samples representing normal gastric mucosa, as well as gastritis, intestinal metaplasia, and adenocarcinoma of the stomach. Using Fisher’s linear discriminant analysis, we identified a series of molecular classifiers that could distinguish between cancer and noncancer samples. Importantly, we also identify a series of intestinal metaplasias of which the gene expression profile resembles that of adenocarcinoma.
|
MATERIALS AND METHODS |
|---|
|
TopABSTRACTINTRODUCTIONMATERIALS AND METHODSRESULTSREFERENCES |
|---|
. Samples labeled as "GF" or "GH" were obtained from surgery, and the majority came from patients with gastric adenocarcinomas. Samples labeled as "BIO" were obtained by endoscopic biopsy, and the majority is from patients with only the indicated pathology.
|
Production of cDNA Arrays.
We constructed a cDNA array with 376 genes including 141 clones of genes described previously as being altered in gastric cancer (12)
and 235 genes known to be altered in human cancers on the basis of available literature (complete gene list is available).5
All of the clones correspond to ORESTES fragments derived from the Fundação de Amparo à Pesquisa do Estado de São Paulo/Ludwig Institute for Cancer Research Human Cancer Genome Project and were sequence verified. Whenever possible, each gene is represented by two clones corresponding to different regions of the complete cDNA and each cDNA clone was printed in triplicates onto nylon membranes (Flexys robot; Genomic Solutions, Ann Arbor, MI) making a total of 2400 spots. Production of the cDNA array, labeling, hybridization, and detection of signals were carried out as described previously (12)
. For each tissue sample, 25 µg of total RNA were radioactively labeled with [
-33P]dCTP (3000 Ci/mmol; Amersham, Piscataway, NJ) and hybridized against a nylon-based cDNA array. Data acquisition was performed with the ArrayVision software (Amersham), using gel files.
Statistical Analysis.
Data analysis was performed using R,6
an open source interpreted computer language for statistical computation and graphics, and tools from the Bioconductor project,7
adapted to our needs. Principal component analysis was performed using TMEV (13)
. After image acquisition and quantification (see above), spots with signal lower or equal to background were identified and excluded from the analysis. Next, background-subtracted spot intensities were normalized by global mean normalization procedure (14
, 15)
. Replica spots representing the same gene were identified, and average signal intensity was determined.
Next, we searched our data for differentially expressed genes in the four clinical conditions. We used a nonparametric test (Mann-Whitney) to determine the P for each individual gene in each pair-wise comparison. For display purpose, we highlighted genes with P 
0.0009 in Figs. 1
and 3
. For all of the pair-wise comparisons (Figs. 1
and 3
;Table 2
), genes that are overexpressed in the second entity have the [-log2(P)] preceded by a minus signal. For clustering, we selected the nonredundant set of 6 genes with lowest P for each comparison and used the resulting 18 genes for clustering samples into four groups using the nonsupervised algorithm k-means. Once clusters were obtained, samples were organized hierarchically, based on their correlation distances (16)
. For classifiers, we use Fisher’s linear discriminant analysis and made exhausted search of the entire dataset for trios of genes such that data points representing signal intensity for all 3 of the genes for each sample would be separated by a plane in a three-dimensional space. More precisely, for a given group of genes, this linear classification method searches for linear combinations of their expressions with large ratios of between-groups to within-groups sum of squares (16)
. This maximal ratio of sum of squares, or its square root, which is denoted here by SVD (singular value decomposition), measures how well separated the two groups are. For the search of trios, the 99 samples dataset was split into two groups, one with 65 samples, used as learning set and a second, independent group, with 34 samples used for validation. Using the learning set of samples (65 samples) we performed an exhaustive search for the best classification trios for each one of the six comparisons of interest among normal, gastritis, metaplasia, and tumor. Trios were ranked according their SVD. In the case of normal versus tumor, where many trios were available, we only considered trios with perfect classification (1044 trios). Next, the best classifiers were tested with the remaining 34 samples with the results presented here.
|
|
|
For the construction of the stomach tissue microarray, new sections were obtained from the representative paraffin blocks, and all of the H&E-stained slides for these cases were reviewed. A slide with representative condition was selected from each case, and an area of one of the studied groups was circled on the slide. The corresponding formalin-fixed, paraffin-embedded blocks were retrieved, and the area corresponding to the selected area on the slide was circled on the block with a felt marker for tissue microarray construction. Using a tissue microarrayer (Beecher Instruments, Silver Spring, MD), the area of interest in the donor paraffin block was cored twice with a 0.6-mm diameter needle and transferred to a recipient paraffin block. Sections of 4 µm were cut from stomach tissue microarray block, deparaffined, dehydrated, and submitted to immunohistochemistry with a polyclonal antibody against metalloproteinase 2 (Oncogene; clone 75–7f7; dilution 1:40). For determining arbitrary units, we gave a score (1, 2, 3, 4) for intensity of staining plus a score (1, 2, 3, 4) for the percentage of positive cells in each tissue spot. Each tissue sample was spotted in duplicate in the slide, and four slides were measured, making a total of eight spots for each tissue sample. All four of the slides were analyzed by two independent pathologists and, for each sample, we sum the two scores (intensity and percentage) for all eight spots gave by each pathologist. Arbitrary units correspond with the average score for each tissue sample (average of 16 determinations for the corresponding eight spots).
|
RESULTS |
|---|
|
TopABSTRACTINTRODUCTIONMATERIALS AND METHODSRESULTSREFERENCES |
|---|
we represent all of the genes and their respective Ps for each pair-wise comparison, indicating in color those with Ps 0.0009. Green color was used for genes with higher expression in the first entity of the comparison, and red color was used for genes with higher expression in the second entity of the comparison. As expected, there were more differentially expressed genes when tumor samples were compared with normal, gastritis, or intestinal metaplasia and, conversely, fewer genes with statistically significant differences could be identified when we compare intestinal metaplasia with normal or gastritis. Finally, no statistically significant differences (P 0.0009) could be observed in the expression profile when we compared normal versus gastritis. If we consider genes with P < 0.05, a larger number of genes differentially expressed (153 genes) could be identified. In Table 2
, we describe the identity of the 42 most differentially expressed genes for each comparison. For convenience, instead of showing the Ps, we present minus the logarithm of the P, which is more directly associated with the significance as it is large for large significance (small Ps), with a minus signal indicating whether the mean expression is higher on the second condition being compared.
We next used a nonsupervised method of clustering to determine whether the 6 genes with lowest Ps for each comparison would be capable of grouping samples based on their expression profiles. Using the 
-means algorithm (16)
, samples were grouped in four clusters on the basis of the expression profile of 18 genes that were nonredundant among the 6 genes with lowest Ps for all of the comparisons except normal x gastritis (Fig. 2)
. In the first group, the majority of samples representing normal gastric mucosa (green labels) or gastritis (blue labels) were clustered together with only one tumor sample (red labels) and four samples representing metaplasia (brown labels). The second cluster is composed by the majority of samples corresponding to metaplasia plus two normal and three gastritis. The third cluster is composed of a mix of all four of the tissue classes in which a higher expression of the c-Myc oncogene and a lower expression of the CDKN1A gene could be detected. The fourth cluster corresponds with the vast majority of tumor samples, either of the intestinal or the diffuse type plus one gastritis and one metaplasia. Interestingly, when we did principal component analysis along all 99 samples using the expression profile of all 370 genes, the group of 6 samples comprising the third cluster, representing all four of the tissue classes, is detached from the remaining 93 samples, additionally corroborating their unique gene expression profile (figure available elsewhere).5
|
, the fourth cluster has the majority of tumors (25 of 28), with the majority of intestinal type tumors at the bottom branch (14 of 18). Because we could observe this dichotomy in the clustering of two types of gastric adenocarcinomas, we searched for genes differentially among them as compared with normal gastric mucosa (Fig. 3)
. Genes with Ps 0.0009 (Mann-Whitney Test) are denoted in green (for only one comparison) or in red (for both comparisons). Among the genes with augmented expression in the diffuse type adenocarcinomas we identified matrix metalloproteinase (MMP2), and its overexpression could contribute to the phenotypic characteristics of this tumor. Of notice, expression of the VHL gene was diminished on intestinal type adenocarcinomas.
Tissue Expression of MMP2.
We found MMP2 to be expressed in higher levels in both intestinal and diffuse types of adenocarcinomas (Figs. 4A)
. To confirm this observation, we generated a tissue array having 75 samples of diffuse and 75 samples of intestinal types of gastric adenocarcinomas, spotted in duplicates, plus 25 samples of normal gastric mucosa, 50 samples of chronic gastritis, and 25 samples of intestinal metaplasia (total of 500 tissue fragments). In agreement with mRNA levels, both diffuse- and intestinal-type adenocarcinomas showed stronger staining for MMP2 as compared with other entities (Fig. 4, B and C)
with a statistically significant difference when we compared normal and all tumor samples (P of 0.036 and 0.034 for Mann-Whitney and t test, respectively). The fact that the majority of the samples used for the stomach tissue microarray were different from those used as source of RNA for gene expression increases the significance of these findings. In the case of diffuse-type adenocarcinomas, we observed that signet ring cells failed to express MMP2 (see panel SRC in Fig. 4B
). A pictorial case stained with a polyclonal antibody against CTNNB1 is available at the provided website.5
This figure clearly demonstrates not only the augmented expression but also changes in cellular localization going from the cell surface in the normal sample to the nucleus in both diffuse and intestinal types of adenocarcinomas (17
, 18)
.
|
0.0009 in the extreme comparison i.e., normal versus tumor of the intestinal type (tumors of the diffuse type were excluded from this comparison). The top 7 genes with increased and the top 7 genes with decreased expression are represented in Fig. 5
. The 7 selected genes with increased expression pattern were COL1A1, FN1, CTSB, COL1A2, Hs.177781, DAF, and VIM. The 7 selected genes with decreased expression pattern were PRPF8, Hs.327751, VHL, LCK, BAD, VEGFB, and POLR2H.
|
. We identified 1,044, 17, and 1 trios of genes that could precisely separate all of the normal, gastritis, and intestinal metaplasia from tumor samples, respectively, with perfect distinction of all of the samples. When we accepted one sample misclassified, we found a larger number of trios, 12,278, 512, and 19, for the same comparisons described above, plus 52 and 4 trios that could classify normal and gastritis from intestinal metaplasia, respectively. We could not identify a single trio that could distinguish normal from gastritis even accepting one sample misclassified.
|
Next, all of the identified trios represented in Table 3
were tested again, in an independent set of 34 samples (validation set), corresponding with an increase of 18 samples in NxT comparison, 15 samples in NxG comparison, 21 samples in NxM comparison, 13 samples in GxT comparison, 16 samples in GxM comparison, and 19 samples in MxT comparison. In Fig. 6
we represent the trio with highest SVD (COL4A1, XBP1, and RPL14) and its performance using the training set of samples (Fig. 6A)
and the separation of the validation set of samples (Fig. 6B
, samples represented by stars) using the same rule defined during training. On the basis of their performance on the learning set of samples, trios were ranked by their SVD score (square root of the ratio of between groups and within groups sum of squares, as explained above). Trios with highest SVD show the least dispersed distribution of samples of a given class and the greater distance between the two classes. We selected the 100 trios with highest SVD score and again determined their performance on the validation set of samples using the rule defined during training (Fig. 6C)
or allowed definition of new rules (Fig. 6D)
. By comparing Fig. 6, C and D
, it can be notice that, when new rules are allowed, classification of samples GF63 improves, but samples BIO124 and GH971 are misclassified by a larger number of trios. The actual numbers for this heatmap is available elsewhere.5
|
Molecular Classification of the 99 Gastric Tissue Samples.
Having identified trios and their respective SVD, we determined the efficiency of these trios to correctly classify all 99 samples. As the number of trios for NxT classification was quite large and very redundant, we selected the top 100 trios for the NxT separation, according to their SVD. For the remaining classification, we used 17, 20, 52, and 4 trios for the GxT, MxT, NxM, and GxM, respectively. The number of genes involved in these trios is 103, 24, 38, 57, and 11, for NxT, GxT, MxT, NxM, and GxM, respectively (the identity of the genes, the structure of the trios, and the identity of misclassified samples can be obtained elsewhere).5
Each of the 99 samples was then classified by all trios of the five comparisons. In Fig. 7
, we represent the performance of four trios that can classify NxT, GxT, and MxT. For the NxT classification, we represent the trio with highest SVD with all 54 normal and tumor samples (Fig. 7A)
and with all 99 samples (Fig. 7B)
. We also represent the best trio that precisely separates all normal and tumor samples (Fig. 7C)
, and the same trio with all 99 samples (Fig. 7D)
. In Fig. 7, E and F
, we represent the trio with highest SVD for GxT and MxT separation with the corresponding samples. The performance of all of the trios against all of the samples is represented in Fig. 8
. The spectrum of colors from green to red was used to denote 100% to 0% classification as the first entity of each comparison (for example, in the first line, representing NxT, 100% means all trios classifying the corresponding sample as N and 0% means all trios classifying the corresponding sample as tumor). Thus, a black bar denotes that a sample was classified by 50% of the trios as one of the two entities.
|
|
, are placed in the third and fourth clusters and, at the principal component analysis, are detached from the remaining samples (figure available).5
Also compatible with the number of genes with significant expression differences observed in Fig. 1
, there was a less clear separation of samples in the two columns representing NxM and GxM comparisons but nevertheless, a larger number of red bars can be observed in samples representing intestinal metaplasia. To determine whether the biological phenomena that were supporting class distinction between tumor and nontumor samples were somehow related, we identified the genes composing the top 45 trios for the NxT distinction plus the genes composing the 17 trios for GxT and the 20 trios for MxT. The number of genes was 50, 24, and 38, respectively, and only 1 gene, CTSB, was at the intersection of the three groups, whereas the sum of the three groups has 100 genes indicating very little redundancy among the genes.
DISCUSSION
During the last 2 years, several groups, including ourselves (12)
, published studies focusing on the expression profile of gastric cancer and the identification of differentially expressed genes (19, 20, 21, 22, 23, 24, 25)
. Whereas the majority of the published data are based on the comparison between normal and tumor tissues, less information concerning the molecular events that could establish a link between gastric cancer and intestinal metaplasia is available (11)
. Likewise, no systematic effort for the construction of molecular classifiers for gastric cancer that could impact on early diagnosis has been reported.
In the present work, we describe the identification of a large set of molecular classifiers that could be used for distinction between normal gastric mucosa and those representing gastritis, intestinal metaplasia, and gastric adenocarcinomas. We have also identified genes of which the gene expression profile can be correlated with the transition stages between normal and intestinal type of adenocarcinomas and genes that are differentially expressed between diffuse and intestinal type of gastric adenocarcinomas.
Before searching for classifiers, we first validated our cDNA array by searching for differentially expressed genes in the various disease states. As expected there were fewer differentially expressed genes when we compared samples representing the nontumor states than comparisons having tumor samples (Fig. 1
and Table 2
). As a way to confirm the correlation between our findings and the biology of the process we are dealing with, we determined the pattern of expression of genes that, based on current knowledge, should be altered in our data set. Alterations in the Wnt pathway have been found in a large set of gastric adenocarcinomas, and overexpression of ß-catenin (CTNNB1) in gastric tumors has been described (12
, 17
, 18
, 23
, 25)
. We observed over expression of CTNNB1 in this collection of samples in both intestinal and diffuse types of gastric adenocarcinomas (figure available).5
It was reported recently that MMP2 mRNA is elevated in both diffuse and intestinal types of gastric adenocarcinomas (11)
, and we have confirmed these observations both at the RNA and at the protein levels (Fig. 4)
. Also, KRT20 is typically augmented in intestinal metaplasia and, in agreement with this notion; we found its expression augmented in relation to all other three of the tissue classes (see Table 2
).
The cDNA microarray technology has been largely applied to cancer research (26)
, and expression profile is being increasingly used for distinction between physiological and disease states, as well as to distinguish between groups of disease samples of which the expression profile can discriminate between clinically or biologically similar entities (27, 28, 29)
. Using a set of 18 nonredundant genes representing the 6 genes with lowest Ps for all six of the pair-wise comparisons, we applied the nonsupervised algorithm k-means to see whether our samples could be grouped accordingly. In Fig. 3
, we can observe that the vast majority of samples representing normal and gastritis samples are grouped together, and most of the samples representing intestinal metaplasia and tumors groups in two other clusters. Interestingly, a small cluster with only 6 samples of all four of the entities could be observed and, by the pattern of their expression profile, it is clear that overexpression of MYC and lower expression of CDKN1A is the hallmark of this group of samples. When we did principal component analysis along all 99 samples, this group of 6 samples is detached from the remaining 93 samples (figure available).5
Also, KRT20 was expressed at higher levels in virtually half of the samples from the cluster where the majority of intestinal metaplasias were grouped (see upper branch). There is a controversy on the literature concerning the pattern of expression of KRT20 in intestinal metaplasias of the gastric esophageal junction and its pattern of expression in gastric mucosa (30, 31, 32, 33, 34, 35)
. Apparently, KTR20 expression in intestinal metaplasias of the gastric mucosa is also variable. Importantly, we repeated the clustering of samples using self-organizing maps, also a nonsupervised algorithm, and again, samples were grouped with comparable results (data not shown).
The phenotypic and biological differences observed between intestinal- and diffuse-type gastric adenocarcinoma should be a consequence of differences in gene expression. The hallmark of diffuse-type gastric adenocarcinoma is the lack of glandular organization with spreading of tumor cells throughout the parenchyma implying the need for extracellular matrix destruction. In agreement with this notion and in agreement with data presented by Boussioutas et al. (11)
, we found MMP2 to be expressed at higher levels in both types of tumor, with diffuse type having even higher mRNA levels than the intestinal type (Fig. 4A)
. Again, these data were confirmed by immunohistochemistry in a collection of 150 tumor samples (Fig. 4, B and C)
. Interestingly, signet-ring cells, frequently observed in diffuse type of adenocarcinomas, did not express MMP2 (Fig. 4B
, signet ring cells or SRc).
To additionally contribute to the understanding of the cascade of events that takes place during the oncogenesis of intestinal type gastric adenocarcinoma, we identified genes of which the expression showed a constant increase or decrease along the cascade from normal to tumor samples with a P 0.0009 between the normal and the tumor samples (Fig. 5)
. Five of the 7 genes with increased expression toward malignancy have functions related to the extracellular matrix (COL4A1, FN1, CTSB, COL1A2, and VIM). Another gene, DAF, also showed augmented expression from normal to gastritis, metaplasia, and tumor. DAF could be involved in escaping from complement, and its increase expression in gastric cancer was also observed by serial analysis of gene expression analysis (SAGE anatomical viewer8
). Among the genes of which the expression decreased from normal to tumor samples, we identified the tumor suppressor gene VHL. There are few reports in the literature where the status of VHL gene in gastric cancer was investigated. Leung et al. (36)
failed to demonstrate hypermethylation of the promoter region of VHL in 5 gastric cancer cell lines and 26 gastric carcinomas. Diminished expression of the VHL gene, assessed by immune histochemistry was found in 27 of 318 samples of gastric carcinomas (37)
. Of notice, there was a diminished expression of BAD, a proapoptotic gene, and such reduced expression could contribute to cell survival. It would be important to see whether there is a correlation between reduced BAD expression and expression of trefoil factor 1, known to antagonizes BAD-induced apoptosis in gastric mucosa (38)
. Finally it is likely that reduced expression of LCK just reflects the reduction of infiltrating lymphocytes, because our samples were dissected to exclude inflammatory cells from tumor samples.
Having compared tumor types and identified genes of which the pattern of expression correlates with disease progression we next searched for genes that could be used for the construction of molecular classifiers. As mentioned before, different clustering approaches were successfully used to distinguish between tumor and nontumor samples (24) , morphologically similar samples (29 , 11) , and determine disease outcome (27 , 39) . Whereas a large group of statistical methods is available for such a task, (reviewed in Ref. 40 ), they are often based on the expression of a large set of genes and, necessarily, require data from the two sample groups to identify the set of genes where similarities and differences can be used to define the clusters. To be routinely applied, these requirements could represent potential pitfalls. An alternative would be the implementation of supervised learning procedures where classification could be based on expression signatures rather then comparative expression profile, as proposed earlier (40, 41, 42) . The major advantage would be the possibility of creating a database against which the test sample would be classified, as demonstrated by Ramaswamy et al. (41) . A known group of samples could be used for training and the resulting classifier used for prediction of an unknown sample (class prediction), as demonstrated by Golub et al. (29) . Support vector machine, an example of supervised learning algorithm, was successfully used for class distinction by Shipp et al. (43) and by us (12) . Several other mathematical methods such as Nearest Neighbors Classifiers and Classification Trees (44) could also be applied to search for groups of genes with different expression patterns or sample signature. It appears to us that Fisher’s linear discriminant analysis (16) attains a good compromise between simplicity and performance, making it a good choice for this investigation. Moreover, this approach to identify expression signatures corresponds to the usual approach to identify differentially expressed (single) genes, based on the t-statistics.
Hence, we decided to apply Fisher’s linear discriminant and, by exhausted search among all 370 genes, we identified trios of genes that could be used for class distinction of all four entities in a pair-wise manner (Table 3)
. First, it was interesting to observe that the closer along the cascade of disease progression two samples are, the lowest the number of trios. Also, many trios that can distinguish between normal and tumor samples can be explained by the fact that four pairs of genes (KIAA0106-COL4A1, CLTC-COL4A1, XBP1-COL4A1, and COL4A1- MCL1) could precisely separate all 38 of the normal and tumor samples used in the training set of samples. Hence, the list of trios for this comparison is highly redundant for trios composed of these four pairs.
It is noteworthy that, according to our strategy to select cDNA fragments for immobilization in the array, many genes were represented by two or even three cDNA fragments representing distinct regions of the same gene. Thus, it was interesting to determine whether classifiers based on the average signal intensity for all spots of different cDNA fragments corresponding to a given gene would be reproduced when signal intensity of replicas for a single cDNA fragment of that given gene was considered. Indeed, there was a strong correlation between classifiers identified by the two strategies.
All of the trios identified as potential classifiers using the training set of samples were first validated against an independent group of samples, the validation set, comprising 34 new samples using the same classification rule defined during training (Fig. 6C)
or by new rules defined during validation (Fig. 6D)
. The performances of all of the trios were ranked by their SVD score. Interestingly, the trio with highest SVD, composed of COL4A1, XBP1, and RPL14, misclassified one normal sample (GF63), whereas a trio with lower SVD (TNFRS6, COL4A1, and Hs.325445) correctly classified all 54 of the normal plus tumor samples (Fig. 7, A and C
, respectively). With the exception of NxT classification, for which we selected only the top 100 trios with highest SVD, the performance of all of the trios for all five of the comparisons using all 99 samples is represented in Fig. 8
. It is noteworthy that, from the 150 genes present in top 45 trios for NxT plus the 17, 20, 52, and 4 trios for the GT, MT, NM, and GM, respectively, a single gene, CTSB, is present in the five groups of classifiers. These data strongly suggest that the biological phenomena that these genes are involved in differ for each comparison.
There is a strong correlation between data from classifiers and from cluster analysis. Samples that are misclassified by the trios are those that were not grouped in the expected cluster in Fig. 2
. For the NxT comparison (Fig. 8
, first column) there are 6 nontumor samples that were classified as tumor by >50% of the trios; GF 63 (N), GF59 (G), BIO133 (M), and BIO136 (M) were all grouped in the third cluster, whereas GH828 (M) and GH880 (G) fall into the fourth cluster, together with the majority of tumor samples. Together, these observations suggest that the expression profile of the 18 genes used for clustering reflects the structure detected by the signatures identified by the trios. It is important to remember that the 100 trios used for NxT classification are composed of 103 genes. Also, it is clear that the closer the samples are in the cascade of disease progression, the higher the number of misclassified samples, probably reflecting a gradual change in the pattern of gene expression. Nonetheless, there is always a predominant distribution of green and red bars for the samples that represents a given comparison.
The classifiers presented here, based on trios of genes defined by Fisher’s linear discriminant analysis, were able to discriminate tissue samples representing the cascade of events related to gastric adenocarcinomas and should now be validated for class prediction. Importantly, it is now imperative to apply these classifiers to a large set of samples representing intestinal metaplasia to determine the correlation between misclassification of these samples as adenocarcinomas and the frequency they transform into malignant disease. This study will require a large collection of samples and a period of follow-up that cannot be precisely anticipated but is now under way, in a multicentric effort.
|
ACKNOWLEDGMENTS |
|---|
|
FOOTNOTES |
|---|
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.
Requests for reprints: Luiz F. L. Reis, Rua Prof. Antonio Prudente 109, Sao Paulo, SP 01509-010, Brazil. Phone: 55-11-33883228; Fax: 55-11-32077001; E-mail: lreis{at}ludwig.org.br
5 Internet address: http://array.ludwig.org.br/publications/meirelescancerresearch2003 
6 Internet address: http://cran.r-project.org/.
7 Internet address: http://www.bioconductor.org.
8 Internet address: http://cgap.nci.nih.gov/SAGE/viewer.
Received 6/23/03. Revised 12/ 8/03. Accepted 12/10/03.
|
REFERENCES |
|---|
|
TopABSTRACTINTRODUCTIONMATERIALS AND METHODSRESULTSREFERENCES |
|---|
This article has been cited by other articles:
|
|
 |
|
  I.-S. Song, N. S. Oh, H.-T. Kim, G.-H. Ha, S.-Y. Jeong, J.-M. Kim, D.-I. Kim, H.-S. Yoo, C.-H. Kim, and N.-S. Kim Human ZNF312b Promotes the Progression of Gastric Cancer by Transcriptional Activation of the K-ras Gene Cancer Res., April 1, 2009; 69(7): 3131 - 3139. [Abstract] [Full Text] [PDF]
|
|
|
|
 |
|
 Y. Qu and S. Xu Quantitative Trait Associated Microarray Gene Expression Data Analysis Mol. Biol. Evol., August 1, 2006; 23(8): 1558 - 1573. [Abstract] [Full Text] [PDF]
|
|
|
|
|
|
L. I. Gomes, G. H. Esteves, A. F. Carvalho, E. B. Cristo, R. Hirata Jr., W. K. Martins, S. M. Marques, L. P. Camargo, H. Brentani, A. Pelosof, et al. Expression Profile of Malignant and Nonmalignant Lesions of Esophagus and Stomach: Differential Activity of Functional Modules Related to Inflammation and Lipid Metabolism Cancer Res., August 15, 2005; 65(16): 7127 - 7136. [Abstract] [Full Text] [PDF]
|
|
|
|
 |
|
 S. Y. Chui, T. M. Clay, H. K. Lyerly, and M. A. Morse The Development of Therapeutic and Preventive Vaccines for Gastric Cancer and Helicobacter pylori Cancer Epidemiol. Biomarkers Prev., August 1, 2005; 14(8): 1883 - 1889. [Abstract] [Full Text] [PDF]
|
|
| ||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||
| 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 | Meeting Abstracts Online |
