| HOME | HELP | FEEDBACK | SUBSCRIPTIONS | ARCHIVE | SEARCH | TABLE OF CONTENTS |
Institut National de la Santé et de la Recherche Médicale (S.D., C.F.-P., S.S.-R., C.B., F.B.-C., B.R.), Unité Mixte de Recherche 369 et Unité Mixte de Recherche 664, and Université Lyon 1 (S.D., C.F.-P., S.S.-R., F.B.-C., B.R.), Lyon F-69372, France; Institut National de la Santé et de la Recherche Médicale (M.E.A., F.B.), Unité 836, Grenoble F-38043, France; Centre Hospitalier Universitaire Lyon (C.F.-P., B.R.), Hôpital Edouard-Herriot, Fédération de Biochimie et de Biologie Spécialisée, Lyon F-69437, France; and Centre Hospitalier Universitaire Lyon (C.F.-P., C.P., N.B.-D., M.D., J.-L.P., C.B., J.O., F.B.-C., B.R.), Hospices Civils de Lyon, Lyon Thyroid Tumor Bank Organization, Lyon F-69229, France
Address all correspondence and requests for reprints to: Professor Bernard Rousset, Unité Mixte de Recherche 664, Institut National de la Santé et de la Recherche Médicale, Faculté de Médecine Laennec, 7 rue Guillaume Paradin, 69372 Lyon, Cedex 08, France. E-mail: rousset{at}sante.univ-lyon1.fr.
 |
Abstract |
|---|
 Top Abstract IntroductionPatients and MethodsResultsDiscussionReferences |
|---|
Objective: The aim of the study was to use a set of genes discriminating benign from malignant tumors, on the basis of their expression levels, to build tumor classifiers and evaluate their capacity to predict malignancy on FNAB.
Design: We analyzed the level of expression of 200 potentially informative genes in 56 thyroid tissue samples (benign or malignant tumors and paired normal tissue) using nylon macroarrays. Gene expression data were subjected to a weighted voting algorithm to generate tumor classifiers. The performances of the classifiers were evaluated on a series of 26 sham FNAB, i.e. FNAB carried out on thyroid nodules after surgical resection.
Results: A series of 19 genes with a similar expression in follicular adenomas and normal tissue and discriminating follicular adenomas+normal tissue from the following: 1) follicular thyroid carcinomas (FTCs), 2) papillary thyroid carcinomas (PTCs), or 3) both FTCs and PTCs. These were used to generate four classifiers, the FTCs, PTCs, common (FTC+PTCs), and global classifiers. In 23 of the 26 sham FNAB, the four classifiers yielded a diagnosis in agreement with the diagnosis of the pathologist used as reference; in the three other cases, the correct diagnosis was given by three of four classifiers.
Conclusions: We developed a procedure of molecular diagnosis of benign vs. malignant tumors applicable to the material collected by FNAB. The molecular test complied with a preclinical validation stage; it must be now evaluated on ultrasound-guided FNAB in a large-scale prospective study.
|
Introduction |
|---|
|
TopAbstractIntroductionPatients and MethodsResultsDiscussionReferences |
|---|
For a rather long period of time, efforts have been concentrated on the study of individual molecules expected to help in tumor classification and diagnosis. This approach did not bring real progress. The development of high throughput technology of gene expression on microarrays has opened true possibilities of identification of markers for diagnosis or prognosis of cancer. Since the pioneering work of Huang et al. (4) on the identification of genes differentially expressed in papillary thyroid carcinomas vs. normal thyroid tissue, numerous studies have identified series of genes exhibiting distinct expression profiles in the different types or subtypes of thyroid tumors (5, 6, 7, 8, 9, 10, 11, 12, 13, 14). Some recent studies have already investigated how gene expression profiling data can be used to predict the type or class of individual tumors particularly tumors with difficult cytological diagnosis such as follicular variant of papillary thyroid carcinomas (PTCs) (15, 16) and follicular thyroid carcinomas (FTCs) (17, 18, 19, 20, 21).
The challenge is now to choose combination(s) of marker genes discriminating benign from malignant thyroid tumors to try to set up molecular diagnostic test(s) in view of an application to FNAB. Because the pan-genomic microarray technologies, now readily available, are expensive and cannot be used on large number of samples, we constructed a dedicated macroarray containing about 200 potentially informative genes, i.e. genes previously found to be differentially expressed in thyroid carcinomas, compared with adenomas and/or normal tissue or between thyroid carcinoma subtypes. To develop this simplified approach, we used a methodology based on the spotting of oligonucleotides on nylon membrane and hybridization of [33P]-labeled cDNA probes.
Here we report data of a two-stage study that consisted of the elaboration of gene expression profile-based thyroid tumor classifiers and an evaluation of their capacity to diagnose benign vs. malignant tumors on material from sham FNAB (FNAB performed on nodules after surgical resection). Stage 1 corresponded to transcript analyses on a series of thyroid samples (normal tissue, benign tumors, and carcinomas) from a tumor bank to identify genes discriminating benign and malignant tumors to generate tumor classifiers using the fewest possible number of genes and exhibiting the highest prediction strength. Stage 2 of the study was devoted to the evaluation of performances of four tumor classifiers on a new set of samples, i.e. sham FNAB. The molecular diagnosis (benign vs. malignant tumor) given by each classifier was compared with the diagnosis given by the pathologist used as reference. We report that the combination of the four classifiers gave the correct diagnosis in 100% of cases.
|
Patients and Methods |
|---|
|
TopAbstractIntroductionPatients and MethodsResultsDiscussionReferences |
|---|
Thyroid tissue samples were taken from the Lyon Thyroid Tumor Bank of the Lyon University Hospital, as previously described (22). Tissue samples were classified according to World Health Organization recommendations. Samples used in this study were follicular adenomas (FAs) (n = 6), FTCs (n = 8), PTCs (n = 14), and the paired normal tissue samples (NT) (n = 28); among PTCs, nine were PTCs of classical form and five follicular variant of PTC. Patients were 25 females and three males with an average age of 45 ± 3 yr. Information about patients and tumors is provided as supplemental data (supplemental Table 1, published as supplemental data on The Endocrine Society’s Journals Online Web site at http://jcem.endojournals.org). Five other samples, three FAs and two FTCs without paired NT were also analyzed on microarrays. This study was approved by the supervision interdisciplinary committee of the tumor bank and performed in accordance with protocols approved by the local human studies committee.
FNAB
Sham FNABs were performed with the standard technique but on nodules after surgical resection. They were obtained from surgical pieces of 25 patients (20 females and five males; mean age 52 ± 3 yr) using a 27-gauge needle; two samples originated from the same patient who had a nodule in each thyroid lobe. The mean size of nodules was 3.2 ± 0.3 cm. The cellular material collected by FNAB was rapidly mixed with a lysis buffer (RLT buffer; QIAGEN S.A., Courtaboeuf, France) and stored at –20 C.
RNA extraction
RNA was isolated from tissue samples using the phenol-chloroform extraction procedure (23) and was subsequently purified on silica column (RNeasy minikit; QIAGEN) according to the manufacturer’s protocol. Potentially contaminating genomic DNA was eliminated by RNase-free DNase I (QIAGEN) treatment. RNA from sham FNAB was directly purified on silica column (RNeasy microkit; QIAGEN) according to the manufacturer’s protocol. RNA was analyzed by microfluidic electrophoretic separation on chips using the Agilent 2100 BioAnalyzer (Agilent Technologies Inc., Santa Clara, CA). The average value of the 28S to 18S ratio of RNA extracted from 56 tumors and NT samples was 1.93. Similar results were obtained with sham FNAB.
RNA amplification
RNA amplification was performed according to Eberwine and colleagues (24). Briefly, cDNAs, generated by reverse transcription using an oligo(dT) primer bearing a T7 promoter, were submitted to in vitro transcription using T7 RNA polymerase to produce antisense mRNA. Amplification was performed from 1 µg RNA using the MessageAmp antisense RNA kit from Ambion (Austin, TX). In the case of FNAB, the amount of RNA was reduced to 100–300 ng.
Oligonucleotide macroarray preparation
An oligonucleotide-based low-density macroarray of about 200 genes was generated on nylon membrane. It was composed of genes reported to be differentially expressed between tumors and normal tissue or between benign and malignant thyroid tumors. The list of genes composing the macroarray is provided as supplemental data (supplemental Table 2). Internal marker genes (tomato plant) were also added (25). Oligonucleotides (70 mers), located in the 3' region of the genes at a maximum distance of 800 bp from the poly(A) tail and having a G/C ratio between 40 and 70%, were designed and synthesized by Eurogentec (Liege, Belgium). Nylon macroarrays were generated by oligonucleotide spotting at the Transcriptome Analysis Platform of Genopole Rhône-Alpes (Grenoble, France) as previously described (25). Quality of spotting was controlled by quantification of oligonucleotides using a T4-polynucleotide-kinase labeling method (26).
cDNA probe labeling
Antisense RNA (2 µg) was reverse transcribed in a 30-µl reaction volume containing 0.6 ng tomato plant mRNA; 4.8 mmol dATP, dTTP, and dGTP; 28.8 pmol dCTP; 160 pmol oligo-dT20V (Invitrogen Corp., Carlsbad, CA); 100 pmol of random hexamers (Amersham, Little Chalfont, UK); 40 U of RNasin ribonuclease inhibitor (Promega Corp., Madison, WI); 30 µCi of 
-[33P] dCTP (PerkinElmer, Waltham, MA); and 200 U of Moloney murine leukemia virus reverse transcriptase RnaseH minus (Promega). After 1 h of incubation at 42 C, 200 U of enzyme were added, and incubation was continued for 1 h. Labeled cDNAs were purified on Microcon YM-50 centrifugal filter devices (Millipore Corp., Billerica, MA); their specific radioactivity was 14.0 x 106 ± 0.6 x 106 cpm/µg aRNA (mean ± SEM, n = 98).
Hybridization and signal acquisition
Nylon membranes were prehybridized in 5x Denhardt’s solution in the presence of 40 µg/ml herring sperm DNA (Promega) for 4 h at 60 C. Labeled probes were denatured and added into the prehybridization solution. Hybridization was continued for 72 h at 60 C. After washings, membranes were exposed to phosphorus imaging plates (Fujifilm Corp., Tokyo, Japan) for 4 d, and signals were detected using a Fuji BAS 8000 Imager (CeCiL, IFR 62; Faculté de Médecine Laennec, Lyon, France) at a 20-µm resolution. ImageGauge and ArrayGauge softwares (Fujifilm) were used to visualize and collect numerized signals, which were then transferred on Excel tables. After correction for background and normalization (for an equal total radioactivity on the macroarray), signal intensity values of duplicate were averaged and subjected to log2 transformation for statistical analyses (27, 28).
Statistical analyses and hierarchical clustering of data
Macroarray data were analyzed by significance analysis of microarray (SAM) software (29). The criteria selected for SAM included a minimum 2-fold difference in gene expression and a false discovery rate (FDR) lower than 5%. Sample classifications were obtained by hierarchical clustering using the uncentered correlation distance and average linkage method provided by Cluster software and visualized using TreeView (30).
Generation of tumor classifiers by supervised analyses
Data from informative genes (according to SAM) were used to build supervised predictors (or classifiers) using GeneCluster 2.0 software (http://www.broad.mit.edu/cancer/software/genecluster2/gc2.html) (31, 32). This software includes procedures for building and testing supervised models using a weighted voting algorithm. Tumor classifiers were built with gene expression data from 56 tissue samples representing the training set. The accuracy of predictions made by the classifiers was tested by the leave-one-out cross-validation method on the training set (31).
Quantitative PCR
PCR was performed on a LightCycler (Roche Diagnostics, Meylan, France). Amplification of cDNAs was carried out as previously described (22). The sequence of primers and the size of amplicons were: CTGCCAGTATGTGACCGAGA and ACAGGAAGGTGATCCCAGTG for ALDOA (214 bp), TGGACCTGGAGACTCTCAGG and ATTCTAATGCCAGAGGCTGG for CDKN1A (276 bp), CTCAACGACCACTTTGTCAG and CTTACTCCTTGGAGGCCATGT for GAPD (103 bp), TGCTTCTGTCATGAAGCACC and GGACCAAGTAATCCGCACG for GOT1 (267 bp), CATCAAGACATATCTCAGTTGGACCT, and TGCAAGCCACCAAATATCAA for SLC26A4 (226 bp), TCCAGCTCTGCTGAGGAGTACG and GCCTGGCAGCAATCACAGCC for TFF3 (268 bp), TAGAGGGACAAGTGGCGTTC and CGCTGAGCCAGTCAGTGTAG for 18S (104 bp). PCR conditions included an initial denaturation of 10 min at 95 C, followed by 40 cycles consisting of 15 sec at 95 C for denaturation, 6 sec at 57 C for CDKN1A, 58 C for GAPD and GOT1, 61 C for ALDOA, 62 C for TFF3, 64 C for SLC26A4 and 18S rRNA for annealing and 8–12 sec at 72 C for the final extension step (depending on amplicon size). Amplicons corresponding to the different mRNA assayed by RT-PCR were cloned into the pGEMT easy vector (Promega). Amounts of plasmid, corresponding to 101 to 106 cDNA copies were included in each PCR assay to generate calibration curves.
|
Results |
|---|
|
TopAbstractIntroductionPatients and MethodsResultsDiscussionReferences |
|---|
The yield of mRNA amplification from tumor or NT RNA was similar; the average amount of antisense mRNA generated from 1 µg total RNA was 27.5 ± 1.5 µg (n = 56). As shown in Fig. 1A
, synthesized antisense mRNA had the same size (between 0.5 and 1.5 kb) whatever the group of samples: NT, FA, FTC, or PTC. The amplification step did not modify gene expression profiles (Fig. 1B). Indeed, transcript contents determined by quantitative RT-PCR from amplified antisense mRNA were proportional to those measured from total RNA. This is illustrated for three distinct genes (ALDOA, GOT1, and TFF3) whose expression level (i.e. transcript copy number) differed by more than 2 orders of magnitude. The regression coefficient calculated from 135 data points (three genes, 45 samples) by the least-square method was 0.92.
|
We first identified genes, the expression level of which differ in carcinomas (FTC or PTC), compared with paired NT, using SAM (fold change > 2 and FDR value < 5%). Analyses of FTC vs. NT revealed 24 genes with a highly statistically significant difference of expression; six were overexpressed and 18 underexpressed in FTC. Among underexpressed genes, CRABP1, TFF3, PGCP, C11ORF8, DPT, APOD, CCL21, TNRS11b, FGFR2, and FGL2 have been quoted by Aldred et al. (5). Sixty-nine genes exhibited a statistically significant difference of expression in PTC vs. paired NT; 21 were overexpressed and 48 underexpressed in tumors. Most of these genes were previously reported to be altered in PTC (4, 12, 15). The list of the differentially expressed genes can be viewed as supplemental data (supplemental Table 3: FTC vs. NT; supplemental Table 4: PTC vs. NT). Only few genes were found differentially expressed in FA compared with paired NT. Genes with a difference of expression in FTC (or PTC) compared with NT allowed a clear-cut and complete separation of FTC (or PTC) and NT samples by hierarchical clustering (see Fig. A1 at http://ifr62.univ-lyon1.fr/users/b_rousset/deCanThyr2007/).
Validation of macroarray data by quantitative RT-PCR
Transcripts corresponding to six genes present on the macroarray were assayed by quantitative RT-PCR. Figure 2 shows the relationship between transcript content measured by RT-PCR and hybridization signal intensity on macroarray for each of the six genes. For the clarity of the figure, only mean and SEM values of each group of samples (NT, FA, FTC, and PTC) are presented. Whatever the gene, there was a very good correlation between expression levels measured by hybridization on macroarray and by RT-PCR; correlation coefficients varied from 0.93 to 1.0. Data in Fig. 2 also show selective changes in the level of expression of the six genes; One can see the following: 1) an increase of ALDO A, GOT1, and GAPD in FTC, 2) an increase of CDKN1A in PTC and FTC and to a lesser extent FA, and 3) a reduced expression of SLC26A4 and TFF3 in FTC and PTC.
|
Gene expression data from NT and FA were joined to form the nonmalignant or benign group of tissues and were compared with gene expression data from FTC and PTC considered either separately or together (as the group of malignant tissues). The few genes differentially expressed in FA, compared with NT, or equally expressed in FA, and malignant tumors have been withdrawn for subsequent analyses. The number of genes differentially expressed in FTC vs. NT+FA and in PTC vs. NT+FA was 26 and 57, respectively; 21 genes were differentially expressed in PTC+FTC (malignant group) vs. NT+FA (benign group). The segregation of malignant from nonmalignant tissue samples from data of the three series of genes is shown in Fig. 3. FTCs were clearly separated from NTs and FAs (Fig. 3A); one FTC was misclassified within the benign group. Similarly, PTC samples were separated from NT and FA with one exception: 212PTC (Fig. 3B). Using 21 genes, PTCs and FTCs segregated from NTs and FAs (Fig. 3C); among the 56 samples, only one FTC and one NT were misclassified.
|
Data corresponding to the three series of 26, 57, and 21 genes (see Fig. 3
for the name of genes) were processed with GeneCluster 2.0 software to build tumor classifiers. Among the numerous possible models differing by the number of genes, we selected the ones giving the minimal error in cross-validation tests (31). The best FTC prediction model built from data of 28 samples (eight FTCs, six Fas, and 14 paired NTs) used nine of 26 genes (Fig. 4). Using the leave-one-out method, this classifier did not make any error of sample assignment in benign vs. malignant tumor classes. The median prediction strength was 0.85; the sensitivity and specificity were equal to 100%. The best PTC prediction model generated from 40 samples used nine of 57 genes (Fig. 4); only two of these genes were present in the FTC classifier. The PTC classifier mishandled two of 40 samples (sensitivity 86%; specificity 100%). The median prediction strength was close to 1 (data not shown). The common (FTC+PTC) classifier, based on genes similarly altered in FTC and PTC, compared with the benign group, used 12 genes. Among the 56 samples, two were wrongly predicted (sensitivity 91%; specificity 100%). The median prediction strength was 0.93 (data not shown). The 19 genes, which composed the FTC, PTC, and common classifiers, are listed in Fig. 4B. Only two genes (ChGn and TFF3) belong to the three classifiers. A fourth classifier named global classifier was built from the set of 19 genes; three of the 56 samples were misclassified. The median prediction strength was 0.92. Data summarized in Fig. 5 show to which extent the level of expression of genes forming the classifiers differ in benign and malignant tumors of the training set. The fold change in gene expression between benign and malignant tumors varies from about 2 for CITED in P classifier or CTSB in C classifier to more than 10 for CHI3L1 or TFF3 in P classifier.
|
|
The amount of RNA extracted from material collected by sham FNAB was 1.4 ± 0.4 µg (mean ± SEM, n = 26). Gene expression profiles of FNAB were subjected to computer analyses and the resulting tumor class predictions made by the classifiers were compared with the diagnosis made by the pathologist (used as gold standard). According to the pathologist, nodules subjected to FNAB were 14 benign tumors and 12 carcinomas (Fig. 6). Twenty-four tumors were correctly predicted by the FTC classifier and two benign tumors were assigned to the malignant class. Tumor class predictions made by the PTC and global classifiers were in agreement with the diagnosis made by the pathologist in 26 of 26 cases. The common classifier also gave the correct diagnosis with one exception; a malignant tumor (FNAB21) was assigned to the benign class. The median prediction strength of the FTC, PTC, common, and global classifiers was 0.47, 0.86, 0.59, and 0.79, respectively. It is worth to notice that the few errors made by F and C classifiers did not occur on the same FNAB. The four classifiers yielded the same and correct diagnosis in 23 of 26 cases, and in the three other cases, the correct diagnosis was given by three of the four classifiers.
|
|
Discussion |
|---|
|
TopAbstractIntroductionPatients and MethodsResultsDiscussionReferences |
|---|
Unlike most previous studies of gene profiling of human thyroid tumors, we used a low density macroarray instead of genome-wide microarrays. Genes introduced on the macroarray were genes known (or suspected) to be differentially expressed in tumors vs. normal tissue or between different tumor types. Our gene expression data are in full agreement with those generated on Affymetrix or other high-density microarrays. A comparison of our data with those of six other studies (4, 10, 11, 12, 15, 16) have been made on a series of 16 genes (see supplemental Table 5). The amplitude of variations in transcript levels (i.e. fold change) between groups of samples tend to be higher in our study based on radioisotope labeling of probes than in the other studies bringing into play fluorescently labeled probes. As expected from the choice of genes introduced on the macroarray, we found a high proportion of genes differentially expressed between groups of samples (up to 39% in PTC, compared with NT). As previously reported from high-throughput microarray data (4, 5, 6, 7, 9, 10, 11, 15), our data, sifted through SAM procedure, have yielded list of genes from which a complete or almost complete distinction between thyroid carcinomas and nonmalignant tissues can be made by hierarchical clustering.
Because the hierarchical clustering approach was not adapted to predict the position of unknown samples within a predetermined tumor classification (33), we used the weighted voting algorithm developed by Golub et al. (31) to generate prediction models and test them as tumor classifiers for thyroid cancer diagnosis. Gene expression profiles of FTCs and PTCs being rather divergent, we hypothesized that separate classifiers, based on FTC or PTC selective molecular features, could have a higher prediction efficiency than a single classifier. The third and fourth classifiers named common and global classifiers were generated from genes exhibiting similar changes in FTCs and PTCs (compared with nonmalignant tissues) and from the 19 genes used by the other classifiers. We found that each classifier had its proper contribution to the molecular diagnosis on FNAB. Indeed, when there was an error of prediction, the error affected only one classifier. Thus, considering that a correct response of the molecular test was a tumor class predicted by at least three of the four classifiers, 100% of tumors were correctly diagnosed by gene expression profiling from FNAB. However, in some cases, especially FNAB 21, the strength of the prediction was rather weak (low prediction strength values). Interestingly, FNAB samples we analyzed appear rather representative of the different types of thyroid tumors including tumors with diagnostic difficulties at the time of cytological examination, such as atypical adenomas.
Classifiers for thyroid tumors have already been built using other modelization methods (15, 16, 20, 21, 34, 35, 36, 37); the accuracy of prediction tested on unknown tumor samples (used as test set) was generally higher than 0.9. Among the studies cited above, only one (36) aimed at predicting malignancy from FNAB. In this study, measurements of the level of expression of six genes by quantitative RT-PCR on FNAB (made at the time of surgery), and application of a scoring model of prediction led to a correct diagnosis in 28 of 31 samples.
In conclusion, our study indicates that an accurate molecular diagnosis of benign or malignant tumor can be made on FNAB by measurements of the transcript level of a limited number of marker genes and processing of data through preestablished tumor classifiers. After validation by large-scale prospective studies, transcriptome analyses of that type, run on low-density macroarray or by real-time PCR, compatible with a routine application, should become the complementary test in case of suspicious or indeterminate cytology.
|
Acknowledgments |
|---|
|
Footnotes |
|---|
Disclosure Statement: The authors have nothing to disclose.
First Published Online January 22, 2008
Abbreviations: FA, Follicular adenoma; FDR, false discovery rate; FNAB, fine-needle aspiration biopsy; FTC, follicular thyroid carcinoma; NT, normal tissue; PTC, papillary thyroid carcinoma; SAM, significance analysis of microarrays.
Received July 16, 2007.
Accepted January 11, 2008.
|
References |
|---|
|
TopAbstractIntroductionPatients and MethodsResultsDiscussionReferences |
|---|
This article has been cited by other articles:
 |
|
 |
|
  S. Durand, C. Ferraro-Peyret, M. Joufre, A. Chave, F. Borson-Chazot, S. Selmi-Ruby, and B. Rousset Molecular characteristics of papillary thyroid carcinomas without BRAF mutation or RET/PTC rearrangement: relationship with clinico-pathological features Endocr. Relat. Cancer, June 1, 2009; 16(2): 467 - 481. [Abstract] [Full Text] [PDF]
|
|
| |||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||
| HOME | HELP | FEEDBACK | SUBSCRIPTIONS | ARCHIVE | SEARCH | TABLE OF CONTENTS |
| Endocrinology | Endocrine Reviews | J. Clin. End. & Metab. |
| Molecular Endocrinology | Recent Prog. Horm. Res. | All Endocrine Journals |
, FA; 
, FTC; 
, PTC.
