Skip to main content

Thank you for visiting nature.com. You are using a browser version with limited support for CSS. To obtain the best experience, we recommend you use a more up to date browser (or turn off compatibility mode in Internet Explorer). In the meantime, to ensure continued support, we are displaying the site without styles and JavaScript.

A general method to derive robust organ-specific gene expression-based differentiation indices: application to thyroid cancer diagnostic

Abstract

Differentiation is central to development, while dedifferentiation is central to cancer progression. Hence, a quantitative assessment of differentiation would be most useful. We propose an unbiased method to derive organ-specific differentiation indices from gene expression data and demonstrate its usefulness in thyroid cancer diagnosis. We derived a list of thyroid-specific genes by selecting automatically those genes that are expressed at higher level in the thyroid than in any other organ in a normal tissue's genome-wide gene expression compendium. The thyroid index of a tissue was defined as the median expression of these thyroid-specific genes in that tissue. As expected, the thyroid index was inversely correlated with meta-PCNA, a proliferation metagene, across a wide range of thyroid tumors. By contrast, the two indices were positively correlated in a time course of thyroid-stimulating hormone (TSH) activation of primary thyrocytes. Thus, the thyroid index captures biological information not integrated by proliferation rates. The differential diagnostic of follicular thyroid adenomas and follicular thyroid carcinoma is a notorious challenge for pathologists. The thyroid index discriminated them as accurately as did machine-learning classifiers trained on the genome-wide cancer data. Hence, although it was established exclusively from normal tissue data, the thyroid index integrates the relevant diagnostic information contained in tumoral transcriptomes. Similar results were obtained for the classification of the follicular vs classical variants of papillary thyroid cancers, that is, tumors dedifferentiating along a different route. The automated procedures demonstrated in the thyroid are applicable to other organs.

Introduction

Differentiation and proliferation are the fundamental processes of multicellular life. Cell proliferation is routinely measured with objective quantitative methods, such as bromodeoxyuridine incorporation, Ki-67 or proliferating cell nuclear antigen (PCNA) immunostaining. However, the differentiation state of a cell type manifests itself in a wide range of parameters—many of them qualitative—that include cellular- and tissue-level morphologies, molecular state and function. It is unclear how typical organ-specific differentiation markers integrate these phenomena and to what extent they reflect the overall degree of differentiation of cells and tissues. For example, thyroglobulin is a canonical thyroid marker and is essential in thyroid hormone synthesis. FRTL-5 cells express thyroglobulin, trap iodine and grow in response to thyroid stimulation hormone (TSH), but they are depolarized, that is, they lack a basic morphological feature of differentiated thyroid cells. Conversely, FRT cells are polarized, but do not express thyroglobulin and other thyroid-specific properties (Zurzolo et al., 1991). Moreover, the evaluation of many differentiation features rest on subjective and qualitative histological observations. Thus, there is currently no method to quantify differentiation objectively. Herein, we propose a procedure to derive quantitative multi-gene gene expression markers for organ-specific differentiation. It is completely automatic and agnostic regarding the biology of organs and which aspects of their differentiation are important.

The thyroid is an interesting organ to assess our method because a single cell type, the follicular thyroid cell, can produce a range of benign and malignant tumors with a range of underlying genetic alterations leading to a range of dedifferentiated phenotypes (Kondo et al., 2006); those include hypofunctioning follicular adenomas (FTAs), which are benign encapsulated tumors, and the malignant carcinomas. These can be further subdivided into follicular or papillary carcinomas. They are still partly differentiated, but may both evolve into poorly differentiated carcinoma or into totally dedifferentiated anaplastic carcinoma. Papillary carcinomas further fall in a number of histological subtypes. The most frequent are the classical and the follicular variants. The latter displays some aspects of the typical follicular morphology of normal thyroid tissues.

The differential diagnosis of FTAs and carcinomas is a major challenge for pathologists—and incidentally for the thyroid-differentiation index derived with our method. Between 4 and 7% of the general population will develop a palpable thyroid nodule (Hegedüs, 2004). Of these, 80% are benign hyperplastic nodules, 10–15% are benign follicular neoplasms and 5% are cancers (Hegedüs, 2004). Fine needle aspirate examination is currently the most direct and sensitive method to detect malignancy. However, it is inconclusive in approximately 20% of nodules, resulting in the need for potentially avoidable diagnostic thyroidectomy. In addition, a second opinion by an independent pathologist leads to a contradictory conclusion in 30–60% of the cases (Baloch et al., 2001; Hegedüs, 2004; Clary et al., 2005). Microarray studies suggest that there is no clear-cut boundary between FTAs and carcinomas (Lubitz et al., 2005).

Because tumor proliferation is in general inversely correlated with differentiation, there is a real possibility that any quantitative measure of differentiation is the trivial mirror image of proliferation measures. However, this inverse correlation does not hold in all biological contexts. For example, the TSH promotes both the proliferation and the differentiation of thyroid cells when combined with insulin. Thus, the thyroid also provides an alternative system to assess the relatedness of our differentiation index with proliferation.

Results

The thyroid index and the meta-PCNA index: unbiased measures of thyroid tissue differentiation and proliferation from gene expression data

We propose an unbiased procedure to compute an organ-specific differentiation index from mRNA expression data, using the thyroid as a test case. It proceeds in three steps. Step #1 selects genes specific of the organ of interest from RNA sequencing (RNA-seq) data. RNA-seq estimates mRNA expression with read counts normalized according to the transcript length, a measure that reflects more accurately the absolute transcription levels than hybridization-based microarray (Wang et al., 2009). Step #2, although optional, optimizes the probe selection for these genes in the gene expression platform in which differentiation is to be measured. Finally, step #3 uses the probes/genes derived in previous steps to compute a differentiation index across samples of a data set of interest. An application to the thyroid is described below. Computational details are available in the Materials and methods section.

First, we extracted a set of genes with thyroid-specific expression in BodyMap 2.0 (http://www.ncbi.nlm.nih.gov/geo), a collection of 16 RNA-seq profiles of healthy human organs, that is, these genes were among the 1000 most expressed genes in the thyroid, but not among the 5000 most expressed genes in the non-thyroid organs. Eight genes, listed in Table 1, fulfiled this criterion. Four were canonical thyroid follicular genes: forkhead box protein E1 (FOXE1, a.k.a. TTF2), thyroglobulin, thyroperoxydase and the TSH receptor. Iodotyrosine deiodinase mutations cause congenital hypothyroidism (Moreno et al., 2008). Interestingly, solute carrier family 26, member 7 (SLC26A7) and cellular retinoic acid binding protein 1 (CRABP1) were thus far unknown thyroid markers.

Table 1 Thyroid index genes

The thyroid-specific expression of seven of these genes was confirmed in two independent normal organ gene expression compendia based on massively parallel signature sequencing (Supplementary Figure S1; Jongeneel et al., 2005), and Affymetrix arrays (Affymetrix, Santa Clara, CA, USA) (Supplementary Figure S2; Ge et al., 2005). The last gene, parathyroid hormone, most probably results from the erroneous inclusion of parathyroid tissues in BodyMap 2.0 thyroid sample. Its high thyroid expression is also observed in the compendium of Jongeneel et al. (2005), but not in the compendium of Ge et al. (2005).

On most microarray platforms, several probes are available to measure the expression of a given gene. However, because of technical shortcomings, such as annotation errors and non-specificity, not all of them are necessarily accurate. In step #2, we selected, for each of the eight genes, the probe that maximizes the thyroid-specific signal in a normal organ-expression compendium obtained using the platform of interest. Genes with no probe delivering a thyroid-specific signal were eliminated. We applied this procedure to the Affymetrix U95av2 and U133v2 platforms using the compendia of Su et al. (2002) and Roth et al. (2006), respectively. Parathyroid hormone was eliminated following our optimization with the U133v2 compendia. Thus, in addition to probe optimization, step #2 factored-in extra information that helps refine the list of measurable thyroid genes. However, parathyroid hormone was not eliminated in the U95av2 optimization. We included it in all subsequent analyses conducted on this platform because our goal is to assess an agnostic, automated method.

Finally, given the gene expression profile of a tissue in a given platform we may compute its thyroid index as the median expression of the probes selected in step #2. Figure 1 depicts this index from two normal human organ compendia not used to select the eight genes (Ge et al., 2005; Jongeneel et al., 2005). The thyroid sample stands out among the other tissues.

Figure 1
figure1

The thyroid index singles out thyroid tissue in two independent normal tissue compendia. The tissues from the compendia of (a) Jongeneel et al. (2005) and (b) Ge et al. (2005) were not used to derive the gene list underlying the thyroid index. Organs are aligned along the y-axis. The x-axis represents the thyroid index (log2 scale in b). The thyroid, depicted with the dark bar, stands out among other organs.

We defined meta-PCNA as the 1% of genes that are most positively correlated with PCNA expression across a gene expression compendium of normal human organs (Venet et al., 2011). In plain language, meta-PCNA genes are consistently expressed when PCNA is expressed in normal tissues and consistently repressed when PCNA is repressed. Meta-PCNA genes include many canonical proliferation markers, including MCM2, MKI67, TOP2A and, of course, PCNA. Similar to the thyroid index, we defined the meta-PCNA index of a tissue as the median expression of the meta-PCNA genes.

The thyroid and meta-PCNA indices are negatively correlated in thyroid cancers, but positively correlated in an in vitro TSH stimulation time course

We generated full genome expression profiles for 11 anaplastic thyroid cancers (ATC), 49 papillary thyroid cancers (PTC) and 45 healthy tissues adjacent to 45 of the PTCs (see Materials and methods), and compared the thyroid and meta-PCNA indices across them. The thyroid index is lower in ATCs than in PTCs (P=10−9) and lower in PTC than in normal tissues (P=10−15). Conversely, the meta-PCNA index is higher in ATC than in PTC (P=10−6) and higher in PTC than in normal tissues (P=10−4). Overall, the two indices are anti-correlated (Spearman's ρ=−0.71, P<10−16; Figure 2a) in agreement with the general notion that dedifferentiated anaplastic tumors proliferate more and are more aggressive than the more differentiated PTCs—ATCs have a 5-year survival rate <10% (Kebebew et al., 2005) and PTCs >90% (Kondo et al., 2006).

Figure 2
figure2

Correlation between the thyroid index and meta-PCNA. Overall, the two indices are anti-correlated in tumors (a and b), but not in the TSH stimulation time course (c). (a) Anaplastic thyroid carcinoma (ATC), papillary thyroid carcinomas (CPTC) and normal tissues (N). Note the absence of a clear-cut boundary between PTCs and normal tissues. However, as shown in Supplementary figure S3, tumors have lower thyroid index and higher meta-PCNA index than their patient-matched contralateral normal tissues. (b) Data from Aldred et al. (2004) and Finley et al. (2004), with CPTC and FVPTC, FTA and carcinoma (FTC), and normal tissues (N). (c) Thyroid index vs meta-PCNA in the TSH-stimulation time course. Individual points at a given time point represent primary cells extracted from different subjects. Because the platform had only 4000 genes, the overlap with the 8-gene signature included only 1 gene. Therefore, we used the alternative 72-gene signature of section ‘Robustness of the organ-specific gene selection’ in panel (c). Related data are available in Supplementary tables S1 to S3.

To check that this relation holds in an alternative data set and with tissues displaying less-contrasted dedifferentiation phenotypes, we compiled healthy tissues and differentiated tumor-expression profiles from earlier publications (Aldred et al., 2004; Finley et al., 2004). A strong anti-correlation of our indices is also observed across this panel (ρ=−0.52, P=10−5; Figure 2b).

Are the two indices redundant or are they capturing independent biological information? We calculated the thyroid and meta-PCNA indices in expression profiles from a TSH stimulation time course of thyroid cells in primary cultures (van Staveren et al., 2006). The indices were positively correlated (ρ=0.73, P=0.01, Figure 2c), they both increased with stimulation time, in agreement with the fact that TSH promotes both follicular cell proliferation and differentiation.

In conclusion, the thyroid and meta-PCNA indices put on a quantitative ground the notion that differentiation inversely correlates with proliferation in thyroid neoplasms. Yet, these indices do measure biologically independent parameters.

Follicular and papillary thyroid tumors dedifferentiate along distinct routes in the gene expression continuum

A single cell type, the thyroid follicular cell, may give rise to distinct tumoral phenotypes, including PTCs and tumors of the follicular family. The latter include FTA, which may dedifferentiate further into follicular carcinomas (FTC). Most PTCs fall into the follicular variant (FVPTC) category, which retain some follicular morphology, or into the classical variant (CPTC) category, which have lost the follicular morphology. Before assessing the thyroid index in these tumors, we compared their global expression profiles.

We compiled from two studies (Aldred et al., 2004; Finley et al., 2004) the expression profiles of 7 healthy thyroid tissues, 17 FTAs, 18 FTCs, 13 FVPTCs and 9 CPTCs, and ran a multidimentional scaling analysis (Figure 3). Multidimensional scaling collapses the high-dimensional full-genome gene expression space into two dimensions while preserving the similarity distances between pairs of samples. Hence, samples lying close on the multidimensional scaling plot have similar gene expression profiles.

Figure 3
figure3

Follicular and papillary tumors dedifferentiate along distinct routes in the gene expression continuums. This multi-dimensional scaling plot reduces the high dimensional full-genome expression data into two dimensions. The ‘stress’ of the transformation, that is, the average distortion between the pair-wise distances between samples in the reconstructed 2D space and in the actual gene expression space is 11%. Units in the 2D space are arbitrary. Samples cluster by tissue types rather than that of lab of origin. Normal tissues are in the center of the plot with follicular and papillary tumors spread on opposite sides. The more dedifferentiated a tumor, the more distant it tends to be from normal tissues. Given the diagnosis uncertainties with these tumors and limits in the 2D reconstruction, some discrepancies are expected, for example, one FTC stands out between CPTCs and FVPTCs.

The samples in Figure 3 are clustered according to tissue types, not the study of origin. Papillary and follicular tumors both radiated from the normal samples, but in opposite directions, that is, they had clearly distinct overall molecular phenotypes. FTAs are closer to normal samples than FTCs. Likewise, FVPTCs are closer to normal tissue than CPTC. As already noted (Lubitz et al., 2005), there is an overlap between FTA and FTC suggesting a continuum rather than discrete tumor categories. We show here that the same observation holds for FVPTCs and CPTCs.

Thus, follicular and papillary thyroid tumors dedifferentiate along very different directions in the gene expression continuums. Yet, we show in the next section that a single quantity, the thyroid index, is a useful differentiation marker for both of them.

The thyroid index discriminates follicular adenomas from carcinomas and the classical from the follicular variants of papillary carcinomas

Can the thyroid index, which was derived exclusively from normal tissue data, discriminate tumor samples according to their histological types? We first focused on the classification of CPTCs and FVPTCs using the data set presented in the previous section. The thyroid index is significantly higher in FVPTCs than in CPTCs (P=0.003, Figure 4a), suggesting that FVPTCs are more differentiated than CPTCs. We further assessed its discriminatory performance by computing the area under the receiver operating characteristics (AUC), a statistic that integrates specificity and sensitivity, both shown in Figure 4b. The AUC was 0.86 (Figure 4b). We verified that this AUC was not explained by chance alone, by rerunning the entire AUC calculation on 10 000 indices on which class labels were assigned randomly, yielding P=0.0003. The thyroid index may non-specifically capture a diagnostic signal omnipresent in the transcriptome as was observed for most prognostic signatures in breast cancer (Venet et al., 2011). To rule out this possibility, we recomputed AUC with 100 000 signatures made of eight probes selected at random on Affymetrix U133v2 arrays, yielding P=0.005 (Supplementary Figure S4).

Figure 4
figure4

Differential diagnosis. Top panels address the CPTC vs FVPTC classification task, lower panels FTA vs FTC. (a) and (d) box plots for the thyroid index. (b) and (e) ROC curves, gray areas represent 95% confidence intervals. (c) and (f) ROC curves for the thyroid index and two supervised machine-learning algorithms that select optimal classifying genes from the entire set of genes spotted on the microarrays. The P-values stand for the difference between the AUCs obtained for the thyroid index and either supervised classifiers. None is significant, that is, the thyroid index, which rests on genes selected in the complete absence of cancer data, is as discriminatory as these classifiers.

The differential diagnosis of FTAs vs FTCs is a notorious cytopathological challenge. Using the data from the previous section, the thyroid index obtained was higher in FTA than in FTC (P=0.0008, Figure 4d) and had an AUC of 0.82 (class label permutations, P=0.0001; random probes control, P=0.02; Figure 4e; Supplementary Figure S4). Taken together, these results demonstrate that the thyroid index is discriminatory for tumors following distinct dedifferentiation routes.

Does the thyroid index capture all the discriminatory information present in the genome-wide expression profiles? Better classifiers may be discovered by applying supervised machine-learning algorithms directly to the cancer data. Starting with all the genes present on the microarrays, we searched for optimal classifiers with two supervised algorithms, linear kernel support vector machine (SVM) and random forests (RF), both of them combined with a feature-selection step (see Materials and methods). The selections of features and algorithm parameters were nested in an inner cross-validation loop to preclude any possibility of parameters and feature-selection biases. None of the classification strategies produced a better classifier than the thyroid index for the FVPTC vs CPTC and FTA vs FTC classification tasks (Figures 4c and f).

The meta-PCNA index was inversely correlated to the thyroid index in the above panel of tumors (P=2 × 10−5) and was significantly higher in FTC than FTA (P=0.01) and in CPTC than FVPTC (P=0.05). Yet, meta-PCNA was not as discriminant as the thyroid index according to the AUC metrics (not shown).

In conclusion, the thyroid index integrates most transcriptional information relevant to the differential diagnosis of thyroid tumors, and does so for tumors arising from different dedifferentiation pathways.

Robustness of the organ-specific gene selection

We defined thyroid-specific genes as among the 1000 most expressed in the thyroid and not among the 5000 most expressed in any of the 15 non-thyroid organs profiled in the BodyMap 2.0. Hence, we choose thyroid genes that have high expression in the thyroid, that is, discarded gene with important thyroid-specific function, but moderate expression. For example, 13 136 genes have higher thyroid expression than the sodium iodide symporter (SLC5A5, a.k.a. NIS) in BodyMap 2.0. In addition, some genes are characteristic of the thyroid, but may also have a role in another organ. For example, paired box gene 8 (PAX8) is a transcription factor involved in thyroid and kidney differentiation.

We investigated a radical departure from the above thyroid-gene definition by selecting genes that are among the 5000 most expressed thyroid genes, but not in the 7000 most expressed genes in at least 14 of the 15 non-thyroid organs of the BodyMap 2.0. This resulted in a list of 72 thyroid genes (Supplementary table S4). It included additional known thyroid genes, for example, NKX2-1 (a.k.a. TTF1), DIO2, DUOX1, DUOX2 and SLC26A4 (a.k.a. pendrin). But, most of the new genes were, as far as we could tell, not known to be characteristic of this organ. Nevertheless, all the results from the previous sections could be reproduced, with slightly better performances, with a thyroid index computed from this gene list (Supplementary Figures S5 to S7). In particular, the AUC for FTA vs FTC increased from 0.82 to 0.86, and that for FVPTC vs CPTC from 0.86 to 0.94. Thus, our method is robust with respect to the definition of thyroid-specific expression, and suggests that many genes with thyroid-specific expression await functional characterization.

Discussion

We derived an index of thyroid differentiation exclusively from normal tissue gene expression compendia. We then showed that (1) the thyroid index brings into a quantitative framework the qualitative belief that tumor proliferation and differentiation are inversely related; (2) the thyroid index nevertheless quantifies a process biologically independent of proliferation, as both are positively correlated in a TSH time course experiment; (3) follicular and papillary thyroid cancers dedifferentiate along two distinct gene expression continuums; (4) the thyroid index distinguishes tumor subtypes within each direction and, (5) it does so as accurately as classifiers derived from a whole genome search for genes distinguishing these subtypes.

We could apply the thyroid index and get biologically meaningful results on in vitro primary cell data and in vivo tumor data generated with Affymetrix U133v2, Affymetrix U95av2 (Aldred et al., 2004; Finley et al., 2004) and on custom cDNA dual channel arrays (van Staveren et al., 2006). Furthermore, the erroneous inclusion of parathyroid hormone in the analysis of U95av2 data did not impair our ability to classify tumors as accurately as possible in this data set. These results support the robustness of the thyroid index across systems and gene expression platforms.

However, all data sets investigated were either generated on single channel arrays, or from sequencing, or from dual channel arrays with mRNA references shared among arrays. Some studies fit none of these setup. The thyroid index could not discriminate FVPTC from CPTC in a data set from our lab (Delys et al., 2007; Detours et al., 2007) based on dual channel arrays on which individual tumors where hybridized together with patient-matched healthy thyroid tissues (data not shown).

Besides its use as a tool for basic research, the thyroid index could be relevant to a range of clinical problems. It could guide pathologists in the complex differential diagnosis of FTA and FTC, perhaps even preoperatively. In addition, the thyroid index is lower in more aggressive tumors. We observed this when comparing it between FTAs and FTCs, and between PTCs and ATCs. FVPTCs and CPTCs patients have similar overall survival, but the latter have more lymph node metastases and more frequent extrathyroidal involvement (Lang et al., 2006; Lin and Bhattacharyya, 2010), in agreement with their lower thyroid index. The thyroid index could also distinguish PTCs from ATCs (Figure 2a). Thus, it could be a useful prognostic marker applicable to several types of thyroid tumors. Measuring the eight genes of the index is manageable in a clinical setup. Large-scale studies are needed to validate the clinical utility of the index.

Rhodes et al. (2004) proposed a gene expression signature believed to correlate positively with dedifferentiation in bladder, brain, breast, prostate, lung and ovarian tumors. This signature shares 23 of the 67 genes of the meta-PCNA index (hypergeometric test yields P=5 × 10−30), including PCNA itself, TOP2A and MCM2 among other proliferation genes. Moreover, the signature of Rhodes et al. is positively correlated with meta-PCNA in tumors (Supplementary Figure S8). Thus, it measures dedifferentiation to the extent that it is correlated to proliferation. By contrast, the thyroid index is organ-specific, it correlates negatively with tumor proliferation, dedifferentiation and aggressiveness, and it measures a quantity biologically independent of proliferation as demonstrated by the reversal of the correlation relationship upon TSH treatment of primary thyroid cells. This does not imply of course that the thyroid index of a tumor does not convey information about its proliferation rate.

The thyroid index is tissue specific, but not the method used to derive it. Could this method be applied to other organs? In addition to fibroblasts and endothelial cells, the thyroid is composed of follicular and C cells, but the follicular cells are far more numerous and are the relevant cells in the diseases investigated in this paper. Several other organs have a more complex cell-type composition or the cell type relevant to medical application is not the dominant cell type. For example, the epithelial cells that potentially give rise to breast cancer are few compared with the mass of adipocytes that constitute the normal non-lactating breast. Thus, although profiling of bulk tissues is adequate to select the thyroid index genes, profiling of carefully isolated cell types may be required for other tissues. On the other hand, the cell types in a complex organ may share common transcriptional characteristics. For example, in two normal tissue compendia with a detailed coverage of brain structures, all brain tissues cluster together apart from other non-brain tissues (Jongeneel et al., 2005). This opens the possibility to derive both region-specific differentiation indices and a generic brain differentiation index. Their usefulness remains to be investigated in specific applications. The same situation applies to white blood cells. Interestingly, the broad classes of hematological cancers can be traced back to the specific differentiation lineages of the cells they affect. The vast amount of gene expression data sets available for these cancers (Kohlmann et al., 2008; Mullighan et al., 2008; Mills et al., 2009), and the fact that carefully sorted cell types were systematically profiled (Novershtern et al., 2011), opens exciting prospects for our method.

Materials and methods

Data

Normal tissue data sets

We obtained the fastq files for the paired-end BodyMap 2.0 data from Illumina (San Diego, CA, USA) (also available from NCBI's GEO http://www.ncbi.nlm.nih.gov/geo, accession number GSE30611). Reads were aligned on the reference human genome hg18 with the Bowtie/Tophat suite (Kim and Salzberg, 2011) in supervised node using the ENSEMBL transcript database (--gtf option) and default parameters otherwise. Expression was then quantified with Cufflink (Roberts et al., 2011). We used merged isoform expression. Expression values of mitochondrial genes, pseudogenes and noncoding transcripts were ignored in all subsequent analysis.

Expression profiles from four additional gene atlases profiling healthy human tissues were retrieved as gcrma-normalized (Wu et al., 2004), gene-annotated matrices from the InSilico database (insilico.ulb.ac.be, Taminau et al., 2011), or from GEO for non-Affymetrix data. These include GSE1747 (Jongeneel et al., 2005; 25 tissues profiled by massively parallel signature sequencing), GDS181 (Su et al., 2002; 39 tissues profiled on Affymetrix U95a), GSE2361 (Ge et al., 2005); 26 tissues profiled Affymetrix U133A) and GSE3526 (Roth et al., 2006; 43 tissues profiled on Affymetrix U133v2). The list of tissue names was then manually standardized across studies. For instance, skeletal muscle tissues of distinct origins, otherwise labeled by their anatomical designation, were aggregated under the skeletal.muscle label and central nervous system tissues in GSE3526, representing nearly half of the profiled tissues in that study, were aggregated under the CNS label in order to avoid biases. After standardization, expression values from tissues with the same label were averaged. Diseased and fetal tissues were subsequently discarded.

ATC–PTC–normal thyroid data set

A group of 11 anaplastic thyroid carcinomas (ATCs), together with 49 papillary thyroid carcinomas (PTCs) paired with 45 adjacent tissues (N), were hybridized onto Affymetrix U133v2 arrays. PTCs were obtained from the Chernobyl Tissue Bank (CTB, http://www.chernobyltissuebank.com). ATCs were obtained from the Jules Bordet Institute (Brussels, Belgium) and the Ambroise Paré Hospital (Paris, France). All tissues were immediately dissected, placed on ice, snap-frozen in liquid nitrogen and stored at –80 °C until processing. The ethics committees of the involved institutions approved the protocol. RNAs were prepared using TRIzol Reagent and RNeasy columns (Qiagen, Venlo, Netherlands), followed by a verification of their quality. Amplification and hybridizations were performed following Affymetrix instructions. CEL files were normalized with RMA (Irizarry et al., 2003) and gene annotations were reconstructed using the Bioconductor package annotate (Gentleman et al., 2004). Data have been deposited in the GEO database (http://www.ncbi.nlm.nih.gov/geo, pending submission).

FTA–FTC–CPTC–FVPTC–normal thyroid data set

Seven normal thyroid samples (N) and nine follicular thyroid carcinomas (FTCs) were profiled by microarray as referenced in Aldred et al. (2004). We downloaded them from the authors’ web site. A total of 17 follicular thyroid adenomas (FTAs), 9 FTCs, 13 FVPTC and 9 classical papillary thyroid carcinomas (CPTCs) were profiled by microarray as described in Finley et al. (2004). These data have been submitted to GEO with accession number GSE29315. CEL files from these 64 samples were normalized altogether with RMA (Irizarry et al., 2003) and gene annotations were reconstructed using the Bioconductor (Gentleman et al., 2004) package annotate.

TSH-time course data set

This data set, published in van Staveren et al. (2006), was downloaded from http://www.ulb.ac.be/medecine/iribhm/microarray/data/ and used without any further processing.

Computational Methods

Software platform

All calculations were performed with the R language for statistics version 2.11.0 (R Development Core Team) and Bioconductor 2.6 (Gentleman et al., 2004) software and annotation packages. Graphical outputs (with the exception of ROC curve analyses) were generated with the R package ggplot2 (Wickham, 2009). All functions were run with default parameters unless specified otherwise.

Gene signatures and indices

Thyroid differentiation signature

To derive a thyroid differentiation signature, we ranked genes according to their FPKM in each tissues and selected genes with rank <1000 in the thyroid and >5000 in all other organs (or <5000 in thyroid, >7000 in 14 non-thyroid tissues for the alternate list).

Platform-specific probe selection

The following was repeated for each gene on both the compendia of Roth et al. (Affymetrix U133v2, GSE3526) and Su et al. (Affymetrix U95av2, GDS181). We first filtered the probes by retaining those whose median expression in all thyroids profiled ranked in the top 10% of the measurements across all tissues. From those left, if any, we then selected the probe with the highest t-statistics when comparing its expression in thyroid vs all other organ expression.

Meta-PCNA signature

The signature can be retrieved from the online supplementary material of Venet et al. (2011).

Indices computation

The thyroid and the meta-PCNA indices were computed on a per microarray basis as the median of the expression of the genes comprised, respectively, in the thyroid differentiation signature and in the meta-PCNA signature.

Statistical tests

Non-parametric Mann–Whitney U tests, as implemented in the R function wilcox.test, were used to test the null hypothesis that the observed distributions of either the thyroid index or the meta-PCNA index between two morphologically distinct cancer types were drawn from the same distribution.

A test for association of paired samples, based on Spearman's rank correlation (as implemented in the R function cor.test, was used to test the correlations between paired observations of the thyroid index and the meta-PCNA index.

The function roc.test from the pROC package was used to compare ROC curves, implementing the method described in DeLong et al. (1988).

Unsupervised analyses

Multidimensional scaling analysis

Multidimensional scaling was performed over a distance matrix of the FTA-FTC-CPTC-FVPTC-N data set obtained by computing the correlation (Pearson method) over all complete pairs of observations between samples. The function isoMDS (with k=2, maxit=1000, tol=1e–20) of the MASS R package (http://cran.r-project.org/) was then used to compute the plot of Figure 3.

Supervised analyses

Machine learning

A customized version of the Bioconductor package MCRestimate (Ruschhaupt et al., 2004) was used to determine the best possible RF and linear kernel SVM\ classifiers to predict the outcome of the FTA vs FTC and CPTC vs FVPTC classification in the FTA-FTC-CPTC-FVPTC-N data set. This package uses a protocol of repeated inner/outer cross-validation to estimate the expected accuracy of the prediction of each of those classifiers on new data.

RF and SVM classifiers were built for following feature selections: thePreprocessingMethods = varSel.highest.t.stat. Classifiers were optimized over large ranges for their specific parameters. For RF classifiers, the following range of values for each parameter was optimized: var.numbers {2, 4, 8, 16, 32, 64, 128, 256, 512}; nodesize {1, 3, 5}; ntree {250, 500, 1000}. For SVM classifiers, the following range of values for each parameter was optimized: var.numbers {2, 4, 8, 16, 32, 64, 128, 256, 512}; cost {0.0001, 0.001, 0.01, 0.1, 1}.

Classifiers were trained on 9/10 of our samples and used to predict the remaining 1/10. They were tuned on 9/10 of the training samples and optimized by comparing the prediction on the remaining 1/10 of the training samples. For each sample, our version of MCRestimate returns for RF a percentage of decision trees votes and for SVM a real score between −1 and 1. The whole loop was repeated 10 times and the results were averaged. Those scores were then used as a diagnostic test to compute ROC curves.

ROC curve analyses

ROC curves and AUC were computed using the R package pROC. AUC P-values were obtained by permuting 10 000 times the class labels and counting the fraction of permutation AUCs greater of equal than those obtained from the original data. We also checked the AUC values against the null hypothesis that any signatures of eight probes is diagnostic (Venet et al., 2011) by evaluating AUCs on non-permuted samples from 100 000 indices computed from eight probes selected at random among all probes printed on the microarrays.

Accession codes

Accessions

GenBank/EMBL/DDBJ

References

  1. Aldred MA, Huang Y, Liyanarachchi S, Pellegata NS, Gimm O, Jhiang S et al. (2004). Papillary and follicular thyroid carcinomas show distinctly different microarray expression profiles and can be distinguished by a minimum of five genes. J Clin Oncol 22: 3531–3539.

    CAS  Article  Google Scholar 

  2. Baloch ZW, Hendreen S, Gupta PK, LiVolsi VA, Mandel SJ, Weber R et al. (2001). Interinstitutional review of thyroid fine-needle aspirations: impact on clinical management of thyroid nodules. Diagn Cytopathol 25: 231–234.

    CAS  Article  Google Scholar 

  3. Clary KM, Condel JL, Liu Y, Johnson DR, Grzybicki DM, Raab SS . (2005). Interobserver variability in the fine needle aspiration biopsy diagnosis of follicular lesions of the thyroid gland. Acta Cytol 49: 378–382.

    Article  Google Scholar 

  4. DeLong ER, DeLong DM, Clarke-Pearson DL . (1988). Comparing the areas under two or more correlated receiver operating characteristic curves: a nonparametric approach. Biometrics 44: 837–845.

    CAS  Article  Google Scholar 

  5. Delys L, Detours V, Franc B, Thomas G, Bogdanova T, Tronko M et al. (2007). Gene expression and the biological phenotype of papillary thyroid carcinomas. Oncogene 26: 7894–7903.

    CAS  Article  Google Scholar 

  6. Detours V, Delys L, Libert F, Weiss Solís D, Bogdanova T, Dumont JE et al. (2007). Genome-wide gene expression profiling suggests distinct radiation susceptibilities in sporadic and post-Chernobyl papillary thyroid cancers. Br J Cancer 97: 818–825.

    CAS  Article  Google Scholar 

  7. Finley DJ, Zhu B, Barden CB, Fahey TJ . (2004). Discrimination of benign and malignant thyroid nodules by molecular profiling. Ann Surg 240: 425–436; discussion 436–437.

    Article  Google Scholar 

  8. Ge X, Yamamoto S, Tsutsumi S, Midorikawa Y, Ihara S, Wang SM et al. (2005). Interpreting expression profiles of cancers by genome-wide survey of breadth of expression in normal tissues. Genomics 86: 127–141.

    CAS  Article  Google Scholar 

  9. Gentleman RC, Carey VJ, Bates DM, Bolstad B, Dettling M, Dudoit S et al. (2004). Bioconductor: open software development for computational biology and bioinformatics. Genome Biol 5: R80.

    Article  Google Scholar 

  10. Hegedüs L . (2004). Clinical practice. The thyroid nodule. N Engl J Med 351: 1764–1771.

    Article  Google Scholar 

  11. Irizarry RA, Hobbs B, Collin F, Beazer-Barclay YD, Antonellis KJ, Scherf U et al. (2003). Exploration, normalization, and summaries of high density oligonucleotide array probe level data. Biostatistics 4: 249–264.

    Article  Google Scholar 

  12. Jongeneel CV, Delorenzi M, Iseli C, Zhou D, Haudenschild CD, Khrebtukova I et al. (2005). An atlas of human gene expression from massively parallel signature sequencing (MPSS). Genome Res 15: 1007–1014.

    CAS  Article  Google Scholar 

  13. Kebebew E, Greenspan FS, Clark OH, Woeber KA, McMillan A . (2005). Anaplastic thyroid carcinoma. Treatment outcome and prognostic factors. Cancer 103: 1330–1335.

    Article  Google Scholar 

  14. Kim D, Salzberg SL . (2011). TopHat-Fusion: an algorithm for discovery of novel fusion transcripts. Genome Biol 12: R72.

    CAS  Article  Google Scholar 

  15. Kohlmann A, Kipps TJ, Rassenti LZ, Downing JR, Shurtleff SA, Mills KI et al. (2008). An international standardization programme towards the application of gene expression profiling in routine leukaemia diagnostics: the Microarray Innovations in LEukemia study prephase. Br J Haematol 142: 802–807.

    CAS  Article  Google Scholar 

  16. Kondo T, Ezzat S, Asa SL . (2006). Pathogenetic mechanisms in thyroid follicular-cell neoplasia. Nat Rev Cancer 6: 292–306.

    CAS  Article  Google Scholar 

  17. Lang BH-H, Lo C-Y, Chan W-F, Lam AK-Y, Wan K-Y . (2006). Classical and follicular variant of papillary thyroid carcinoma: a comparative study on clinicopathologic features and long-term outcome. World J Surg 30: 752–758.

    Article  Google Scholar 

  18. Lin HW, Bhattacharyya N . (2010). Clinical behavior of follicular variant of papillary thyroid carcinoma: presentation and survival. Laryngoscope 120 (Suppl 4): S163.

    Article  Google Scholar 

  19. Lubitz CC, Gallagher LA, Finley DJ, Zhu B, Fahey TJ . (2005). Molecular analysis of minimally invasive follicular carcinomas by gene profiling. Surgery 138: 1042–1048; discussion 1048–1049.

    Article  Google Scholar 

  20. Mills KI, Kohlmann A, Williams PM, Wieczorek L, Liu W-min, Li R et al. (2009). Microarray-based classifiers and prognosis models identify subgroups with distinct clinical outcomes and high risk of AML transformation of myelodysplastic syndrome. Blood 114: 1063–1072.

    CAS  Article  Google Scholar 

  21. Moreno JC, Klootwijk W, van Toor H, Pinto G, D'Alessandro M, Lèger A et al. (2008). Mutations in the iodotyrosine deiodinase gene and hypothyroidism. N Engl J Med 358: 1811–1818.

    CAS  Article  Google Scholar 

  22. Mullighan CG, Miller CB, Radtke I, Phillips LA, Dalton J, Ma J et al. (2008). BCR-ABL1 lymphoblastic leukaemia is characterized by the deletion of Ikaros. Nature 453: 110–114.

    CAS  Article  Google Scholar 

  23. Novershtern N, Subramanian A, Lawton LN, Mak RH, Haining WN, McConkey ME et al. (2011). Densely interconnected transcriptional circuits control cell states in human hematopoiesis. Cell 144: 296–309.

    CAS  Article  Google Scholar 

  24. R Development Core Team, R: A Language and Environment for Statistical Computing. 1: ISBN 3-900051-07-0.

  25. Rhodes DR, Yu J, Shanker K, Deshpande N, Varambally R, Ghosh D et al. (2004). Large-scale meta-analysis of cancer microarray data identifies common transcriptional profiles of neoplastic transformation and progression. Proc Natl Acad Sci USA 101: 9309–9314.

    CAS  Article  Google Scholar 

  26. Roberts A, Pimentel H, Trapnell C, Pachter L . (2011). Identification of novel transcripts in annotated genomes using RNA-Seq. Bioinformatics 27: 2325–2329.

    CAS  Article  Google Scholar 

  27. Roth RB, Hevezi P, Lee J, Willhite D, Lechner SM, Foster AC et al. (2006). Gene expression analyses reveal molecular relationships among 20 regions of the human CNS. Neurogenetics 7: 67–80.

    CAS  Article  Google Scholar 

  28. Ruschhaupt M, Huber W, Poustka A, Mansmann U . (2004). A compendium to ensure computational reproducibility in high-dimensional classification tasks. Stat Appl Genet Mol Biol 3: Article37.

    Article  Google Scholar 

  29. van Staveren WCG, Solís DW, Delys L, Venet D, Cappello M, Andry G et al. (2006). Gene expression in human thyrocytes and autonomous adenomas reveals suppression of negative feedbacks in tumorigenesis. Proc Natl Acad Sci USA 103: 413–418.

    CAS  Article  Google Scholar 

  30. Su AI, Cooke MP, Ching KA, Hakak Y, Walker JR, Wiltshire T et al. (2002). Large-scale analysis of the human and mouse transcriptomes. Proc Natl Acad Sci USA 99: 4465–4470.

    CAS  Article  Google Scholar 

  31. Taminau J, Steenhoff D, Coletta A, Meganck S, Lazar C, de Schaetzen V et al. (2011). inSilicoDb: an R/Bioconductor package for accessing human Affymetrix expert-curated datasets from GEO. Bioinformatics [Internet]. Available from:http://www.ncbi.nlm.nih.gov/pubmed/21937664.

  32. Venet D, Dumont JE, Detours V . (2011). Most random gene expression signatures are significantly associated with breast cancer outcome. PLoS Comput Biol 7: e1002240.

    CAS  Article  Google Scholar 

  33. Wang Z, Gerstein M, Snyder M . (2009). RNA-Seq: a revolutionary tool for transcriptomics. Nat Rev Genet 10: 57–63.

    CAS  Article  Google Scholar 

  34. Wickham H . (2009). ggplot2: Elegant Graphics for Data Analysis 2nd edn. Springer (http://cran.r-project.org/).

  35. Wu Z, Irizarry RA, Gentleman R, Martinez-Murillo F, Spencer F . (2004). A model-based background adjustment for oligonucleotide expression arrays. J Am Stat Assoc 99: 909–917.

    Article  Google Scholar 

  36. Zurzolo C, Gentile R, Mascia A, Garbi C, Polistina C, Aloj L et al. (1991). The polarized epithelial phenotype is dominant in hybrids between polarized and unpolarized rat thyroid cell lines. J Cell Sci 98 (Part 1): 65–73.

    PubMed  Google Scholar 

Download references

Acknowledgements

This research was partially funded by the Brussels-Capital IRSIB project ICT-impulse 2006, In Silico Wet Lab. GT is supported by the Wallonie–Bruxelles International grant (7450/AMG/VDL/IN,WBI/doh/2009/21649). MT is supported by a FRIA fellowship from FNRS.

Author information

Affiliations

Authors

Corresponding author

Correspondence to V Detours.

Ethics declarations

Competing interests

The authors declare no conflict of interest.

Additional information

Supplementary Information accompanies the paper on the Oncogene website

Supplementary information

Rights and permissions

Reprints and Permissions

About this article

Cite this article

Tomás, G., Tarabichi, M., Gacquer, D. et al. A general method to derive robust organ-specific gene expression-based differentiation indices: application to thyroid cancer diagnostic. Oncogene 31, 4490–4498 (2012). https://doi.org/10.1038/onc.2011.626

Download citation

Keywords

  • gene expression
  • differentiation
  • proliferation
  • cancer
  • thyroid

Further reading

Search

Quick links