Thyroid nodules are endemic in iodine deficient areas, like Europes alpine regions, where they have a prevalence of 10-20%. They are classified by their histology into five main classes: the benign types Nodular Goiter (NG) and Follicular Thyroid Adenoma (FTA), and the malignant entities Follicular Thyroid Carcinoma (FTC), Papillary Thyroid Carcinoma (PTC), Medullary Thyroid Carcinoma (MTC) and Anaplastic Thyroid Carcinoma (ATC). Only approximately 5% - 10% of thyroid nodules are malignant [1], the majority of which are papillary carcinomas. Conventionally, discrimination between benign and malignant thyroid nodules is attempted by fine needle aspiration biopsy (FNAB) followed by cytological assessment. Despite many advances in the diagnosis and treatment of thyroid nodules and thyroid cancer, these methods have a well known low specificity [2], resulting in an 'indeterminate' or 'suspicious' diagnosis in 10% - 20% of cases. These patients usually undergo surgery, although in only 20% of these cases the nodules are actually malignant [3, 4]. This leads to a number of patients unnecessarily treated for malignant disease.

In other types of cancer it has been shown that gene expression profiling can add substantial value to the discrimination of the different clinically relevant tumor-entities [5, 6]. To date, many studies have tried to classify the different entities of thyroid carcinoma on the basis of their gene expression profiles. Each study has published a gene list which they believe discriminates between benign and malignant thyroid nodules or between different tumor entities. However, the lists have no or very few genes in common and applying a classifier from one study to the data from another study generally yields poor classification results. A notable exception to this are the studies from Jarzab et al. and Eszlinger et al. [7, 8] who established a 20-gene signature for PTC which they were able to apply to another study and classify all samples correctly.

Here we focus on Papillary Thyroid Carcinoma (PTC), which is the most common and therefore most extensively studied form of thyroid cancer. We hypothesised, that feature selection based on a larger sample cohort will be more robust than using single studies, so we decided for a meta analysis approach which allows us to analyse all publicly avail-able datasets (99 samples) for PTC with one common analysis approach. From this dataset we were able to identify SERPINA1 as a single gene which allows to discriminate between PTC and benign nodules or healthy tissue with 99% accuracy (one misclassification). To validate these findings and to quantify the discriminatory power of SERPINA1 for the detection of PTC, we performed RT-qPCR experiments on an independent set of thyroid nodules (instantaneous sections) measuring the mRNA levels of SERPINA1. In contrast to the meta analysis data, we included not only PTC and benign nodules but also all other histological entities of thyroid nodules plus rare histological variants (follicular variant and tall cell PTC). Despite the marked overrepresentation of these difficult to diagnose histological variants, the classification accuracy was still as high as 93%.

Encouraged by the high classification accuracy in the independent validation data, showing the validity of our meta analysis approach, we went on to perform Gene Set Enrichment Analysis (GSEA) on the meta-analysis microarray data to elucidate some of the specific mechanisms involved in papillary thyroid tumorigenesis.

The present microarray meta analysis makes use of the latest methods for microarray data integration and classification. Nevertheless, meta-analysis of microarray data still poses a challenge, mainly because researchers ask at least partly different questions and hence use different experimental designs. Moreover, the number of thyroid tumor microarray data available to researchers to date is still comparably low (compared to breast cancer, e.g.). Therefore, when doing meta analysis one is forced to use all data available, even if the patient cohorts represent a rather heterogeneous and potentially biased population. More specifically, it is difficult to obtain a homogenous collection of control material (from healthy patients). These are usually taken from patients who were operated for other thyroid disease which is in turn very likely to cause a change in gene expression as measured on microarrays. In other studies, healthy tissue surrounding the nodule was takes as control material. The generation of homogeneous patient cohorts is further hampered by limited availability of patient data like age, gender, genetic background, etc.

Nevertheless we believe to have shown that there still is a great deal of information to be discovered in the wealth of DNA microarray data deposited in databases like ArrayExpress and GEO [11, 12]. As shown in the present and other papers [9, 13], methods for harvesting this information are available. Using approaches like DWD for data integration followed by classification and validation, it is possible to analyse a large number of samples at minimal cost. For the discovery of mRNA-based markers, a part of the scientific community has already moved on to other, more accurate, even higher parallel and even more expensive methods like mRNA sequencing using next-generation sequencing technologies. We (see Table 1) and others [14] have shown that feature selection in highly multivariate data like mircoarray data is not very robust, due to high feature numbers and low sample numbers. Applying these findings to next-generation sequencing data, it is likely that the feature selection problems will be even more pronounced, since feature numbers are even higher and even fewer people will be able to afford the analysis of sample sizes which account for the heterogeneous nature of clinical samples. This is why we believe that low cost/high sample number techniques like the one utilised here can provide valuable additional information.

This is also supported by another meta analysis on papillary thyroid carcinoma data which has been conducted by Fujarewicz et al. [15]. They used 3 datasets, two of which were used in our study as well (Jarzab and He). They skipped data integration and used a bootstrap strategy instead, which consisted of iterative construction of Support Vector Machine (SVM) classifiers (on features selected for each iteration) based on randomly selected sets of specimen and testing the classifiers on the remaining samples. Through this sampling scheme, they select genes which exhibit little study bias and good discrimination between tumor subtypes. They achieved 98.5% classification accuracy between PTC and benign nodules on a 20 gene classification rule.

When doing meta analysis of microarray data, researchers have also based their approach on modeling of inter-study variation [16] or on comparing gene lists from published studies. A meta analysis and meta review by Griffith et. al. [17] has summarised genes with a diagnostic potential in the context of thyroid disease. They published lists of genes which appeared in more than one high-throughput study (Microarray, SAGE) analysing thyroid disease and applied a ranking system. In their analysis SER-PINA1 scored the third highest. Six of our top 20 most significant genes appear in the list of Griffith et al. Their approach is very useful, as one can include all studies in the analysis and is not limited to the studies where raw data is available. However, the studies generally follow very different analysis strategies, some more rigorous than others. It is not under the control of the meta-analyst how the authors arrived at the gene lists. It is also not possible to assign a measure of confidence at the gene level. However, given the overlap in the marker genes identified by these two methods, they seem to complement each other well for in-silico marker discovery. ANOVA-based approaches are very useful for identifying differentially regulated genes but to our knowledge there is no method to test these genes in the meta analysis data for their discriminatory power.

Most of these lists were generated from microarray analysis. However, even when comparing the genes in the classifiers to gene lists generated with independent technologies, like cDNA library generation as performed by Kaserer et al. [18], there is substantial overlap. SERPINA1 appears in their lists as well as seven of our top 20 genes.

SERPINA1 has been reported before as a potential diagnostic marker for papillary thyroid carcinoma in a study investigating its protein expression in thyroid biopsy tissue by Immunocytochemistry and Western Blot [19]. In line with our results, they found SERPINA1 immunoreactivity in nine of ten papillary biopsies while the surrounding tissue showed no such reaction.

In GSEA, the most apparent feature of PTC is an upregulation of extracellular activities, like cell communication and adhesion and extracellular matrix receptor interactions. While the gene lists and the derived classification rules of the 4 studies under investigation here show little over-lap, we found the same pathways to be overrepresented when each of the single studies is being analysed separately. Upregulation of ECM microenvironment genes have been described as an effect of BRAF mutations through phospho-ERK1/2 signalling, which in turn may be involved in the up-regulation of some ECM remodeling genes like TSP-1 [20]. While ECM has been seen as mere scaffolding for many years, recently its role in tumor progression and invasion through ECM remodeling and stiffening of the tissue stroma became evident [21]. Similar Pathways have been reported to be implicated in the progression from simple ductal hyperplasia to atypical ductal hyperplasia and further to ductal carcinoma in situ in ER+ sporadic breast cancer by Emery et al. [22]. Like us, they found, amongst other pathways, overrepresentation of the ECM-receptor interaction, cell communication and p53 signaling pathways, but none of the downregulated metabolic pathways which we identified in papillary thyroid carcinoma. Their data also suggests, that in breast cancer progression, dysregulation of the ECM-epithelial cell interactions is an early event, even before any evidence of invasion becomes visible. Similarly, Birnie et al. [23] reported Cell Adhesion and ECM receptor interactions pathway dysregulated when comparing prostate cancer stem cells to their normal and differentiated counterparts .

The power of the meta analysis approach adopted here is demonstrated by a 99% loocvaccuracy (97.9% weighted average accuracy in the study crossvalidation) for the distinction between papillary thyroid carcinoma and benign nodules. One sample was classified wrongly, and although it is not possible to correctly map the samples from this analysis to the original analysis [7], the misclassified sample is from the same group (PTC, validation group) as the sample which was wrongly classified in the original analysis. According to Jarzab et. al. the sample was an outlier because it contained only ≈ 20% tumor cells.

In the qPCR data we achieved a prediction accuracy of 93%. We chose the threshold to allow for false positives rather than false negatives. For this threshold 6 samples were false positives and none false negative. Three follicular nodules were false positive, one of them was a recurrent FTC nodule and one was an oxyphilic FTA nodule. The third was a classical FTC. Two MTC and one ATC nodules were falsely positive. The latter two entities are rather easy to distinguish by alternative methods like cytology or serum calcitonin testing (MTC). However, the raw, unnormalised data (not shown) suggests, that the slightly elevated levels in the MTC and ATC group may not be down to elevated SER-PINA1 levels, but to a different cellular turnover in these cells which leads to a problem in the normalisation procedure. A more thorough search for appropriate normalisation genes may circumvent this problem.

Gerhard et al. [24] studied the intra- and inter-observer variability in the cytological assessment of fine needle aspirates from thyroid nodules. They found 24% (n = 97) disagreement between two observers, both of them experienced pathologists. While most of these disagreements were between different follicular nodules, there was still an inter-observer disagreement of 6.3% for the PTC nodules. This highlights the need for objective diagnostic methods which include the possibility to assign a measure of confidence in the final diagnosis. While PTC shows the most distinct cytological features and is therefore relatively easy to diagnose, a molecular diagnostic test could also provide the possibility to include molecular markers for tumor entities which are harder to diagnose by cytology.

These results are very encouraging, however, more studies need to be carried out in order to be able to transfer these findings into clinical applications for routine diagnostics. Firstly, markers with sufficient sensitivity and specificity for the other histological subtypes need to be found. There are few studies reporting molecular classifiers for follicular thyroid disease [25, 26], and these classifiers cannot be applied to data from other studies (data not shown). To date, microarray studies for medullary and anaplastic thyroid carcinoma are largely missing. Secondly, in all of these studies, frozen thyroid nodules e.g. from instantaneous section pathology have been used. However, molecular markers for thyroid carcinomas will only be useful for routine diagnostics, if they exhibit the same discriminative power in fine needle aspiration biopsy (FNAB) samples. Microarray data from a study on thyroid FN-ABs [27] and a study on FNABs of various different tumor types [28] show good correlation between the aspirate and the tissue, but more detailed data is needed.