Skip to main content

Molecular profiling of thyroid cancer subtypes using large-scale text mining



Thyroid cancer is the most common endocrine tumor with a steady increase in incidence. It is classified into multiple histopathological subtypes with potentially distinct molecular mechanisms. Identifying the most relevant genes and biological pathways reported in the thyroid cancer literature is vital for understanding of the disease and developing targeted therapeutics.


We developed a large-scale text mining system to generate a molecular profiling of thyroid cancer subtypes. The system first uses a subtype classification method for the thyroid cancer literature, which employs a scoring scheme to assign different subtypes to articles. We evaluated the classification method on a gold standard derived from the PubMed Supplementary Concept annotations, achieving a micro-average F1-score of 85.9% for primary subtypes. We then used the subtype classification results to extract genes and pathways associated with different thyroid cancer subtypes and successfully unveiled important genes and pathways, including some instances that are missing from current manually annotated databases or most recent review articles.


Identification of key genes and pathways plays a central role in understanding the molecular biology of thyroid cancer. An integration of subtype context can allow prioritized screening for diagnostic biomarkers and novel molecular targeted therapeutics. Source code used for this study is made freely available online at


Thyroid cancer (TC) is the most common endocrine malignancy [1] and its incidence increase has been significant in recent years despite some controversies about the extent [2]. Many possible factors causing thyroid cancer have been reported including exposure to ionising radiation, iodine-deficiency and heredity [3]. Conventional treatment strategies include surgical resection, radiation therapy (especially radioactive iodine therapy), chemotherapy and thyroid hormone therapy [4]. However, the understanding of the underlying molecular mechanisms is still incomplete.

Thyroid tumours are usually classified into multiple subtypes according to their histopathological characteristics, and treatments are selected depending on the subtype and stage of thyroid cancer. The main subtypes include papillary thyroid cancer (PTC), follicular thyroid cancer (FTC), anaplastic thyroid cancer (ATC) and medullary thyroid cancer (MTC) [4]. PTC and FTC are also sometimes collectively referred to as differentiated thyroid cancer (DTC) or well-differentiated thyroid cancer (WDTC), while ATC can also be referred to as undifferentiated thyroid cancer. In addition, a number of rare subtypes have been described. Cellular origins and some of the known molecular mechanisms differ for each subtype [5], which may include subtype-specific alterations in DNA methylation patterns [3] and have led to new therapeutic approaches based on the molecular signature of the tumours.

These "targeted therapeutics" of thyroid cancer are being rapidly developed [6, 7]. Several potential drugs are currently in preclinical testing or in clinical use [8, 9]. However, lack of systematic studies of underlying molecular mechanisms can lead to a high risk for thyroid cancer patients, who might suffer from unexpected side effects. For instance, RET has been shown to be an oncogene in thyroid cancer but is considered as a potential tumour suppressor gene in colorectal cancer [10]. Consequently, studies focusing on one or a few genes are likely to miss the molecular context that could be vital for a comprehensive understanding of the disease.

For systematic studies, a major challenge is to efficiently utilise the myriad of knowledge and information from unstructured scientific literature. PubMed, one of the most widely used systems for biomedical literature search [11], returns over 50,000 results with the search query 'thyroid cancer'. The number is increasing rapidly, with over 2,000 articles published annually in recent years, as illustrated in Figure 1. This trend has made it extremely difficult for scientists to identify, retrieve and assimilate all relevant publications.

Figure 1
figure 1

Number of PubMed articles returned by the 'thyroid cancer' query. The size of thyroid cancer related literature is increasing rapidly, with over 2,000 articles published annually in recent years.

To support the identification and retrieval of relevant articles, MeSH (Medical Subject Headings, terms are used in PubMed to label the topic of each document, along with additional headings known as Supplementary Concepts (SCs). The MeSH term used for thyroid cancer is 'thyroid neoplasms', and the major subtypes are represented as SCs. However, the thyroid cancer SCs were only introduced in August 2010, which means the majority of the thyroid cancer literature does not have a subtype label indicated by SC.

Text mining techniques have been extensively used to support access to the biomedical literature [12, 13]. A number of tasks can now be successfully fulfilled by text mining, including gene mention recognition and normalization [14, 15], pathway mention recognition [16] and event extraction [17, 18]. Those efforts enable the extraction of structured and explicit data, and this further targets the exploration and investigation of specific biological questions.

In this paper we focus on identifying genes and pathways that are reported in the thyroid cancer literature, grouped by subtypes. We then compare our systematic effort to some established databases and demonstrate how our results could boost the understanding of thyroid cancer and promote comprehensive investigation for potential therapeutic targets.

Related work

There are several curated databases that contain molecular information of thyroid cancer. For example, the Thyroid Cancer and Disorder Gene Database (TCGDB, contains information about genes and miRNAs involved in different thyroid cancer subtypes. The data in TCGDB have been extracted through manual literature review of a few selected articles. However, details of these articles are not revealed, so it is not clear how systematic and comprehensive the database is.

MalaCards is a generic integrated compendium for diseases and associated annotations [19], with some information related to TC. It provides disease-specific information from multiple sources, including related publications, genes and pathways. However, there are several limitations: firstly, the publications associated with a disease are obtained by title search only, which could miss out many relevant articles; secondly, pathways are provided by affiliated genes rather than their mentions in the literature. Finally, the coverage of this data set is limited: for example, a search for "anaplastic thyroid cancer" returns no affiliated genes.

Gene2Pubmed [20] is a generic database maintained by NCBI that provides a set of manually added links from PubMed articles to genes. Gene2Pubmed integrates information from a number of public databases and the links are not limited to articles specifically defining the function of a given gene. However, similarly to other resources, its coverage is limited: for example, we found quite a few examples where TC-related genes were missing from the Gene2Pubmed annotation (see Results).

Several databases contain information about pathways and their associations to diseases. KEGG [21], for example, contains pathways for different diseases, but its coverage is often limited, in particular for diseases that are not widely studied (e.g. there is a single pathway diagram for thyroid cancer). Even for databases that focus on specific diseases (e.g. AlzPathway [22] is a database of pathways associated to Alzheimer's disease) it is often difficult to keep up with new findings through manual curation, and automated text mining approaches need to be used to ensure a wider coverage [16].

There are various efforts to automatically extract relationships between cancer and its molecular bases from the literature. For instance, a maximum entropy-based named entity recognizer and relation recognizer were applied to find relations between prostate cancer and genes [23]. Other relation finding tools can be applied to cancer related studies as well, including FACTA [24], CoPub [25], DigSee [26] and OncoSearch ( Both FACTA and CoPub support searching for co-occurring biomedical concepts (genes/diseases), facilitated by indexing through dictionary matching or regular expression matching. DigSee and OncoSearch can scan the literature to identify whether a gene is up-regulated or down-regulated in a particular type of cancer. The systems look for genes that are reported together with a biological event and cancer in the same sentence. While the results are therefore often highly accurate, they may not be comprehensive. However, no systems could support extraction of information specific to thyroid cancer subtypes.


Pipeline for generating the molecular profiling

The pipeline we engineered to generate the molecular profiling of thyroid cancer subtypes is depicted in Figure 2. It starts with the corpus construction and is followed by the text mining tasks implemented as annotators within TextPipe [18], a framework designed for large-scale, modular text processing and data integration. Text mining results are then used to find the association between genes/pathways and thyroid cancer subtypes. These stages are explained in the following subsections.

Figure 2
figure 2

The pipeline to generate the molecular profiling of thyroid cancer subtypes. The pipeline starts with corpus construction, which is followed by text mining annotators implemented within the TextPipe framework. The text mining results are then used to find genes/pathways association with thyroid cancer subtypes, which lead to a comprehensive molecular profiling of primary thyroid cancer subtypes.

Data - a corpus of thyroid cancer from MEDLINE

All the documents in our corpus were retrieved from MEDLINE. We performed a PubMed search with the query "(((thyroid neoplasms[majr] AND human[mh] AND english[la]) OR thyroid[ti]) AND (cancer OR carcinoma OR malignant OR malignancy))", as suggested by the National Cancer Institute for retrieving the TC-relevant literature (see It should be noted that the query limits the results to human studies in English. A total of 38,572 PubMed IDs (PMIDs) were fetched from PubMed (accessed on 13/01/2014). The complete list of PMIDs is available in Additional File 1. We retrieved abstracts from PubMed using those PMIDs and constructed the thyroid cancer corpus (TC_TEXT) using PyPubmedText, which is freely available from

For the evaluation, we created the SC_GOLD corpus, which consists of all 1,213 documents with specified subtype SCs since their introduction in 2010 (see Additional File 2).

Subtype classification using subtype scoring

For subtype classification, we aimed to label the articles in the TC_TEXT corpus with the major TC subtype(s) discussed in the associated abstract. The subtypes we investigated included PTC, ATC, FTC and MTC. We considered this as a document classification task and have developed an efficient classification method based on a subtype-scoring scheme. Although machine-learning methods are popular for document classification [27], we did not employ them here due to insufficient training data for each of the TC subtypes: while the recent articles with SCs can be used as the training data, the number of documents is still limited (1,213 in total, much fewer for each subtype). Instead, we used a flexible dictionary matching method as described below.

For each article in the corpus, we calculate a vector of scores corresponding to each TC subtype through the following steps:

1. Subtype names are often mentioned partially and in abbreviated forms, due to language economy and flexibility (e.g. "...Inflammatory infiltrates could increase the risk of papillary cancer in patients with autoimmune lymphocytic thyroiditis..." (PMID: 21042739)). We therefore dissociate each subtype name into three parts: (1) subtype keyword (SK); (2) anatomy keyword (AK); (3) malignancy keyword (MK). For instance, in "papillary thyroid cancer", the SK is "papillary", the AK is "thyroid" and the MK is "cancer". We then match those parts against text separately, maintaining the three separate keyword lists for matching: a subtype keyword list, an anatomy keyword list and a malignancy keyword list. For each list, possible variants are also included. For instance, "anaplastic thyroid cancer" is also named as "undifferentiated thyroid cancer", which can be written as "un-differentiated thyroid cancer". So "anaplastic", "un-differentiated" and "undifferentiated" will all be mapped to "anaplastic". The lists are included in the additional files. For matching, we applied the open-source package LINNAEUS [28].

2. Each input document is split into sentences. For each sentence, we apply dictionary matching to calculate the subtype relevance scores of that sentence calculated as follows: (1) if SK co-occurs within that sentence with both AK and MK, then the sentence has a relevance score 1 to the subtype SK; (2) if SK co-occurs with only AK or MK, then the sentence has a relevance score 0.5 to the subtype SK; (3) if SK appears alone without the presence of either AK or MK, the sentence has a relevance score 0.25 to the subtype SK. We defined these rules to cover the situations where subtype names are simplified, as illustrated above.

3. The vector of subtype scores S i = S i P , S i A , S i F , S i M for article d i is calculated by weighted accumulation of the vectors of subtype scores for each sentence (the four elements in S i correspond to PTC, ATC, FTC and MTC respectively). The weights of different sentences are assigned in the following way: the title of a document is considered as the most important element and is assigned a weight of 4; the first sentence in the abstract usually mentions the main topic of the document and the last sentence usually concludes the article, and both are assigned a weight of 2; the second and the penultimate sentence can be quite important as well, and are both assigned a weight of 1; other sentences in the abstract are given a weight of 0.5 in order to weaken bypassing mentions of subtype names that are not the major scope of the article.

For classification, we set threshold values for each subtype and assign the corresponding label to the article if the associated subtype score is above a pre-set value. We assigned slightly different thresholds to different subtypes, with the PTC's percentage threshold slightly higher than other subtypes, given that PTC occurs more frequently in the literature (over 50%).

Gene recognition and normalisation

A number of tools are available for identifying mentions of genes in the literature and normalising them to database identifiers. We utilized two open source libraries, Moara [14] and GNAT [15], which have been successfully applied in other studies [18, 26]. For gene name recognition, Moara utilizes the CBR-tagger [29], which treats the recognition problem as a binary classification on each token; GNAT employs dictionary-expanded regular expressions together with BANNER [30]. For the normalisation of recognized gene names, both tools map mentions to the Entrez Gene database [20], adopting similar methods for disambiguation. We wrapped Moara and GNAT as TextPipe annotators that output normalised gene mentions including document ID (PubMed ID), positions in the text, normalised Entrez Gene ID, original text and the applied tool (Moara or GNAT). If the two tools report overlapping mentions, we created a new one covering both mentions (union); if the overlapping mentions have been assigned different Entrez Gene identifiers, the priority is given to GNAT as its reported performance is higher.

Pathway mention recognition

We utilised PathNER for pathway mention recognition from the literature [16]. PathNER is implemented using soft dictionary matching and manually created rules. For this study, we recompiled the pathway dictionary using the data from the 2013 update of the ConsensusPathDB interaction database [31]. The outputs from this annotator are biological pathway mentions including document ID (PubMed ID), positions in the text, pathway database identifier (if available) and original text.

Association between genes/pathways and subtypes

To establish the association between text-mined information and thyroid cancer subtypes, we performed the following steps for each text-mined entity (gene or pathway, referred to as E thereafter):

  1. 1.

    Get all the documents D E = { d i | i [ 1 , n ] } that mention E according to the text-mining results;

  2. 2.

    For each document d i D E , get the subtype score vector S i = S i P , S i A , S i F , S i M as described above;

  3. 3.

    Sum up all S i to give the subtype relevancy vector of E:

    S E = i = 1 n S i = i = 1 n S i P , i = 1 n S i A , i = 1 n S i F , i = 1 n S i M = S E P , S E A , S E F , S E M

For a given subtype j { P , A , F , M } , we generate a list of entities L j , which is composed of entities that have at least one document labelled as subtype j in the classification stage; we then rank the entities in L j = E 1 , E 2 , , E k in a descending order by the value of S E j for each entity E in L j .

Results and discussion

Subtype classification evaluation

For performance assessment against the gold standard, we adopted the standard metrics: Precision (P), Recall (R) and F1-score (F1) defined by the following equations:

P= T P T P + F P ,R= T P T P + F N ,F1=2 P R P + R

Here, TP is the number of true positives, FP is the number of false positives and FN is the number of false negatives. Those metrics were calculated for each subtype. For documents with multiple subtype labels, the labels were evaluated separately.

The evaluation of the classification performance against the SC_GOLD corpus for each TC subtype is listed in Table 1. The performance for PTC, ATC and MTC shows good precision and moderate recall, with F1-score over 85%. This demonstrates the effectiveness of our scoring scheme.

Table 1 TC subtype classification performance on the SC_GOLD corpus.

Unexpectedly, the precision for FTC was low (17.5%). We therefore performed error analysis to investigate the reason for that. We found that more than half of all errors (54 of the 99 FPs for FTC) are from articles about DTC or WDTC, referring to both PTC and FTC [32]. However, it seems that the PubMed SCs typically only annotate an article with PTC if it is about DTC. This is probably because the SC for FTC was introduced in August 2012, while the SC for PTC was introduced two years earlier. If those 54 cases are considered TPs then the performance is obviously improved (F1 is 72.5%; listed in Table 1 as FTC*). We also found a few other variants such as "follicular variant papillary carcinoma", "mixed medullary-follicular carcinoma of the thyroid", "follicular variant of PTC", etc. We checked the frequency of those variants by dictionary matching and we found them in 31 documents in SC_GOLD. As our classification "dissects" thyroid cancer subtype names into several parts, follicular variants will score and get incorrectly classified as FTC in those situations. The FTC precision is thus more affected as "follicular" is a word commonly used in PTC articles, since PTCs are derived from follicular cells [33].

In Table 1, we also listed the performance of a baseline method for comparison. The baseline method performs classification based on subtype name occurrences in titles and abstracts by dictionary matching using LINNAEUS [28]. The dictionary includes all available synonyms for the TC subtype names. The performance results demonstrate that our method outperforms the baseline for all subtypes.

Application to TC_TEXT

We ran our classification method on the whole TC_TEXT corpus (38,572 abstracts). The complete classification results (with subtype relevancy scores and predicted labels) can be found in Additional File 3. The number of articles labelled with each subtype is listed in Table 2, which also shows the incidence rate of each subtype. We can see that the amount of literature somewhat reflects the incidence rate, with PTC as the most frequent and ATC as the least frequent among the four subtypes studied here. The results also show that it is relatively uncommon to discuss more than one subtype in an article (see Figure 3), with the exception of PTC and FTC, which are frequently studied together as they are both derived from follicular cells [33].

Table 2 Number of articles for each subtype in the TC_TEXT corpus.
Figure 3
figure 3

Overlap of documents related to different TC subtypes. The Venn diagram illustrates the overlap between documents associated to different subtypes.

Gene normalization results and gene-subtype association

We ran gene recognition and normalization on the TC_TEXT corpus and a total number of 2,833 unique genes were detected. We first compared the resulting list against the gene list generated from the Gene2Pubmed database [20]. Out of 914 genes identified in Gene2Pubmed, 762 (83.4%) are found by our text mining approach. On the other hand, 2,071 out of 2,833 text-mined genes (73.1%) are not listed in the Gene2Pubmed database (see Figure 4). We randomly sampled 50 genes from that difference set and manually checked whether they are associated with TC or not. Table 3 shows these genes sorted by their document-level frequency, according to text mining results. If a gene is related to TC, an example of relevant document is given. Overall, 42 genes out of 50 are found to be reported as associated with thyroid cancer. This highlights the importance of using text mining for a comprehensive retrieval of information, complementing manually annotated data, which may potentially miss over 1,700 TC-related genes (84% of 2,071). We also note that even genes with low document-level frequencies can be relevant: for instance, WNT7A (Entrez Gene ID: 7476) was only detected once by text mining, but is reported to be over expressed in both PTC and its aggressive variant. Therefore, we did not use any filters based on document-level frequency in this study.

Figure 4
figure 4

Comparison between text-mined TC genes and Gene2PubMed. This Venn diagram compares the text-mined genes (TM) to a manually annotated database Gene2PubMed (PM). The figure demonstrates a good coverage by text-mining and it also shows that text mining could be more comprehensive.

Table 3 Results of TC-relevance for 50 sampled text-mined genes not in Gene2PubMed.

With the subtype classification, we generated ranked gene lists for each subtype as discussed in the Methods section (see Table 4 for the statistics). The top 20 genes for each subtype are given in Table 5 (the full results are provided in Additional File 4). The genes are ranked in descending order by their subtype scores (as described in Methods section). To validate the results, we looked at the most recent review on thyroid cancer [34]. Almost all well-observed mutations (see Table 1 in [34]) are identified by our top text mining results, including BRAF, RET, TP53, TRK (NTRK1), RAS (RASA1), PAX8-PPARG, RET/PTC (PTCH1), PTEN and AKT. The only missing gene PIK3CA is ranked 30th and 32nd in the lists for FTC and ATC, respectively. This provides solid evidence for the capability of our method in assisting systematic acquisition of knowledge of the molecular biology of thyroid cancer.

Table 4 Number of genes extracted for each TC subtype.
Table 5 Top 20 genes for the four thyroid cancer subtypes as extracted from the literature.

We also compared the subtype gene lists generated by text mining to the manually created TCGDB database (see Table 6): only a small number of genes for each subtype were missed by our method. The majority of the genes in the subtype lists from TCGDB have been successfully retrieved by our method, along with a number of candidates for further manual curation.

Table 6 Text-mined genes compared with the TCGDB database.

Figure 5 shows a Venn diagram of gene lists shared between the four subtypes. Compared to Figure 3, we observe a much higher degree of overlap among subtypes. This indicates that a considerable number of common genes are investigated across different subtypes but in separate studies. Meanwhile, each subtype is still characterized by a significant number of unique genes.

Figure 5
figure 5

Overlap of genes associated to different TC subtypes. This figure shows a Venn diagram of gene lists for the four subtypes and their overlap.

Pathway mention recognition results

Table 7 presents the top 20 pathways for each TC subtype (For the complete text mining results for TC related pathways, please refer to Additional File 5). The most frequently mentioned pathways are typical biological processes involved in most cancers, including apoptosis, angiogenesis, cell cycle, etc.

Table 7 Top 20 pathways for the four thyroid cancer subtypes as extracted from the literature.

To validate these results, we compiled a list of pathways from the most recent review on the molecular biology of thyroid cancers mentioned above [34]. The list includes five pathways: the mitogen-activated protein kinase (MAPK) pathway, the phosphatidylinositol 3-kinase (PI3K)/Akt signaling pathway, the Wnt/beta-catenin pathway, the NF-kB pathway and the Hypoxia-Inducible Factor (HIF)-1α pathway. Except for the HIF-1α pathway, all other pathways are present in the top 20 pathways extracted for subtypes by text mining (see Table 8). The HIF-1α pathway was reported to be a potential therapeutic target for thyroid cancer [35] and has been experimentally verified in a couple of cell lines [36]. However, this pathway is absent from the text-mined results, even though the reporting paper [36] was included in TC_TEXT corpus. Our method did detect the HIF-1α pathway mention and found its association with thyroid carcinomas. However, the subtype relevancy scores for that abstract (PMID: 19808899) are low. The reason is that the cell lines used in the study were actually subtype-specific but this information was only available in the full-text. This example suggests the importance and potential of the analysis of full text articles as opposed to abstracts, which is an objective for future work.

Table 8 Reviewed TC related pathways and text-mined evidence for each subtype.

For other pathways in Table 7 that have not been mentioned in the above review, we have further investigated whether they are associated with the corresponding subtypes. For instance, the Notch signalling appears in the top 20 pathways for PTC, ATC and MTC but was not mentioned in [34]. We however found direct literature evidence that links this pathway to TC: the Notch pathway has crosstalk with the MAPK pathway and affects the PTC proliferation [37], and the activation of the Notch signalling has been identified as the potential therapeutic strategy for ATC [38]. We followed this procedure for each pathway in Table 7: an example PMID is given in the "Evidence" column, which represents an article that reports the association between the subtype and the pathway (if applicable). Almost all pathways in the top 20 list are supported by literature evidence: only for five pathways in the FTC list we were not able to find support in the literature. This demonstrates the potential of our work for the systematic and comprehensive acquisition of knowledge of thyroid cancer.


In this paper we presented an approach to finding molecular information associated with different subtypes of thyroid cancer. We developed a method for subtype classification and performed text mining to identify genes and biological pathways associated with each subtype. The generated gene and pathway lists provide a comprehensive compendium of the key molecular information related to thyroid cancer subtypes.

The data and results from our study form the basis for a further comprehensive analysis of the molecular biology of thyroid cancer, enriched by subtypes. Based on the molecular differences between subtypes revealed in our study, biologists can look for better diagnostic biomarkers. In addition, the molecular information can be used to build molecular networks of thyroid cancer subtypes, which would enable further systems biology analyses and stimulate the development of targeted therapeutics.


The code and data files are available at



Named Entity Recognition


PubMed ID


Thyroid Cancer


Papillary Thyroid Cancer


Follicular Thyroid Cancer


Anaplastic Thyroid Cancer


Medullary Thyroid Cancer


Differentiated Thyroid Cancer


Well-Differentiated Thyroid Cancer


Medical Subject Headings


Supplementary Concept


Thyroid Cancer and Disorder Gene Database


Subtype Keyword


Anatomy Keyword


Malignancy Keyword


True Positives


False Negatives


False Positives


  1. Nix P, Nicolaides A, Coatesworth AP: Thyroid cancer review 1: presentation and investigation of thyroid cancer. International Journal of Clinical Practice. 2005, 59: 1340-1344. 10.1111/j.1368-5031.2005.00671.x.

    Article  CAS  PubMed  Google Scholar 

  2. Ito Y, Nikiforov YE, Schlumberger M, Vigneri R: Increasing incidence of thyroid cancer: controversies explored. Nature Reviews Endocrinology. 2013, 9: 178-184. 10.1038/nrendo.2012.257.

    Article  CAS  PubMed  Google Scholar 

  3. Rodríguez-Rodero S, Fernández AF, Fernández-Morera JL, Castro-Santos P, Bayon GF, Ferrero C, Urdinguio RG, Gonzalez-Marquez R, Suarez C, Fernández-Vega I, Fresno Forcelledo MF, Martínez-Camblor P, Mancikova V, Castelblanco E, Perez M, Marrón PI, Mendiola M, Hardisson D, Santisteban P, Riesco-Eizaguirre G, Matías-Guiu X, Carnero A, Robledo M, Delgado-Álvarez E, Menéndez-Torre E, Fraga MF: DNA methylation signatures identify biologically distinct thyroid cancer subtypes. J Clin Endocrinol Metab. 2013, 98: 2811-2821. 10.1210/jc.2012-3566.

    Article  PubMed  Google Scholar 

  4. Lalami Y, Awada A: Recurrent thyroid cancer: a molecular-based therapeutic breakthrough. Curr Opin Oncol. 2011, 23: 235-240. 10.1097/CCO.0b013e3283448206.

    Article  CAS  PubMed  Google Scholar 

  5. Xing M, Haugen BR, Schlumberger M: Progress in molecular-based management of differentiated thyroid cancer. Lancet. 2013, 381: 1058-1069. 10.1016/S0140-6736(13)60109-9.

    Article  PubMed Central  CAS  PubMed  Google Scholar 

  6. Sipos JA, Shah MH: Thyroid cancer: emerging role for targeted therapies. Ther Adv Med Oncol. 2010, 2: 3-16. 10.1177/1758834009352667.

    Article  PubMed Central  CAS  PubMed  Google Scholar 

  7. Antonelli A, Fallahi P, Ferrari SM, Ruffilli I, Santini F, Minuto M, Galleri D, Miccoli P: New targeted therapies for thyroid cancer. Curr Genomics. 2011, 12: 626-631. 10.2174/138920211798120808.

    Article  PubMed Central  CAS  PubMed  Google Scholar 

  8. Nikiforova MN, Nikiforov YE: Molecular genetics of thyroid cancer: implications for diagnosis, treatment and prognosis. Expert Rev Mol Diagn. 2008, 8: 83-95. 10.1586/14737159.8.1.83.

    Article  CAS  PubMed  Google Scholar 

  9. Russo D, Damante G, Puxeddu E, Durante C, Filetti S: Epigenetics of thyroid cancer and novel therapeutic targets. J Mol Endocrinol. 2011, 46: R73-R81. 10.1530/JME-10-0150.

    Article  CAS  PubMed  Google Scholar 

  10. Mulligan LM: RET revisited: expanding the oncogenic portfolio. Nat Rev Cancer. 2014, 14: 173-186. 10.1038/nrc3680.

    Article  CAS  PubMed  Google Scholar 

  11. Hunter L, Cohen KB: Biomedical language processing: what's beyond PubMed?. Mol Cell. 2006, 21: 589-594. 10.1016/j.molcel.2006.02.012.

    Article  PubMed Central  CAS  PubMed  Google Scholar 

  12. Ananiadou S, Kell DB, Tsujii J-I: Text mining and its potential applications in systems biology. Trends in Biotechnology. 2006, 24: 571-579. 10.1016/j.tibtech.2006.10.002.

    Article  CAS  PubMed  Google Scholar 

  13. Ananiadou S, Pyysalo S, Tsujii J, Kell DB: Event extraction for systems biology by text mining the literature. Trends in Biotechnology. 2010, 28: 381-390. 10.1016/j.tibtech.2010.04.005.

    Article  CAS  PubMed  Google Scholar 

  14. Neves ML, Carazo J-M, Pascual-Montano A: Moara: a Java library for extracting and normalizing gene and protein mentions. BMC Bioinformatics. 2010, 11: 157-10.1186/1471-2105-11-157.

    Article  PubMed Central  PubMed  Google Scholar 

  15. Hakenberg J, Gerner M, Haeussler M, Solt I, Plake C, Schroeder M, Gonzalez G, Nenadic G, Bergman CM: The GNAT library for local and remote gene mention normalization. Bioinformatics. 2011, 27: 2769-2771. 10.1093/bioinformatics/btr455.

    Article  PubMed Central  CAS  PubMed  Google Scholar 

  16. Wu C, Schwartz J-M, Nenadic G: PathNER: a tool for systematic identification of biological pathway mentions in the literature. BMC Syst Biol. 2013, 7: S2-

    Article  PubMed Central  PubMed  Google Scholar 

  17. Bjorne J, Ginter F, Pyysalo S, Tsujii J, Salakoski T: Complex event extraction at PubMed scale. Bioinformatics. 2010, 26: i382-i390. 10.1093/bioinformatics/btq180.

    Article  PubMed Central  PubMed  Google Scholar 

  18. Gerner M, Sarafraz F, Bergman CM, Nenadic G: BioContext: an integrated text mining system for large-scale extraction and contextualisation of biomolecular events. Bioinformatics. 2012, 28: 2154-2161. 10.1093/bioinformatics/bts332.

    Article  PubMed Central  CAS  PubMed  Google Scholar 

  19. Rappaport N, Nativ N, Stelzer G, Twik M, Guan-Golan Y, Stein TI, Bahir I, Belinky F, Morrey CP, Safran M, Lancet D: MalaCards: an integrated compendium for diseases and their annotation. Database (Oxford). 2013, 2013: bat018-

    Article  Google Scholar 

  20. Maglott D, Ostell J, Pruitt KD, Tatusova T: Entrez Gene: gene-centered information at NCBI. Nucleic Acids Res. 2011, D52-7. 39 Database

  21. Kanehisa M, Goto S, Sato Y, Furumichi M, Tanabe M, Kanehisa M, Goto S, Sato Y, Furumichi M, Tanabe M: KEGG for integration and interpretation of large-scale molecular data sets. Nucleic Acids Res. 2011, 40: D109-D114.

    Article  PubMed Central  PubMed  Google Scholar 

  22. Mizuno S, Iijima R, Ogishima S, Kikuchi M, Matsuoka Y, Ghosh S, Miyamoto T, Miyashita A, Kuwano R, Tanaka H: AlzPathway: a comprehensive map of signaling pathways of Alzheimer's disease. BMC Syst Biol. 2012, 6: 52-10.1186/1752-0509-6-52.

    Article  PubMed Central  PubMed  Google Scholar 

  23. Chun H-W, Tsuruoka Y, Kim J-D, Shiba R, Nagata N, Hishiki T, Tsujii J: Automatic recognition of topic-classified relations between prostate cancer and genes using MEDLINE abstracts. BMC Bioinformatics. 2006, 7: S4-

    Article  PubMed Central  PubMed  Google Scholar 

  24. Tsuruoka Y, Tsujii J, Ananiadou S: FACTA: a text search engine for finding associated biomedical concepts. Bioinformatics. 2008, 24: 2559-2560. 10.1093/bioinformatics/btn469.

    Article  PubMed Central  CAS  PubMed  Google Scholar 

  25. Frijters R, Heupers B, van Beek P, Bouwhuis M, van Schaik R, de Vlieg J, Polman J, Alkema W: CoPub: a literature-based keyword enrichment tool for microarray data analysis. Nucleic Acids Res. 2008, 1-5. 36 Web Server

  26. Kim J, So S, Lee H, Park JC, Kim J-J, Lee H: DigSee: Disease gene search engine with evidence sentences (version cancer). Nucleic Acids Res. 2013, W510-7. 41 Web Server

  27. Sebastiani F: Machine learning in automated text categorization. ACM Computing Surveys (CSUR). 2002, 34: 1-47. 10.1145/505282.505283.

    Article  Google Scholar 

  28. Gerner M, Nenadic G, Bergman CM: LINNAEUS: a species name identification system for biomedical literature. BMC Bioinformatics. 2010, 11: 85-10.1186/1471-2105-11-85.

    Article  PubMed Central  PubMed  Google Scholar 

  29. Neves M, Chagoyen M, Carazo JM, Pascual-Montano A: CBR-Tagger: a case-based reasoning approach to the gene/protein mention problem. 2008, Association for Computational Linguistics

    Chapter  Google Scholar 

  30. Leaman R, Gonzalez G: BANNER: an executable survey of advances in biomedical named entity recognition. Pac Symp Biocomput. 2008, 652-663.

    Google Scholar 

  31. Kamburov A, Stelzl U, Lehrach H, Herwig R: The ConsensusPathDB interaction database: 2013 update. Nucleic Acids Res. 2013, D793-800. 41 Database

  32. Busaidy NL, Cabanillas ME: Differentiated thyroid cancer: management of patients with radioiodine nonresponsive disease. Journal of Thyroid Research. 2012, 2012: 618985-

    Article  PubMed Central  PubMed  Google Scholar 

  33. Grande E, Díez JJ, Zafon C, Capdevila J: Thyroid cancer: molecular aspects and new therapeutic strategies. Journal of Thyroid Research. 2012, 2012: 847108-

    Article  PubMed Central  PubMed  Google Scholar 

  34. Omur O, Baran Y: An update on molecular biology of thyroid cancers. Crit Rev Oncol Hematol. 2014, 90: 233-252. 10.1016/j.critrevonc.2013.12.007.

    Article  PubMed  Google Scholar 

  35. Burrows N, Babur M, Resch J, Williams KJ, Brabant G: Hypoxia-inducible factor in thyroid carcinoma. Journal of Thyroid Research. 2011, 2011: 762905-

    Article  PubMed Central  PubMed  Google Scholar 

  36. Burrows N, Resch J, Cowen RL, Wasielewski von R, Hoang-Vu C, West CM, Williams KJ, Brabant G: Expression of hypoxia-inducible factor 1 in thyroid carcinomas. Endocrine Related Cancer. 2010, 17: 61-72. 10.1677/ERC-08-0251.

    Article  PubMed Central  CAS  PubMed  Google Scholar 

  37. Yamashita AS, Geraldo MV, Fuziwara CS, Kulcsar MAV, Friguglietti CUM, da Costa RB, Baia GS, Kimura ET: Notch pathway is activated by MAPK signaling and influences papillary thyroid cancer proliferation. Transl Oncol. 2013, 6: 197-205. 10.1593/tlo.12442.

    Article  PubMed Central  PubMed  Google Scholar 

  38. Yu X-M, Jaskula-Sztul R, Ahmed K, Harrison AD, Kunnimalaiyaan M, Chen H: Resveratrol induces differentiation markers expression in anaplastic thyroid carcinoma via activation of Notch1 signaling and suppresses cell growth. Mol Cancer Ther. 2013, 12: 1276-1287. 10.1158/1535-7163.MCT-12-0841.

    Article  PubMed Central  CAS  PubMed  Google Scholar 

  39. Sippel RS, Kunnimalaiyaan M, Chen H: Current management of medullary thyroid cancer. Oncologist. 2008, 13: 539-547. 10.1634/theoncologist.2007-0239.

    Article  PubMed  Google Scholar 

Download references


CW is co-funded by UK Biotechnology and Biological Sciences Research Council (BBSRC), the University of Manchester and the Chinese Scholarship Council (CSC) for his PhD study in the Doctoral Training Centre for Integrative Systems Biology at the University of Manchester. GN acknowledges support from the Serbian Ministry of Education and Science (projects III44006; III47003) and the Health eResearch Centre (HeRC).


Publication of this article was funded by the UK Biotechnology and Biological Sciences Research Council (BBSRC) through the Doctoral Training Centre for Integrative Systems Biology at the University of Manchester.

This article has been published as part of BMC Medical Genomics Volume 7 Supplement 3, 2014: Thirteenth International Conference on Bioinformatics (InCoB2014): Medical Genomics. The full contents of the supplement are available online at

Author information

Authors and Affiliations


Corresponding author

Correspondence to Goran Nenadic.

Additional information

Competing interests

The authors declare that they have no competing interests.

Authors' contributions

CW developed the system for the detection on the subtype-genes/pathways association, and drafted the manuscript. GN provided support and guidance from the text mining perspective and JMS from the systems biology perspective. GB provided expertise on thyroid cancer. GN and JMS conceived and supervised the project. CW, JMS, GB and GN wrote the manuscript. All authors read and approved the final manuscript.

Electronic supplementary material


Additional File 1: TC_TEXT PMID list. The PMIDs in the TC_TEXT corpus are listed in this .txt file, one PMID per line. (TXT 359 KB)


Additional File 2: SC_GOLD PMID list. The .txt file is the gold standard derived from the MeSH Supplementary Concepts for thyroid cancer. It consists of 1,213 PMIDs and their corresponding thyroid cancer SCs. The values in the file are tab separated. (TXT 57 KB)


Additional File 3: Classification results with scores and subtype labels. The .xlsx (Excel) file contains the classification results on the TC_TEXT corpus with subtype relevancy scores and predicated subtype labels. If a document is not related to TC at all, it is given the label "NON"; if a document is related to TC but without a specific focus on one or multiple subtypes, it is given the label "TC"; if a document is assigned multiple subtype labels, the labels are separated by '|'. (XLSX 2 MB)


Additional File 4: Text-mined gene results. The .xlsx (Excel) file contains the text mining results for TC-related genes. The file contains a mapping between TC-related documents and their associated genes, categorized by subtypes. (XLSX 558 KB)


Additional File 5: Text-mined pathway results. The .xlsx (Excel) file contains the text mining results for TC-related pathways. The file contains a mapping between TC-related documents and their associated pathways, categorized by subtypes. (XLSX 117 KB)

Rights and permissions

Open Access  This article is licensed under a Creative Commons Attribution 4.0 International License, which permits use, sharing, adaptation, distribution and reproduction in any medium or format, as long as you give appropriate credit to the original author(s) and the source, provide a link to the Creative Commons licence, and indicate if changes were made.

The images or other third party material in this article are included in the article’s Creative Commons licence, unless indicated otherwise in a credit line to the material. If material is not included in the article’s Creative Commons licence and your intended use is not permitted by statutory regulation or exceeds the permitted use, you will need to obtain permission directly from the copyright holder.

To view a copy of this licence, visit

The Creative Commons Public Domain Dedication waiver ( applies to the data made available in this article, unless otherwise stated in a credit line to the data.

Reprints and permissions

About this article

Check for updates. Verify currency and authenticity via CrossMark

Cite this article

Wu, C., Schwartz, JM., Brabant, G. et al. Molecular profiling of thyroid cancer subtypes using large-scale text mining. BMC Med Genomics 7 (Suppl 3), S3 (2014).

Download citation

  • Published:

  • DOI: