Text-mining in cancer research may help identify effective treatments

Global cancer incidence has rapidly increased over the past few decades. In 2018, around 18.1 million people were diagnosed with cancer worldwide (1). With advanced cancer research and healthcare systems, five-year survival trends are generally increasing. Still, cancer is the second leading causes of death and counts for 9.6 million deaths globally (2). Systematical curations of cancer-related studies are beneficial to researchers as they seek to discover novel treatments for curing cancer.

In the biological aspect, cancer is known to be a type of genetic disease, that is, specific changes to genes may increase one’s predisposition to cancer (3). Intensive clinical research has been devoted to identifying biomarkers, such as genes and proteins, that serve as the diagnostic, predictive, prognostic, or even therapeutic molecules for precision medicine (4-6). Both web-lab experiments and bioinformatical approaches have been developed to investigate potential biomarkers for specific cancer types (7), and the rapid accumulation of such scientific literature provides an enormous number of resources to clinical researchers. More than 950,000 publications dated from 2009 to 2019 are available in PubMed by searching abstracts (8). However, it is time-consuming to evaluate such a mountain of biomedical texts manually and challenging to unveil the underlying relationships among these studies or what knowledge is new.

Text-mining technology has emerged as a powerful approach to extract information and knowledge from a wealth of scientific literature. Typically, four main stages are involved in this technology: (I) information retrieval (IR), a process in which biomedical domains such as PubMed are searched for subjects of interest; (II) named entity recognition (NER), which evaluates the occurrence of keywords from collected texts; (III) information extraction (IE), which detects links among the recognized keywords in the text; and (IV) knowledge discovery (KD), which identifies concepts that are new, based on known facts derived from detected relationships (9). Finally, visualization of the extracted information, such as literature networks and word clouds, is also an important step for interpreting the results correctly and efficiently, guiding researchers to generate hypotheses and start follow-up studies (10).

To accurately decode the information from the text based on human Natural Language Processing (NLP), four different levels of language structure must be known: (I) words, which are the units of language; (II) syntax, which presents how to group words into phrases and ultimately sentences; (III) semantics, which captures the logical meaning of sentences; and (IV) pragmatics, which seeks to understand the relationships of sentences in larger context to determine meaning (11). Aside from that, the understanding of biological terminology is also essential to avoid ambiguity and redundancy, enabling effective IE. Many biomedical corpora, such as MeSH (Medical Subject Headings) (12), and databases, such as UniProt, RefSeq (NCBI Reference Sequences) and KEGG (the Kyoto Encyclopedia of Genes and Genomes) (13-15), are established and regularly updated to accelerate the recognition process of biomedical terms and their variations.

IE is an essential text-mining step, for the current data structure is becoming multifaceted, and the extracted results can directly reflect the topic of the text. Hence, many algorithms have been proposed to make this process more efficient and the output more accurate for further analysis. TF-IDF (term frequency-inverse document frequency) is a simple and efficient approach for calculating a value that represents the word’s frequency in both the short text and a given collection of documents in order to determine the importance of a keyword in a short text (16). The algorithm TextRank is a graph-based model, built on an iterated random walk strategy, to rank the importance of words in the text. Additionally, it provides unsupervised text summarization from a set of textual resources (17). For broader texts, topic models, including latent semantic analysis (LSA), latent semantic indexing (LSI) and latent Dirichlet allocation (LDA), are widely used (18-20). Generally, these algorithms rely on probabilistic generative models within the framework of Bayesian statistical inference. These developments greatly accelerate the process of the automatic organization and classification of digitized information.

Many limitations exist in the current algorithms, including in the three described above. First, the weighted relations from multiple text collections remains difficult to correctly interpret; this error may result from incorrect linking of texts due to different biological entities sharing the same name or abbreviation (11). Second, topic models like LDA suffer from some conceptual flaws in practical use; for example, lack of robustness for the number of input texts and free setting of parameters, which causes discrepancies in context. Furthermore, these Bayesian-based approaches ignore the justification for the use of prior knowledge (21). Lastly, cancer involves a complex molecular mechanism; that is, different genotypes or gene expression profiles in the same pathway or network may be triggered even within the same cancer phenotype (22). Analyzing different levels of studies such as molecules, motifs and pathways at the same time to discover new knowledge with a hierarchal view may still not feasible.

In the last section, we use real studies to illustrate how text-mining strategy plays an important role in cancer-related biomarker identification and cancer drug discovery or repositioning (23,24). Previous studies have applied the text-mining approach to systematically discover the relationship between mutations and cancer (25). In addition to discovering a mutation-cancer association, a literature-based mining website extracts sentences from the different sections of the paper to identify whether the gene is a novel driver, oncogene or tumor suppressor (26). This tool can also identify well-supported genes with a dual role as oncogenes or tumor suppressors in different types of cancer. For example, ATF3 is found as an oncogene in breast cancer, but it is a tumor suppressor in prostate cancer. For individual cancer phenotypes, text-mining has been applied to extract the relationship between breast cancer and candidate genes, and candidate association words were found to point to the relationship between breast cancer and related genes using pattern clustering (27). In another example, clinical concepts and genes associated with colorectal cancer were explored through literature and statistical analysis of clinical information and candidates (28). That study confirmed 51 genes from the mined results involving the colorectal cancer pathway. For example, among them, KRAS and CTNNB1 are important oncogenes in colorectal cancer.

Aside from identifying cancer biomarkers, text-mining technology also accelerates the process of cancer drug discovery and repurposing. DrugQuest is a website that applies text-mining technologies and the TextQuest algorithm to mine a publicly accessible database to identify biologically significant terms and group these words based on the textual content (29). Another cancer-related study demonstrated using text-mining and pathway analysis tools is a prospective method for exploring potential drugs targeting the genes or pathways related to cutaneous squamous cell carcinoma (cSCC) (30). In this study, 121 genes relevant to cSCC were identified using a text-mining approach, and 55 drugs were revealed that could target 10 pathways through gene enrichment analysis. Among these identified drugs, 49 have not been tested for this cancer type, suggesting these may be novel, targeted therapies for cSCC treatment. For the drug repositioning study, the importance of all words in cancer-related abstracts in PubMed were represented using word embedding technology. Further, the drug-disease associations were identified, and repurposed drugs were found by classifying candidate genes (31).

In conclusion, advanced text-mining technology may shed light on not only the identification of predictive biomarkers for multiple cancers but also protentional drugs, including novel drug discovery and existing drug repositioning, for effective treatments in the future.

Ethical Statement: The authors are accountable for all aspects of the work in ensuring that questions related to the accuracy or integrity of any part of the work are appropriately investigated and resolved.