|Home | About | Journals | Submit | Contact Us | Français|
To systematically explore the molecular mechanism for hepatocellular carcinoma (HCC) metastasis and identify regulatory genes with text mining methods.
Genes with highest frequencies and significant pathways related to HCC metastasis were listed. A handful of proteins such as EGFR, MDM2, TP53 and APP, were identified as hub nodes in PPI (protein-protein interaction) network. Compared with unique genes for HBV-HCCs, genes particular to HCV-HCCs were less, but may participate in more extensive signaling processes. VEGFA, PI3KCA, MAPK1, MMP9 and other genes may play important roles in multiple phenotypes of metastasis.
Genes in abstracts of HCC-metastasis literatures were identified. Word frequency analysis, KEGG pathway and PPI network analysis were performed. Then co-occurrence analysis between genes and metastasis-related phenotypes were carried out.
Text mining is effective for revealing potential regulators or pathways, but the purpose of it should be specific, and the combination of various methods will be more useful.
Hepatocellular carcinoma (HCC) is the second cause of death from malignancy tumor and the sixth most prevalent cancer worldwide . Like many other types of malignance tumor, metastasis was considered the primary cause for treatment failure and HCC-associated mortalities [2, 3]. Understanding the molecular mechanisms involved in HCC metastasis would facilitate to develop novel strategies, such as personalized therapies and molecular-targeted drugs, which may help to improve the rate of survival.
A large number of genes and proteins have been examined for HCC metastasis. However, since metastasis is thought to be a multi-step process regulated by sophisticated molecular network, it is necessary to systematically assess the significance of each gene and find out the most important regulators.
Nowadays, text mining (TM) technology is increasingly applied for data analysis. In this study, we employed text mining technology and other bioinformatics methods to perform systematic analysis toward published articles, in order to figure out the critical genes and pathways for HCC metastasis, and to profile the unique regulations for HCCs with different etiologies such as HBV and HCV.
According to text mining and frequency analysis, 1116 genes were identified within 8218 abstracts. The top 20 genes and their frequencies were listed in Table Table1.1. Among these genes, VEGFA, AFP, CDH1, MMP2 and MMP9 were mentioned more than 150 times, while MAPK1, TGFB1, AKT1, CTNNB1 and other genes were also widely studied.
KEGG pathway analysis was carried out with identified genes. The top 20 pathways were listed in Table Table2,2, including focal adhesion, adherens junction, regulation of actin cytoskeleton, cytokine-cytokine receptor interaction and so on.
Critical regulators usually work as hub proteins in regulation network, so the PPI network among these proteins was generated and illustrated in Figure Figure1.1. The degree of nodes (the number of proteins interacting with it) was demonstrated in Table Table3.3. The degree of UBC (ubiquitin C) is much higher than others’, which may partly be attributed to its function in protein degradation. EGFR (degree = 157), MDM2 (degree = 153), TP53 (degree = 152), APP (degree = 149), HSP90AA1 (degree = 149) and other proteins are also predicted as core nodes among the network.
Chronic HBV or HCV infections now have been recognized as the major risk factors for the development HCC , but they take very different strategies for tumorigenesis. HBV DNA is integrated into the host genome and increases the HCC risk through several approaches such as increased levels of HBV proteins, transactivation of transcription factors, disruption of chromosomal stability and so on . Instead, HCV has no integration into host DNA, neither direct oncogenic activity of its genes. Subsequent HCC always develops following liver fibrosis and cirrhosis [4, 6]. So HCCs caused by HBV or HCV may differ in metastasis mechanisms.
To verify this hypothesis, 286 genes were identified in Pubmed retrieves, 78 of which were shared by HBV and HCV papers, 165 particularly for HBV and 43 particularly for HCV. Genes uniquely mentioned with HBV or HCV were demonstrated in Figure Figure2,2, and their KEGG results were listed in Table Table4.4. Besides the overlapping processes, HBV-particular genes were involved in MAPK pathway, tight junction and adherens junction. HCV-particular genes, by contrast, participated in more extensive signaling cascades, including TGF-beta, Jak-STAT, cell cycle, ECM-receptor interaction and gap junction, though the number of them was much less (43 VS 165).
According to classical theory , metastasis can be roughly divided into several steps. Cancer cells have to depart from the original position, invade the extracellular matrix and enter vascular system, where they travel to other sites of body and eventually form new colonization [8, 9]. That is a real tough experience. Reasonably, if a gene can simultaneously affect more than one metastasis related phenotypes, its mutation or de-regulation is more likely to facilitate metastasis.
To identify these genes, four metastasis-related phenotypes were extracted from metastasis cascades. They were “adhesion”, “migration”, “invasion” and “angiogenesis”. The co-occurrence between these words and genes were examined, and the results were demonstrated in Figure Figure3.3. The mostly concerned phenotype was “invasion” (freq = 2786), and then “migration” (freq = 1826). “Adhesion” (freq = 696) and “angiogenesis” (freq = 561) had relatively low co-occurrence with genes. Go along with the four typical phenotypes, PTK2 (with “adhesion”, freq = 64), CDH1 (with “migration”, freq = 49), MMP9 (with “invasion”, freq = 104) and VEGFA (with “angiogenesis”, freq = 98) were the most popular genes, respectively.
To highlight the genes that may affect multiple phenotypes, genes co-occurred with 2 or more phenotypes (threshold of frequency was 7) were listed and cluster analysis was performed in Figure Figure4.4. Several genes, such as VEGFA, PI3KCA, MAPK1 and MMP9, were obviously involved in almost all steps, and they should be preferentially regarded as potential regulators for HCC metastasis.
As an effective method , text mining has been used to explore regulation mechanism for several types of cancers, including glioblastoma , endometrial cancer , prostate cancer [13, 14], and breast cancer . Actually, the relationship between gene and cancer is specific: under specific environment, through specific pathways, and affect specific processes or phenotypes of tumor biology, such as transformation, proliferation, metastasis, recurrence, drug resistance and so on. To get more valuable information, the objective of text mining should be clearly defined. For our study, we focus on “HCC metastasis”, and the relationship between genes and different steps of metastasis were calculated respectively. With specific aim and multiple methods, the association between genes and HCC metastasis could be more precisely addressed.
The etiology of HCC is another crucial issue. As mentioned above, HBV and HCV infection have different mechanisms of tumorigenesis, so they may exert different approaches for HCC metastasis. Exploration for such etiology-specific pathways may shed promise for targeted drugs and personalized treatments.
As a matter of fact, the most popular genes are not always the most important ones. Meanwhile, some other genes though crucial for HCC metastasis, may not get due attention. The purpose of our paper is not just to gather what have been done by now, but also to reveal what should be focused in the future. Genes with moderate frequency, but involved in multi-processes or interacting with principal molecules, might be fertile land for novel discoveries.
Each method used in this research has its own advantages and disadvantages. Frequency calculation is an effective method for text mining, but not enough for functional analysis. PPI network analysis is useful to find out hub regulators, but the specificity may be impaired with too much inputs. Co-occurrence analysis with phenotypes or biological functions is an improvement of word frequency analysis, where the universality of information and the specificity of the results are simultaneously emphasized, but the feature words have to be identified in advance. So these methods should be carefully selected and results should be extensively considered.
There are still some limitations in this study. First, ABNER is a biomedical entity recognition software based on statistical machine learning . Although it has been optimized, not all genes will be identified. Second, it takes a long time to check gene names, symbols and alias. Third, text mining can calculate the frequencies of specific words and calculate their relationships, but they cannot actually “understand” literatures. However, text mining is still helpful for us to observe molecular biology achievements of HCC metastasis on a macro level, and quantitatively assess the roles and relationships for genes with multidimensional perspectives.
We search PubMed with the statement (HCC OR “hepatocellular carcinoma”) AND metastasis, and 8218 literatures were found out (to June 15th, 2016). Similarly, “HBV AND (HCC OR “hepatocellular carcinoma”) AND metastasis” and “HCV AND (HCC OR “hepatocellular carcinoma”) AND metastasis” were applied to collect genes involved in HCC metastasis resulting from HBV or HCV. All abstracts were downloaded from PubMed document retrieval system. Genes and proteins among abstracts were identified with ABNER (Version 1.5) [16, 17]. Gene symbols were normalized manually based on the Entrez Gene Database. Word frequency analysis was performed with Microsoft Excel 2010. To reflect the relationship between genes and metastasis, several phenotypes were selected, such as “metastasis”, “adhesion”, “migration”, “invasion” and “angiogenesis”. Only the genes that co-appeared with these words in the same sentence will be counted. If a gene appeared several times in one sentence, it would be treated once. KEGG pathway analysis was performed on GATHER (Gene Annotation Tool to Help Explain Relationships, http://gather.genome.duke.edu/) , and the threshold of P is 0.05.
The protein-protein interaction (PPI) network was integrated with BisoGenet , and interaction data come from BIND [20, 21], BioGrid , DIP , MINT , IntAct  and HPRD . All interactions were based on experiments. The PPI network was illustrated with Cytoscape (version 3.4.0)  and analyzed with CytoNCA , a plugin for Cytoscape. Co-occurrence analysis between genes and phenotypes was illustrated with Gephi (Version 0.8.2 beta) . Hierarchical cluster analysis based on maximum-linkage (similarity metric with Euclidean distance) was performed with HemI (Version 1.0) .
CONFLICTS OF INTEREST
The authors have no conflicts of interest to declare.