Abstract
Introduction
Many studies have been dedicated to efficient screening and identification of causative genes and pathways implicated in cancerous phenotypes. Much attention has been drawn to gene prioritization that helps to select from a large number of candidate genes a subset of those that are biologically most relevant and are worth further validation experiments.
To understand the pathophysiological mechanisms, it is insufficient to investigate only at the level of single-nucleotide polymorphisms (SNPs) or copy number variations (CNV) as in genome-wide association study (GWAS). This is partly due to the lack of statistical strength that plagues GWAS, 1 which stems from the need to correct for multiple testing, and makes it difficult to implicate causative genes. On the other hand, cancers that involve complex and epistatic genetic traits cannot be adequately explained by additive genetic models. Moreover, cancer phenotypes necessarily involve the full scope of Omics including genomics, transcriptomics, epigenomics, proteomics, and metabolomics, as well as their interaction and correlation with phenotypes. Rather, levels in the Omic hierarchy need to be considered simultaneously and interactively. This is because the genes, RNAs, proteins, and epigenetic factors interplay in the form of signaling networks, metabolic networks, regulatory networks, etc. Moreover, biological networks further interact and exchange information among each other, as shown in Figure 1.
The Omic hierarchy on the left, biological networks on the right, and their interactions. TF stands for transcription factor, The Figure shows some typical network interaction scenarios such as: a signaling network activates transcription factors for a regulatory network; transcription factor complexes that control a regulatory network may be formed through protein interactions (eg, binding); a metabolic network may produce energy (through catabolism) and amino acids (through anabolism) that are necessary for other functional networks; and enzymes that catalyze many metabolic networks are in fact proteins and are produced and regulated by other biological networks. Note that regulatory networks often have participants from multiple levels of the Omic hierarchy.
The importance of the integrated Omic view toward cancer is increasingly realized, and drives collaborative efforts toward development of shared expression, functional annotation, protein–protein interaction (PPI) network, experimental Omic data sources, literature, etc. Most data sources are structured and have predefined schemas for prioritization tools to follow. In fact, gene prioritization tools have been progressively refining their customized components for analyzing structured data sources; see reviews.2–5 In contrast, text mining methods on literature data source and text descriptions in structured data sources have seen slow progress. Thus unlike previous reviews, this review focuses on the text mining methodology used in gene prioritization tools and recent developments on the emerging direction of pathway prioritization.
Heterogeneous Data Sources
With recent advances in the Omic era, large volumes of data are gathered from multiple levels of the Omic hierarchy including genome, transcriptome, epigenome, proteome, and metabolome. These data sources may focus on different subsets of the Omic hierarchy. These data sources also present themselves in various formats such as sequence reads, numerical expression levels, and narrative text in the scientific literature. The increasing availability of high throughput technology such as next-generation sequencing (NGS) not only calls for a changing perspective on the Omic model, but also calls for upgrading hypotheses, algorithms, and software that integrate the evolution of knowledge. In particular, in addition to subsequent analysis on GWAS results to accumulate evidence by statistical procedures, 6 gene prioritization methods also need to integrate an increasing number of heterogeneous data sources, in terms of both content and data formats.
There are a multitude of data sources that encode different types of information regarding the Omic hierarchy and information exchange. These data sources have been conventionally categorized based on their purposes, for example, functional annotation, ontologies, and sequence data. Here, we try to summarize the currently active data sources that have been used by gene prioritization tools as well as by cancer biologists and clinicians in their routine research and practice, grouping them according to their primary utility categories. Table 1 lists the summarized data sources. From the table it can be seen that many data sources contain narrative text or literature, suggesting the broad applicability of text mining to gene prioritization. Thus closer understanding and more improvements on text mining methods in gene prioritization is likely to make a general impact.
Data sources for gene and pathway prioritization according to their primary utility.
Text Mining in Gene Prioritization
This section provides an extensive review of computational tools for gene prioritization. Because we are focusing on text mining in such tools, we categorize them according to the characteristics of their text mining components. In each category, prioritization methods are ordered chronologically.
Prioritization without text mining
Despite the prevalence of literature and narrative text segments in many data sources, earlier gene prioritization methods may leave out the text mining component and simply rely on mining structured data.
POCUS 80 relies on statistics composed of term IDs in functional annotation databases that are shared by gene–disease pairs. PROSPECTR 81 uses alternating decision trees to classify whether the candidate genes and known disease genes share similar sequence patterns. Gentrepid 82 is directly based on biological data including protein domain similarity, protein interaction, and pathway membership. PhenoPred 83 is a supervised method based on experimental PPI networks, protein–disease association, protein sequences, and protein function annotations. PhenoPred encodes gene features using shortest-path distances to all disease genes or genes with known function annotation, as well as sequence and function features using physicochemical or predicted structural properties (real-valued) and GO terms and PROSITE patterns. Principal component analysis (PCA) 84 was used to reduce dimensionality and support vector machine (SVM) 85 was used to predict individual disease–gene correlation.
Prioritization with keyword search text mining
Prioritization methods falling in this category rely on keyword matching when retrieving literature or text segment in response to query terms such as phenotype descriptions or gene names. Indexing is frequently performed to preprocess the narrative text in data sources in order to categorize the literature and reduce the search time.
GeneSeeker 86 depends on gene positional data and expression data, as well as phenotype data extracted using keyword search from MEDLINE and OMIM. GeneSeeker uses synonym lists compiled from Swiss-Prot and the now inactive Human Genome Database 87 to unify nomenclature across different data sources. Prioritizer 88 constructs the gene network using Bayesian integration to account for both manually curated protein interaction and pathway information as well as co-expression data, yeast-two-hybrid (Y2H) interaction data, etc. To initialize the network, known contributing genes are identified by text mining on the OMIM disease entries. CANDID 89 asks the users to input keywords related to the traits of interest and uses them to query the literature and protein domain descriptions (keyword matching). CANDID also uses other data sources to generate source-specific numeric scores, such as those on phylogenetic conservation, gene expression, protein interactions, linkage, SNP association, and user provided gene scores.
All source-specific scores are merged into a final score by user provided weights. PGMapper 90 uses keywords to search phenotype databases such as OMIM and PubMed, and combines linkage, expression, and sequence data sources to search candidate genes for 16 species including human. GeneProspector 91 keeps an in-house literature database by first using text mining to screen PubMed, and then having curators review and manually index screened abstracts. Indexing tasks include associating the publications with gene symbols, pre-defined categories, and study types, in order to calculate a heuristic score for each gene. MaxLink 92 relies on clues based on connectivity of candidate genes to the known disease genes in a PPI network to rank genes and uses keyword searches to provide known gene sets.
Prioritization with vector space model based text mining
Vector space model represents text document and segments as vectors of identifiers, eg, index terms. Compared with keyword search, vector space model can take into consideration counts and weights of index terms. Vector space model based text mining is frequently explored in gene prioritization tools.
G2D93–95 performs text mining on MEDLINE to associate MeSH phenotype terms with MeSH chemistry terms and to associate MeSH chemistry terms with GO terms. For both associations, co-occurrence statistics are calculated. Association between MeSH phenotype terms and GO terms is then calculated transitively from the two associations, based on fuzzy set theory. G2D then searches the RefSeq and STRING databases and assigns gene score by averaging across GO terms matching that gene. SNPs3D 96 uses text profiling on MEDLINE abstracts to rank candidate genes. More specifically, the authors collected nouns and adjectives as keywords, tallied the raw counts of keywords and gene–keyword pairs within the corpus, devised heuristic scores for gene–gene interaction and disease–gene, and applied a rule-based approach to assign the final rank. MimMiner 97 uses the anatomy and the disease sections of MeSH to extract terms from OMIM records and builds a vector space model, weighting the vector element according to the MeSH hierarchy and inverse document frequency. The authors reported correlations between phenotype similarity and gene sequence similarity, correlations between phenotype similarity and protein interaction, correlations between phenotype similarity and gene functions in pathways, and the implication of phenotype grouping on the modular nature of human disease genes, all derived from text mining. Endeavour98,99 trains separate models on different data sources including literature, functional annotation, gene expression, protein domains, protein interactions, pathway membership, cis-regulatory modules, transcriptional motifs, sequence similarity, and user provided data. Endeavour then ranks the target genes against individual models and pools the individual ranks to get an overall rank, using a customized Q-statistic. CAESAR 100 accepts as input a text excerpt containing previously implicated genes, performs text mining (vector-space model) against phenotypic, anatomic, and genetic ontologies and extracts ontology terms. CAESAR then maps such terms to candidate genes in expression, interaction, and pathway data sources and generates data-source specific gene scores that are in turn aggregated into a final score. ToppGene101,102 uses co-citation counts in PubMed as an indication of gene relationship. Using data sources covering gene, protein, and pathway; ToppGene implements two gene prioritization systems: one based on function annotation and the other based on interaction network. CIPHER 103 is a network based prioritization method integrating PPI networks, disease phenotype similarity, and known gene–phenotype associations. CIPHER uses the text mining component of MimMiner 97 to calculate phenotype similarity. GeneDistiller 104 uses literature co-occurrence statistics to filter the candidate genes, besides integrating data sources on genes, proteins, and pathways. PRINCE 105 is a network based algorithm that not only predicts gene associations but also infers protein complex associations with diseases. The construction of the network begins with scanning causal genes of known diseases that share phenotype similarity to the target disease, where the authors used the MeSH based phenotype similarity metric from MimMiner. 97 The network is then initialized with prior knowledge and gene scores are computed using an iterative network propagation algorithm. The web tool PolySearch 106 searches PubMed abstracts as well as gene, protein, pathway, and metabolite data sources to integrate their text content. PolySearch manually curates a corpus consisting of terms and synonyms from dictionaries compiled from gene function databases. PolySearch also adopts a heuristic scoring system to rank disease queries and matching gene terms where scoring is based on a discretized sentence relevancy. GeneWanderer 107 applies random walk and diffusion kernel on PPI networks and compares the PPI domain knowledge collected with and without text mining. The authors pointed out that integrating literature data may lead to better performance on retrospective studies than prospective settings, a phenomenon referred to as “knowledge contamination” as retrospective knowledge is already present in the literature. GPsy 108 profiles candidate genes using data sources on gene sequence, expression, function annotation, and gene–disease association, augmented with orthologous genes extracted from HomoloGene 77 and OMA (Orthologous MAtrix). 78 Text mining is used to extract phenotype annotation based on co-occurrence statistics in the biomedical literature.
Prioritization with text mining using ontology structure
In addition to using vector space model and using ontology only as a terminology mapping tool, gene prioritization methods in this category further explore the ontology structure (eg, the is-a relation between parent and child nodes in an ontology), in order to better quantify the semantic similarity between concepts.
Tiffin et al. 109 use the eVOC anatomical ontology to combine literature text mining and gene expression data. The eVOC terms serve as a bridge to connect the PubMed literature (via association frequency between eVOC term and disease names) and RefSeq genes (via frequency of Ref-Seq genes annotated with eVOC terms). A heuristic ranking score is then used to prioritize the bridged disease–gene pair. SUSPECTS 110 uses four data sources including gene sequence patterns (obtained from PROSPECTR), gene expression, shared rare protein domains, and semantic similarity of the associated GO terms between the candidate and seed genes. The calculation of semantic similarity is done using the metric of information content proposed by Lord et al. 111 , which is based on counting how many times a GO term occurs in the Swiss-Prot database. MedSim 112 uses GO terms as the function annotation of a gene and experiments with multiple configurations on whether to include GO terms of ortholog genes, interacting genes, or semantically similar GO terms. The semantic similarity is calculated with the simRel score, taking into account the differences, commonalities, and specificities of GO terms. 113
Prioritization with statistical text mining
Recently, statistical text mining in gene prioritization has gained traction, where a pre-defined distribution or a Bayesian model is used to ft the data in order to boost the prediction power.
GRAIL
114
is a text mining approach that resembles correlating gene pathways to diseases in that it identifies not only the disease genes but also their biological relationships. GRAIL takes seeds from the implicated SNPs identified by GWAS to anchor seed chromosomal regions, so it is less susceptible to knowledge contamination. A gene is represented as a feature vector by using a term frequency-inverse document frequency (tf–idf) weighted vector space model to analyze the PubMed corpus, from which the relatedness between genes can be calculated and connections between candidate genes and seed regions (containing implicated genes) can be made. By fitting a Poisson distribution over the connections between candidate genes and seed regions, a
There are also efforts toward comparing and integrating computational tools for correlating candidate genes, such as gene prioritization portal 117 , which links to 19 previous computational solutions. In Ref. 118, the authors compared algorithms that use PPI networks including network neighbor analysis, unsupervised clustering, semi-supervised clustering, network flow with prior information, and random walks, and reported superior performance of random walk on network data. Although both works did not have text mining as their focus, they can potentially shed lights on best practice when integrating analysis on structured data with text mining in gene prioritization. They also serve to illustrate the imbalance between the advances in mining structured data and mining text in gene prioritization, which is readily evidenced by the complexity of models employed. We summarize the text mining components in the above gene prioritization methods in Table 2, grouped using the same categorization as we introduced these methods. We also briefly comment the advantages and disadvantages of the text mining components.
Summarization of text mining components in gene prioritization methods.
Challenges of Text Mining in Gene Prioritization
From the progression of gene prioritization methods over time as summarized in the previous section, it becomes clear that the application of text mining has seen slow methodology advance. In our opinion, the evidence from the literature has been largely treated in an oversimplified fashion. This section lays out some of the challenges encountered by text mining in gene prioritization, and the next section enumerates several promising directions emerged lately.
Noisy, contaminated and biased knowledge
Representation of gene–disease relations extracted by current text mining in gene prioritization is often noisy because of the insufficient level of detail in co-occurrence counts or simple statistics. Mined relations are also biased toward genes and phenotypes that are already present in the literature and may offer only limited insights to new discoveries. This effect of knowledge contamination is said to make the prioritization task easier for retrospective discovery than for novel discovery. In fact, this problem is not specific to text mining only, because after publication in literature, knowledge is also quickly integrated into structured data in sources such as KEGG, STRING, and InterPro. Thus cross-validation based methods for checking generalizability may not guarantee good performance on novel genes. Previous works usually tried to reduce such overestimation by sub-setting data prior to the real discovery of the cross-validated genes. However, this does not eliminate another, more subtle form of knowledge contamination, as literature based text mining may still present bias toward existing concepts about the disease. GRAIL addressed this problem by not using known pathways and genes at all, but increased the risk of reinventing the wheel for known discoveries. A finer statistical approach should be devised to utilize existing knowledge and eliminate or offset the potential bias.
Evaluation of text mining components in gene prioritization methods
Ideally, we need objective evaluation for this ever expanding array of text mining components in gene prioritization methods. In typical text mining settings, a ground truth needs to be established as a gold standard to perform intrinsic evaluation and compare different text mining approaches as isolated systems. For example, in a related task of extracting mentions of genes from documents, experts need to first manually annotate the corpora and mark the gene mentions. With such a ground truth produced, evaluation metrics such as precision, recall, and
Problems with guilt-by-association
Network based gene prioritization methods integrate evidence from text mining and similarity profiling, as well as function and interaction annotations. The assumption shared by those methods is the so called Guilt-by-Association (GBA), which states that functionally related genes will give rise to similar mutational phenotypes. GBA has been used to statistically infer previously unknown functions of a gene from prior knowledge about known genes and association data such as PPI and co-expression. However, recent studies have shown that GBA may not in general hold true. 119 , 120 Further, most prioritization tools do not distinguish the types of gene relations, hence they are “black-box” in nature. The black-box nature is partly responsible for the poor performance of GBA. Typically, the construction of functional networks is done by aggregating context independent data such as physical PPI and context specific data such as co-expression 121 without distinguishing them. This is especially true for gene and protein interactions extracted from literature using text mining with simple co-occurrence or co-citation statistics. Separating context specific evidence may lend better insights to improve the robustness of the results. 122 Thus taking finer semantic information into text mining consideration is likely to improve the accuracy of the mined relations. For example, distinguishing “gene A promotes expression of gene B”, “gene A inhibits gene B,” and “gene A inhibits gene B under the co-expression of gene C” will produce more context specific relations and networks than “gene A and gene B are related”. Further, improving explanatory power can also help elucidate molecular mechanisms of major pathways that underlie cancer development.
Heterogeneous data sources
In gene prioritization, multiple heterogeneous data sources often need to be taken into account, such as expression level, protein interaction, domain interaction, sequence reads, and literature data. Although previous works strive for increasingly comprehensive coverage for those data sources, to date none covers all data sources. Moreover, the inter-operative information exchanges between data sources in previous gene prioritization approaches are largely based on customized ontology and keyword term matching. On the other hand, most gene prioritization methods analyze different types of evidence separately and only pool scores produced by individual types of evidence at a later stage. In keeping with the integrative Omic view in Figure 1, data integration should happen at an earlier stage so that reasoning can be on a more integrative level: “enrichment in expression of gene A” (expression evidence) and “protein A binds protein B” (PPI evidence) together implicate a phenotype C. This level of integration is yet to be achieved and there is still no principled way to construct a unified reasoning platform across data sources.
Cross-species insights
There are genetic knowledge bases describing thoroughly studied model organisms, such as gene–phenotype associations in mouse, worm, fly, and yeast in the STRING database. They can be leveraged to draw insights into human pathways that are not as heavily investigated, through orthologous genes, 79 or through cross-species annotation transfer of interolog PPIs. 123 In fact, some gene prioritization methods have exploited ortholog and interolog data sources to cover the understudied human cancer phenotypes or genes, using either manually curated lists (eg, GPsy 108 ) or text mining on the literature (eg, MedSim, 112 Genie 115 ). However, transferring knowledge across species remains a significant challenge as biological context and environmental factors also contribute to the ways genes and proteins interact. Thus, there is no guaranteed mapping from sequence similarity to functional similarity. For this reason, we also expect that taking finer semantic information into text mining consideration to provide context specificity to ortholog and interolog gene relations is likely to help the overall prioritization task.
Emerging and Future Directions
This section presents several recent advances that may provide partial solutions to some challenges raised in the analysis of the previous section.
Converting literature and clinical records to networks
The task of automatically annotating biomedical text with semantic information is an area of active research of medical Natural Language Processing (NLP). There are existing methods that extract named entities such as genes and proteins, as well as their relations (see related work section of 124 for an overview). Previous gene prioritization methods have partially used these methods to identify gene and protein names, but not as much their relations, when constructing gene and protein networks. This is partly due to the fact that most NLP based semantic relation extraction tools identify only binary relations that are too coarse for gene prioritization tasks so that they do not offer much more information compared to simple co-occurrence statistics. Recently, Luo et al. 124 proposed an algorithm that translates text into a network representation, where the nodes of the network may be nominal concepts such as genes and proteins or relational concepts such as a verb specifying an interaction. The edges are syntactic dependency links. We give an example sentence–network translation in Figure 2, which shows the network representation for a sentence from the first paragraph in an example paper, 125 along with two of its sub-networks.
(
Given the text translated networks, useful sub-networks can be harvested under predefined criteria (eg, frequency,
Pathway prioritization
Most gene prioritization methods only focus on using network analysis to rank genes’ relatedness to diseases. A particular gene often participates in multiple pathways, yet maybe only some of those pathways are implicated in a cancer phenotype. Moreover, existing pathways may only have part of themselves involved in a disease process. To this end, associating and prioritizing genetic networks or sub-networks instead of genes can be both more discriminative and more informative, to identify new pathways. Finding cancer correlated pathways is mostly carried out on a case-by-case basis, targeting specific cancer types (eg, 126,127). Moreover, identifying such pathways often starts from implicating candidate genes, and known functional pathways are then filtered (sometimes manually) based on prior knowledge. In reality, cancer pathogeneses range from monogenic and oligogenic disorders to complex and epistatic disorders. Direct identification of carcinogenesis pathways may provide more intuition about the underlying molecular mechanisms and hence be more amenable for experimental biologists to start with. On the other hand, gene prioritization methods on direct identification of pathways involved in carcinogenesis have been studied little. GRAIL can be seen as an initial step toward this goal in that it identifies biological relationships among the genes in addition to ranking them. However, GRAIL does not identify what types of relations hold between genes. Hence the predicted genes plus their relationships do not form a pathway as they lack putative mechanism information. The aforementioned text-to-network translation tool may be used for producing context specific gene associations with detailed relation types in order to “build the bridge to the pathway”.
Conclusions
In this review, we focused on text mining in gene prioritization approaches, which is demonstrably an important component because of the large volume and the prevalence of the biomedical literature and narrative text across many heterogeneous data sources. We reviewed and categorized gene prioritization text mining methods according to progressively advanced models employed, and pointed out that they have seen slower advances compared to the other components in gene prioritization that analyze structured data. We discussed several key challenges to text mining in gene prioritization. We also identified promising future directions including text–network translation to provide finer semantic information and pathway prioritization to offer more mechanism information.
Author Contributions
YL wrote the first draft of the manuscript. GR contributed to writing the section of data sources for gene and pathway prioritization. PSZ contributed to the analysis and categorization of text mining methods in gene and pathway prioritization. All authors reviewed and approved of the final manuscript.