In this paper, we present a large biomedical term resource automatically compiled from the terminology of a selection of biomedical databases. The resource has a very simple and intuitive format and therefore can be easily embedded into a system for biomedical text mining and used as a linguistic resource. It is continuously updated and a user interface makes it possible to compile a new term resource according to individual requirements by selecting specific databases to be included. We present statistics for each included biomedical entity type separately as well as in the context of the combined terminology.
Discovering entities such as genes, chemicals, diseases, species, etc. in the written text of research articles is an important part of biomedical text mining. This task is commonly known as “named entity recognition” (NER). Given a word from a text, it consists in deciding if this word is the name of an entity of interest. However, in the biomedical domain, the main focus of interest are not only the words in the text that refer to specific entities, but also under which database identifiers these entities are registered. The main purpose of using unique database identifiers is to provide unique conceptual referents for entities. Since biomedical entities tend to be highly ambiguous this disambiguation process is crucial. Furthermore, identifiers establish a reference to a database. This can be used either to retrieve additional information about the specific entity or to add related information to the database.
There are three major approaches for tackling the task of NER. A rule-based approach uses hand-crafted rules capturing structures of the word itself as well as its context. A machine learning approach extracts different lexical and contextual features from annotated corpora and applies a sophisticated statistical analysis. The third approach simply uses a dictionary look-up for discovering entity names which are already known and present in a given database.
Whereas the first and the second approach are possibly able to discover novel entity names which have never been seen in the literature before, the third approach is restricted to the dictionary, which means that only previously seen entity names can be discovered. The third approach, on the other hand, has the big advantage that by applying a dictionary look-up, a reference to a database can be established immediately.
In practice usually a combination of approaches is used. Combined approaches typically include a dictionary look-up as second step after using machine learning or rules for identifying entity mentions in the text. This second step provides database identifiers for the entity mentions discovered in the first step. Therefore, terminological resources are an important component of most systems for biomedical text mining.
There are a range of available biomedical databases containing terminological information for one or more entity types. Most of these databases are not designed to meet the needs of biomedical text mining and are typically available in very different formats. Therefore, when building a text mining system, it can be timeconsuming to extract their terminology and bring it into a format which can be easily used for term look-up.
Furthermore, different biomedical entity types have differences in their lexical properties. For instance genes tend to have a high degree of ambiguity. Chemicals tend to have many synonyms as their official name can be very long and complex, depicting different characteristics of the respective molecule. As a result of this, they are typically replaced by a shorter version in practice.
The aim of this paper is, on the one hand, to present a terminological resource in a very simple format which allows for easy integration as a dictionary in text mining systems, and, on the other hand, to analyze its term content regarding the lexical properties of the different terminologies that have been integrated, as well as the entity types in focus.
In the following sections of this paper we first give a very short introduction to the setting and purpose of biomedical text mining in general. Then we describe known properties of the different entity types. In the main part of the paper, we give a listing of a selection of databases containing terminological information about entities. We describe how we process these databases in order to transfer them into a simple and intuitive format. We consider the characteristics of the different entity types involved, namely genes and proteins, chemicals, diseases, species and cell lines. Finally, we present statistics of the terminology taken from each database on its own, as well as in the context of the whole combined term resource. 2
It is essential to keep databases in the domain of life-sciences and biomedicine up to date in order to make knowledge easily accessible to researchers and support them in their daily work flow. New findings in the domain are typically published in the format of scientific articles. In the last decades the task of entering these findings into the database has still mainly relied on human manual work as an expensive and time-intensive process. Nowadays it becomes increasingly impossible for these specially trained curators alone to keep up with the increasing rate of publications in the domain. Automatizing this process, does not only help save money and time, but with the increasing rate of research and published papers it becomes the only way of coping with the huge inflow of information.
Biomedical text mining can be used to partially
automate the process of biomedical literature
curation by using computational power for
discovering biomedical entities together with interaction
and events in which they participate
Compared to NER for other text genres,
biomedical NER is known to be especially challenging.
This is mainly due to the lexical properties of
biomedical entities. The most prominent of these
properties is the variability of entity names.
Despite the existence of terminology
recommendations by nomenclature organizations, authors are
still free to use whatever variant of an entity name
they prefer. Therefore, every author of the
domain tends to use his or her own variant, often
due to orthographic variations
The biomedical entity type that has been most investigated in the past are genes and gene products, i.e. proteins. Usually, genes and proteins are treated as one group in biomedical NER, as their names are frequently used in place of each other. It is characteristic of genes/proteins to have a high level of ambiguity. One reason for this is that genes are typically treated as different concepts, depending on the species for which the gene is described. Even though the concept is understood as different, often the same entity name is used. For example the gene p53 (tumor suppressor gene) is found in a range of species, among them humans, mice and rats, and for all of them the same gene name is used. Another reason why gene names are so ambiguous is the fact that gene names can be very arbitrary, either because the discovering researcher chose an unusual name or because they are named according to their function, position on a gene or relation to other genes1. Furthermore, they can be built up of letters, numbers, punctuation, stopwords or non-alphabetical characters and frequently they are multi-word units (e.g. daughters against decapentaplegic, short stop, cheap date).
Chemicals have gained the reputation to
belong to the most challenging entity names of the
biomedical domain. They are highly
heterogeneous as they can include generic names (e.g.
water, alcohol or cigarette smoke), brand names (e.g.
Aspirin), IUPAC (International Union of Pure and
Applied Chemistry) names (2-(Acetyloxy)benzoic
Acid) to name just a few
From all biomedical named entity types,
dis1http://www.curioustaxonomy.net/gene/fly.html
ease names, as for example HIV, Back Pain or
Breast Cancer, are possibly presenting the least
difficulties to NER. In the past they have been
shown to have less variability than named
entities of genes and chemicals. This is mainly due
to the reason that disease names are highly
standardized throughout the literature
In biomedical NER, names of organisms and
species are usually treated together under the type
“species”. Species names can also have a high
level of ambiguity, often the same species name
is used to refer to several different entities (C.
elegans can be used to refer to up to 41
different species in the NCBI taxonomy
An overview of cell line nomenclature has been
given by Sarntivijai et al. (2008). Cell lines names
share the characteristics of other biomedical entity
names: there is no obligation to use standardized
names and authors are free to use whatever
variants they like. One more reason for ambiguity of
cell lines is the scientific experimental setting: cell
lines mutate or become subject to contamination,
which also bring along a change of concept
With the aim of building a terminological resource for biomedical text mining, we decided to focus on a selection of commonly used databases, each of which contains terminological information about one or more of the most typical biomedical entity types. In this section, we first describe the structure of the resulting combined terminological resource before we give details about the databases from which we take the terminological information, which files we use, how we process them and the general architecture that we apply.
The terminological resource which we compile using terminological data extracted from the selected databases is contained in one single file. This file has the very simple format of comma separated values (csv). The fieldnames, defining the contents of each of the six columns are the following: ‘oid’, ‘resource’, ‘original id’, ‘term’, ‘preferred term’, ‘entity type’ (Table 1).
For each ID in each original database, an internal Base36 identifier ‘oid’ is generated. Base36 is a binary-to-text encoding scheme2, using digits and all letters (in this case capital) of the English alphabet. A five digit sequence may e.g. encode over 60 million decimal values, while taking up only five bytes as opposed to eight. Synonyms are assigned the same oid as the main term. As such, the oid is not a unique identifier. Hence, the primary row key of the output file is a combination of the oid and the (synonymous) term.
The contents of the term field are matched in the text by the dictionary look up. The preferred term is the most standard term for a concept, which is preferred over other term variants.
Finally, the entity type field contains the type of entity, in this case normalized to the following entity types: gene/protein, chemical, disease, species and cell line.
By restricting the database to these fields, we focused on the most important information from the selected databases. The intention is to exclude any redundant information from the term resource and keep it as lightweight as possible by only focusing on information that is absolutely necessary for its application in a biomedical text mining system.
We decided to include terminology from the databases described in the current section, sorted by entity types. This selection is only made for illustrative purpose and to cover a sample of the 2https://en.wikipedia.org/wiki/Base36 most commonly used databases. However, as mentioned before, the format of the terminological resource is flexible and allows for easy integration of terminology from further databases.
NCBI Gene NCBI Gene
edgebase (UniProtKB) of the Universal Protein Resource provides various functional information on proteins. Only the section “UniProtKB/SwissProt” is considered, because it is manually annotated and reviewed, whereas the (much larger) UniProtKB/TreMBL resource is automatically curated and not manually reviewed. UniProt uses a mnemonic identifier and one or more accession numbers. Both identifiers and accession numbers are considered in our work. The identifier has the quality of being a humanreadable mnemonic code, unlike the accession number5. In order to reduce redundancy, only the first accession number is considered, although a separate mapping is maintained for accession numbers that refer to the same gene.
MeSH diseases MeSH (Medical Subject
Headings)6 is a controlled vocabulary maintained by
the United States National Library of Medicine
(NLM) and updated annually. It contains
keywords used to manually annotate PubMed
abstracts and NLM’s book database with the aim of
facilitating search by providing the subjects of a
text. These so called subject headings are
available in the format of descriptors which are
hierarchically sorted. A connected tree number defines
the position in the hierarchy and, at the same time,
3http://www.ncbi.nlm.nih.gov/gene
4ftp://ftp.ncbi.nih.gov/gene/DATA/GENE INFO/
All Data.gene info.gz
5UniProt KB User Manual
6https://www.nlm.nih.gov/mesh/
oid
resource
original id
term
preferred term
entity type
provides information about the entity type
described by a descriptor. The tree branches that we
considered recursively with all their subbranches
are those for chemicals and drugs (branch D),
diseases (branch C) and organisms (branch B).
Apart from MeSH descriptors we also included
MeSH supplementary records. Supplementary
records (SCRs) are updated weekly and contain
additional terms which might occur in the
literature as concept data
drug branch (branch D) of MeSH. Its structure and method applied are the same as described above for diseases.
ChEBI ontology ChEBI (Chemical Entities of Biological Interest)7 is a “freely available dictionary of molecular entities focused on ‘small’ chemical compounds”. ChEBI’s terminology follows IUPAC (International Union of Pure and Applied Chemistry) and NC-IUBMB (Nomenclature Committee of the International Union of Biochemistry and Molecular Biology) but establishes its own unique and stable identifiers. Data is not only manually curated by ChEBI curators but also integrated from different sources, such as IntEnz, KEGG and PDBeChem.
7https://www.ebi.ac.uk/chebi/
NCBI Taxonomy The NCBI Taxonomy is a
“curated classification and nomenclature”8 of
species referenced in other Entrez databases. At
the time of writing this paper, the coverage of
the taxonomy is 10% of all described species of
life
MeSH Organisms The organisms branch (branch C) of MeSH (description above).
Cellosaurus Cellosaurus9 is a thesaurus of cell lines. It is described as a “controlled vocabulary of cell lines” that is used in biomedical research. The resource is freely available under the Creative Commons Attribution-No Derivs License 3.0.
All components of the program were implemented in CPython 2.7.
All resources are downloaded automatically using a standalone downloader. The script queries a text file specifying source URLs and, if present, loads timestamps from previous downloads from a log file. For each FTP or HTTP URL specified, the script attempts to query the modification timestamp from the server. If it deviates from the timestamp recorded in the log file, the file is downloaded and, if compressed, extracted from its archive. Finally, the updated log file is written. 8http://www.ncbi.nlm.nih.gov/taxonomy 9http://web.expasy.org/cellosaurus/
For resources that are provided in a form not ready for automatic processing, a preprocessing step is triggered. In the Entrez Gene gene info file, the header row is reformatted to correspond to the tab-separated format and only columns relevant to term resource are kept. In the NCBI taxonomy names listing, only names for records with the taxonomic rank species are kept.
For each resource, a parser was implemented, tailored to extract the identifiers and terms from each original database. For each term in the database, the parser produces a row hash with name-value pairs for each field. If a record specifies synonyms or multiple IDs, additional hashes are created for each synonymous term or additional ID. The main term is specified as the preferred term for each synonym.
In order to keep the memory footprint to a minimum, most parsers process resources in an iterative way, row by row. An exception is the MeSH resource, which depends on two separate, interlinked files, hence, the entire resource is loaded into memory.
Finally, a resource builder iterates over all specified resource parser objects and writes each row hash to a row in the output TSV.
Uniprot and Cellosaurus parser Based on a parser described on Mannheimia goes programming10 Uniprot sequence entries and Cellosaurus cell lines are specified as lists of space-separated key value pairs, separated by delimiter lines. Keys may be mandatory or optional and can occur once or multiple times. Multiple occurrences of the species key are concatenated, as these constitute continuations of previous values, not additional values (see Uniprot knowledgebase). Relevant information (terms and accession numbers) are mapped to the row hashes mentioned above. Some keys may have multiple values, each separated by a semicolon. For the Uniprot resource, a copy of the row hash is created for each accession number indicated in a sequence entry.
vided in a pipe-separated TSV-like format. For each id, a list of terms is provided, each paired 10http://mannheimiagoesprogramming.blogspot.ch /2012/04/uniprot-keylist-file-parser-in-python.html with a name class, specifying the type of term. The entry with the name class “scientific term” is used as preferred name. All name classes are considered with the exception of “authority”, as these terms specify not only terms but authorship and publication dates. For obscure reasons, entries of the name class “synonym” also occasionally cite authorship. Regular expressions are used to detect and remove citations and additional information in parenthesis, as well as to strip double quotes from these entries. A small subset of entries specify a unique name. For these cases, additional row hashes are generated.
Entrez Gene gene info parser The gene info format is a standard TSV file. A preprocessor converts the non-standard headers and generates a reduced file containing only relevant columns (containing the gene ID, symbol and all synonyms). Synonyms are specified as pipe-separated sequences. For each synonym, an additional row hash is generated.
MeSH XML parser MeSH is provided in two separate files, a descriptor record set and an associated supplement record set. First, both files are parsed into memory. Only the tree structures Organisms [B], Diseases [C] and Chemicals and Drugs [D] are considered. Additionally, a look-up table is generated from the descriptor record set, mapping the ID of a descriptor record to IDs of all trees, in which it occurs. Each supplement is mapped to its descriptor record using the look-up table. Finally, for each descriptor and supplement record, a row hash is generated.
CHEBI OBO parser The CHEBI OBO parser wraps the OBO parser from the Orange Bioinformatics add-on for the open-source data mining utility Orange11. A row hash is generated for each term and, if present, for each synonym. Placeholder synonyms (containing only periods as terms) are discarded.
The combined resource is generated and accessed through a web interface12. The interface is controlled by a Python Common Gateway Interface (CGI) script, which allows direct control of the creation pipeline. Visitors to the website 11http://orange.biolab.si/ 12pub.cl.uzh.ch/purl/biodb can select the desired resources through checkboxes. Additionally, the user may provide pattern/replacement pairs for changing how the resources and their entity types are labeled. The reason for this lies in the extensive labeling of certain resources. This is particularly true for MeSH, where labels like mesh desc(Anatomy), mesh desc(Chemicals and Drugs) etc. give a detailed description of origin. By replacing the regular expression mesh desc.* with MeSH, for example, the number of resource labels can be reduced.
After submitting the request, the resource creation process is started in the background. As the creation process may take several minutes, the user is provided an individual link that allows downloading the resource file as soon as it is ready. 5
Graphs depicting frequency distributions for the terms in the term resource, for each entity type as well as for the whole term resource, can be found in the appendix13. Each graph shows the ambiguity of a term (how many IDs per each term), and the reverse property, i.e. how many terms are available for each ID. For example, while cell lines and organisms are mostly unambiguous (the vast majority of them shows a 1:1 correspondence between IDs and terms), chemicals show a much higher degree of ambiguity. Since common matching strategies in dictionary look-up are to lowercase the terms, or to strip them of nonalphanumeric characters with the aim of increasing recall, we also show the additional ambiguity generated by this process. 6
Recently, the focus of research on biomedical named entity recognition has rather been on machine learning approaches than on dictionary based approaches, however, assigning unique database identifiers is often a necessity and is frequently used as a second step after the application of a machine learning system.
The task of retrieving database IDs for named
entities is promoted by shared tasks, such as the
BioCreative workshop which includes
normalization sub-tasks by asking participants to provide
13The graphs were generated using plotly (https://plot.ly/)
database identifiers for the named entities in the
text. BioCreative II included a gene
normalization task encouraging the development of systems
which are able to assign an Entrez Gene identifier
to genes found in PubMed articles
Thom
Kaljurand et al. (2009) have also compiled a term resource from different database terminologies but focuses mainly on the identification of protein mentions. Furthermore, the current work considers a richer set of terminology, and presents more detailed statistics.
One system, that has successfully used a dictionary look up for gene and protein NER has been described by Hanisch et al. (2005). The lexical properties of the gene ontology (which we have not used in our work so far) have been explored by McCray et al. (2002).
Compared to these related resources, the combined term resource presented in this paper has several advantages. It is continuously updated with the newest version of all included databases, which keeps it up to date. A user can produce a customized file with selected databases, according to specific needs of a project. Apart from this, it can be easily extended with further resources, which we are planning to do in the future. Last but not least the format is very simple and therefore allows for easy integration into any kind of text mining system. Furthermore, also due to its simplicity, the format is more lightweight, taking up less memory than other formats commonly used for terminological databases. 7
In this paper we presented a new terminological resource for biomedical text mining. The resource is compiled of terminology extracted from a selection of biomedical databases. Its very simple format facilitates integration into any system for biomedical text mining in order to perform a dictionary look-up. The novelty of our approach is that the resource is continuously updated with new terms found in the respective databases. Furthermore, a user interface provides a user with a choice of different databases and entity types so that an individual resource can be compiled. Various statistics for the compiled termfile give insight to the lexical properties of the terms contained in the resource. This sheds light on differences between the lexical properties of different entity types. Additional terminological resources will be included in future versions of the system. Proceedings of the conference Terminology and Artificial Intelligence 2015 (Granada, Spain) 49