Keywords
amino acids, DNA, deletions
amino acids, DNA, deletions
As the cost of genomic sequencing continues to fall, the amount of data being collected and studied for the purpose of understanding the genetic basis of disease is increasing dramatically. There are large-scale efforts to catalog the results of this research in structured databases, including in the Online Mendelian Inheritance in Man (OMIM) database1 and the Human Gene Mutation Database (HGMD)2. Much of the source information relevant to such efforts is available only from unstructured sources such as the scientific literature, and significant resources are expended in manually curating and structuring the information in the literature. As such, there have been a number of systems developed to target automatic extraction of mutations and other genetic variation from the literature using text mining tools3–9, inter alia. Such tools have been shown to perform well, benefiting from a well-defined target vocabulary (nucleic and amino acids), the availability of reference sequences for position validation, and increasing adoption of standard nomenclature such as the Human Genome Variation Society (HGVS) format10. The natural language descriptors of genetic variation are fairly consistent, and lend themselves well to automated processing.
In previous work11,12, we assessed one of these tools, the Extractor of Mutations (EMU) tool6, for its ability to identify the genetic variant information that had been manually curated in the COSMIC13 and the International Society for Gastro-intestinal Hereditary Tumours (InSiGHT)14 databases from targeted literature sources. That work found very low recall for the text mining tool when considering the narrative content of publications alone, and identified processing of the supplementary material associated with publications as a critical component of an approach to automated genetic variant curation from the literature.
In this work, we perform a broad survey of the existing publicly available tools for extraction of genetic variants from the scientific literature. We consider not just one tool but a number of different tools, individually and in combination, and apply the tools in two scenarios. First, they are compared in an intrinsic evaluation context, where the tools are tested for their ability to identify specific mentions of genetic variants in a corpus of manually annotated papers, the Variome corpus15. Unlike previous test corpora, this corpus was not designed exclusively for the purpose of testing mutation extraction tools and hence is a better test of real-world applicability than prior corpora. Second, they are compared in an extrinsic evaluation context based on our previous study with COSMIC and InSiGHT. Our results demonstrate that several of the tools have complementary coverage and can be used together effectively. This study suggests several directions for the improvement of text mining tools for genetic variant extraction from the literature.
Text mining of mutations in the scientific literature has been addressed by several tools, including MutationMiner3, MarkerInfoFinder16, EMU (Extractor of Mutations)6, MutationFinder4, tmVar9, and SETH17. A summary of previous work can be found in Naderi and Witte (2012)7. These tools have been shown to achieve a performance over 0.90 in F1 measure, and in some cases almost perfect Precision/Recall, on intrinsic evaluations. There are also several corpora that are publicly available to support intrinsic evaluation of mutation extraction tools4,6,16,18–20. On the other hand, cross-comparison of these tools has been limited. The most commonly used is the corpus provided with the Mutation Finder tool, which covers protein variants. Some of these tools have also been used to reproduce the information curated in existing databases about genetic variants, allowing for extrinsic evaluation of the mutation extraction tools.
In the following sections, we present the tools for genetic variant extraction that we have considered in our study. Each is a publicly available tool. The tools are introduced and their published results in intrinsic evaluation are presented. Results are presented in terms of precision (P), recall (R) and F1 measure (F).
MutationFinder (MF) was one of the earliest tools developed for extraction of mutations. This tool performs point mutation extraction based on a set of regular expressions4. The coverage of this tool is thus limited compared to the other ones. A corpus, the MutationFinder corpus, was established to guide the construction of the patterns. The development data set is made up of 605 point mutation mentions in 305 abstracts selected randomly from primary citations in PDB. The evaluation data set is made up of 910 point mutation mentions in 508 abstracts annotated by two of the authors, not involved in the development of the system. Mean pairwise interannotator agreement, calculated on the fifty overlapping abstracts, was 94%4. Performance of MF in the extracted mutations is P: 0.984 R: 0.817 F: 0.893.
Open Mutation Miner (OMM) is a tool7 that extracts protein mutation mentions and maps them to their properties in the OMM impact ontology, which covers protein mutation types (insertion, deletion, point mutation), protein types and the impact of the mutation. The extraction of mutations component couples grammar rules with a normalization step, e.g. transformation into single-letter format. The system is combined with MutationFinder to identify variants in natural language. OMM has been developed under the GATE platform and it is available as a JAPE-based mutation tagging component. Mutation extraction was evaluated on a set of 11 full text articles, with an average performance of P:0.99, R:0.96 and F:0.977.
The Extractor of Mutations (EMU) tool6 was designed to capture a broader range of mutations than other tools available when it was developed and hence is a better fit for the variants we might expect to find. It identifies protein and DNA point mutations, Single Nucleotide Polymorphism Database (dbSNP) identifiers21 (RSIDs), and DNA insertions and deletions. In addition, it links the mutations to the proteins and genes that appear in text and performs sequence verifications using existing sequence databases to increase the precision of the annotations. EMU has been shown to have a performance of 0.92 F1 measure on an intrinsic evaluation6, i.e., it has high recall and high accuracy.
tmVar9 (http://www.ncbi.nlm.nih.gov/CBBresearch/Lu/pub/tmVar) is a recently released mutation extraction tool based on a conditional random field model used with a special set of features, which has shown to perform better than MutationFinder on the MutationFinder corpus. tmVar has been trained on 334 newly annotated citations, different from the MutationFinder corpus, and evaluated using the MutationFinder test set (P: 0.9880, R: 0.8962, F: 0.9398) and using their own data set made up of 134 manually annotated test citations (P: 0.9138, R: 0.9140, F: 0.9139)9.
SNP Extraction Tool for Human Variations (SETH) (http://rockt.github.io/SETH) implements an Extended Backus–Naur Form (EBNF) grammar proposed by Laros et al.22 to identify mentions of mutation that obey the HGVS nomenclature. Since mentions in text might not follow the HGVS nomenclature, SETH integrates MutationFinder to extend its coverage. SETH returns whether a mutation is a DNA or protein variant and the type of variant (e.g. deletion).
SETH has been evaluated on several corpora, as reported on the SETH web site. Among other corpora, SETH evaluation has been performed on the MutationFinder test set (Precision 0.97 Recall 0.83 F: 0.89), the tmVar corpus (P: 0.94, R: 0.81, F: 0.87), the Thomas et al. corpus23 (P: 0.95, R: 0.58, F: 0.72) and the Osiris corpus24 (P: 0.98, R: 0.85, F: 0.91).
Intrinsic evaluation: annotated corpora. There are several text corpora that have been made available for the evaluation of mutation extraction tools. There are corpora that focus on protein mutations alone, on protein and DNA mutations or on normalizing the mentions to dbSNP identifiers.
Several corpora are available for the evaluation of protein mutation extraction tools. As presented above, the developers of MutationFinder4 made available a data set (http://mutationfinder.sourceforge.net) of 305 abstracts annotated with point mutations that was used for system development and 508 abstracts available for evaluation. The developers of OMM7 (http://www.semanticsoftware.info/open-mutation-miner) performed experiments on 11 full text articles annotated manually with protein mutations, although these documents are not publicly available for distribution, the manual annotations are available with the tool download. Other corpora exist annotated with protein residue information either manually annotated18,19 or prepared using automatic methods20.
There are corpora available that contain both protein and DNA mutations. The EMU corpus6 (http://bioinf.umbc.edu/EMU/ftp) was developed for annotation of mutations related to prostate cancer. This data set was developed by querying MEDLINE for the medical subject heading (MeSH) Mutation and selecting citations relevant for prostate cancer based on MetaMap annotation. It contains 500 manually annotated abstracts with 95 mutations in 55 abstracts. The tmVar9 (http://www.ncbi.nlm.nih.gov/CBBresearch/Lu/pub/tmVar) system also comes with its own annotated set. The set comprises 500 abstracts manually annotated from which 334 were used for training tmVar while the remaining 166 were used for testing it. These citations were recovered from PubMed selecting English abtracts containing novel human mutations, targeting formulaic mentions. The Variome corpus15 (http://www.opennicta.com/home/health/variome) is an annotated biomedical textual resource pertaining to human genetic variation and its relation to diseases and other entity types. At present, the corpus comprises ten double annotated full text journal publications on inherited colorectal cancer, which were selected on the basis of their relevance to the genetics of the Lynch Syndrome to support the curation of the InSiGHT database.
There are two corpora to evaluate the normalization of extracted mutations to dbSNP identifiers. OSIRIS24 (https://sites.google.com/site/laurafurlongweb/databases-and-tools/corpora) was collected from MEDLINE covering English abstracts for human mutation, limited to 2004 and 2005 and focused on specific project criteria. The annotation is performed focused on evaluation of entity recognition and of disambiguation of variation entities to dbSNP identifiers. The final version of the corpus contains 105 abstracts available with 109 normalized variants and 155 unnormalized ones. The second corpus, developed by Thomas et al.23 (http://www.scai.fraunhofer.de/snp-normalization-corpus.html), consists of 296 MEDLINE abstracts annotated with 527 mutationdbSNP id pairs. The corpus was annotated initially with MutationFinder and then manually annotated for completion and normalized to dbSNP. Mutations without a valid dbSNP identifer were removed. Only the annotations are available, while the abstracts can be downloaded from PubMed using NCBI’s E-utilities (http://www.ncbi.nlm.nih.gov/books/NBK25500).
Table 1 summarizes the results of the tools presented in the previous section as reported on available corpora. We can see that not many tools have been evaluated with the same corpus, other than the MutationFinder test set. The different tools evaluated on the MutationFinder corpus show that the performace in precision is generally very high across the tools with some differences in recall. However, the MutationFinder corpus covers only a limited set of protein mutations. In this work, we perform an intrinsic evaluation of the tools using the Variome corpus as the common reference set, producing a broader comparison of the different mutation extraction tools. We will introduce this corpus in detail in the Methods section.
Extrinsic evaluation: curated databases. In addition to intrinsic evaluation, there have been several efforts in trying to reproduce the information curated in mutation databases through information extraction from the literature. A broader summary of extrinsic evaluation is available in11.
There have been several efforts with varying success in recovering the curated information from variants databases. Krallinger et al.5 extracteded mutations from literature for the kinase domain from abstracts and full text showing different levels of coverage of KinMutBase25, the Swissprot Variant database26, SAAPdb27 and the COSMIC database28, for which only 6% of the mutations were recovered. Schenck et al.29 worked on a small set of articles curated in COSMIC that they annotated, recovering up to 30% of their annotated mutations. Caporaso et al.30, Nagel et al.18 and Verspoor et al.31 tried to recover information about protein mutations and residues annotated in PDB protein records with limited coverage when using abstracts and with larger, but still very limited, coverage when using full text. On the pharmacogenomics side, Rance et al.32 and Hakenberg et al.8 tried recovering variants and related drugs to reproduce the data in PharmGKB with different coverage depending on the target genes.
In a previous study11, we evaluated the ability of a mutation extraction tool to recover the curated mutations in the COSMIC and InSiGHT databases using the articles that were curated in these databases. We found, as in previous studies, that the recall considering the mutations extracted from the abstracts was very low. When considering the full text of these articles the recall increased but was still low. We found that many of the missing mutations were extracted by EMU from tables and supplementary material. In our current work, we have expanded this study by performing an intrinsic evaluation of several mutation extraction tools using the Variome corpus and performing a coverage evaluation of the tools when recovering the curated mutations from COSMIC and InSiGHT.
We have performed the evaluation and comparison of mutation extraction tools intrinsically, using the Variome corpus15, developed in collaboration with the Human Variome Project (http://www.humanvariomeproject.org), as a reference set. For the extrinsic study, we required a curated database that includes mutations and specific links to the literature (with PubMed identifiers [PMIDs] included for each mutation). We selected the COSMIC and InSiGHT databases for our investigation. These databases are used as reference sets; the information extracted from the corresponding scientific literature is compared directly to the information curated from those articles in the databases. We normalize extracted mutation mentions to Human Genome Variation Society (HGVS) format10.
The Variome corpus15 (http://www.opennicta.com/home/health/variome) is an annotated resource of biomedical texts pertaining to human genetic variation and its relation to diseases and other related entity types. At present, the corpus comprises ten double annotated full text journal publications on inherited colorectal cancer, which were selected on the basis of their relevance to the genetics of the Lynch Syndrome to support the curation of the InSiGHT database. The annotation schema covers thirteen relations, such as gene-has-mutation, mutation-has-size and disease-related-to-bodypart; and eleven entity types, such as genomic categories (e.g., gene, mutation), phenotypic categories (e.g., disease, body-part), categories related to the occurrence of mutations in a disease (e.g., age, ethnicity), and a characteristic category for the eventual addition of relevant information. Compared to other variation corpora, the Variome corpus not only annotates mutation mentions but also other entity types, providing a larger set of relevant entities. In addition, it contains annotations for relations between the entities, which provide a more exhaustive context for the training and evaluation of text mining tools supporting the curation of genetic variant databases.
The mutation entity type captures mentions of mutations which specify changes in the protein or DNA sequence as well as mutation terms which refer to general properties of a mutation (e.g. somatic mutation) or terms specifying a mutated gene (e.g. APC+). Current mutation extraction tools are only concerned with the first type, thus extracting mentions of protein or DNA changes. We have manually catalogued the annotated mutations and identified 118 mutation instances that are annotated in the corpus. From this set, 52 are DNA mutations and 66 are protein mutations.
In this work, we expand our previous study11 and retain the original data sets for ease of comparison.
COSMIC13 (http://www.sanger.ac.uk/cosmic) contains comprehensive, curated, information on somatic mutations in human cancer. We used version v62 (from 29th of November 2012) available from COSMIC’s FTP site (ftp://ftp.sanger.ac.uk/pub/CGP/cosmic/), including mutation information curated from 9,950 unique PubMed articles, as well as Cancer Genome Project (CGP) (http://www.sanger.ac.uk/genetics/CGP) studies and international system screens (e.g. International Cancer Genome Consortium (ICGC) (http://dcc.icgc.org/web)). We identified 7,898 publications associated to mutation information in this resource. cDNA and protein mutation information is already available in HGVS format. Genes are referenced by name and by HGNC (HUGO Gene Nomenclature Committee) identifier.
InSiGHT14 maintains a database of genetic variants for both Lynch Syndrome and Familial Adenomatous Polyposis. The current database only has curated mutations for four genes: MLH1, MSH2, MSH6 and PMS2. The original database was established in the 1990s with mutations reported by individual laboratories. Reports manually extracted from published literature currently comprise the majority of entries in the InSiGHT database (approximately 75%, according to the database curator), with the balance direct submissions from clinics.
We accessed the InSiGHT database on 02 January 2013 to establish our data set. The data includes variants with curated associations linked to 809 PubMed citations. The database contains information about the variants in the fields Variant/DNA and Variant/Protein. The amino acids in protein variants have been normalized to single letter amino acid abbreviation form.
There are 41 articles that have been curated both in COSMIC and InSiGHT databases. Unfortunately, none of the mutations in the overlapping articles has been curated by both databases because COSMIC is focused on somatic mutations, while InSiGHT is focused on germline mutations in only four genes.
An abstract for each PMID was retrieved from MEDLINE using NCBI’s E-utilities. Abstracts were downloaded in XML format and XML escaped characters were converted to their text characters (e.g., A–>T becomes A–>T). In the case of the COSMIC database, 17 articles did not seem to be available when querying PubMed.
A small portion of PubMed is available as full text articles through the Open Access collection in PubMed Central (PMC-OA). From the 9,950 PMIDs available from the COSMIC set only 563 were available from PMC-OA. From the 809 citations for InSiGHT only 12 were available through the full text PMC-OA. This represents less than 10% of the overall set referenced by both databases.
In addition to narrative text, we have used the mutation extration tools with further content linked to the papers, which includes the tables and supplementary material and is representative of the broader full text literature33. We collected articles from COSMIC and InSiGHT that are available in the open access part of PubMed Central (PMC), since it already contains the tables in the XML of the article and there are pointers to the supplementary material. For the set of 13 articles in the InSiGHT database that could be found in PMC, InSiGHT contains 252 mutation triples. COSMIC associates 33,814 mutation triples to the 563 articles in PMC.
We extracted the tables and table captions from the full text PMC articles. The COSMIC database references 394 PMC articles with tables; 197 of these were identified as having mutations in the tables. From the InSiGHT database there are only 8 articles with tables, of which 4 contain mutations. In these articles, no mutations were found in the abstract or full text.
Supplementary material was also identified from links within the PMC articles and downloaded. The InSiGHT set contains a limited number of supplementary material files (in 1/12 articles), while COSMIC has a larger number linked to the papers (in 138/563 articles). In contrast to PMC articles, available in XML following a consistent DTD (Document Type Definition), supplementary material appears in a variety of file formats. The most frequent types of supplementary material in this corpus, shown in Table 2, are, in order of frequency: MS Word documents, MS Excel spreadsheet, PDF documents, TIFF images and MS Powerpoint documents. Text from the supplementary material was extracted with Apache Tika 1.3 (http://tika.apache.org/1.3). No image processing was performed.
Set | COSMIC | InSiGHT | |
---|---|---|---|
Files | PMIDs | ||
MS Word MS Excel PDF documents MS Powerpoint CSV files Images | 176 111 82 34 1 101 | 87 57 70 17 1 36 | 1 0 0 0 0 0 |
Total | 505 | 138 | 1 |
During the extraction of tables and supplementary material, we realized that some PMC articles do not contain the full text in XML format but a link to a PDF version of it. From the InSiGHT collection, 4 papers out of the 13 contained only the abstract with a link to the full text in PDF format. In the COSMIC collection, the proportion was 76 papers out of 563. The PDF version for these papers has been downloaded from the European PMC mirror (http://europepmc.org), which offers a straightforward link to download the PDF files.
We have considered several state of the art mutation extraction tools in this study, as introduced in the Text mining tools for genetic variant extraction section above. We normalized the mutation mentions identified by these tools to follow the HGVS nomenclature, to be comparable to the information in the COSMIC and InSiGHT databases. This normalization required considering the specificities of each tool, thus a normalization program was prepared for each tool. Missense mutations, mutations in the DNA that result in a protein change, are normalized to amino acid (wild type), position, amino acid (mutated), using single letter amino acid abbreviations. Thus, a mutation identified with wild type amino acid Ala, position 140 and mutated amino acid Thr is converted to A140T. DNA mutations identified by any tool are normalized to the format “c.[position][wild type nucleotide]>[mutated nucleotide]”. In the case of insertion and deletions, given position ranges, hyphens are replaced by the underscore character (e.g. c.597-598delGA to c.597_598delGA).
We ignored EMU’s Genome category since genome variants do not appear in COSMIC or InSiGHT. We did not filter out mutation mentions based on sequence validation, so we consider all the extracted mutations from EMU. When EMU identifies a dbSNP identifier, the dbSNP API is queried to obtain further details about the mutations, identifying all available candidates for DNA and protein mutations associated with each ID. There were some mentions in which the position of the DNA or protein substitution mutation was provided as exon/intron number or a codon position. The codon positions were converted to the three candidate nucleotide positions. Exon and intron mentions were removed since no precise position could be derived.
In the curated databases, the mutations are linked to the genes or proteins where they happen. In addition to the extraction of mutations, we have annotated and normalized the genes in the documents based on a dictionary developed from the NCBI Gene database, using only the human genes. We followed the procedure in Jimeno Yepes (2013)34 and removed duplicates and filtered out certain misleading or ambiguous gene names, such as those ending with disease, syndrome, or susceptibility, and removed terms from a standard stopword list. Based on observations from previous work12, we have added the following variations to the genes related to the InSiGHT database. These terms are variations of the original gene term but prefixing the letter h to indicate that it is a human gene. Thus we have added hMSH2 for MSH2, hMSH6 for MSH6 and hPMS2 for PMS2.
We used our own dictionary because this has shown to be effective34,35 and human genes are not as ambiguous compared to other species. Since our objective is to investigate the coverage of current approaches when dealing with the curation of existing databases, we can have more control on the false positives and false negatives. In addition, we are considering resources in addition to MEDLINE citations and full text documents, for which current methods based on machine learning approaches do not need to perform as expected. We have used ConceptMapper36 (http://uima.apache.org/d/uima-addons-current/ConceptMapper/ConceptMapperAnnotatorUserGuide.html) as the dictionary tagger tool reusing the configuration prepared for the BioCreative 2013 CTD track35, which does not make case distinction, tokens have to be matched in the same order and only the longest match is considered and tokens must be adjacent to each other. The identified genes are related to the mutations based on document co-occurrences.
We have performed two types of evaluation. An intrinsic evaluation of the variant extraction methods on an annotated corpus developed for the purpose of variant curation and an extrinsic evaluation based on the ability of the methods to recover the mutations from the articles.
Intrinsic evaluation of the mutation extraction tools is shown in Table 3 in terms of precision, recall and F1 score. The results are estimated based on two matching schemas: exact matching, in which the annotated entities must match exactly the span of the entities in the reference set, and partial matching, in which the annotated entities may have any overlap with the entities in the reference set. The partial matching relaxes annotation boundaries, so entities with differences like DNA or protein variant prefixes c. and p. respectively are not considered as errors.
Considering the partial matching, the precision of each tool is over 90%, which is in agreement with previously reported work on different corpora. On the other hand, the tools show different recall values. EMU and SETH show the best performance, with SETH showing better results than EMU in exact matching and in partial matching. MutationFinder has a significantly lower coverage due to its focus on point mutations. tmVar and OMM show similar partial matching performance, even though OMM extracts only protein variants. tmVar has lower performance than expected based on its previously reported performance. This might be due to the fact that tmVar was trained on abstracts, while the Variome corpus consists of full text articles. If we combine the annotations of the tools by simple merge, when evaluating the results based on partial matching, the performance is higher than any single system (precision = 0.9725, recall = 0.9060 and F1 = 0.9381), in particular due to an increase in recall.
The Variome corpus contains not only abstracts but also full text. These results show that mutation extraction tools that were developed based on MEDLINE abstracts, once applied to narrative literature still have high precision, and that their combination provides a high precision and recall solution.
Since Open Mutation Miner and MutationFinder explicitly only deal with protein mutations, we have divided the results into DNA and protein mutation subsets and estimated recall for each subset, shown in Table 4 and Table 5 respectively. Unsurprisingly, OMM and MutationFinder did not recover any DNA mutation. The result on protein mutations show that MutationFinder has a very low performance, due to its coverage of only point mutations. Open Mutation Miner has a high recall, over 96%, as well as high precision as previously reported. tmVar has quite a high recall in the protein mutation set compared to the DNA mutation set. SETH has the overall highest performance, but its recall is below EMU for protein mutations. This is because many mutations in text do not exactly follow the HGVS nomenclature. Considering DNA mutations, we find that except for SETH, the performance of the other tools is lower. This is explained because there are specific types of DNA variants, e.g. deletions, that are not as well covered by the other tools as by SETH. tmVar performance is low for DNA mutations compared EMU and SETH, with a recall of 9.62%. As can be seen in Table 5, tmVar has much stronger coverage of protein mutations as compared to DNA mutations.
Exact | TP | FN | Recall |
---|---|---|---|
EMU OMM MF tmVar SETH | 20 0 0 1 33 | 32 52 52 51 19 | 0.3846 0.0000 0.0000 0.0192 0.6346 |
Partial | TP | FN | Recall |
EMU OMM MF tmVar SETH | 33 0 0 5 37 | 17 52 52 47 12 | 0.6600 0.0000 0.0000 0.0962 0.7551 |
Exact | TP | FN | Recall |
---|---|---|---|
EMU OMM MF tmVar SETH | 46 7 7 29 48 | 20 59 59 37 18 | 0.6970 0.1061 0.1061 0.4394 0.7273 |
Partial | TP | FN | Recall |
EMU OMM MF tmVar SETH | 55 64 16 61 50 | 11 2 50 6 16 | 0.8333 0.9697 0.2424 0.9104 0.7576 |
Table 6 and Table 7 contain the frequency of false positives and false negatives made by each tool, grouped by type of genetic variation. Types of genetic variation were manually annotated. There are 8 DNA deletions, 2 DNA insertions/deletions, 42 DNA substitutions and 66 protein substitutions. In addition to a substitution that EMU identified incorrectly, all two other false positives are annotations that should be corrected in the Variome corpus. Specifically a protein substitution (V600E) and a DNA deletion (c.2700_2701delTC). EMU false negatives include deletions c.3927_3931del AAAGA, substitutions such as c.1852_1853AA>GC and mentions surrounded by parentheses. Considering the false negatives, MutationFinder failed at extracting mutations with three-letter amino acids (e.g., p.Pro622Thr) or single-letter amino acids without the p. prefix as M23A. OMM has only two false negatives that are protein mutations that appear within parentheses in text. tmVar has many DNA false negatives, which as indicated before, shows the low coverage of this variant type. SETH fails with noncompliant HGVS mutations, e.g. C1668 C > T that should be c.1668C > T. All DNA mutation extraction tools fail with some expressions like codon 41: A→G, which might require additional regular expressions.
Results on the mutation extraction are available in Table 8 for the COSMIC database and in Table 9 for the InSiGHT database. The tables show results for different representations of the articles and different mutation extraction tools used. For each representation group, there is an additional row showing the result of combining the mutations extracted by each method. Each row shows the total number of mutations in the reference set, the number of mutations matched and the recall, which is just the proportion of the matched mutations with respect to the mutations in the reference set. Matching requires matching the complete triple {PMID, gene, mutation}.
For the COSMIC database, most of the mutations are found in the supplementary material, while for InSiGHT we find that most of the mutations are spread between tables and supplementary material. Low recall is obtained from the mutations extracted from the articles’ abstracts and full text representations. This is in accordance with previously published work11,12, in which a similar effect is observed.
We explored the effect of combining the tested tools together, since it is apparent several have complementary scope. The combination was implemented simply by merging the results of the systems. The combination of all systems improves the previously published text mining coverage. Coverage increases from 45.63% recall to 62.30% in the case of the InSiGHT database and from 62.30% recall to 70.56% in the case of the COSMIC database.
The coverage of the COSMIC database is larger than the InSiGHT database. In InSiGHT, a large proportion of the information is found in tables in expressions that involve an intron position, i.e. IVS17(+5)G>C, which are not properly identified by the mutation extraction systems.
In addition to these results, we have relaxed the gene matching requirements, so only the PMID and mutation are required to match. In the result tables, these results are shown in the fields “M NG” (NG=No Gene) for the number of matched mutations and the “Rec NG” for the proportion of mutations covered from the reference set. We find that there is just a small increase in the case of the COSMIC database. There is no difference for the InSiGHT database and hence this additional data is not shown for the InSiGHT database. The InSiGHT database focuses on only four genes and the dictionary seems to cover all their possible gene name variations.
Considering the tools individually, MutationFinder has low coverage of the curated mutations. This result follows the observations from the intrinsic evaluation performed on the Variome corpus. OMM is focused on protein mutations, thus it also suffers from lower recall. Given its excellent performance in intrinsic protein variant extraction, this suggests that when its performance is low compared to other methods it means that DNA variants are more common, for instance in the supplementary material in the InSiGHT database. The coverage of SETH depends largely on the compliance with HGVS nomenclature and explains why in some cases its recall is low compared to other methods. EMU provides a more robust coverage of mutations overall. However, the combination of different methods shows an increase compared to previous published work based only on this tool. This is partially explained by the coverage provided by OMM of protein mutations but also by the performance of SETH in the extraction of more complex deletions from the InSiGHT database.
During the analysis of the results, we realized that in a small number of cases PMC makes reference to the tables of the article without including their content. In the InSiGHT database this happens only with the PMID:12373605. To mitigate this problem we downloaded all the PDF files for all the PMC documents and converted them to plain text. This is given as pdf.all in the result tables. We find that this set contains more mutations than the full text or the table sets, because full text and tables are contained in the PDF of the articles.
There are articles for which no mutation extraction tool could recover any mutation. We have performed an analysis of the coverage of the mutation extraction tools for the articles in which at least one mutation can be identified by any mutation extraction tool and at least one mutation is in the reference set, referred to as the common set. The results on the common set are available from Table 10 and Table 11. Generally the coverage of the common set is higher. This difference is most dramatic when considering the citations for which mutations can be identified in the abstract, with 23% and 29% recall in COSMIC and InSiGHT respectively.
When considering tables, the coverage of COSMIC seems to increase only slightly compared to the InSiGHT database. This might mean that many of the mutations are in tables in the InSiGHT database, while this is not the case for the COSMIC database.
Table 12 and Table 13 show the overlap of the mutations extracted by each tool using all the data sources from the articles. The results in the tables show the complementarity of the mutation extraction tools. MutationFinder has the lowest overlap with the mutations extracted by other systems, while EMU has the best coverage. The overlap of MutationFinder and OpenMutationMiner with other tools is lower in the InSiGHT database, which might indicate that there are proportionately more DNA mutations in this set compared to COSMIC.
COSMIC | EMU | OMM | MF | SETH | tmVar |
---|---|---|---|---|---|
EMU OMM MF SETH tmVar | - 0.1830 0.0266 0.3315 0.1682 | 0.7939 - 0.1623 0.7840 0.4325 | 0.5114 0.9783 - 0.0498 0.5056 | 0.7546 0.6065 0.0064 - 0.3423 | 0.6962 0.6987 0.1334 0.7077 - |
InSiGHT | EMU | OMM | MF | SETH | tmVar |
---|---|---|---|---|---|
EMU OMM MF SETH tmVar | - 0.1797 0.1459 0.0712 0.5162 | 0.1463 - 0.7134 0.0488 0.0976 | 0.1367 1.0000 - 0.0000 0.0171 | 0.3300 0.2700 0.0000 - 0.2700 | 0.7249 0.2169 0.0106 0.1429 - |
Intrinsic results show that Open Mutation Miner and SETH have the best performance for protein and DNA mutations respectively. Table 11 shows that the combination recovers a large proportion of the mutations for the COSMIC database. On the other hand, the tools still fail to identify some genetic variants, mainly when they do not follow the HGVS format, as is the case in the supplementary material in InSiGHT. When we add EMU, as shown in Table 15, the coverage increases for InSiGHT, being close to the coverage obtained by combining the different tools.
The Results section presented results for two types of experiments. The first one compares the coverage of current mutation extraction tools on a data set intended to support the curation of the InSiGHT database. The second type of experiment looks at the performance of the tools in the context of mutation database curation.
The result on mutation extraction shows that performance is quite high using partial matching, with a result over 85 in F1 measure by SETH. We see a big change from the performance of the mutation extraction tools between exact and partial matching regimes, mainly due to the differences in boundaries due to annotation of the prefixes “c.” and “p.” in the gold standard, while not included by the extraction tools.
The split between DNA and protein variants shows the varying scope of the tools. OMM shows a high recall of protein mutations, only missing the mentions M23A and V600E, which were sourrounded by parentheses. MutationFinder missed, in addition to the OMM examples, protein variants using three letter amino acid abbreviations, e.g. p.Pro622Thr. EMU missed protein variants surrounded by parenthesis too, and mutations for which the amino acid substitution was specified by an X, e.g. p.Arg226X. tmVar only missed protein variants surrounded by parenthesis. SETH missed any mention not following the HGVS nomenclature strictly, including variants without the p. protein variant prefix.
Considering the DNA variant annotation performed by EMU, tmVar and SETH, we find that all the tools missed expressions such as codon 33: C>A, expression that resembles the protein variants A1796 and very complex expressions. For example, c. 423 -6delAAATAGGTinsGAAGCAAGATCAG in PMID:18433509. EMU missed DNA variants that seem more complex, with ranges in the substitution c.1852_1853AA>GC or other expressions c.4236del8ins13. tmVar missed a large number of the DNA mutations. Many were substitutions with spaces within the text, e.g. c.5465A > T, expressions in parenthesis and verbose expressions, e.g. 1799T to A. SETH, in addition to the examples mentioned above, missed expressions that do not exactly follow the HGVS nomenclature, e.g. C1668C > T, even though it recovered the larger set of DNA variants.
Furthermore, the remaining mutation entities without change and location information include terms that are related to mutation. In some cases, there are terms that denote the location of the mutation but not the specific change, e.g. exon 10, mention the change without specifying the position, e.g. G:C to A:T transition, name a mutated gene APC+, and terms that describe the type of mutation somatic mutation. The variety of information covered by these terms might require the use of different techniques for each type. The first three types could be annotated using more general regular expressions, while a dictionary approach might be suitable for the terms describing the mutation types.
We processed the mutation types with the NCBO annotator37 (http://bioportal.bioontology.org/annotator), which uses a large set of terminologies and ontologies including the NCI thesaurus38 and the Sequence Ontology39. Some of the terms are properly annotated with concepts from these resources, e.g. somatic mutation or germline mutation while others are not covered by any of these resources, e.g. truncating mutation, including resources such as the Unified Medical Language System (UMLS)40. Some terms are partially annotated but some simple rules could be considered to fully annotate them, for example, pathogenic mutation where pathogenic and mutation are annotated with different concepts or large mutation where only mutation is annotated.
Results on the recovery of curated mutations s how two things. First, recall of curated mutations from the narrative part of the documents is very low. Second, a large number of them can be recovered from tables and supplementary material. These results echo our previous results obtained using the EMU system alone11,12, and demonstrate that the complete set of material associated with a publication is commonly considered in curation of mutation information.
There are mutations not extracted by EMU that have been found by other mutation extraction methods. Most of the annotations missed by EMU are deletions and duplicates that seem to be partly covered by EMU but are better covered by tools like SETH. There are many protein mutations missed by SETH, because of the lack of an explicit p. prefix, that are covered by Open Mutation Miner.
The most significant increase in recall of the COSMIC database happens in the supplementary material. The recall of the combination increases from 0.5671 to 0.6730. The overall recall, which was around 0.52 for EMU alone, increases to 0.7053 for the combined outputs.
In the InSiGHT database, the most significant increase in recall takes place when adding the mutations extracted from the tables. The main reason for this is the SETH tool’s identification of deletions in the tables. The overall recall, which was previously around 0.45 using EMU, increases to 0.6230 when combining the output of all the mutation tools.
We have annotated the variants available in each of the databases with a mutation type, using SETH. SETH, as mentioned in the methods section, annotates mutations based on grammar defined for HGVS22 and produces a mutation type for each string it recognizes. The list of mutation types is: SUBSTITUTION, DELETION, DELETION_INSERTION, DUPLICATION, INSERTION, FRAMESHIFT. To this set, we have added the type UNKNOWN in the case SETH does not identify a specific mutation type, typically caused by underspecified phrases such as c.? and p.?. For a few cases like the substitution p.H776_C777>QS SETH also does not return any mutation type. There are many cases in which a DNA mutation cannot be mapped to a protein mutation and frameshift mutations like L280FfsX4 are not covered by SETH.
The mutations in the InSiGHT and COSMIC databases are already in HGVS format, and so SETH can be used directly to classify each mutation in the databases by variant type. The analysis by type is available in Table 16 and Table 17. We can see that a large proportion of the variants are substitutions. The analysis of missed variants by type is available in Table 18 and Table 19.
In COSMIC, the most common type of missed mutations are DNA substitutions, mostly from two papers. These articles are PMID:21720365 with 5,589 variants and PMID:18772890 with 1,112 variants. The variants appear spread within supplementary material and in tables. We had already observed this previously11, although here we perform more detailed annotation of the entities. Most of the DNA substitutions result in a known protein mutation, although there are around 805 mutations for which the effect on the protein is unknown.
In contrast to the large number of substitutions missing from the COSMIC database, DNA deletions are the most common variant type missing in the InSiGHT database. In some cases this is due to the failure of the text mining approaches. For instance, in PMID:15655560, only the substring as 1705delAG of the deletion c.1704_1705delAG is identified by the combination of text mining tools. In addition, many of the deletions are not in a form usually expected by the mutation tools. This is the case of deletions expressed in non-standard nomenclature, e.g. del exon 3, as well as substitutions or deletions that require transformations, such as IVS17(+5)G>C, found in a table in PMID:14970868. The mutation extraction tools would need to take a closer look at these examples and incorporate the appropriate patterns.
There are mutations that are extracted by the mutation tools that do not appear in the curated databases, as found by Schench et al.29. This is because these mutations are not of the interest to the databases. This also explains the low number of matches between InSiGHT and COSMIC within the 40 overlapping articles. COSMIC focuses on somatic mutations while InSiGHT collects germline mutations related to Lynch Syndrome for just four genes. In addition, some of the extracted mentions are not functional or significant for the disease, as previously described11. For instance in PMID:10469011, the mutation Ala140Thr is extracted but the article states this mutation … is known to be functionally silent, and hence was excluded from the database.
We have looked at the types of mutations annotated by each tool, using SETH to identify the mutation type. These are available from Table 20 and Table 21. Similarly to existing results, most of the mutations are substitutions and in lower number, deletions and duplications. The tools do not reliably identify insertions, although a large number of variants that could not be annotated by SETH, and were labelled as UNKNOWN, are actually insertions. As expected, OMM and MF do not annotate DNA variants and annotate protein substitutions.
Supplementary material for the COSMIC database and tables for the InSiGHT database contribute a large number of mutations. On the other hand, processing these type of resources poses new challenges. Examination of the supplementary files show that many of the mutations listed in MS Excel files and MS Word files are in a tabular format, thus the problem could be reduced to mutation extraction from tables. Table processing for mutation extraction has been already explored by Wong et al.41. The scope of that work was to classify the type of information in each field in a table, based on a machine-learned model. This work could be extended to properly identify the gene associated to a mutation, even if it appears in the caption of the table, in a column header, and/or as a field in the same row as the mutation. This requires additional mechanisms to extract additional fields and post-process the extracted information in order to normalize the mutation mentions.
We have performed a broad assessment of the state-of-the-art performance of existing tools for extraction of genetic variants from the published biomedical literature, considering and comparing five publicly available extraction tools on two complementary evaluation tasks. We have proposed combining multiple mutation extraction tools together by merging their results, and have shown that their combination results in substantially improved recall of mutations with minimal impact on precision, providing evidence of the complementary nature of the tools.
Our results show that current tools have a very good performance on the narrative parts of published articles, and demonstrate that earlier performance claims for MEDLINE abstracts extend to full text. On the other hand, the excellent extraction performance on this narrative content contrasts with substantially lower recall in the context of database curation, even when all results of all considered tools are merged together. Our results demonstrate that only a small fraction of the curated mutations are available from the narrative part of the articles, and that most of the information is available only in tables and supplementary files associated to the articles. When the tools are deployed against this additional material, we are able to substantially increase recall. This increase is particularly evident when the tools are used together.
We have further examined in detail the performance of these tools on different types of genetic variants, considering not only the distinction between protein variation and DNA variation, but also contrasting performance on different types of variation, e.g. substitutions, deletions, and insertions. We demonstrate that the coverage of different tools is quite complementary with respect to these distinctions, providing an explanation for the performance benefit obtained by merging their results.
Future work involves integrating our results into a genetic variant database curation tool. Before achieving this goal, there are several improvements to perform. As we have seen, the combination of mutation extraction tools recover a large part of the mutations curated in existing databases, and therefore combining several tools together is a viable strategy for genetic variant extraction, but there are still variants that are not covered. Our error analysis shows that a particular gap is mutations that do not include an explicit location or that imply a variation in a specific region in a gene (e.g. Del exon 3). Coverage of DNA insertions is low, and could be a particular target for improvement.
Furthermore, special processing is required to recover information from tables. To address this, we plan to extend work previously done by Wong et al.41. Finally, a curation tool must consider the differing scopes of different genetic variant databases, i.e. COSMIC is interested in somatic mutations while InSiGHT is interested in germline mutations. Further extension to this work would therefore include classifying the variants into these two categories.
AJ designed and carried out the experiments, participated in the development of the methods and drafted the manuscript. KV designed the experiments, participated in analysis of the results, and drafted the manuscript.
National ICT Australia (NICTA) is funded by the Australian Government as represented by the Department of Broadband, Communications and the Digital Economy and the Australian Research Council through the ICT Centre of Excellence program.
The funders had no role in study design, data collection and analysis, decision to publish, or preparation of the manuscript.
We thank the InSiGHT database curator, John-Paul Plazzer of the Royal Melbourne Hospital, for sharing the InSiGHT data and helping us to interpret the database fields. We also thank the COSMIC team for helpful details about their database. We would like to thank the developers of SETH and tmVar for making their tools available. Finally, we would like to thank Andrew MacKinlay for his support in preparing the gene normalization procedure.
Views | Downloads | |
---|---|---|
F1000Research | - | - |
PubMed Central
Data from PMC are received and updated monthly.
|
- | - |
Competing Interests: No competing interests were disclosed.
Competing Interests: No competing interests were disclosed.
Competing Interests: I am one of the authors of the benchmarked tools (i.e., SETH).
Alongside their report, reviewers assign a status to the article:
Invited Reviewers | |||
---|---|---|---|
1 | 2 | 3 | |
Version 2 (revision) 10 Jun 14 |
read | read | |
Version 1 21 Jan 14 |
read | read | read |
Provide sufficient details of any financial or non-financial competing interests to enable users to assess whether your comments might lead a reasonable person to question your impartiality. Consider the following examples, but note that this is not an exhaustive list:
Sign up for content alerts and receive a weekly or monthly email with all newly published articles
Already registered? Sign in
The email address should be the one you originally registered with F1000.
You registered with F1000 via Google, so we cannot reset your password.
To sign in, please click here.
If you still need help with your Google account password, please click here.
You registered with F1000 via Facebook, so we cannot reset your password.
To sign in, please click here.
If you still need help with your Facebook account password, please click here.
If your email address is registered with us, we will email you instructions to reset your password.
If you think you should have received this email but it has not arrived, please check your spam filters and/or contact for further assistance.
Comments on this article Comments (0)