|Home | About | Journals | Submit | Contact Us | Français|
Negated biomedical events are often ignored by text-mining applications; however, such events carry scientific significance. We report on the development of BioNT, a database of negated sentences that can be used to extract such negated events.
Currently BioNT incorporates ≈32 million negated sentences, extracted from over 336 million biomedical sentences from three resources: ≈2 million full-text biomedical articles in Elsevier and the PubMed Central, as well as ≈20 million abstracts in PubMed. We evaluated BioNT on three important genetic disorders: autism, Alzheimer's disease and Parkinson's disease, and found that BioNT is able to capture negated events that may be ignored by experts.
The BioNT database can be a useful resource for biomedical researchers. BioNT is freely available at http://bionot.askhermes.org/. In future work, we will develop semantic web related technologies to enrich BioNT.
In the biomedical domain, a large amount of published literature is available in electronic format, spurring the development of several text-mining applications that can process the available literature to automatically extract information such as protein-protein interaction and gene-disease association. Unfortunately, the text mining community tends to focus on positive events only. Many text-mining applications either ignore sentences containing negation or do not process negation at all, a situation that could lead to negated events being identified as positive events. We argue that negated events provide valuable information and may help researchers formulate research hypotheses.
A use case for extracting negated events can be seen in the case of genomic incidentalomes . As genomic medicine develops to offer genome-level screening tests, it is important to identify genes that were earlier believed to be associated with a disease, but eventually were found not to be associated with the disease. Such genes should be removed from the array of genomic tests a patient undergoes since failure to do so will result in the patient being subjected to unnecessary tests, causing additional morbidity, and hence, increasing the cost of genomic medicine substantially. Finding reported instances of a gene not being associated with a disease is difficult, which is why our goal in this study is to develop a text mining application that can identify such negated relations.
In this study we attempt to fill the gap created due to the absence of text mining applications that extract negated events. Our long-term goal is to expand the existing BioNT system to identify biomedical named entities (e.g., gene and protein names), and therefore enable BioNT to capture negative relations between named entities. Here we report the development of a database called BioNT, which contains negated sentences from three sources: abstracts of articles indexed by PubMed, full-text of articles in the PubMed Central Open Access Subset, and full-text of articles published by Elsevier publisher. We have indexed the sentences in BioNT and made them available online through a search engine, available at http://bionot.askhermes.org/.
Detection of negation in biomedical literature is an important task. As evidence, the BioNLP'09 Shared Task on Event Extracted included negation detection as one of the tasks. Several text mining applications exploring negation identification have been developed.
In the clinical domain, rule-based approaches have been developed for negation detection. For example, Chapman et al.  developed the NegEx system to identify negation of target findings and diseases in narrative medical reports. The current version of NegEx uses 272 rules, which are matched by using regular expression patterns. The reported recall of the system was 95.93%, precision was 93.27%, and accuracy was 97.73%. A similar system, Negfinder, was developed to identify negated concepts in medical narratives . The system first identifies negation markers in the sentence by using regular expression patterns. These words are then passed to a parser that uses a single-token look-ahead strategy to identify negated concepts. The reported recall and precision of the system were 95.27% and 97.67%, respectively. Along the same lines, Elkin et al.  developed a system to identify the negation of concepts in electronic medical records. The system was built by identifying textual cues for negation in 41 clinical documents. The reported recall and precision of the system was 97.2% and 91.2%, respectively. A hybrid approach that classifies negations in radiology reports based on the syntactic categories of the negation signal and negation patterns was developed by Huang and Lowe . Thirty radiology reports were manually inspected to develop the classifier and the classifier was validated on a set of 470 radiology reports. Evaluation was conducted on 120 radiology reports and the reported recall and precision were 92.6% and 98.6%, respectively.
Supervised machine-learning approaches have also been developed for negation detection. Averbuch et al. developed an algorithm to automatically learn negative context patterns in medical narratives . The algorithm uses information gain to learn negative context patterns. Aramaki et al. developed a negative event recognition module for their medical text summarization system . The module was based on a supervised machine-learning algorithm, Support Vector Machines, which uses syntactic information to detect negation.
In the genomics domain, a rule-based system was developed by Sanchez-Graillet and Poesio to detect negated protein-protein interactions in the biomedical literature . The system was built using a full dependency parser. Hand-crafted rules were then used to detect negated protein-protein interaction. An example rule reads as follows: if cue verb, such as 'interact,' is an object of 'fail,' 'Protein A' is subject of fail, and 'Protein B' is object of interact, then there is no interaction between 'Protein A' and 'Protein B.' Evaluation was conducted on 50 biomedical articles and the best recall and precision reported were 66.27% and 89.15%, respectively.
If a negation is reported in a sentence, it might not apply to the entire sentence. For example, in the sentence, ''While there was no difference in overall growth between BRCA1+ and BRCA1 wt cells, BRCA1+ cells showed a marked reduction in survival following STS treatment.'' the negation marker 'no' negates the observation 'difference in overall growth between BRCA1+ and BRCA1 wt cells following STS treatment.' The observation 'BRCA1+ cells showed a marked reduction in survival following STS treatment' is positive and not modified by the negation marker. Hence, it is important to identify the scope of negation as well. The following studies identify the scope of negation in a sentence.
Morante and Daelemans  developed a two-phase approach to detect the scope of negation in biomedical literature. In the first phase, negation cues were identified by a set of classifiers. In the second phase, another set of classifiers was used to detect the scope of the negation. The system performed better than the baseline in identifying negation signals in text and the scope of negation. The percentages of correct scope for abstract, full-text and clinical articles were 66.07%, 41.00% and 70.75%, respectively.
We developed a negation scope detection algorithm called NegScope previously . NegScope was developed by training supervised machine-learning algorithm conditional random field (CRF)  using words and parts of speech as features. The CRF models were trained on the BioScope dataset , which contains more than 20,000 manually annotated sentences from clinical notes and published biological articles. In each sentence, the scope of negation and hedging has been annotated. On evaluating NegScope, we found that it predicted the negation status of a biological sentence with 99.86 % accuracy and 96.5 % F1-score. The F1-score was calculated as the harmonic mean of precision and recall. NegScope correctly identified the scope of negation in 81% of biological sentences. To our knowledge, NegScope is the only open-source package that can detect scope of negation in biological text and one of two open-source packages for clinical notes. Moreover, we found that it performed better than other negation or negation scope detection algorithms when tested on biological sentences .
BioContrasts  was developed to detect and search contrastive relations between proteins. In this system, contrastive information was extracted using manually curated patterns such as 'A but not B,' where A and B were restricted to protein names from Swiss-Prot entries. A total of 41,471 contrast relations were identified by the system from 2.5 million Medline abstracts.
As mentioned earlier, BioNT is a searchable database of negated biomedical sentences. We obtained these sentences from three sources - (1) abstracts of Medline articles (≈ 19 million abstracts; ≈ 101 million sentences), (2) full-text of PubMed Central Open Access Subset (≈ 167,000 articles; ≈ 20 million sentences) and (3) full-text of articles published by Elsevier publisher (≈1.9 million articles; ≈ 215 million sentences). We split articles for sentences using the NaCTeM sentence splitter .
Many text-mining applications make use of sentences to extract information from literature. These sentences often contain multiple entities. If a negation is reported in such sentences, it might not apply to all entities in the sentence. To identify the negated entities, it is important to identify the scope of negation. As mentioned in the Related Works, the NegScope algorithm that we developed can identify the scope of negation; hence, we used it to build BioNT.
To prepare the BioNT database, we tagged each extracted sentence. The previous and following sentences were stored to provide contextual information. For the search engine, we indexed all negated sentences and the sentences preceding and following the negated sentence using the open-source Apache Lucene package . The preceding and following sentences are displayed along with the search results. We also indexed the negation scope in the sentence. When the user enters a query, we retrieve relevant sentences based on the terms in the query. Here, a term is an individual word in the query. When searching for negated events, we applied the following heuristic - if a single term is used, it can be present anywhere in the negated sentence; if more than one term is used, all terms should appear in the negated sentence and at least one of those terms should appear within the scope of negation. Note that since the scope of negation is a part of the sentence, terms appearing in the scope of negation appear in the sentence as well.
As described in the Background section, in case of incidentalomes, it is important to identify genes that were earlier believed to be associated with a disease, but eventually were found not to be associated with the disease. Sometimes literature is published indicating that the association is not held in certain circumstances. With BioNT, our goal is to develop a searchable database that can be used by researchers to identify such negated relationships. To test the utility of BioNT, we evaluated it on the detection of knowledge of three important genetic diseases: Autism, Alzheimer's disease, and Parkinson's disease. Several genes have been thought to be associated with the manifestation of these diseases. We consulted published reviews to identify genes thought to be associated with these diseases and found 26 putative genes for autism (see Table Two in ), 10 for Alzheimer's disease (see Table Two in ), and 6 for Parkinson's disease (see Table Two in). Using BioNT, we searched for each disease and its putative gene as query. We also searched the index by replacing the disease name with related keywords; for example, autism was replaced with keywords 'ASD' (Autism Spectrum Disorder) and 'autistic.' We manually analyzed the sentences that were returned for autism to analyze the errors in our system.
To build the BioNT database, we analyzed a total of 336 million sentences, out of which 32 million sentences had negation; hence, 9.53% of all sentences contained negation in them (Table (Table1).1). It should be noted that these sentences contained some form of negation, and do not necessarily indicate negation between biomedical entities.
We searched BioNT for negated sentences containing a potential autism, Alzheimer's disease or Parkinson's disease-related gene (list of genes obtained from [16-18]) and the disease name. We found negated relation evidence for 12 out of 26 autism-related genes (Table (Table2),2), 8 out of 10 Alzheimer's disease-related genes (Table (Table3),3), and 3 out of 6 Parkinson's disease-related genes (Table (Table44).
We manually analyzed the sentences that were returned for autism. A total of 141 sentences for 20 genes were obtained when we searched the BioNT database with autism and genes thought to be associated with autism. NegScope correctly identified negation in 137 out of 141 sentences, which was consistent with the 97% F1-score observed for negation cue detection with NegScope. On the other hand, we found that 81 out of the 137 sentences did not establish a negated relation between the designated gene and disease. We therefore consider that a total of 85 sentences (81 + 4) were false positives. The remaining 56 sentences were considered to be true positives. Our results show that the precision for detecting a negated relation between a gene and a disease is 40%. The 56 true positive sentences show negated relations for 12 genes (out of the 20 genes returned by BioNT).
We further analyzed the 85 false positive sentences and found that they can be grouped into three categories: (1) No negated relation (31 sentences), (2) Ambiguous negated relation (17 sentences), and (3) Ambiguous terms (37 sentences) (see false positive sentences for categories (1), (2) and (3) in Additional File 1 Additional File 2 and Additional File 3 respectively). The following list shows example false positive sentences (a), (b) and (c) for categories (1), (2) and (3), respectively -
(a) Because deletions encompassing OXTR have not been observed in other studies characterizing structural variation in autism (58,64,65) such events appear to be rare.
(b) A scan of the NRXN1 coding sequence in a cohort of ASD subjects, relative to non-ASD controls, revealed that amino acid alterations in neurexin 1 are not present at high frequency in ASD.
(c) None of them met, or had ever met, the diagnostic criteria for autism.
The four sentences for which NegScope did not correctly identify negation were classified as category 1 false positives. In another four category 1 false positive sentences, the sentence boundaries were not correctly identified, leading to a false positive relation. All of the category 3 false positives were caused due to the same gene, MET, which is also a common English word.
In this study, we report the development of BioNT, a publicly available database of 32 million negated sentences taken from three major literature resources: PubMed, PubMed Central, and Elsevier. BioNT is currently the only database available that reports negated events reported in biomedical literature. Our study found that almost 10% of sentences published in biomedical literature incorporated negated information. The statistics indicate that negated events are abundant in biomedical literature and therefore BioNT can be an important resource for biomedical scientists.
After evaluating negated sentences for autism, Alzheimer's disease, and Parkinson's disease, we found many genes that are thought to be relevant by experts incorporate biomedical evidences suggesting the opposite.
Despite its utility, BioNT has several limitations. Although extensive, it is not comprehensive as there are several full-text articles that were not analyzed by BioNT. BioNT relies on NegScope to identify and mark negation scope; hence, errors in NegScope's predictions could result in certain negated cases being missed by BioNT. Moreover, we used the heuristic that an event is negated if all entities in the query are present in the same sentence and at least one of them is within the scope of negation. However, given the nature of discourse, this situation may not always be true. For example, in the following sentence, the negation scope is marked in boldface, and it can be seen that the genes FMR1, TSC1, TSC2, NF1 and MECP2 are not negated; however, BioNT marked the association between these genes and autism as negative - To date, genome scans, linkage and association studies, chromosomal rearrangement analyses and mutation screenings have identified: (i) genomic regions likely to contain autism susceptibility loci on human chromosomes 1 q, 2 q, 5 q, 6 q, 7 q, 13 q, 15 q, 17 q, 22 q, Xp and Xq; (ii) genes whose mutations represent a rare cause of non-syndromic autism (NLGN3 and NLGN4) or yield syndromic autism (FMR1, TSC1, TSC2, NF1 and MECP2); and (iii) candidate vulnerability genes, with potential common variants enhancing risk but not causing autism per se (Table (Table1).1). Finally, BioNT is not aware of the semantic category of the target entities, which can lead to false positives. For example, gene MET is thought to be associated with autism because several irrelevant sentences have the word 'met' in them but it is not used as a gene name.
Our results show that a long way still remains before negated events can be incorporated for genetic diagnosis. Additional semantic information may benefit the task, including complete or incomplete penetrance, gene expression, and molecular functions.
We plan to address some of the above mentioned limitations as future work. First, we plan to mark the semantic categories of words in the negated sentences. Specifically, we plan to mark entities such as genes, diseases, drugs, cells, chemicals, species and other biomedical entities within these sentences. This approach would help avoid false positives when one of the target entities is also a common English word or when an acronym is ambiguous. Marking semantic information would also help to identify cases when synonyms of entities might have been used. We will also explore heuristics that can better identify if the relationship between two entities is negated or not.
Although often ignored, negated sentences contain valuable information. To capture this information, we have automatically identified negated sentences from various published repositories and built a database of negated sentences called BioNT. Currently, BioNT comprises 32 million sentences. To make the identified sentences publicly available, we have built a search engine that is available online. We showed that our system can be used to find negated relation between genes and diseases by identifying negated relation between three genetic disorders: autism, Alzheimer's disease and Parkinson's disease, and genes thought to be associated with these disorders.
Besides identifying negated gene-disease relationship, our system can be used to identify published negated events between chemicals, drugs, diseases, cells, and other biomedical entities. Although currently our system is currently text-based, in the future, we plan to identify various biomedical entities and normalize them to improve the performance of this system.
BioNT can be freely accessed online at http://bionot.askhermes.org from any modern web-browser.
SA developed the system and conducted evaluation of the BioNT system. HY and IK provided guidance. All authors read and approved the final manuscript.
No negated relation between gene and disease. This file lists the false positive associations caused because a negated association did not exist between the gene and disease.
Ambiguous negated association between the gene and disease. This file lists the false positive associations caused because the negated association between the gene and disease was ambiguous.
Ambiguous terms. This file lists the false positive associations caused because either the gene name or the disease name was ambiguous.
The authors acknowledge the support from the National Library of Medicine, grant numbers 5R01LM009836 to Hong Yu and 5R01LM010125 to Isaac Kohane. Any opinions, findings, or recommendations are those of the authors and do not necessarily reflect the views of the NIH. We would like to thank Dr. Marjorie Piechowski and Stephanie Nicholle Hill for proofreading the manuscript.