Dataset Information

Building deep learning models for evidence classification from the open access biomedical literature.

ABSTRACT: We investigate the application of deep learning to biocuration tasks that involve classification of text associated with biomedical evidence in primary research articles. We developed a large-scale corpus of molecular papers derived from PubMed and PubMed Central open access records and used it to train deep learning word embeddings under the GloVe, FastText and ELMo algorithms. We applied those models to a distant supervised method classification task based on text from figure captions or fragments surrounding references to figures in the main text using a variety or models and parameterizations. We then developed document classification (triage) methods for molecular interaction papers by using deep learning mechanisms of attention to aggregate classification-based decisions over selected paragraphs in the document. We were able to obtain triage performance with an accuracy of 0.82 using a combined convolutional neural network, bi-directional long short-term memory architecture augmented by attention to produce a single decision for triage. In this work, we hope to encourage biocuration systems developers to apply deep learning methods to their specialized tasks by repurposing large-scale word embedding to apply to their data.

SUBMITTER: Burns GA

PROVIDER: S-EPMC6449534 | biostudies-literature | 2019 Jan

REPOSITORIES: biostudies-literature

ACCESS DATA

Publications

Building deep learning models for evidence classification from the open access biomedical literature.

Burns Gully A GA Li Xiangci X Peng Nanyun N

Database : the journal of biological databases and curation 20190101

We investigate the application of deep learning to biocuration tasks that involve classification of text associated with biomedical evidence in primary research articles. We developed a large-scale corpus of molecular papers derived from PubMed and PubMed Central open access records and used it to train deep learning word embeddings under the GloVe, FastText and ELMo algorithms. We applied those models to a distant supervised method classification task based on text from figure captions or fragm ...[more]

PMID: 30938776

Similar Datasets

Project description:BACKGROUND: There is a growing need for efficient and integrated access to databases provided by diverse institutions. Using a linked data design pattern allows the diverse data on the Internet to be linked effectively and accessed efficiently by computers. Previously, we developed the Allie database, which stores pairs of abbreviations and long forms (LFs, or expanded forms) used in the life sciences. LFs define the semantics of abbreviations, and Allie provides a Web-based search service for researchers to look up the LF of an unfamiliar abbreviation. This service encounters two problems. First, it does not display each LF's definition, which could help the user to disambiguate and learn the abbreviations more easily. Furthermore, there are too many LFs for us to prepare a full dictionary from scratch. On the other hand, DBpedia has made the contents of Wikipedia available in the Resource Description Framework (RDF), which is expected to contain a significant number of entries corresponding to LFs. Therefore, linking the Allie LFs to DBpedia entries may present a solution to the Allie's problems. This requires a method that is capable of matching large numbers of string pairs within a reasonable period of time because Allie and DBpedia are frequently updated. RESULTS: We built a Linked Open Data set that links LFs to DBpedia titles by applying key collision methods (i.e., fingerprint and n-gram fingerprint) to their literals, which are simple approximate string-matching methods. In addition, we used UMLS resources to normalise the life science terms. As a result, combining the key collision methods with the domain-specific resources performed best, and 44,027 LFs have links to DBpedia titles. We manually evaluated the accuracy of the string matching by randomly sampling 1200 LFs, and our approach achieved an F-measure of 0.98. In addition, our experiments revealed the following. (1) Performances were similar independently from the frequency of the LFs in MEDLINE. (2) There is a relationship (r2 = 0.96, P < 0.01) between the occurrence frequencies of LFs in MEDLINE and their presence probabilities in DBpedia titles. CONCLUSIONS: The obtained results help Allie users locate the correct LFs. Because the methods are computationally simple and yield a high performance and because the most frequently used LFs in MEDLINE appear more often in DBpedia titles, we can continually and reasonably update the linked dataset to reflect the latest publications and additions to DBpedia. Joining LFs between scientific literature and DBpedia enables cross-resource exploration for mutual benefits.

Project description:Currently, there is a growing interest in ensuring the transparency and reproducibility of the published scientific literature. According to a previous evaluation of 441 biomedical journals articles published in 2000-2014, the biomedical literature largely lacked transparency in important dimensions. Here, we surveyed a random sample of 149 biomedical articles published between 2015 and 2017 and determined the proportion reporting sources of public and/or private funding and conflicts of interests, sharing protocols and raw data, and undergoing rigorous independent replication and reproducibility checks. We also investigated what can be learned about reproducibility and transparency indicators from open access data provided on PubMed. The majority of the 149 studies disclosed some information regarding funding (103, 69.1% [95% confidence interval, 61.0% to 76.3%]) or conflicts of interest (97, 65.1% [56.8% to 72.6%]). Among the 104 articles with empirical data in which protocols or data sharing would be pertinent, 19 (18.3% [11.6% to 27.3%]) discussed publicly available data; only one (1.0% [0.1% to 6.0%]) included a link to a full study protocol. Among the 97 articles in which replication in studies with different data would be pertinent, there were five replication efforts (5.2% [1.9% to 12.2%]). Although clinical trial identification numbers and funding details were often provided on PubMed, only two of the articles without a full text article in PubMed Central that discussed publicly available data at the full text level also contained information related to data sharing on PubMed; none had a conflicts of interest statement on PubMed. Our evaluation suggests that although there have been improvements over the last few years in certain key indicators of reproducibility and transparency, opportunities exist to improve reproducible research practices across the biomedical literature and to make features related to reproducibility more readily visible in PubMed.

Project description:BackgroundAn open challenge in translational bioinformatics is the analysis of sequenced metagenomes from various environmental samples. Of course, several studies demonstrated the 16S ribosomal RNA could be considered as a barcode for bacteria classification at the genus level, but till now it is hard to identify the correct composition of metagenomic data from RNA-seq short-read data. 16S short-read data are generated using two next generation sequencing technologies, i.e. whole genome shotgun (WGS) and amplicon (AMP); typically, the former is filtered to obtain short-reads belonging to a 16S shotgun (SG), whereas the latter take into account only some specific 16S hypervariable regions. The above mentioned two sequencing technologies, SG and AMP, are used alternatively, for this reason in this work we propose a deep learning approach for taxonomic classification of metagenomic data, that can be employed for both of them.ResultsTo test the proposed pipeline, we simulated both SG and AMP short-reads, from 1000 16S full-length sequences. Then, we adopted a k-mer representation to map sequences as vectors into a numerical space. Finally, we trained two different deep learning architecture, i.e., convolutional neural network (CNN) and deep belief network (DBN), obtaining a trained model for each taxon. We tested our proposed methodology to find the best parameters configuration, and we compared our results against the classification performances provided by a reference classifier for bacteria identification, known as RDP classifier. We outperformed the RDP classifier at each taxonomic level with both architectures. For instance, at the genus level, both CNN and DBN reached 91.3% of accuracy with AMP short-reads, whereas RDP classifier obtained 83.8% with the same data.ConclusionsIn this work, we proposed a 16S short-read sequences classification technique based on k-mer representation and deep learning architecture, in which each taxon (from phylum to genus) generates a classification model. Experimental results confirm the proposed pipeline as a valid approach for classifying bacteria sequences; for this reason, our approach could be integrated into the most common tools for metagenomic analysis. According to obtained results, it can be successfully used for classifying both SG and AMP data.

Project description:Background. Open access (OA) journals allows access to research papers free of charge to the reader. Traditionally, biomedical researchers use databases like MEDLINE and EMBASE to discover new advances. However, biomedical OA journals might not fulfill such databases' criteria, hindering dissemination. The Directory of Open Access Journals (DOAJ) is a database exclusively listing OA journals. The aim of this study was to investigate DOAJ's coverage of biomedical OA journals compared with the conventional biomedical databases. Methods. Information on all journals listed in four conventional biomedical databases (MEDLINE, PubMed Central, EMBASE and SCOPUS) and DOAJ were gathered. Journals were included if they were (1) actively publishing, (2) full OA, (3) prospectively indexed in one or more database, and (4) of biomedical subject. Impact factor and journal language were also collected. DOAJ was compared with conventional databases regarding the proportion of journals covered, along with their impact factor and publishing language. The proportion of journals with articles indexed by DOAJ was determined. Results. In total, 3,236 biomedical OA journals were included in the study. Of the included journals, 86.7% were listed in DOAJ. Combined, the conventional biomedical databases listed 75.0% of the journals; 18.7% in MEDLINE; 36.5% in PubMed Central; 51.5% in SCOPUS and 50.6% in EMBASE. Of the journals in DOAJ, 88.7% published in English and 20.6% had received impact factor for 2012 compared with 93.5% and 26.0%, respectively, for journals in the conventional biomedical databases. A subset of 51.1% and 48.5% of the journals in DOAJ had articles indexed from 2012 and 2013, respectively. Of journals exclusively listed in DOAJ, one journal had received an impact factor for 2012, and 59.6% of the journals had no content from 2013 indexed in DOAJ. Conclusions. DOAJ is the most complete registry of biomedical OA journals compared with five conventional biomedical databases. However, DOAJ only indexes articles for half of the biomedical journals listed, making it an incomplete source for biomedical research papers in general.

Dataset Information

Building deep learning models for evidence classification from the open access biomedical literature.

Publications

Building deep learning models for evidence classification from the open access biomedical literature.

Similar Datasets

OmicsDI is part of the ELIXIR infrastructure

Tweets