Tagger: BeCalm API for rapid named entity recognition

Abstract

Abstract. Most BioCreative tasks to date have focused on assessing the quality of text-mining annotations in terms of precision of recall. Interoperability, speed, and stability are, however, other important factors to consider for practical applications of text mining. The new BioCreative/BeCalm TIPS task focuses purely on these. To participate in this task, I implemented a BeCalm API within the real-time tagging server also used by the Reflect and EXTRACT tools. In addition to retrieval of patent abstracts, PubMed abstracts, and PubMed Central open-access articles as required in the TIPS task, the BeCalm API implementation facilitates retrieval of documents from other sources specified as custom request parameters. As in earlier tests, the tagger proved to be both highly efficient and stable, being able to consistently process requests of 5000 abstracts in less than half a minute including retrieval of the document text.Keywords. Named Entity Recognition; Application Programming Interface; Web Service; Performance Evaluation IntroductionBioCreative and other shared tasks in the biomedical text-mining community have over the years played a key role in progressively improving text-mining methods, in particular for named entity recognition (NER). Most BioCreative tasks have focused purely on evaluating the precision and recall (1,2), with the BioC interoperability task (3) and the interactive annotation task (IAT) (4) being notable exceptions. However, as illustrated by the latter two tasks, whereas precision and recall are obviously important factors, they are far from the only factors that matter when using text mining in practice. Interoperability, speed, and stability are other very important factors; the new Technical In-. CC-BY 4.0 International license not peer-reviewed) is the author/funder. It is made available under a The copyright holder for this preprint (which was . http://dx.doi.org/10.1101/115022 doi: bioRxiv preprint first posted online Mar. 8, 2017; Proceedings of the BeCalm challenge evaluation workshop teroperability and Performance of annotation Servers (TIPS) task set out to evaluate just that.I participated in the BioCreative V IAT (4) with the interactive annotation tool, EXTRACT, which helps curators find and extract standardcompliant terms for annotation of metagenomic records and other samples (5). Behind its web-based user interface, the system makes use of the same real-time tagger for NER as the augmented browsing tool Reflect (6). The core NER engine was designed from the ground up with speed in mind and is capable of tagging thousands of PubMed abstracts per second per CPU core (7). This makes it ideally suited for largescale and real-time applications, such as the TIPS task.Here, I present a BeCalm API for the NER tagger underlying the EXTRACT (5) and Reflect (6). The system delivered a total turnaround time of about 1 second for small requests, and was able to process approximately 5,000-10,000 abstracts per minutes for larger batch requests. Notably, the vast majority of this ...

modEnrichr: a suite of gene set enrichment analysis tools for model organisms

“…We then converted these lists of PMIDs into gene sets for each submitted term using either GeneRIF (25) or AutoRIF. AutoRIF is an alternative version of GeneRIF that we created using Tagger (26) to associate genes with PMIDs for each organism. To further enrich these newly created libraries with predicted genes that may be associated with the terms, we used gene–gene co-expression data created from ARCHS4 Zoo (32) to substitute the original genes within each gene set with those genes that collectively are most correlated with the original genes in the set based on co-expression correlations (see ‘Materials and Methods’ section for details).…”

“…This was accomplished by first downloading organism-specific Ensembl ID—PMID association data from the Jensen Lab website (https://jensenlab.org/). These associations were generated using Tagger (26). Protein IDs were converted to gene symbols for each organism using the STRING v10.x version_mapping files, and STRING display names files, for each organism.…”

“…However, there are other sources and resources that can be organized into gene set libraries for model organisms. These include gene expression signatures from transcriptomics and proteomics studies, gene sets created from protein interaction databases, gene sets generated from resources such as the gene reference into function (GeneRIF) (25), and gene sets generated from gene-PubMed ID (PMID) associations determined using Tagger (26), a named entity recognition (NER) application. In general, sources for non-mammalian model organism gene sets are sparser compared to those available for human and mouse.…”

The COVID-19 Gene and Drug Set Library

“…Predictions of additional genes potentially associated with these terms were also added to the COVID-19 gene set library. These predictions were based on the literature-associated genes using each of five strategies: Co-occurrence via AutoRIF, GeneRIF, Enrichr ( 19 ), Tagger ( 20 ), and co-expression using data from ARCHS4 ( 21 ).…”

The COVID-19 Drug and Gene Set Library

“…Predictions of additional genes potentially associated with the genes directly co-mentioned with these terms were also added to the database. These predictions were based on five strategies: co-occurrence via AutoRIF, GeneRIF, 37 Enrichr, 38 or Tagger, 39 and co-expression using data from ARCHS4. 40 …”