The literature presented in PLATIPUS is collected from original publishers in collaboration with the COVIDScholar project at the University of California, Berkeley/Lawrence Berkeley National Laboratory [5]. Articles in COVIDScholar are sourced by a system of dedicated web scrapers, document parsers, databases, and machine learning models that process papers and metadata into a standardized format that is amenable for text mining. The data in COVIDScholar includes a culmination of 19 sources, presented in Textbox 1, and consists of academic preprints, peer-reviewed research papers, book chapters, patents, clinical trial descriptions, and data sets, all of which have been made openly available by the original publishers to advance COVID-19 research. COVIDScholar updates their data multiple times per day and PLATIPUS queries the COVIDScholar database and reingests new articles once a day.

PLATIPUS uses a tool called Automated Analytics and Integration of Data (AAID) to assist in the data ingestion and advanced analytic processing of the COVIDScholar data set. AAID uses multiple algorithms to identify key sources of information while taking into account how the meaning of words change based on the context [14]. AAID uses natural language processing methodologies, specifically entity recognition, machine learning, and human-in-the-loop, to augment the data with additional queryable tags [15]. In PLATIPUS, this means augmenting the COVIDScholar data set with tags such as locations, organizations, diseases, diagnostics and analysis, countermeasures, species, and additional context. AAID uses the NiFi data ingestion and processing pipeline that contains a variety of natural language processing methods such as time-weighted penalized logistic regression models, recursive regex, binary bag of words models, and recurrent neural network models, which is described in detail in Multimedia Appendix 1 Figure S1. The vectorization of the text was based on a bag of words approach. For the clustering visualizations, a k-means default method was used. The analytic capabilities of the AAID pipeline continue to grow to use transformer deep learning classifiers and implement methods to identify anomalies and abnormal characteristics [16].

As of May 2021, there are 159,797 articles that are parsed into various filters. At the top level are authors (n=564,845), categories (n=7), context (n=41), countermeasures (n=28), diagnostics and assay (n=19), disease (n=265), journal (n=11,412), locations (n=365), tags (n=7), species (n=76), and chemicals (n=175). Authors are associated with the publications, and therefore, there are hundreds. For selection purposes, the authors are sorted in order from the most to least prevalent. There are presently seven core categories (treatment, prevention, mechanism, diagnosis, epidemic forecasting, transmission, and case report). Under context, there are 41 groupings associated with the primary context of the article (eg, disease severity or transmission event). Countermeasures are approaches taken against the disease (eg, treatment, vaccine, or awareness campaign). The diagnostics and assay groupings contain the platforms associated with the article, such as transcriptomics or x-rays. Disease is again a broad category where the most prevalent is a categorization of human or animal disease but other specific associated syndrome or special notes are captured here. Journal, similar to author, is a large group of the virtual location of the publication online. Location is a physical location at which the research or case study is conducted for publication, which are extracted using resources from the National Geospatial-Intelligence Agency and United States Geological Survey [17,18]. There are 76 species, the most prevalent being human, rodents, and swine, and 175 chemicals captured that are associated with the manuscripts.

PLATIPUS is built on top of the SERBERUS application, which is an end-to-end software solution that rapidly builds visual analytic web applications (Figure 1). Powered by the Scalable Reasoning System (SRS) [19] on the back end and a flexible user interface toolkit on the front end, and drawing from expertise from a user experience and design team, this system is designed for custom solutions that can be readily constructed to support data exploration, discovery, and understanding.

The PLATIPUS application provides the end user with a variety of ways to search and filter over 100,000 COVID-19 publications. Since PLATIPUS is built on top of SRS and Slykit, PLATIPUS will continue to evolve and grow with new visualizations and features as SRS and Slykit advances. As of May 2021, PLATIPUS allows the user to filter on locations, categories, authors, organization, disease, diagnostics and analysis, countermeasures, species, and additional context as well as a timeline. The visualizations that are currently available are circle pack, cluster pack, donut graphs, edge-based graph, line chart, matrix, metrics, paracord, table, text clusters, treemap, and timeline described in Textbox 2. The first 10 of these visualizations are at the center of the dashboard and can be assembled based on user choice (one, two, three, etc) all in the view. The timeline visualization is maintained across the top of the user interface. At any time during the filtering and searching process, the user can access a high-level overview of an individual publication, which includes the abstract, information about the authors, tags and categories, and the journal where it was published as well as a direct link to the full publication. Once the user filters down to a subset of publications of interest, they can export the list of publications as a CSV file.

The primary manner the scientific community interacts with scientific literature has, up until recently, not changed in decades. COVID-19 has brought to the forefront of research the challenge of mining literature versus identification of potential articles of interest to a user by keyword searches. To date, PLATIPUS has performed text analytics and clusters, and has visualized nearly 160,000 articles related to COVID-19, and it automatically updates as new documents are added to COVIDScholar. The application uses state-of-the-art natural language processing (AAID) to provide insight and unique ways to filter and understand the data. PLATIPUS aims to decrease time spent looking through pages of articles by providing the user with multiple ways to search, filter, and view the data. The PLATIPUS application focuses on taking the large amount of literature related to COVID-19 and displaying keywords, categories, and other metadata to allow a user to quickly find relevant information captured by COVIDScholar.

PLATIPUS was designed to assist in searching a multitude of COVID-19 publications efficiently, so the user can either find their answer using the visualizations, searching, and drill down capabilities or find a document that will assist in their search. Therefore, PLATIPUS does not support saving views or searches, as it was designed to be a visual analytics search engine and visual table of contents. Additional limitations include the suggestion of the optimal visualization based on a query. PLATIPUS allows the users to toggle through visualizations and select those that are of the most utility. Additions to PLATIPUS in the future may be a more guided visualization experience based on the size and complexity of the literature returned from a query. As of March 2021, PLATIPUS does not support finding similar articles to a single selection, but we expect this feature will be available in the future.