[Paper Review] COVID-19 Literature Knowledge Graph Construction and Drug Repurposing Report Generation

Qingyun Wang, Manling Li+25 more|arXiv (Cornell University)|Jul 1, 2020

Topic ModelingComputer Science60 references36 citations

TL;DR

The paper introduces COVID-KG, a multimedia knowledge-graph framework that extracts text and image-based biomedical knowledge from COVID-19 literature to enable QA and generate drug repurposing reports with evidence and subgraphs.

ABSTRACT

To combat COVID-19, both clinicians and scientists need to digest vast amounts of relevant biomedical knowledge in scientific literature to understand the disease mechanism and related biological functions. We have developed a novel and comprehensive knowledge discovery framework, COVID-KG to extract fine-grained multimedia knowledge elements (entities and their visual chemical structures, relations, and events) from scientific literature. We then exploit the constructed multimedia knowledge graphs (KGs) for question answering and report generation, using drug repurposing as a case study. Our framework also provides detailed contextual sentences, subfigures, and knowledge subgraphs as evidence.

Motivation & Objective

• Address the knowledge bottleneck and quality issues in COVID-19 literature by constructing structured, multimedia knowledge graphs from papers and figures.
• Enable effective question answering and evidence-backed drug repurposing reports.
• Provide detailed contextual sentences, subfigures, and knowledge subgraphs as evidence for retrieved results.
• Support incremental updates from major literature sources to scale knowledge discovery.

Proposed method

• Coarse-grained Information Extraction for four entity types (Gene, Disease, Chemical, Organism) with MeSH CTD linking and 133 relation types and 13 event types.
• Fine-grained entity extraction (CORD-NER) with 75 entity types, using distantly- and weakly-supervised Learning.
• Image processing with figure extraction, subfigure segmentation, OCR, and cross-media grounding to attach visual content to KG entities.
• Knowledge Graph semantic visualization via semantic tag clouds and heatmaps, built on a Kibana-based dashboard.
• Knowledge-driven QA combining KG matching and distributional semantic matching across documents with multi-hop capability.
• Evidence mining with sentence-level evidence via BioBert-based representations and EvidenceMiner for meta-symbol queries.
• Drug repurposing report generation by answering 11 typical questions and assembling evidence, subgraphs, and image analyses for each candidate drug.

Experimental results

Research questions

• RQ1 How can a multimedia knowledge graph be constructed from COVID-19 literature by integrating text and image content?
• RQ2 Can the constructed KGs support accurate, evidence-grounded question answering and drug repurposing reporting?
• RQ3 What is the effectiveness of coarse- and fine-grained entity extraction and cross-media grounding for building comprehensive knowledge graphs?
• RQ4 How well does the system scale with literature growth and provide actionable insights for drug repurposing?

Key findings

• As of June 14, 2020, the authors collected 140K papers and constructed a KG with 7,230 Diseases, 9,123 Chemicals, and 50,864 Genes, including 1,725,518 Chemical–Gene links, 5,556,670 Chemical–Disease links, and 77,844,574 Gene–Disease links.
• Coarse-grained IE achieves 83.6% F-score for node extraction and 78.1% F-score for link extraction on 186 documents with expert annotation.
• CORD-NER provides 75 fine-grained entity types with up to 93.95% F-score, surpassing SciSpacy in a sample set.
• The visual IE subsystem segments nearly half a million subfigures and grounds visual content to KG entities, enriching cross-modal knowledge.
• The QA system supports multi-hop queries, presents salient paths as subgraphs, and provides evidence sentences and sources.
• Drug repurposing reports for multiple drugs (e.g., Benazepril, Losartan, Amodiaquine) are generated with evidence and knowledge subgraphs; the authors report informative and valid outputs after expert review.