Source-linked AI summary
COVID-19 Literature Knowledge Graph Construction and Drug Repurposing Report Generation
Qingyun Wang, Manling Li, Xuan Wang, Nikolaus Parulian, Guangxing Han, Jiawei Ma, Jingxuan Tu, Ying Lin, Haoran Zhang, Weili Liu, Aabhas Chauhan, Yingjun Guan, Bangzheng Li, Ruisong Li, Xiangchen Song, Yi R. Fung, Heng Ji, Jiawei Han, Shih-Fu Chang, James Pustejovsky, Jasmine Rah, David Liem, Ahmed Elsayed, Martha Palmer, Clare Voss, Cynthia Schneider, Boyan Onyshkevych
TL;DR
COVID-19 researchers face a rapidly growing body of literature with dispersed and sometimes conflicting findings, complicating access to reliable evidence and insight into drug mechanisms. COVID-KG builds multimedia knowledge graphs from scientific literature and uses them to answer drug-repurposing questions with evidence. Reviewers found most examined outputs informative and valid, while the authors note that benchmark data bias may affect generalizability and fairness.
Problem
Researchers need to connect and assess dispersed, sometimes conflicting COVID-19 findings to identify reliable evidence and understand drug mechanisms.
Source: §1 · pp. 1–2Method
COVID-KG extracts entities, relations, and events from literature text and images into multimedia knowledge graphs, then answers drug-repurposing questions with evidence.
Source: §1 · p. 2, §3 · p. 5, §4.1 · p. 5Results
Clinicians and medical students reported that most reviewed drug-repurposing outputs were informative and valid, and the framework provides answers with evidence and knowledge subgraphs.
Source: §4.2 · p. 6Takeaways & Limitations
The authors say COVID-KG can help researchers and clinicians focus on hypothesis testing and prioritize analysis of candidate exploration directions.
Source: §6 · p. 7Reported component performance relies on specific benchmark datasets that may be affected by data bias, so generalizability and fairness require careful consideration.
Source: Limitations of System Performance and… · p. 8
Abstract
from arXiv · showhide
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.
1 Introduction
COVID-19 research is growing rapidly while findings are dispersed, redundant, complementary, or conflicting, making relevant knowledge difficult to access and assess. COVID-KG proposes to organize literature into multimedia knowledge graphs and use them to answer drug-repurposing questions with evidence.
1 Introduction
- Rapid literature growth and dispersed, sometimes conflicting findings make it difficult for researchers to access and connect reliable COVID-19 knowledge.The introduction identifies both quantity and quality as barriers to digesting the research literature.Source: §1 · p. 1
- Drug repurposing is a pre-clinical search for new uses of approved drugs, but its complexity and the volume of candidates and reports make discovery difficult.Clinicians and scientists need reliable ranked drug lists with evidence and insight into molecular mechanisms and conditions that may affect COVID-19 severity and mortality.Source: §1 · pp. 1–2
- COVID-KG extracts entities, relations, and events from text and images into multimedia knowledge graphs, then uses them to answer natural-language questions.The framework enriches the graphs with path ranking and evidence mining.Source: §1 · p. 2
- The drug-repurposing case study uses 11 expert questions and generates a comprehensive report for each candidate drug.The stated goal is to help researchers and clinicians connect research discovery with clinical testing.Source: §1 · p. 2
2 Multimedia Knowledge Graph Construction
COVID-KG constructs biomedical knowledge graphs through coarse- and fine-grained text extraction, figure and subfigure processing, and semantic visualization. Its methods connect textual entities with visual structures and make relation patterns explorable.
2.1 Coarse-grained Text Knowledge Extraction
- Experiments on 186 documents containing 12,916 sentences manually annotated by domain experts show the method achieved 83.6% F-score for node extraction and 78.1% for link extraction.It also extracts 133 relation subtypes and 13 event types.Source: §2.1 · p. 3
2.2 Fine-grained Text Entity Extraction
- CORD-NER enriches the KG with 75 fine-grained entity types; its annotation quality surpasses SciSpacy, reaching up to 93.95% F-score and over 10% higher F1 on a sample set.The approach uses distant and weak supervision without expensive human annotation.Source: §2.2 · p. 3
2.3 Image Processing and Cross-media Entity Grounding
- The visual IE pipeline extracts figures from CORD-19 papers and segments them into nearly half a million isolated subfigures.It detects subfigure regions, recognizes text within figures, and uses marker locations to associate subfigures with captions and referring context.Source: §2.3 · p. 3
- A classifier identifies subfigures containing molecular structures, and drug names found in figure text link those structures to drug entities in the caption-built KG.This produces an expanded cross-modal KG connecting images of specific molecular structures with drug entities.Source: §2.3 · p. 4
2.4 Knowledge Graph Semantic Visualization
- Semantic visualization provides interactive views of selected relation subtypes, while parameter-reduced dense heatmaps represent multiple chained relations with their types.Figure 7 shows relations between proteins and implicated diseases, including proteins that down-regulate TNF implicated in obesity.Source: §2.4 · p. 4, Figure 7 · p. 4
- Interactive Kibana dashboards let users click or search visualizations; selecting a protein constrains the protein type in modifications and updates the other visualizations.The dashboards are web applications built from indices of KG knowledge elements.Source: §2.4 · p. 4
3 Knowledge-driven Question Answering
COVID-KG answers questions across documents by combining knowledge-graph matching with distributional semantic matching. It supports entity-connection queries and entity-type or pattern queries, with related textual evidence.
3 Knowledge-driven Question Answering
- COVID-KG combines knowledge-graph matching and distributional semantic matching across documents, and supports multi-hop queries over KGs and related texts.This differs from question-answering methods targeting single documents.Source: §3 · p. 5
- For queries connecting two entities, the system generates salient KG paths ranked by the number of papers covering their knowledge elements and supplies related sentences and source information.BioBERT sentence and query embeddings are compared to identify relevant evidence sentences.Source: §3 · p. 5
- For entity-type or pattern queries, EVIDENCEMINER retrieves textual evidence from a COVID-19 background corpus using natural-language statements or meta-symbol relationships.Examples of meta-symbols include CHEMICAL and PROTEIN.Source: §3 · p. 5
4 A case study on Drug Repurposing Report Generation
The case study frames drug-repurposing reports around 11 questions and examines three queried drugs. It reports construction of a large biomedical knowledge graph and review of selected outputs.
4.1 Task and Data
- Drug-repurposing reports address 11 questions, spanning indication, structure, mechanism, screening, preclinical and clinical data, toxicity, funding, and sources.Answers to questions about who is studying a drug and relevant sources come from paper metadata; the others use the KG and knowledge-driven QA.Source: §4.1 · p. 5
- The case study queries Benazepril, Losartan, and Amodiaquine for links to COVID-19-related chemicals and genes.Source: §4.1 · p. 5, Figure 8 · p. 5
4.2 Results
- From 25,534 peer-reviewed papers, COVID-KG built a KG with 7,230 diseases, 9,123 chemicals, and 50,864 genes.The KG also contains 1,725,518 chemical-gene links, 5,556,670 chemical-disease links, and 77,844,574 gene-disease links.Source: §4.2 · p. 6
- Clinicians and medical students reviewed reports for three drugs and KGs linking 41 drugs to COVID-19-related chemicals and genes; they found most output informative and valid.Reviewers also identified results worth further investigation, including a Losartan–p53 connection related to lung cancer.Source: §4.2 · p. 6
5 Related Work
Prior work extracted biomedical entities, relations, and events, while recent work also addressed COVID-19 literature and biomedical question answering. COVID-KG extends this landscape with fine-grained, cross-media knowledge extraction and KG matching for QA.
- Biomedical literature extraction research has addressed entities, relations, and events, with more recent work focused on COVID-19 literature.Source: §5 · pp. 6–7
- Biomedical QA systems and search engines were already available, including systems driven by the BioASQ initiative.Source: §5 · p. 7
- COVID-KG extends prior multimedia extraction by adding finer-grained knowledge types, image analysis, cross-media grounding, and KG matching for QA.Source: §5 · p. 7
6 Conclusions and Future Work
COVID-KG turns scientific literature into structured, actionable knowledge graphs and uses them for drug-repurposing question answering. The authors describe its intended research use and outline ontology and inference extensions.
- COVID-KG transforms massive scientific literature into structured, actionable KGs and uses them to answer drug-repurposing questions.Source: §6 · p. 7
- The framework is intended to help researchers and clinicians focus on hypothesis testing and prioritize analysis of candidate exploration directions.Source: §6 · p. 7
- Ongoing work creates an ontology with 77 entity subtypes and 58 event subtypes and builds a neural information-extraction system following it.Source: §6 · p. 7
- Future work plans to predict new links to automate hypothesis generation and create a multimedia semantic space to improve cross-media grounding and inference.Source: §6 · p. 7
Ethical Considerations
COVID-KG is intended to guide pre-clinical research, not direct clinical use, and its leads require human review and source verification. Its performance and data have limitations, including extraction errors and potential bias in training data, benchmarks, sources, and user queries.
Required Workflow for Using Our System
- COVID-KG provides investigative leads for pre-clinical research; researchers must investigate candidate utility through the usual research steps before clinical studies.The tool is not intended for direct clinical applications on human subjects.Source: Required Workflow for Using Our System · p. 7
- Users should verify source information and discovered leads before launching expensive experimental studies.The system provides source information and evidence sentences for KG nodes and links.Source: Required Workflow for Using Our System · p. 7
- Planned investigations involving human subjects require stakeholder IRB approval before experimental studies.The IRB oversees study safety and the role of COVID-KG.Source: Required Workflow for Using Our System · p. 7
Limitations of System Performance and Data Collection
- Extraction components produce about 6%-22% false alarms and misses, so domain experts validate connections and answers against their sources.The system is intended to target drugs of interest for pre-clinical research.Source: Limitations of System Performance and… · p. 8
- Weak distant supervision from external knowledge bases and ontologies can introduce training-data bias; the output is intended for human interpretation, not unvalidated automatic decisions.The input documents are peer-reviewed, publicly available scientific articles; data should be legally and ethically obtained.Source: Limitations of System Performance and… · p. 8
- Benchmark-specific evaluations may be affected by data bias, so generalizability and fairness require care; the authors recommend caution pending more complete expert evaluation.They describe expanding annotated data, adding user correction and validation, and building new benchmarks for retraining and more systematic evaluation.Source: Limitations of System Performance and… · p. 8
- System outputs may inherit biases from peer review; the authors plan to add fact-checking across sources and provide up-to-the-minute information..Source: Limitations of System Performance and… · p. 8
- User-provided candidate-drug and protein/gene lists may be biased, a limitation outside the project’s scope; users are directed to examine source details and evidence.Suggested checks include author, publication date, contextual sentences, and associated documents.Source: Limitations of System Performance and… · p. 8