TL;DR
This paper introduces TWIRLS, a novel knowledge-mining technology that rapidly analyzes vast amounts of biomedical literature to uncover potential mechanisms of disease. Applied to SARS-CoV-2, TWIRLS identified a key role for ACE2 in host pathological changes, suggesting a link between viral binding, Renin-Angiotensin System (RAS) imbalance, and cytokine storms, and hinting at the potential therapeutic benefit of ARBs.
What Problem Does This Paper Address?
The rapid spread of novel infectious diseases like COVID-19 presents a monumental challenge: an overwhelming deluge of scientific literature that human researchers struggle to process in a timely and effective manner. This information overload, coupled with the inherent complexity of biological systems, hinders our ability to quickly understand disease mechanisms and develop effective treatments. The authors aim to bridge this gap by developing an automated system that can efficiently extract, organize, and infer knowledge from this massive biomedical corpus, accelerating scientific discovery and decision-making.
Methods
The core of this work is the TWIRLS (Topic-wise inference engine of massive biomedical literatures) system. TWIRLS is a machine-based approach designed to automatically acquire, organize, and classify information from biomedical texts. Its key steps include:
- Data Interface and Corpus Organization: Utilizing PubMed as the primary data source, TWIRLS searches for articles related to "coronavirus," downloads the text, and organizes it into specific corpuses. A control corpus with "public health" as the keyword is also created for comparative analysis.
- Gene and Entity Identification: TWIRLS identifies "Coronavirus Study-Specific Host Genes" (CSHG) by comparing gene distributions in the coronavirus corpus against a control corpus, using odds ratios to determine significance. Similarly, "Coronavirus Study-Specific Entities" (CSSE) are identified using Z-scores to measure their distribution significance. The system also handles synonyms and plural/singular forms of words.
- Knowledge Graph Construction and Classification: TWIRLS constructs a "coronavirus" knowledge graph. Identified entities are then classified into categories (32 in this study) based on their co-distribution with CSHGs. This classification helps in understanding broader biological themes and pathways.
- Pathway and Interaction Analysis: The system integrates with functional data sources, including protein-protein interaction (PPI) databases and KEGG pathways. This allows for the inference of relationships between genes and pathways, and the identification of potential regulatory networks and signaling cascades.
- Time-Series Analysis: TWIRLS analyzes the temporal distribution of entities to understand research trends and the evolution of knowledge over time.
The system is developed using a combination of MatLab and Python, with performance optimizations in C/C++.
Datasets
- PubMed: The primary dataset used is the text from biomedical literature indexed in PubMed, specifically focusing on articles containing the keyword "coronavirus." The authors analyzed over 3 million words from more than 14,000 literature articles.
- Human Gene Nomenclature Committee (HGNC) database: Used to identify official gene symbols.
- KEGG database: Utilized for pathway enrichment analysis.
- Protein-Protein Interaction (PPI) databases: Employed to construct gene interaction networks.
- Clinical Data: Medical records of 92 patients with COVID-19 were collected from Chongqing Public Health Medical Center and Chongqing Southwest Hospital. This data included demographic information, clinical characteristics, and blood indices.
Key Insights and Ideas
- Automated Knowledge Discovery: The most significant insight is the demonstration that a machine-learning approach like TWIRLS can rapidly sift through massive amounts of scientific literature to identify critical biological players and propose mechanistic hypotheses, significantly accelerating the research process.
- ACE2 as a Central Player: TWIRLS pinpointed Angiotensin-Converting Enzyme 2 (ACE2) as a crucial factor in host pathological changes following coronavirus infection. This aligns with existing knowledge but the system's ability to rapidly highlight this from a vast literature base is key.
- RAS Imbalance and Cytokine Storms: The analysis suggests that viral binding to ACE2 leads to an imbalance in the Renin-Angiotensin System (RAS), which in turn triggers a cytokine storm, a hallmark of severe COVID-19. This provides a potential mechanistic link between viral entry and the severe inflammatory response.
- Therapeutic Potential of ARBs: The study's preliminary analysis of clinical data from COVID-19 patients with hypertension suggested that Angiotensin II Receptor Blockers (ARBs) might be associated with less severe illness. This is a significant hypothesis that warrants further investigation.
- Bridging Text Mining and Biological Networks: The integration of text-mined entities with PPI networks and pathway databases allows for a more comprehensive understanding of biological mechanisms, moving beyond simple entity extraction to inferring functional relationships.
Key Results
- Speed of Analysis: TWIRLS analyzed over three million words from more than 14,000 literature articles in just 4 hours.
- Identification of Key Genes: The system identified 123 coronavirus study-specific host genes (CSHG) and 623 coronavirus study-specific entities (CSSE).
- Categorization of Knowledge: Entities were classified into 32 categories, helping to organize and understand the vast information landscape.
- Mechanistic Hypothesis: TWIRLS proposed a mechanism involving ACE2/AT2R-RAS-Cytokine signaling leading to cytokine storms.
- Clinical Data Suggestion: Analysis of 92 COVID-19 patients indicated that patients taking ARBs had a lower risk of developing severe lung damage compared to non-ARB users.
Limitations and Future Work
- Validation of Hypotheses: The authors acknowledge that the mechanisms proposed by TWIRLS require further support from pathological evidence and experimental validation.
- Data Granularity: While TWIRLS can identify entities and their associations, understanding the precise biological functions of some identified genes (like IFITMs) requires more detailed information.
- Machine Learning vs. Human Expertise: The paper emphasizes that TWIRLS is a tool to assist human researchers, not replace them. The goal is to combine human expertise with algorithmic insights for optimal outcomes.
- Future Directions: The authors suggest that more rigorous statistical methods, further discussion with scientists, and the generation of more histopathology-related data are needed to fully validate their findings. They also highlight the potential for TWIRLS to be applied to other complex scientific problems.
Summary updated