CoVEffect: interactive system for mining the effects of SARS-CoV-2 mutations and variants based on deep learning

Area of Science:

• Bioinformatics
• Genomics
• Virology

Background:

• SARS-CoV-2 variants' effects are documented across numerous research articles.
• Information on variant effects is fragmented, limiting integration with large datasets like viral sequences.
• A need exists to systematically extract and structure data on variant impacts.

Purpose of the Study:

• To develop a framework for extracting variant/mutation effects from scientific literature.
• To categorize effects into epidemiological, immunological, clinical, or viral kinetics terms.
• To label these effects as higher or lower compared to the nonmutated virus.

Main Methods:

• Utilized the COVID-19 Research Dataset (CORD-19) abstracts.
• Employed a GPT2-based prediction model for identifying mutation/variant effects.
• Developed the CoVEffect web application for assisted, semiautomated data labeling.

Main Results:

• Successfully predicted mutations/variants, their effects, and levels.
• Enabled batch annotation of CORD-19 abstracts and on-demand annotation via the CoVEffect app.
• The CoVEffect interface allows user correction to refine the training dataset.

Conclusions:

• The CoVEffect framework facilitates the extraction of structured data on SARS-CoV-2 variant effects.
• Curated datasets can be downloaded for integration and analysis pipelines.
• The approach is adaptable for similar unstructured-to-structured text translation tasks in biomedicine.