CoVEffect

CoVEffect extracts and annotates effects of SARS-CoV-2 mutations and variants from scientific literature using deep learning to support structured, literature-derived effect data for downstream analysis.

Key Features:

• GPT-2-based prediction model: Uses a GPT-2-based model to identify and annotate effects associated with mutations and variants.
• Text mining on CORD-19: Applies text mining techniques to COVID-19 literature sourced from the CORD-19 corpus, focusing on abstracts.
• Effect categorization: Classifies effects into epidemiological, immunological, clinical, and viral kinetics impacts and annotates them as higher or lower relative to the nonmutated (wild-type) virus.
• Annotation modes: Supports batch annotation of curated CORD-19 abstracts and on-demand annotation of selected abstracts.
• Semiautomated data labeling: Enables semiautomated labeling with expert review and corrections that feed back into the training dataset.
• Curated training strategy: Prototype trained on a minimal but highly diversified curated sample pool to improve predictive robustness.
• Dataset export: Produces curated annotated datasets available for download and integration into analysis pipelines.
• Framework adaptability: Adaptable to other unstructured-to-structured text translation tasks in biomedical domains.

Scientific Applications:

• Mutation/variant effect extraction: Extracts structured information on the effects of SARS-CoV-2 mutations and variants from literature.
• Epidemiological and immunological analysis: Provides literature-derived annotations usable for epidemiological and immunological impact assessments.
• Clinical and viral kinetics assessment: Supplies clinical and viral kinetics effect annotations for downstream research and modeling.
• Integration with sequence datasets: Facilitates integration of literature annotations with large SARS-CoV-2 sequence datasets (including millions of sequences) for combined analyses.
• Curated dataset generation: Generates labeled datasets for downstream bioinformatics workflows and model training.

Methodology:

Performs text mining on CORD-19 abstracts and applies a GPT-2-based prediction model to identify and classify mutation/variant effects into epidemiological, immunological, clinical, and viral kinetics categories, annotating effects as higher or lower relative to the nonmutated (wild-type) virus; training used a curated, diversified sample pool and expert corrections are incorporated into the training dataset.