BioSift: A Dataset for Filtering Biomedical Abstracts for Drug Repurposing and Clinical Meta-Analysis.

Area of Science:

• Biomedical Informatics
• Computational Biology
• Drug Discovery

Background:

• Drug repurposing accelerates the identification of new therapeutic uses for existing drugs.
• Efficiently screening and labeling scientific literature is a bottleneck in drug repurposing.
• The Patient-Intervention-Comparator-Outcome (PICO) framework is essential for meta-analysis.

Purpose of the Study:

• Introduce BioSift, a novel, human-annotated dataset for document classification.
• Facilitate the initial selection and labeling of studies for drug repurposing research.
• Provide a benchmark for evaluating automated document classification methods.

Main Methods:

• Curated 10,000 PubMed abstracts with human annotations.
• Annotated abstracts with up to eight PICO-related attributes.
• Ensured data quality through multiple annotators and senior review.

Main Results:

• Demonstrated that neither PubMed advanced filters nor current state-of-the-art classification models can replace human annotation.
• Established robust benchmark results on the BioSift dataset.
• Provided data statistics including reviewer agreement and label co-occurrence.

Conclusions:

• BioSift presents a pivotal yet challenging document classification task for expediting drug repurposing.
• Human annotation is currently indispensable for accurate study selection in drug repurposing.
• The publicly available BioSift dataset will drive research in document classification algorithms for drug discovery.