SciLinker: a large-scale text mining framework for mapping associations among biological entities.

Area of Science:

• Biomedical Informatics
• Natural Language Processing
• Computational Biology

Background:

• The rapid growth of biomedical literature necessitates automated methods for extracting relationships between biological entities.
• Manual exploration of millions of abstracts is infeasible for identifying comprehensive biological associations.

Purpose of the Study:

• To develop an efficient and modular natural language processing pipeline for analyzing the entire PubMed abstract corpus.
• To extract and quantify associations between biomedical entities, including genes, diseases, cell types, and drugs.

Main Methods:

• Utilized open-source libraries and pre-trained named entity recognition models to identify and normalize biological entities to the Unified Medical Language System (UMLS).
• Implemented a scoring schema for statistical significance of entity co-occurrences.
• Applied a fine-tuned PubMedBERT model for gene-disease relationship extraction.

Main Results:

• Analyzed over 30 million association sentences, identifying more than 1.25 million unique gene-disease associations.
• Demonstrated SciLinker's utility in extracting specific gene-disease relationships, using osteoporosis as a case study.
• Showcased how co-occurrence data can construct disease-specific networks for biological insights and target identification.

Conclusions:

• SciLinker is a novel text mining approach for extracting and quantifying biomedical entity associations from PubMed abstracts.
• Its modular design allows for expansion to new entities and corpora, supporting drug discovery.
• The tool transforms unstructured biomedical data into actionable insights for researchers.