Knowledge mining of brain connectivity in massive literature based on transfer learning

Area of Science:

• Neuroscience
• Computational Biology
• Bioinformatics

Background:

• Mapping intricate brain networks is crucial for understanding neurological function.
• Existing methods struggle with diverse representations of brain regions and connectivity data.
• Text mining advancements offer potential for whole-brain connectivity analysis.

Purpose of the Study:

• To develop an advanced model for extracting comprehensive brain connectivity information from biomedical literature.
• To improve the accuracy of named entity recognition and relation extraction for brain regions and their connections.
• To facilitate large-scale analysis of neural circuits and enhance understanding of brain connectivity.

Main Methods:

• Proposed BioSEPBERT, a biomedical pre-trained model utilizing BERT and start-end position pointers.
• Integrated specialized identifiers and enhanced self-attention for brain region context.
• Applied the model to a large corpus of scientific abstracts and full-text articles.

Main Results:

• Achieved optimal F1 scores of 85.0% (NER), 86.6% (relation extraction), and 86.5% (directional relation extraction).
• Surpassed state-of-the-art models in accuracy for these tasks.
• Extracted 22.6 million standardized brain regions and 165,072 directional relations.

Conclusions:

• BioSEPBERT significantly improves the extraction of brain connectivity data from text.
• The model enables rapid knowledge acquisition regarding neural circuits.
• Facilitates enhanced comprehension of brain connectivity across diverse brain regions.