Automating literature screening and curation with applications to computational neuroscience

Area of Science:

• Computational Neuroscience
• Artificial Intelligence in Science
• Bioinformatics

Background:

• ModelDB is a key platform for computational neuroscience models, but current submission methods limit its scope.
• An estimated two-thirds of NEURON models are not yet included in ModelDB.
• A comprehensive understanding of computational neuroscience research requires better identification of relevant works and their metadata.

Purpose of the Study:

• To develop and evaluate methods for more complete identification and metadata extraction of computational neuroscience research.
• To enhance the ModelDB platform's discoverability and standardization capabilities.

Main Methods:

• Utilized known ModelDB entries and PubMed-queried neuroscience literature.
• Employed SPECTER2 for document pre-screening.
• Applied GPT-3.5 and GPT-4 for identifying computational neuroscience work and extracting metadata.

Main Results:

• GPT-4 achieved 96.9% accuracy in identifying computational neuroscience work.
• Instruction-tuned GPT-3.5 improved identification accuracy from 54.2% to 85.5%.
• GPT-4 demonstrated high potential for extracting relevant metadata annotations.

Conclusions:

• Natural language processing and large language models can enhance ModelDB's model discovery.
• Further improvements can be achieved by refining prompts and incorporating more paper details.
• These AI techniques contribute to a more standardized and comprehensive framework for domain-specific resources.