Building deep learning models for evidence classification from the open access biomedical literature

Area of Science:

• Biomedical informatics
• Computational biology
• Natural language processing

Background:

• Biocuration involves classifying text from biomedical literature to extract evidence.
• Automating these tasks is crucial for managing the growing volume of scientific publications.
• Deep learning offers potential for improving the efficiency and accuracy of biocuration.

Purpose of the Study:

• To investigate the application of deep learning for text classification in biocuration.
• To develop and evaluate deep learning models for classifying molecular interaction papers.
• To encourage the adoption of deep learning methods in biocuration systems.

Main Methods:

• Developed a large-scale corpus of molecular papers from PubMed and PubMed Central.
• Trained deep learning word embeddings using GloVe, FastText, and ELMo algorithms.
• Applied models to distant supervised classification tasks and developed document triage methods using attention mechanisms with convolutional neural network and bi-directional long short-term memory architectures.

Main Results:

• Achieved 0.82 accuracy in document classification (triage) for molecular interaction papers.
• Demonstrated the effectiveness of deep learning, particularly attention mechanisms, in aggregating classification decisions.
• Successfully repurposed large-scale word embeddings for specialized biocuration tasks.

Conclusions:

• Deep learning, especially combined architectures with attention, significantly enhances biocuration efficiency.
• The developed methods provide a robust approach for classifying and triaging biomedical literature.
• Repurposing pre-trained word embeddings is a viable strategy for advancing automated biocuration.