Mining microbe-disease interactions from literature via a transfer learning model.

Area of Science:

• Biomedical Informatics
• Computational Biology
• Text Mining

Background:

• Microbe-disease interactions are vital for biomedical research but largely undocumented in literature.
• Existing structured databases for these interactions are limited.
• Automated methods are needed to extract this information efficiently.

Purpose of the Study:

• To construct a large-scale, automatically curated database of microbe-disease interactions.
• To develop and validate a deep learning-based text mining framework for this purpose.
• To provide researchers with a web-based tool for accessing and querying microbe-disease data.

Main Methods:

• Developed a text mining framework using a pretrained deep learning model (BERE).
• Utilized named entity recognition to identify microbe and disease mentions.
• Fine-tuned the BERE model with a manually curated sliver-standard corpus (SSC) for improved accuracy.
• Created a user-friendly website (MDIDB) for data navigation and searching.

Main Results:

• The deep learning model achieved an average F1-score of 73.81%, outperforming baseline methods.
• Fine-tuning with SSC reduced error rates by approximately 10%.
• Manual validation of 1000 predicted interactions showed a 73% accuracy rate.
• The MDIDB website provides browsing, custom searching, and batch downloading capabilities.

Conclusions:

• The developed text mining approach effectively extracts microbe-disease interactions.
• The MDIDB database and web interface offer a valuable resource for researchers.
• The method demonstrates superior performance compared to traditional rule-based approaches.