Asking the right questions for mutagenicity prediction from BioMedical text

Area of Science:

• Computational chemistry
• Bioinformatics
• Drug discovery

Background:

• Assessing chemical mutagenicity is crucial for drug development.
• Existing mutagenicity databases require laborious manual curation and are difficult to update.
• Automating the extraction of mutagenicity data from scientific literature is needed.

Purpose of the Study:

• To propose and address the problem of predicting chemical mutagenicity directly from textual information in scientific publications.
• To develop a model that can predict mutagenicity based on natural language descriptions of chemical evidence.
• To establish the utility of large language models for constructing structured knowledge bases from unstructured scientific text.

Main Methods:

• Construction of a gold-standard dataset for chemical mutagenicity.
• Development of MutaPredBERT, a prediction model fine-tuned on BioLinkBERT.
• Formulation of the prediction task as a question-answering problem.
• Leveraging transfer learning with large transformer-based models.

Main Results:

• Achieved a Macro F1 score greater than 0.88.
• Demonstrated high prediction accuracy even with a relatively small fine-tuning dataset.
• Successfully applied a question-answering approach to mutagenicity prediction.

Conclusions:

• The MutaPredBERT model effectively predicts chemical mutagenicity from scientific literature.
• Large language models are valuable tools for automating the creation of structured knowledge bases.
• This approach offers a more practical and scalable alternative to manual database curation for mutagenicity information.