Chemical-gene relation extraction using recursive neural network.

Area of Science:

• Computational Biology
• Bioinformatics
• Natural Language Processing

Background:

• Extracting relationships between chemical compounds and genes is crucial for pharmacological and clinical research.
• Limited research exists on automated extraction of these relationships from scientific literature.
• The CHEMPROT task in BioCreative VI challenges the development of such text mining systems.

Purpose of the Study:

• To develop and evaluate advanced text mining systems for extracting chemical-gene relationships.
• To improve the performance of relation extraction models using recursive neural networks.

Main Methods:

• Tested three recursive neural network approaches, including tree-Long Short-Term Memory (tree-LSTM) networks.
• Incorporated features like position and subtree containment into the tree-LSTM model.
• Applied an ensemble method and additional pre-processing steps.
• Evaluated a Stack-augmented Parser Interpreter Neural Network (SPINN) model.

Main Results:

• The initial tree-LSTM model achieved an F-score of 58.53% in the BioCreative VI challenge.
• The enhanced tree-LSTM model with pre-processing reached an F-score of 63.7%.
• The SPINN model achieved a competitive F-score of 64.1%.

Conclusions:

• Recursive neural network models, particularly tree-LSTM and SPINN, show significant promise for chemical-gene relation extraction.
• Advanced features and pre-processing steps can substantially improve model performance.
• These advancements contribute to automated literature analysis in drug discovery and clinical research.