A Multi-View Deep Neural Network Model for Chemical-Disease Relation Extraction From Imbalanced Datasets.

Area of Science:

• Biomedical Informatics
• Computational Biology
• Artificial Intelligence in Healthcare

Background:

• Manual extraction of chemical-disease relations (CDR) from biomedical literature is labor-intensive and time-consuming.
• Developing automated tools for CDR identification is crucial for advancing biomedical research and drug discovery.
• Existing methods often struggle with imbalanced datasets, limiting their effectiveness.

Purpose of the Study:

• To propose a novel multi-view deep neural network model for automated chemical-disease relation extraction.
• To address the challenge of imbalanced datasets in CDR tasks using a new loss function.
• To evaluate the model's performance against established benchmarks and state-of-the-art techniques.

Main Methods:

• A multi-view deep learning approach was developed by training conceptually different neural networks to generate diverse feature representations.
• The model integrates Convolutional Neural Networks (CNN), Bidirectional Long Short-Term Memory (Bi-LSTM), and Multi-Layer Perceptrons (MLP).
• A novel generic loss function, "Penalized LF", was introduced to mitigate issues arising from imbalanced datasets.

Main Results:

• The proposed model achieved the highest F1-score across individual classes on the "chemicals-and-disease-DFE" dataset, demonstrating superior performance in handling class imbalance.
• Validation on the BioCreative V dataset and two Protein-Protein Interaction Identification (PPI) datasets (AiMed, BioInfer) showed competitive or superior results compared to state-of-the-art models.
• The multi-view strategy effectively generated distinct abstract features, enhancing the model's ability to capture complex relationships.

Conclusions:

• The developed multi-view deep neural network model offers an efficient and accurate solution for automated chemical-disease relation extraction.
• The "Penalized LF" loss function is effective in addressing class imbalance, a common challenge in biomedical text mining.
• The model's versatility is confirmed by its strong performance on diverse biomedical datasets, highlighting its potential for real-world applications.