Exploring a deep learning neural architecture for closed Literature-based discovery.

Area of Science:

• Computational Biology
• Bioinformatics
• Artificial Intelligence

Background:

• The exponential growth of scientific literature necessitates advanced methods for knowledge discovery.
• Interdisciplinary research is hindered by the increasing specialization and separation of publications.
• Literature-based discovery (LBD) aims to bridge knowledge gaps by connecting disparate scientific information.

Purpose of the Study:

• To introduce and evaluate a deep learning neural network-based approach for literature-based discovery.
• To investigate the impact of different term-to-concept representations and feature scaling on LBD performance.
• To analyze methods for representing model output and their effect on evaluation and generalizability.

Main Methods:

• Developed a deep learning neural network model for LBD.
• Explored various term representation techniques, including feature scaling.
• Investigated two distinct approaches for representing model output (full concept set vs. subset).
• Evaluated the model on five hallmarks of cancer datasets using closed discovery.

Main Results:

• The choice of input representation significantly impacts LBD evaluation performance.
• Feature scaling of input representations enhanced performance and reduced training epochs for model generalization.
• Reducing model output to a subset of concepts improved evaluation but decreased generalizability.
• The proposed method demonstrated suitability for LBD compared to random concept relations.

Conclusions:

• Deep learning models offer a promising avenue for advancing literature-based discovery.
• Optimizing input representation and feature scaling are crucial for effective LBD.
• Balancing evaluation performance and model generalizability requires careful consideration of output representation.