Towards self-learning based hypotheses generation in biomedical text domain.

Area of Science:

• Biomedical Informatics
• Computational Biology
• Text Mining

Background:

• The vast volume of biomedical literature hinders the identification of crucial connections.
• Existing Literature Based Discovery (LBD) methods face scalability challenges due to complex biological interconnections and lack of edge-formation insights.

Purpose of the Study:

• To develop a scalable and efficient framework for Literature Based Discovery (LBD).
• To address the limitations of current LBD methodologies by incorporating implicit edge-formation processes.
• To enhance the discovery of novel biomedical hypotheses from large-scale literature.

Main Methods:

• Formulating the LBD problem as a collaborative filtering task.
• Leveraging word-vectors to model the implicit edge-formation process between biomedical concepts.
• Employing a single-class classifier to prune redundant and irrelevant hypotheses, optimizing search space.

Main Results:

• The proposed framework prunes up to 90% of hypotheses while maintaining high recall.
• Achieved significant efficiency gains, enabling the exploration of higher-order hypotheses.
• Validated concordance between system-generated data structures and expert opinions, ensuring interpretability.

Conclusions:

• The novel approach significantly improves the efficiency and scalability of biomedical hypothesis generation.
• The method provides interpretable results by understanding the edge-formation process without manual intervention.
• The framework supports both open and closed discovery, demonstrating its versatility.