Generating Biomedical Hypothesis With Spatiotemporal Transformers.

Area of Science:

• Biomedical Informatics
• Computational Biology
• Artificial Intelligence

Background:

• Biomedical hypothesis generation requires identifying implicit associations within vast scientific literature.
• Temporal graph-based methods model term-pair relationships over time but neglect inter-term covariation and direct temporal dependencies.
• Existing approaches using Recurrent Neural Networks (RNNs) independently model temporal evolution, limiting their ability to capture complex interactions.

Purpose of the Study:

• To propose a novel Spatiotemporal Transformer-based Hypothesis Generation (STHG) method.
• To address limitations of existing methods by integrating spatial covariation and temporal progression.
• To enhance the discovery of implicit biomedical relationships from scientific literature.

Main Methods:

• Developed a Spatiotemporal Transformer-based Hypothesis Generation (STHG) framework.
• Interleaved spatial covariation and temporal progression within a unified model.
• Constructed direct connections between term-pairs and modeled temporal relevance between timesteps.

Main Results:

• The STHG method demonstrated superior performance compared to state-of-the-art approaches.
• Experiments were conducted on three distinct biomedical relationship datasets.
• STHG effectively captured both spatial covariation and temporal dynamics in term relationships.

Conclusions:

• The proposed STHG method offers a significant advancement in biomedical hypothesis generation.
• Integrating spatiotemporal modeling enhances the uncovering of implicit scientific associations.
• STHG provides a more comprehensive approach to analyzing temporal evolution of term-pair relationships.