Modeling temporal self and interactive evolution for biomedical hypothesis generation.

Area of Science:

• Biomedical informatics
• Computational biology
• Data science

Background:

• Hypothesis generation (HG) accelerates innovation in drug discovery and disease treatment by uncovering hidden relationships in scientific literature.
• Existing HG methods capture term-pair evolution but struggle with intricate spatio-temporal dynamics.
• Accurate modeling of dynamic term-pair relations is crucial for advancing biomedical research.

Purpose of the Study:

• To propose a novel Temporal Self and Interactive Evolution (TSIE) method for accurately characterizing complex dynamics of term-pair relations in HG.
• To enhance the understanding of temporal relation inference by learning temporal interactive difference features.
• To improve the prediction of future term-pair connectivity in biomedical literature.

Main Methods:

• The TSIE method employs Gated Recurrent Unit (GRU) to model Temporal Self-evolution (TSE) and Temporal Interactive Evolution (TIE) for each term pair.
• TSE Embeddings (TSE_emb) and TIE Embeddings (TIE_emb) are generated to capture distinct evolutionary aspects.
• A dual-tower Transformer architecture models temporal dependencies of embeddings, integrated via a gated fusion layer for final prediction.

Main Results:

• Experiments on Immunotherapy, Virology, and Neurology datasets demonstrate TSIE's effectiveness.
• TSIE successfully captures complex evolutionary patterns in biomedical hypothesis generation.
• The proposed method achieves state-of-the-art performance in HG tasks.

Conclusions:

• The TSIE method offers a novel approach to learning temporal interactive features for enhanced HG.
• By modeling both TSE and TIE, TSIE effectively captures the dynamic relationships between scientific terms.
• The dual-tower Transformer architecture further refines the modeling of temporal dependencies, improving relation inference.