Unifying temporal knowledge Graph and LLMs for Biomedical Hypothesis generation via temporal soft Prompting.

OBJECTIVE: Biomedical Hypothesis Generation (HG) is a critical task for discovering latent yet scientifically meaningful associations from continuously evolving literature or knowledge bases, aiding in disease diagnosis and treatment decisions-making. Most existing studies model Temporal Knowledge Graphs (TKGs) for capturing dynamic relation evolution based on temporal explicit and structured knowledge. However, TKG based methods are hard to generate new facts and represent unseen knowledge since the inherent incompleteness and limited scale of TKG. While, Large Language Models (LLMs) are pretrained by enriching large-amount of unlabeled corpus with semantic knowledge of the real world. Therefore, unifying LLMs and TKG together could leverage precise temporal facts and rich semantic knowledge simultaneously. METHODS: This paper proposes a novel HG framework which synergizes TKGs with LLMs via Temporal Soft Prompts, named TLHG. Firstly, a Spatial Graph Encoder (SGE) is used to capture the spatial structure of temporal subgraphs at the current timestep for generating temporal soft prompts. Then these prompts are integrated with query-related historical facts by an LLM-based Semantic Encoder (LSE) for facilitating a deep interaction between structural information and rich semantic knowledge. Finally, a Temporal Transformer Encoder (TTE) is employed to model long-term dependencies from the historical entity representation sequence to obtain accurate temporal representations for generating novel hypotheses. RESULTS: Extensive experiments on multiple biomedical benchmarks demonstrate that our approach significantly outperforms state-of-the-art baselines. CONCLUSION: The TLHG method addresses Biomedical HG by synergizing TKGs with LLMs through Temporal Soft Prompts. This facilitates a deep interaction between structural information and rich semantic knowledge while modeling long-term dependencies to yield accurate temporal representations for discovering meaningful implicit associations.