关键词: 番茄, 多组学知识图谱, 大语言模型, 实体链接, 知识表示

Abstract: To address the issues of knowledge fragmentation, inconsistent knowledge representation, and difficulties in cross-source entity alignment in tomato multi-omics research, an automated construction method for a tomato multi-source multi-omics knowledge graph based on large language models (LLMs) was proposed, providing methodological support for the organization and integration of multi-source heterogeneous knowledge in the agricultural biology domain. A knowledge graph schema layer was designed for the tomato multi-source multi-omics scenario to uniformly define core entities, such as genes, proteins, metabolites, phenotypes, and environmental conditions, along with their relationship types. Eight structured databases and tomato literature data were integrated to construct a knowledge skeleton, followed by literature data preprocessing. Under schema constraints, LLMs were employed to perform generative knowledge extraction from the literature. During the knowledge fusion stage, a local context-driven entity information completion strategy was introduced, combined with a "candidate recall-LLM-based discrimination" mechanism to achieve cross-source entity alignment, while unsuccessfully linked entities were merged through long-tail entity clustering. The resulting tomato multi-omics knowledge graph contained 81 409 entity nodes and 436 976 relationship edges, forming a knowledge network covering genes, proteins, metabolites, reactions, phenotypes, and experimental conditions. In the knowledge extraction evaluation, the F1 scores of the two categories of core relationships were both 0.84, indicating good performance of the proposed method in relation extraction from tomato multi-omics literature. In the entity linking evaluation, after introducing entity information completion, the linking precision for different entity types reached 0.92~0.98, indicating that the proposed method could improve the standardized mapping of complex entities in agricultural biology literature and provide a basis for multi-source knowledge fusion. The proposed method enabled the automated acquisition, structured representation, and cross-source integration of tomato multi-omics knowledge, providing a reference for tomato multi-omics knowledge management, agricultural knowledge graph construction, and related knowledge analysis applications.

Key words: tomato, multi-omics knowledge graph, large language model, entity linking, knowledge representation