SAGS-DynamicBio: Integrating Semantic-Aware and Graph Structure-Aware Embedding for Dynamic Biological Data with Knowledge Graphs

Abstract

Accurate prediction of drug-target interactions (DTIs) is critical for drug design and optimization in pharmacology. Existing models face challenges such as data sparsity and lack of contextual information, resulting in poor accuracy. Knowledge graphs (KGs) provide a solution by representing relationships in biological data. However, current KG-based DTI methods are limited to static graphs that require time-consuming retraining when knowledge is updated. In this paper, we propose SAGS-DynamicBio, an efficient dynamic embedding model for biological data that integrates semantics and graph structure information. We first generate KGs for the biological knowledge base, representing drugs and targets as entities and interactions as relations. Using KG embedding techniques, we convert each entity and relation into a vector representation. To effectively handle dynamic data, we introduce a semantic perception module based on the attention mechanism, which uses information from neighboring nodes to generate initial representation vectors for new data. Furthermore, we apply graph structure-based representation learning to these initial vectors to satisfy KG’s structural constraints and improve prediction accuracy. To evaluate the effectiveness of our method, we conduct experiments comparing SAGS-DynamicBio with existing KG-based DTI prediction models and generic KGE models. The experimental results show that our method significantly improves the embedding efficiency, reducing the embedding time by 41.5% on average, while maintaining a high prediction accuracy, which proves the effectiveness of our method. SAGS-DynamicBio is able to efficiently adapt to the dynamic data updates without retraining the whole graph, thus providing a promising solution for DTI prediction in real-time scenarios.