Biological Knowledge Graph-Enhanced Cancer State Prediction Network with Adjustable Connections

With recent advances in oncology research, effectively utilizing patient genomic data to predict cancer status has become a significant challenge. Although previous deep learning methods based on biological knowledge have made progress, they still face limitations, such as the restricted scope and imprecision of the biological knowledge covered by the biological functional datasets used, as well as difficulties in adapting to data variations. To address these challenges, this paper proposes a novel deep learning neural network model based on a biological knowledge graph (namely BKGNet) and an adjustable connection mechanism—biological knowledge graph-enhanced cancer state prediction network with adjustable connections. Specifically, we construct a predictive neural network that leverages the richer and more precise structured biological information from the knowledge graph, providing the model with more comprehensive and accurate biological background support, thereby enhancing its ability to model complex biological relationships. We also introduce an adjustable connection mechanism that, while ensuring the rationality of biological relationships, transforms fixed biological connections into learnable connection strengths. This allows the model to flexibly adjust the interactions between biological entities. Experimental results demonstrate that BKGNet outperforms traditional machine learning and deep learning baseline models in terms of prediction accuracy, highlighting the advantages of its network architecture. Further ablation experiments validate the effectiveness of both the knowledge graph and the adjustable connection mechanism.