SLdGNN: Synthetic Lethality Prediction Based on Multi-Source Data and Dual-Graph Neural Networks

Synthetic lethality prediction has become an important strategy for identifying therapeutic targets in precision oncology. However, large-scale experimental screening of SL interactions remains costly and time-consuming. Although deep learning has emerged as an effective alternative, many existing methods mainly rely on graph structural information while overlooking intrinsic biological attributes of genes. Motivated by these gaps, we introduce SLdGNN. This architecture builds upon the SLGNN foundation by incorporating protein sequence embeddings with dual-graph structural information. The pre-trained ESM 2-650 M model is utilized for the distillation of evolutionary and semantic attributes from protein sequences, which are further fused with a heterogeneous knowledge graph and an SL association graph. Graph attention mechanisms and multi-layer graph convolution operations are used to learn integrated gene representations. Experimental results show that SLdGNN achieves competitive performance across multiple evaluation metrics, demonstrating the effectiveness of combining sequence semantics with biological network information for SL prediction.