LSAGSL: A Layer Self-Attention-based Graph Neural Network for Synthetic Le-thality in Human Cancers
Qinglong Zhang, Bowei Yan, Mingyuan Ma, Rui Xiao · 2024
Synthetic lethality (SL) possesses the potential to specifically eradicate cancer cells without impacting normal cells, thus offering a promising approach to expand the range of targeted gene combinations in drug therapies. In recent years, an increasing number of researchers have employed computational methodologies to forecast SL gene pairs. Nevertheless, the majority of algorithms exhibit a deficiency in extracting interaction features between gene pairs and disregard the correlation between feature representations, ultimately resulting in limited generalization capability. In order to effectively capture the interaction mechanism of SL gene pairs, we focus on the correlation between different extracted features and propose a GNN algorithm named LSAGSL based on multi-layer self-attention mechanism. LSAGSL enhances the detection of hierarchical correlation within individual genes and the overall correlation between gene pairs through the incorporation of a dual-layer attention mechanism within single gene and a fused self-attention mechanism between gene pairs, so as to extract crucial information regarding the correlation between gene pairs for the purpose of SL prediction. Based on the experimental findings, LSAGSL surpasses existing SL prediction models in three distinct scenarios and offers a certain level of interpretability for the obtained results.