Abstract
In cancer therapeutics, the elucidation of synthetic lethality principles introduces transformative concepts for devising novel treatment paradigms. Computational methods to predict synthetic lethal (SL) gene pairs have potential to markedly enhance the precision and efficacy of cancer interventions. Despite the array of predictive methodologies proposed in extant research, many overlook pivotal attributes such as protein sequences, three-dimensional configurations, and protein-protein interaction (PPI) networks. This investigation introduces Struct2SL, a predictive framework for SL gene pairs that integrates protein sequences, PPI networks, and three-dimensional protein structures. By initiating at the protein feature stratum, Struct2SL offers a novel vantage point to refine the feature representation of gene interactions, thereby enabling more accurate predictions of prospective SL pairs. Struct2SL encompasses four distinct phases: Initially, protein three-dimensional structures, sequence characteristics, and interaction network attributes are extracted utilizing approaches such as Alphafold2 for predicting protein tertiary structures. Subsequently, the preliminary embedding of genes is derived by consolidating information via the protein-gene mapping relationships. Thereafter, an SL graph is constructed to attain the ultimate gene embedding. Ultimately, a multilayer perceptron is employed for the prediction of SL interactions. The outcomes indicate that Struct2SL outperforms four SOTA methods, as gauged by the evaluation metrics. This implies that Struct2SL is more efficacious in predicting SL gene pairs. This study furnishes a new and efficient computational approach for the prediction of SL gene pairs in cancer therapy, anticipated to catalyze advancements in the field of oncological treatment. We also developed a webserver (Synthetic Lethality Query Server, http://struct2sl.bioinformatics-lilab.cn) to present cancer synthetic lethal genetic interactions, which is designed to provide researchers with an accessible tool for predicting synthetic lethality gene pairs.