SLGNNCT: Synthetic Lethality Prediction Based on Knowledge Graph for Different Cancers Types

Jingru Chen,Jianyong Pan,Yan Zhu,Junyi Li

DOI: https://doi.org/10.1007/978-981-97-5689-6_14

2024-01-01

Abstract:Exploiting synthetic lethality could help researchers to expand the arsenal of potential cancer drug targets greatly and using computational methods to predict cancer synthetic lethal genes has become a highly regarded area of research. Although numerous methods are currently dedicated to the prediction of SL gene pairs, these methods generally adopt a hybrid prediction strategy that fails to adequately account for the heterogeneity of cancers. We propose a model named SLGNNCT, which is based on knowledge graph, to predict SL interactions categorized by cancer types. The model categorizes each SL gene pair by the types of human cancers it may cause, and divides the SL database into eight subsets for prediction separately. The main body of the SLGNNCT contains three parts: first, we consider the combinations of relationships in the gene-related knowledge graph as the key factors affecting SL interactions and model them; subsequently, we generate an initial embedding representation for genes through an explicit message aggregation mechanism; finally, we construct an SL graph based on known SL gene pairs; and finally, we construct an SL graph based on known SL genes. And model the key factors; subsequently, through the explicit message aggregation mechanism in the knowledge graph, we generate initial embedding representations for the genes; finally, we construct an SL graph based on known SL gene pairs and utilize a factor-based message aggregation approach, which leads to the derivation of the final gene embedding representations. By comparing the results of our model with those of the ELISL model for single-cancer prediction, SLGNNCT performs much better on three cancer predictions and is surprisingly prominent in colon cancer (COAD) prediction.