Variable screening plays a crucial role in ultra-high-dimensional data analysis. In this article, we establish a sure independence screening procedure based on the multiscale graph correlation (MGC), a new frame to generalize distance correlation, which can identify the monotonous or non monotonic relationship between predictors and response, and also enjoy the same sure screening property as the DC-SIS. Besides, we extend the method to right-censored survival data in two ways, the Kaplan-Meier estimator and composite quantile, respectively, and build the corresponding sure screening properties. Through numerical simulation, the results show that MGC-based screening methods have better performance than other methods when complicated non linear relationships exist for both complete data and right-censored data. Furthermore, we apply the proposed methods to two real datasets to examine the ranking ability and model prediction accuracy.