Predicting the multiple plausible future trajectories of the surroundings vehicles in the complex traffic environments is crucial for the roll out of self-driving cars. It is still challenging because of the social interaction with the other vehicles and the multimodal characteristic of future. Previous motion prediction work has employed various methods, including pre-defined maneuver, generative model or multiple regression. However, these methods has not been successful to jointly model social interaction and multimodal characteristic. In this work, we presents a graph convolutional network based multimodal vehicle trajectory prediction network (MGCN). We utilize spatial-temporal graph to extract social interaction feature while designing a variational autoencoder (VAE) for endpoint to generate multimodal prediction. We compare our MGCN against many baselines on the benchmark highD and inD. The experiment results demonstrates that our method achieves the state-of-the-art performance and significantly improves both the variousness and precision of predicted trajectories.