Dongting Lake is an important lake in the middle reach of the Yangtze River, and accurately modelling the runoff corresponding relationships of its various input and output stations is crucial for regional ecological protection and flood control and drought defense. To address the complex runoff response relationships in the Dongting Lake basin, this study proposes a multiple-input and multiple-output runoff response model based on graph neural networks. Firstly, the model utilizes the basin topological spatial structure of the Yangtze River, Dongting Lake and Sishui to transform the original observation sequences at each station into graph-structured data to characterize the spatial characteristics of the basin. Secondly, through the mutual correlation analysis method, the time lag relationship between the observed variables at each station is identified to determine the input feature step of the model. Finally, graph neural networks are employed to aggregate and update the features to capture the complex spatial and temporal dependencies among the control station, and to realize the runoff simulation at multiple stations. The results show that in the flood event, compared with the backpropagation neural network (BP) and the long-short term memory neural network (LSTM) models, the graph neural network (GNN) model can achieve the improvement rates over 5% for Nash-Sutcliffe efficiency coefficient (NSE) and mean absolute error (MAE) indicators, and the correlation coefficients (R^2) is more than 0.97, while in the dry water cutoff events, the True Positive Rate (TPR) and Precision are generally more than 0.96. GNN has significant advantages in the simulation of hydrological events such as floods and droughts, which can provide a scientific support for the ecological protection of Dongting Lake and its flood control and drought resistance.