Protein-protein interaction (PPI) is a biochemical event in which two or more proteins bind together to act as enzymes, transport molecules, perform metabolisms, etc. In this paper, we propose an improved graph neural network (GNN) based method for interaction prediction (IGMIP) in PPI networks. The proposed method uses three variants of GNN: graph convolutional network (GCN), message passing neural network (MPNN) and relational graph convolutional network (RGCN). The performance of the IGMIP method is evaluated on two benchmark datasets: Human and Saccharomyces cerevisiae Yeast (SCYeast) and the RGCN model outperforms both the MPNN and the GCN models, achieving an accuracy of \(98.18\%\) , a sensitivity of \(98.99\%\) , a specificity of \(95.93\%\) , an F1 score of \(98.76\%\) and a Matthews correlation coefficient (MCC) of \(95.31\%\) . Although substantial improvements are observed in these metrics, the gains in area under the precision recall curve (AUPRC) and area under the receiver operating characteristic curve (AUROC) remain comparatively modest due to the inherent ranking difficulty in the prediction of PPI and the imbalance present in the datasets.