Protein-protein interactions (PPIs) are essential to various biological processes, including cell signaling and metabolic regulation, making their accurate prediction vital for advancing drug discovery and therapeutic development. The exploration of PPIs has led to the development of several AI-based models that leverage recent advancements in artificial intelligence, primarily focusing on features extracted from diverse sources of protein data. In our study, we focus on PPIs by calculating the amino acid composition of the relevant proteins. We construct a PPI network represented as a graph, where each node signifies a protein pair, and an edge connects nodes if they share a common protein. Each node is linked to a feature vector encapsulating significant information about the proteins. We then employ the GraphFormer model to extract feature embeddings directly from the protein sequences, capturing intricate patterns within the data. To enhance our predictions, we implement a fusion technique that combines the amino acid composition features with the embeddings generated by the GraphFormer model using a Feature Attention Network. This network assigns weights to the features, allowing us to emphasize the most critical information from both the amino acid composition and embedding features. By integrating these two types of data, we leverage both sequence-level information, which describes the biochemical properties of proteins, and structural information from the graph embeddings, which illustrates how proteins interact. Finally, we evaluate our model's performance using key metrics such as accuracy, precision, recall, and F1 score. Our approach shows improved results compared to existing methods, effectively merging both sequence-level and structural information essential for understanding protein interactions.