Abstract: Heterogeneous graphs provide a powerful framework for modeling and analyzing complex real-world problems. These graphs, by representing different types of nodes and their relationships, enable the integration and interpretation of diverse data. Learning effective representations of the components of heterogeneous graphs is one of the fundamental challenges in the development of machine learning and deep learning algorithms, since these representations play a key role in improving prediction accuracy and discovering hidden patterns. In the field of bioinformatics, traditional methods and homogeneous graphs do not perform well due to their limitations in representing the diversity of biological data. In contrast, heterogeneous graphs, by leveraging complex structural information, are capable of more effectively modeling biological relationships. In this paper, we present a novel approach, BioGraph2vec, for learning representations of biomedical entities using heterogeneous graphs. This method combines protein sequence data and biological interaction information to create graphs that include host and pathogen proteins and the relationships between them. To extract precise node representations, attention mechanisms and message passing techniques are utilized to identify the important features of each node within the network. The obtained representations are then provided to machine learning models to predict potential interactions between proteins. Evaluation of this method on diverse datasets shows that integrating biological data in the form of heterogeneous graphs, together with advanced representation learning techniques, can lead to improved bioinformatics data analysis and better identification of complex patterns in biological systems.