Proteins perform an innumerable number of functions within living organisms, and the understanding of their interactions is a primary challenge for medicine and biology. The most natural way to represent their structure and their complex interactions are graphs, a powerful data structure to model objects (vertex) and their relations (edges). At the same time, Machine Learning is a powerful tool to extract knowledge and perform complex tasks on proteins. But, to perform machine learning on graphs its first essential to transform the information expressed as a graph into a structure that can be easily exploited by a Machine Learning model. Traditional approaches for learning representations relies on hand-crafted specialized heuristics to extract meaningful information about the entities of a graph, but hand-engineering features can be expensive and time-consuming. In recent years, vertex embedding methods have proven their potential in automatically representing graphs information as feature input for machine learning models. Existing methods can be divided into different groups, such as random-walk based algorithms or graph neural networks. These methods also differ as some approaches take into account just topological information and others can leverage additional features contained in the graph. Within this work, we show how graph embeddings can excel in the real-world task of protein role classification. We also prove how it is possible to combine embeddings from unsupervised models to match or exceed state-of-the-art results obtained by single supervised models, a promising direction of research to obtain protein embeddings able to generalize to a much wider array of tasks.