In this paper, we present a case study of predicting topics of scientific papers using a co-authorship graph. Co-authorship graphs constitute a specific view on bibliographic data, where scientific publications are modelled as a graph’s nodes, and two nodes are linked by an undirected edge whenever the two corresponding papers share at least one author. We apply a simple collective classification algorithm based on relaxation labelling to the ILPnet2 bibliographic database. The approach is based on the assumption that papers in the same neighbourhood of the co-authorship graph tend to be on the same topics, and that the predicted topic for one node in the graph depends on the actual or predicted topics of the nodes linked to it. We evauate the performance of this method on the ILPnet2 data in terms of ROC analysis, and explain the results in terms of the co-authorship graph and the position and properties of papers on a certain topic in the graph.