Protein-protein interactions (PPIs) are fundamental to biological processes, and computational prediction of PPIs is important for supplementing gaps in experimental data coverage. However, the quality of PPI predictions critically relies on the distributional characteristics of the training data. Here we investigate how data obtained from two widely-used experimental methods for detecting PPIs, Co-Immunoprecipitation (Co-IP) and Yeast Two-Hybrid (Y2H), differ in graph-theoretic and functional aspects. We document substantial differences between the two modalities, and find each assay type to be significantly more predictive than the other at specific function and network-associated tasks. We accordingly provide concrete recommendations on assay choice for a range of downstream tasks. Our work emphasizes the need for careful curation of PPI data based on the downstream task and underscores the importance of accounting for subtle but critical variations within biological training data.