The challenge of predicting protein-protein interactions is at the core of most research efforts in biology and therapeutics. Because of the combinatorial nature of molecular interactions, a prediction approach based on a straightforward application of the principles of physics is often impractical—at least for any proteome-scale investigation or for the purpose of screening large numbers of protein designs. Recent progress in machine learning has led to the discovery of entirely new representations for protein sequences and structures—often abstract vectors in high-dimensional spaces. While these representations provide useful cues about a protein’s propensity to form an interaction, they exhibit little connection to the physical concepts usually invoked to explain such interactions. This chapter discusses the general problem of predicting protein-protein interactions and explains how machine learning models designed to generate protein representations can be used to solve it. It discusses in more detail how these representations can be tied to certain physical priors that make them more interpretable and that make any prediction more explainable.