The native conformation of a protein is generally assumed to be the one with the lowest free energy (1). The successful prediction of protein structure depends on the surmounting of three subproblems: (1) choosing a representation of protein conformation that includes structures similar to the correct conformation but limits the search space; (2) formulating a scoring function that relates a particular protein conformation to its free energy; and (3) devising a method to combine the first two elements in a search through conformational space for the state with the globally optimum score. These three requirements apply to the major classes of protein structure prediction: homology modeling, threading (fold recognition), and ab initio folding. In this chapter, we focus on the second of the three subproblems, that of developing energy functions, and place an emphasis on functions tailored for ab initio folding, although much of the discussion will also apply to threading.