The goal of the inverse protein folding problem is to identify amino acid sequences that stabilize a given target protein conformation. Methods that attempt to solve this problem have proven useful for protein sequence design. Here we show that the same methods can provide valuable information for protein fold recognition and for ab initio protein structure prediction. We present a measure of the compatibility of a test sequence with a target model structure, based on computational protein design. The model structure is used as input to design a family of low free energy sequences, and these sequences are compared with the test sequence by using a metric in sequence space based on nearest-neighbor connectivity. We find that this measure is able to recognize the native fold of a myoglobin sequence among different globin folds. It is also powerful enough to recognize near-native protein structures among nonnative models.K nowing the structure of a protein is most useful for predicting, analyzing, and modifying its function. As it is not feasible to determine experimentally the structure of every protein, structure prediction has become central to the field of structural biology and more specifically to structural genomics. On the basis of their study of ribonuclease A (1), Anfinsen and coworkers provided the first clues that all of the information required for folding a protein is to be found in its sequence. Not long after this discovery, people took on the challenge of discovering the rules that allow the protein to fold. This problem is far from simple and has not yet been solved (2). Three major routes are usually considered paths to the solution: homology modeling, threading, and ab initio prediction. To study a protein with unknown conformation C, the first two methods follow the same scheme: a similar protein whose three-dimensional structure is known is identified, and this protein is used as a scaffold to generate a model for C. When the sequences of the two proteins are homologous (i.e., when they have an obvious common ancestry), sequence similarity is assumed to infer structural similarity (3, 4), and the method is then referred to as ''homology modeling.'' When the two sequences show no obvious evolutionary relationship, the method is referred to as ''fold recognition,'' which works by assessing the compatibility of the target sequence with each member of a library of known structures (5).Ab initio structure prediction methods try to build a model for the target protein structure without using a specific template protein. Most of these methods proceed by first generating a large collection of possible conformations (decoys), which are then searched with a scoring function to identify native or, more realistically, near-native conformations (6-9). This second step resembles the fold recognition problem, with the major difference that the library of folds considered includes computergenerated models instead of naturally occurring protein folds. In this paper, we show how recent developments of threading...