Statistical potentials can be applied to both protein folding and inverse protein folding problems, and aspects of both types of problem will be covered. Firstly, the different levels of difficulty in fold recognition will be discussed. The hardest problems are those where there is apparently no evolutionary relationship between the proteins concerned (i.e., analogous folds). An example of this level of similarity is between, say, an immunoglobulin domain and one of the copper binding proteins such as plastocyanin. The somewhat easier problems are those which involve proteins that are likely to be evolutionarily related (i.e., homologous proteins), but which have very little sequence similarity, e.g., the different c-type cytochromes. This second class of problem is also particularly interesting in that both the structure and the function of a protein can be predicted.A method will be described for evaluating pairwise sequence alignments between these very distantly related proteins using a novel combination of traditional sequence analysis, a set of potentials similar to those used for full optimal sequence threading, and a neural network based expert system. This very quick approach to fold recognition, whilst not being capable of recognizing folds in the harder class of problem, is very successful in the second class, and can be applied to entire genomes and sequence data banks. The results of applying this method to all 1680 predicted ORFs in the Haemophilus influenzae genome and all 486 predicted ORFs in the Mycoplasma genitalium genome will be discussed.In terms of inverse protein folding, a method for automaticde novo protein design will be outlined, and results presented on the synthesis and characterization of a small protein designed using this method.