An empirical relation between the amino acid composition and three-dim
ensional folding pattern of several classes of proteins has been deter
mined. Computer simulated neural networks have been used to assign pro
teins to one of the following classes based on their amino acid compos
ition and size: (1) 4alpha-helical bundles, (2) parallel (alpha/beta)8
barrels, (3) nucleotide binding fold, (4) immunoglobulin fold, or (5)
none of these. Networks trained on the known crystal structures as we
ll as sequences of closely related proteins are shown to correctly pre
dict folding classes of proteins not represented in the training set w
ith an average accuracy of 87%. Other folding motifs can easily be add
ed to the prediction scheme once larger databases become available. An
alysis of the neural network weights reveals that amino acids favoring
prediction of a folding class are usually over represented in that cl
ass and amino acids with unfavorable weights are underrepresented in c
omposition. The neural networks utilize combinations of these multiple
small variations in amino acid composition in order to make a predict
ion. The favorably weighted amino acids in a given class also form the
most intramolecular interactions with other residues in proteins of t
hat class. A detailed examination of the contacts of these amino acids
reveals some general patterns that may help stabilize each folding cl
ass.