DETERMINATION OF PROTEIN TERTIARY STRUCTURE CLASS FROM CIRCULAR-DICHROISM SPECTRA

Fifty-three circular dichroism (CD) spectra consisting of the spectra of 46 native proteins, 3 denatured proteins, and one oligopeptide (the spectra of two denatured proteins and oligopeptide were taken at two different temperatures) were investigated in order to examine the corr elation between the shape of the CD spectrum and the tertiary structur e class of the protein. Five classes were considered-all - alpha, all - beta, alpha + beta, alpha/beta, and denatured proteins. Spectra from 190 to 236 nm with 2 nm interval were described as points in 24-dimen sional hyperspace, where coordinates were values of ellipticities at f ixed wavelengths. This allows the spectra to be treated as patterns an d subsequently analyzed using pattern recognition algorithms. Cluster analysis, which does not need predefined information about protein str ucture, divides spectra into several compact groups or clusters with g ood correlation with tertiary structure class. To visualize these resu lts, orthogonalization procedures were imposed on the original data se t in 24-dimensional space. The new 3-dimensional coordinate system dem onstrated well-separated all-beta class and denatured proteins. Region s corresponding to all - alpha and especially alpha + beta and alpha/b eta proteins were not as well resolved. The following approach was the n applied to the original data set to obtain an objective mathematical algorithm for the determination of a protein's tertiary structure cla ss from its CD spectrum. Regions in 24-dimensional hyperspace correspo nding to all of the tertiary structure classes were found by calculati ng the decision functions, or equations of hyperplanes, which separate groups of spectral patterns of different classes. The class represent ing the region which involves the pattern of a protein spectrum can be interpreted as a tertiary structure class of this protein. The accura cy of the method was checked by removing one of the proteins from the training set, finding all the decision functions, and determinating th e class of the excluded protein. This test gives 100% accuracy for all - alpha, alpha/beta, and denatured proteins; 85% for alpha + beta and 75% for all - beta proteins. (C) 1994 Academic Press, Inc.