Prediction of protein secondary structure from circular dichroism spectra: an attempt to solve the problem of the best-fitting reference protein subsets.

In least-squares fitting of protein circular dichroism (CD) spectra using basis CD spectra for the respective secondary structure components, as given by reference proteins of known structural composition, good fits of the CD spectrum do not necessarily correspond to appropriate fits of the underlying structural composition of a test protein. In an attempt to overcome this problem, CD similarity measures were used to construct a subset of five reference CD spectra which permitted three-component fitting of the CD and prediction of the relative magnitude of total beta-sheet (antiparallel + parallel) and total other structure (beta-turn + remainder), relative to helix, in the test protein. A backpropagation neural network (BPN) was also trained to make this prediction. In subsequent five-component fitting of the CD spectrum, using a Monte Carlo method to generate subsets of reference proteins from the working data set, only those secondary structure fits which conformed to the consensus prediction of the CD similarity measures and the BPN were accepted. The method enhanced the fitting of antiparallel beta-sheet and beta-turn for 16 proteins, compared to the variable selection method of P. Manavalan and W. C. Johnson (1987, Anal. Biochem. 167, 76-85). Some other proteins were less well fitted. Improvement in results is expected with a larger representation of feasible basis CD spectra in the working reference proteins.