Accurate prediction of protein secondary structural class with fuzzy structural vectors

The prerequisites for accurate prediction of protein secondary structural class (all-alpha, all-beta, alpha+beta, alpha/beta or multidomain) were studied, and a new similarity-based method is presented for the prediction of the secondary structural class of a protein from its sequence. The new method uses representatives of nuclear families as a learning set. For the sequence to be predicted, the method produces a vector of certainty factors called a fuzzy structural vector. Validation with independent test sets shows that the prediction accuracy of the proposed method has clear dependency on the representativity of the learning set. The representatives obtained from the nuclear families of the Brookhaven Protein Data Bank (PDB) were shown to give accurate predictions for PDB proteins, whilst the amino acid composition-based methods used previously achieve their maximum predictability with relatively limited learning sets, and they remain inaccurate even with highly representative learning sets. The usability of the new method is increased further by the fuzzy structural vectors, which substantially reduce the risk of misclassification and realistically describe vague secondary structural tendencies.