A new method for the prediction of protein structural classes is constructed based on MVP (Maximum variance projection) algorithm, which is a manifold learning-based data mining method. DC (Dipeptide Composition) and PseAA (Pseudo Amino Acid) are used as conditional attributes for the construction of decision system. A DR (Dimensionality Reduction) algorithm, the so-called MVP is introduced to reduce the decision system, which can be used to classify new objects. Experimental results thus obtained are quite encouraging, which indicate that the above method is used effectively to deal with this complicated problem of protein structural classes.