A method predicting protein secondary structure from sequence information could be assessed for its real efficiency by applying it to a number of proteins which lie completely outside a given data set. This type of test is performed for the three methods of Chou and Fasman (Adv. Enzymol. 47 (1978) 45-148), Robson and co-workers (J. Mol. Biol. 120 (1978) 97-120) and Lim (J. Mol. Biol. 88 (1974) 873-894) by using data of 19 proteins for the former two methods and 11 proteins for the method of Lim. The prediction abilities of these methods turn out to be of almost the same level, but unexpectedly low: their average scores are commonly less than 55% measured by the three-state assessment (alpha, beta and coil) or less than 45% measured by the four-state assessment (alpha, beta, turn and coil). This level of accuracy is more than 20% lower than that of current expectations as summarized by Schulz and Schirmer (Principles of Protein Structure (1979) Ch. 6, Springer, New York). A joint prediction attempted with the simultaneous usage of the three prediction methods did not improve the results. Causes and implications of the unsatisfactory results are discussed. In this study, computer programs were prepared for the methods of Chou and Fasman and of Robson and co-workers. While difficulties arose in the course of the computerization of the Chou-Fasman method, the prediction algorithm was arranged in a fully automatic form with optimization of the original rules as well as introduction of a modified treatment for solving the overlap among initially predicted regions of the secondary structures. Large discrepancies observed between the original results and those obtained by the computerized method are examined.