MOTIVATION: With the enlargement of protein structure databases, it is hoped that a method to classify proteins automatically will be developed. Although the classification criterion proposed by Nakashima et al. ( J. Biochem., 1986, 99, 153-162) was widely used in the literature, it leads to some inconsistencies with the classification databases currently available in the class assignment of protein structures. To improve their work, a new classification criterion is proposed relying on statistical analysis of the secondary structure contents of more than 200 proteins with well-known structural classes. The Fisher linear discriminant algorithm is used to derive the new classification criterion. RESULTS: Three cross-validation tests are performed to evaluate the new criterion. In the jackknife test, of the 210 proteins used to derive the criterion, 206 are correctly classified with an accuracy of 98.10%. Of the 16 proteins of purely intermediate structure (i.e. structures lying near borderlines between two classes) in the first test set, 15 are correctly classified with an accuracy of 93.75%. For the second test set which consists of 200 proteins selected randomly from SCOP, a testing accuracy of 94.00% is obtained. For comparison, the criterion of Nakashima et al. is also used to classify the 210, 16 and 200 proteins, respectively. Consequently, accuracies of 94.76%, 62.50% and 91.50% are obtained, respectively. On average, the accuracy of the new classification criterion is 4% higher than that of Nakashima et al. AVAILABILITY: The program is available on request from the first author. CONTACT: [email protected]