Quantitative molecular variation may be used for the development of methods for tumor classification. We used the statistical concept of principal component analysis to type ovarian tumors. We purified tumor cells from ovarian tumors and subjected them to two-dimensional gel electrophoresis. Using a data set derived from the quantitation of 170 polypeptides, we established a model (learning set) with 22 tumors for classification into three groups (benign, borderline and malignant) and then used 18 tumors to test the model. We correctly classified six of eight carcinomas and three of four borderline tumors. Of six benign lesions, two were correctly classified, three were classified as borderline and one was classified as a carcinoma. It may be possible to classify tumors by their constitutive gene expression profile using multivariate analysis.