Partial least squares modelling in latent variables is applied to the analysis of short-term test genotoxicity data obtained by testing 42 chemicals, known carcinogens and non-carcinogens, in 35 different assay systems that are deemed to be potential indicators of carcinogencity. Results of a preliminary analysis of all data, and of a second analysis on a reduced data base, are presented. The latter analysis provided a model explaining 73% of variance, whereby no false negative or positive was predicted among 38 chemicals, in spite of qualitative predictions indicating five false positives and two false negatives. Only nine out of 35 assays appear to be relevant to the description of carcinogenicity/non-carcinogenicity of chemical substances. Copyright © 1988 John Wiley & Sons, Ltd.