Résumé : Some antipsychotic agents have been found to produce agranulocytosis and aplastic anemia. The oxidation phenomena and/or the formation of free radicals has been suggested to be causally related to various hematological disorders, e.g., agranulocytosis. Using five experimental conditions, we tested the oxidative potential of compounds with and without a history of hematological side effects, e.g., agranulocytosis and aplastic anemia. A statistical analysis was undertaken for each experimental condition and a multivariate analysis combining all results was performed. Two peroxidase-induced free radical models did not successfully discriminate between drugs with and without a history of causing hematologic problems (<70%). The lipid peroxidation system provided even less satisfactory discrimination, with only 56.25% correct classification. However, an 87.5% correct classification was obtained when using the oxidation potentials of these drugs determined at pH 4.7 and at pH 7.4. A multivariate analysis taking into account the five variables provided 87.5% success in classification. The two clusters were better discriminated in terms of a "distance coefficient." In a second analysis, the putative antipsychotic pyridobenzodiazepine analogues (JL5, JL8, JL18, and JL25) were classified in the cluster of toxic compounds, while the oxa- and thiazepine analogues (JL2, JL3, and JL13) were classified as nontoxic compounds. On the other hand, a few metabolites of clozapine and fluperlapine were classified in the toxic compound group. The procedure described herein is, to our knowledge, the first which classifies molecules of different structures as well as different pharmacological profiles according to their hematotoxic potential. Such a procedure could be used to predict drug-induced hematological side effects.