The rapid discovery of sequence information from the Human Genome Project has exponentially increased the amount of data that can be retrieved from biomedical experiments. Gene expression profiling, through the use of microarray technology, is rapidly contributing to an improved understanding of global, coordinated cellular events in a variety of paradigms. In the field of toxicology, the potential application of toxicogenomics to indicate the toxicity of unknown compounds has been suggested but remains largely unsubstantiated to date. A major supposition of toxicogenomics is that global changes in the expression of individual mRNAs (i.e., the transcriptional responses of cells to toxicants) will be sufficiently distinct, robust, and reproducible to allow discrimination of toxicants from different classes. Definitive demonstration is still lacking for such specific “genetic fingerprints,” as opposed to nonspecific general stress responses that may be indistinguishable between compounds and therefore not suitable as probes of toxic mechanisms. The present studies demonstrate a general application of toxicogenomics that distinguishes two mechanistically unrelated classes of toxicants (cytotoxic anti-inflammatory drugs and DNA-damaging agents) based solely upon a cluster-type analysis of genes differentially induced or repressed in cultured cells during exposure to these compounds. Initial comparisons of the expression patterns for 100 toxic compounds, using all ∼ 250 genes on a DNA microarray (∼ 2.5 million data points), failed to discriminate between toxicant classes. A major obstacle encountered in these studies was the lack of reproducible gene responses, presumably due to biological variability and technological limitations. Thus multiple replicate observations for the prototypical DNA damaging agent, cisplatin, and the non-steroidal anti-inflammatory drugs (NSAIDs) diflunisal and flufenamic acid were made, and a subset of genes yielding reproducible inductions/repressions was selected for comparison. Many of the “fingerprint genes” identified in these studies were consistent with previous observations reported in the literature (e.g., the well-characterized induction by cisplatin of p53-regulated transcripts such as p21waf1/cip1 and PCNA [proliferating cell nuclear antigen]). These gene subsets not only discriminated among the three compounds in the learning set but also showed predictive value for the rest of the database (∼ 100 compounds of various toxic mechanisms). Further refinement of the clustering strategy, using a computer-based optimization algorithm, yielded even better results and demonstrated that genes that ultimately best discriminated between DNA damage and NSAIDs were involved in such diverse processes as DNA repair, xenobiotic metabolism, transcriptional activation, structural maintenance, cell cycle control, signal transduction, and apoptosis. The determination of genes whose responses appropriately group and dissociate anti-inflammatory versus DNA-damaging agents provides an initial paradigm upon which to build for future, higher throughput-based identification of toxic compounds using gene expression patterns alone.