ene expression profiling with microarrays is a promising method to identify groups of genes, which are likely to differentiate between complex diseases, such as severe malignancies.1, 2 One problem, besides the well-known sample preparation and hybridization challenge3 is the abundance of experimental results (typically thousands of oligonucleotide probes) for a small number of clinical samples (typically 10 to 100). Since traditional statistics require large data sets and are often tedious, I looked for a simpler and widely applicable filtering approach.The method is based on observations obtained with artificial neural nets (ANN), which we are using for the classification of tumor genes in a similar way as described by Khan, et al.2 An ANN in its simplest structure calculates the relationships between input patterns and their respective output classes using weight factors (Fig. 1). The higher the weight, the higher is the contribution of an input signal to the observed output pattern.