Optimized weight matrices defining four major eukaryotic promoter elements, the TATA-box, cap signal, CCAAT-, and GC-box, are presented; they were derived by comparative sequence analysis of 502 unrelated RNA polymerase II promoter regions. The new TATA-box and cap signal descriptions differ in several respects from the only hitherto available base frequency Tables. The CCAAT-box matrix, obtained with no prior assumption but CCAAT being the core of the motif, reflects precisely the sequence specificity of the recently discovered nuclear factor NY-I/CP1 but does not include typical recognition sequences of two other purported CCAAT-binding proteins, CTF and CBP. The GC-box description is longer than the previously proposed consensus sequences but is consistent with Sp1 protein-DNA binding data. The notion of a CACCC element distinct from the GC-box seems not to be justified any longer in view of the new weight matrix. Unlike the two fixed-distance elements, neither the CCAAT- nor the GC-box occurs at significantly high frequency in the upstream regions of non-vertebrate genes. Preliminary attempts to predict promoters with the aid of the new signal descriptions were unexpectedly successful. The new TATA-box matrix locates eukaryotic transcription initiation sites as reliably as do the best currently available methods to map Escherichia coli promoters. This analysis was made possible by the recently established Eukaryotic Promoter Database (EPD) of the EMBL Nucleotide Sequence Data Library. In order to derive the weight matrices, a novel algorithm has been devised that is generally applicable to sequence motifs positionally correlated with a biologically defined position in the sequences. The signal must be sufficiently over-represented in a particular region relative to the given site, but need not be present in all members of the input sequence collection. The algorithm iteratively redefines the set of putative motif representatives from which a weight matrix is derived, so as to maximize a quantitative measure of local over-representation, an optimization criterion that naturally combines structural and positional constancy. A comprehensive description of the technique is presented in Methods and Data.