ABSTRACT: BACKGROUND: The accurate determination of transcription factor binding affinities is an important problem in biology and key to understanding the gene regulation process. Position weight matrices are commonly used to represent the binding properties of transcription factor binding sites but suffer from low information content and a large number of false matches in the genome. We describe a novel algorithm for the refinement of position weight matrices representing transcription factor binding sites based on experimental data, including ChIP-chip analyses. We present an iterative weight matrix optimization method that is more accurate in distinguishing true transcription factor binding sites from a negative control set. The initial position weight matrix comes from JASPAR, TRANSFAC or other sources. The main new features are the discriminative nature of the method and matrix width and length optimization. RESULTS: The algorithm was applied to the increasing collection of known transcription factor binding sites obtained from ChIP-chip experiments. The results show that our algorithm significantly improves the sensitivity and specificity of matrix models for identifying transcription factor binding sites. CONCLUSION: When the transcription factor is known, it is more appropriate to use a discriminative approach such as the one presented here to derive its transcription factor-DNA binding properties starting with a matrix, as opposed to performing de novo motif discovery. Generating more accurate position weight matrices will ultimately contribute to a better understanding of eukaryotic transcriptional regulation, and could potentially offer a better alternative to ab initio motif discovery.