Transcription factor binding sites are short fragments in the upstream regions of genes, to which transcription factors bind to regulate the transcription of genes into mRNA. Computational identification of transcription factor binding sites remains an unsolved challenging problem though a great amount of effort has been put into the study of this problem. We have recently developed a novel technique for identification of binding sites from a set of upstream regions of genes, that could possibly be transcriptionally co-regulated and hence might share similar transcription factor binding sites. By utilizing two key features of such binding sites (i.e. their high sequence similarities and their relatively high frequencies compared to other sequence fragments), we have formulated this problem as a cluster identification problem. That is to identify and extract data clusters from a noisy background. While the classical data clustering problem (partitioning a data set into clusters sharing common or similar features) has been extensively studied, there is no general algorithm for solving the problem of identifying data clusters from a noisy background. In this paper, we present a novel algorithm for solving such a problem. We have proved that a cluster identification problem, under our definition, can be rigorously and efficiently solved through searching for substrings with special properties in a linear sequence. We have also developed a method for assessing the statistical significance of each identified cluster, which can be used to rule out accidental data clusters. We have implemented the cluster identification algorithm and the statistical significance analysis method as a computer software CUBIC. Extensive testing on CUBIC has been carried out. We present here a few applications of CUBIC on challenging cases of binding site identification.