The most informative annotation of genes and proteins classically derives from small-scale, biological experimentation, often resulting in the cloning of a molecular sequence of known function or phenotype. In the genome era, massive cloning and sequencing come first, followed by computational annotation based on similarity to classically annotated sequences. Both types of annotation are subject to continual revision based on new experimental and computational results, resulting in an unstable foundation on which to build interpretations of large-scale gene expression profiles. Furthermore, one's ability to efficiently interpret the data is severely limited by the necessity of manually exploring the annotation spaces. We have been experimenting with semi-automated methods to annotate clusters of genes based on data mining and summarization of information in textual databases using document 'neighbouring' and other techniques. The potential power of this approach may be limited by the structure and consistency of information in archival databases. Expression profiles (and other types of functional genomics data) may themselves be the best hope for a new type of holistic, relational annotation in the post-genome era.