The functional state of an organism is determined largely by the pattern of expression of its genes. The analysis of gene expression data from gene chips has primarily revolved around clustering and classification of the data using machine learning techniques based on the intensity of expression alone with the time-varying pattern mostly ignored. In this paper, we present a pattern recognition-based approach to capturing similarity by finding salient changes in the time-varying expression patterns of genes. Such changes can give clues about important events, such as gene regulation by cell-cycle phases, or even signal the onset of a disease. Specifically, we observe that dissimilarity between time series is revealed by the sharp twists and bends produced in a higher-dimensional curve formed from the constituent signals. Scale-space analysis is used to detect the sharp twists and turns and their relative strength with respect to the component signals is estimated to form a shape similarity measure between time profiles. A clustering algorithm is presented to cluster gene profiles using the scale-space distance as a similarity metric. Multi-dimensional curves formed from time series within clusters are used as cluster prototypes or indexes to the gene expression database, and are used to retrieve the functionally similar genes to a query gene profile. Extensive comparison of clustering using scale-space distance in comparison to traditional Euclidean distance is presented on the yeast genome database.