There are far fewer classes of three-dimensional protein folds than sequence families but the problem of detecting three-dimensional similarities is NP-complete. We present a novel heuristic for identifying 3-D similarities between a query structure and the database of known protein structures. Many methods for structure alignment use a bottom-up approach, identifying first local matches and then solving a combinatorial problem in building up larger clusters of matching substructures. Here, the top-down approach is to start with the global comparison and select a rough superimposition using a fast 3-D lookup of secondary structure motifs. The superimposition is then extended to an alignment of Ca atoms by an iterative dynamic programming step. An all-against-all comparison of 385 representative proteins (150,000 pair comparisons) took 1 day of computer time on a single R8000 processor. In other words, one query structure is scanned against the database in a matter of minutes. The method is rated at 90 % reliability at capturing statistically significant similarities. It is useful as a rapid preprocessor to a comprehensive protein structure database search system.