Various methods for the prediction of secondary structure from amino acid sequence can consistently achieve on average 60% accuracy when tested for several proteins. Improvement on this value has proved difficult, despite increasing the size of the data set and refining predictive techniques. The difficulty almost certainly derives from the influence of long-range interactions and the restrictions required to attain favourable protein topologies. We describe here a novel approach to structure prediction from amino acid sequence based on the recognition of super-secondary structure. The structure we initially consider is the beta alpha beta unit, which consists of two parallel beta-strands connected by an alpha-helix. From an analysis of all known beta alpha beta units, an ideal secondary structure sequence was derived. This was used as a template to locate probable beta alpha beta sequences in a standard secondary structure prediction. The method correctly predicted the location of 70% of the beta alpha beta units in 16 beta/alpha type proteins. This led to a 7.5% average improvement over the original secondary structure prediction.