The extent to which the contemporary dataset of protein structures can be segregated into four structural "classes" as originally defined by Levitt & Chothia in 1976 is examined and a simple method presented for the assignment of protein domains into these classes. Assignments are based on known three-dimensional structures, and for successful assignment it was found that helix/sheet content, contacts between secondary structures and their sequential order had to be used. The procedure attempts to maximise the automatic separation into classes for a dataset of 197 manually classified, non-homologous domains. It was found that approximately 90% of the structures were classified automatically; the remainder were borderline and were left for manual inspection. The method was then applied to a test set of 43 protein domains with similar results. The data support the concept of distinct classes of protein structure, although a few intermediate structures are found, demonstrating that it is possible to define relatively simple parameters complying with commonly accepted nomenclature that automatically define 90% of protein domains with essentially 100% accuracy. However, re-examination of the data also suggested that the previously separate alpha/beta and alpha + beta classes show considerable overlap and are more naturally represented as a single alpha beta class. This large alpha beta class can then be most easily subdivided by consideration of whether the sheets are mainly parallel, antiparallel or mixed. The correlation between structural class and function is discussed, together with the conservation of class within a sequence superfamily. This represents the first step in an automated phenetic description of protein structure complementing the usual phylogenetic approach to protein structure classification.