A method is described for identifying and evaluating regions of significant similarity between two sequences. The notion of a 'homology domain' is employed which defines the boundaries of a region of sequence homology containing no insertions or deletions. The relative significance of different potential homology domains is evaluated using a non-linear similarity score related to the probability of finding the observed level of similarity in the region by chance. The sensitivity of the method is demonstrated by simulating the evolution of homology domains and applying the method to their detection. Several examples of the use of homology domain identification are given.