Predicting the structural fold of a protein is an important and challenging problem. Available computer programs for determining whether a protein sequence is compatible with a known 3-dimensional structure fall into 2 categories: (1) structure-based methods, in which structural features such as local conformation and solvent accessibility are encoded in a template, and (2) sequence-based methods, in which aligned sequences of a set of related proteins are encoded in a template. In both cases, the programs use a static template based on a predetermined set of proteins. Here, we describe a computer-based method, called iterative template refinement (ITR), that uses templates combining structure-based and sequence-based information and employs an iterative search procedure to detect related proteins and sequentially add them to the templates. Starting from a single protein of known structure, ITR performs sequential cycles of database search to construct an expanding tree of templates with the aim of identifying subtle relationships among proteins. Evaluating the performance of ITR on 6 proteins, we found that the method automatically identified a variety of subtle structural similarities to other proteins. For example, the method identified structural similarity between arabinose-binding protein and phosphofructokinase, a relationship that has not been widely recognized.Keywords: dynamic programming; inverted protein structure prediction; local environment template; multiple sequence template; profile; protein structureThe number of distinct protein structural folds is thought to be relatively small, perhaps less than 1,000 (Chothia, 1992). If so, many of the 6 0 , O O O proteins in existing sequence databases must adopt conformations similar to an already known structure. However, it remains a challenging problem to place proteins known only by their amino acid sequences into the correct structural family.Various computer programs have been developed for using a fixed template for detecting structurally related proteins. Broadly speaking, they fall into 2 categories:1. Structure-bused methods. In this approach, a template encoding information about a specific structure is used to search the protein database to find other sequences compatible with the structure. The first such approach involved classifying each position in a structure according to its solvent accessibility, local conformation, and polarity (Bowie et al., 1991). Subsequent generalizations have included other types of environments defined according to the nature of contacts made by each residue