Automatic Pattern Embedding in Protein Structure Models

sequence-to-model alignments. Although DSMs are generally represented as hidden Markov models (HMMs), they are designed around a set of rules rather than trained on a large representative set, as most HMMs are. The DSM approach to protein structure modeling has many advantages: flexibility, generality, a high degree of abstraction, and the possibility to build models for hypothetical structures. A few experts—using their understanding of the protein structure families— designed the originally proposed models. However, with the number of deposited protein structures in the protein databank (PDB) surpassing 15,000 and the number of classified protein folds approaching 600, updating DSM libraries through this approach has become difficult. Our method constructs DSMs automatically from determined protein structures. It inherits many merits from the original DSMs, such as dividing structural states into several classes and using expert prior structural knowledge for model design to reduce the model’s free parameters.2 The DSM design’s modularity lets us combine a functional family’s sequence motif with the structural model.3 Ample evidence shows that including sequence information considerably improves fold recognition in other threading approaches.4 Most methods of combining sequence and structural information use some measure of sequence similarity over the entire length of the sequence and the structural model. Combining a minimal (few residues) functional pattern with a structural DSM can provide a sensitive method to identify functional homologs that have similar structures but no detectable overall sequence similarity.2,3 Previously, we had to select and embed those functional patterns in the models manually. In this article, we present a set of rules for automated minimal-pattern selection and embedding.