Computational approaches to modeling the conserved structural core among distantly homologous proteins

Abstract

Modem techniques in biology have produced sequence data for huge quantities of proteins, and 3-D structural information for a much smaller number of proteins. We introduce several algorithms that make use of the limited available structural information to classify and annotate proteins with structures that are unknown, but similar to solved structures. The first algorithm is actually a tool for better understanding solved structures themselves. Namely, we introduce the multiple alignment algorithm Matt (Multiple Alignment with Translations and Twists), an aligned fragment pair chaining algorithm that, in intermediate steps, allows local flexibility between fragments. Matt temporarily allows small translations and rotations to bring sets of fragments into closer alignment than physically possible under rigid body transformation. The second algorithm, BetaWrapPro, is designed to recognize sequences of unknown structure that belong to specific all-beta fold classes. BetaWrapPro employs a "wrapping" algorithm that uses long-distance pairwise residue preferences to recognize sequences belonging to the beta-helix and the beta-trefoil classes. It uses hand-curated beta-strand templates based on solved structures. Finally, SMURF (Structural Motifs Using Random Fields) combines ideas from both these algorithms into a general method to recognize beta-structural motifs using both sequence information and long-distance pairwise correlations involved in beta-sheet formation. For any beta-structural fold, SMURF uses Matt to automatically construct a template from an alignment of solved 3-D structures.

(cont.) From this template, SMURF constructs a Markov random field that combines a profile hidden Markov model together with pairwise residue preferences of the type introduced by BetaWrapPro. The efficacy of SMURF is demonstrated on three beta-propeller fold classes.