The task of recognizing sequence-structure compatibility (or threading) holds promise for detecting similarities that can not be seen with sequence analysis alone. One of the chief limitations to current threading methods is their relative insensitivity to the correct alignment. We have created a new threading evaluation function that (1) uses a full atomic representation of structure (rapidly instantiating a plausible set of sidechains), (2) reflects the idiosyncrasies of particular environments within a fold family, and (3) is very sensitive to shifts away from the correct alignment. We illustrate these characteristics in the context of the globin family. We created a statistical model of globin environments using a structural alignment of 13 globin (and globin-like) molecules. We tested these with cross-validation, and found that our method reliably recognizes the correct alignments with sequence identities in the range of 11% to 18% identity, as long as the backbone upon which the sequence is threaded (the template backbone) is within 2.9 A RMSD of the actual backbone. When the backbone is more than 3.0 A from the template backbone, we are unable to detect the correct alignment. Our method shows sensitivity to both systematic and random shifts in the alignment-even with shifts of a single residue. These results suggest that threading methods using atomic detail and statistical descriptions of structural environments can be effective. Such methods, however, make the search for sequence-structure compatibilities more difficult, since they create narrow, steep optima.
Notes: Submitted to the Pacific Symposium of Biocomputing ’98
Keywords:
Full Paper: Full paper available in PDF here (0.29 megabytes) .
PDF readers are available.
Type of Publication: Report
Reference: L. Wei & R. B. Altman. Recognizing Protein Binding Sites Using Statistical Descriptions of their 3D Environments. 1997.
Abstract: We have developed a new method for recognizing sites in three-dimensional protein structures. Our method is based on our previously reported algorithm for creating descriptions of protein microenvironments using physical and chemical properties at multiple levels of detail (including features at the atomic, chemical group, residue, and secondary structural levels). The recognition method takes three inputs: a set of sites that share some structural or functional role, a set of control nonsites that lack this role, and a single query site. The distribution of properties for the query site is compared to the distributions for both sites and nonsites to determine the group with which it is most similar. A log-odds scoring function, based on Bayes’ Rule, computes a score that indicates the likelihood that the query region is a site of interest. In this paper, we apply the method to the task of identifying calcium binding sites in proteins. Cross-validation analysis shows that this recognition approach has high sensitivity and specificity. We also describe the results of scanning four calcium binding proteins (with the calciums removed) using a three-dimensional grid of probe points at 2 A spacing. The probe points that have high scores cluster around the true calcium binding sites, with the highest scoring points at or near the binding sites. The method fails to find a calcium binding site created by four proteins in the crystal lattice, and thus not recognizable within the crystallographic asymmetric unit. Our results show that property-based descriptions can be used for recognizing protein sites in unannotated structures.
Full Paper: Available in PDF PDF readers are available.
