Identification of structural domains in proteins by a graph heuristic

Abstract

A novel automatic procedure for identifying domains from protein atomic coordinates is presented. The procedure, termed STRUDL (STRUctural Domain Limits), does not take into account information on secondary structures and handles any number of domains made up of contiguous or non-contiguous chain segments. The core algorithm uses the Kernighan-Lin graph heuristic to partition the protein into residue sets which display minimum interactions between them. These interactions are deduced from the weighted Voronoi diagram. The generated partitions are accepted or rejected on the basis of optimized criteria, representing basic expected physical properties of structural domains. The graph heuristic approach is shown to be very effective, it approximates closely the exact solution provided by a branch and bound algorithm for a number of test proteins. In addition, the overall performance of STRUDL is assessed on a set of 787 representative proteins from the Protein Data Bank by comparison to domain definitions in the CATH protein classification. The domains assigned by STRUDL agree with the CATH assignments in at least 81% of the tested proteins. This result is comparable to that obtained previously using PUU (Holm and Sander, Proteins 1994;9:256-268), the only other available algorithm designed to identify domains with any number of non-contiguous chain segments. A detailed discussion of the structures for which our assignments differ from those in CATH brings to light some clear inconsistencies between the concept of structural domains based on minimizing inter-domain interactions and that of delimiting structural motifs that represent acceptable folding topologies or architectures. Considering both concepts as complementary and combining them in a layered approach might be the way forward.

Toward Consistent Assignment of Structural Domains in Proteins

“…The choice of training dataset determines, to a great extent, the tendencies of the algorithm, particularly in ambiguous cases, in which the expert-defined domain is not a single, autonomous, compact structure. Different methods use different datasets to tune their performance: some methods compare the results to AUTHORS, 19 others to CATH, 22 and yet others to manually annotated datasets. 16 Each algorithm then reflects the "philosophical" approach and associated bias towards the training dataset used.…”

Improvement of domain linker prediction by incorporating loop‐length‐dependent characteristics

“…However, although the concept of domains is widely accepted, no consensual, specific, and objective definition of a domain is yet available. Hence, the various automated methods of defining domain regions on the basis of protein structures have few common features, except their nonreliance on the visual inspection of protein structures 20–23. Because our project focuses on the prediction of structural domains, we constructed an algorithm specifically aimed at identifying structural domains that are highly likely to fold in isolation.…”

Mining protein family specific residue packing patterns from protein structure graphs

“…This approach recognizes natural nearest neighbor residues by finding four-tuples that define tetrahedra with an empty sphere property [7]; this tessellation is dual to the Voronoi diagram. The Delaunay tessellation graph (DT) has been used to analyze protein packing [24,29] and structure [22,26,34,32,28] The third representation, derived from the recently-defined almost-Delaunay edges [1], expands the set of Delaunay edges to account for perturbation or motion of point coordinates, controlled by a parameter . Thus, we actually define a family of graphs, AD( ), that interpolates between the CD and DT representations so that DT ⊆ AD( ) ⊆ CD for all ≥ 0.…”