Automated main-chain model building by template matching and iterative fragment extension.

Terwilliger TC. · 2003

Read Automated main-chain model building by template matching and iterative fragment extension. by Terwilliger TC. (2003) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/europmc-pmc2745878.txt.

Introduction

Terwilliger described in 2003 an automated method for building protein models. It builds the main chain into an X-ray electron-density map. It works by template matching and iterative fragment extension. Model building is the first reading of the map in terms of atom positions, and at low resolution errors are hard for later refinement to fix. Manual building can take a day or far longer when maps are unclear.

The method finds likely helices, strands, and other fragments by matching templates against the map, then extends fragments step by step to trace the chain. It builds on earlier tracing and fragment-fitting methods but automates the search decisions. Faster automated building speeds the whole structure determination process, especially for difficult, low-resolution cases.

Catalog record

CREATORTerwilliger TC.
YEAR2003
COLLECTIONPAPERS
FORMATTXT
FILE SIZE28 KB
LICENSECC BY 4.0
RIGHTSCC BY
SOURCEEurope PMC

Source: https://europepmc.org/articles/PMC2745878

Browse the full library · About the Tower of Babel

Full text

The opening of the work, from the library's hosted text:

Automated main-chain model building by template matching and iterative fragment extension By Thomas C. Terwilliger Published in: Acta Crystallographica Section D: Biological Crystallography DOI: 10.1107/S0907444902018036 ABSTRACT A method for automated macromolecular main-chain model building is described. INTRODUCTION Model building is a key and often time-consuming step in macromolecular structure determination. This step is important because model building is the initial interpretation of the experimental electron-density map in terms of the locations of atoms in the structure. If the resolution of the X-ray data is high (<2 Å), then atomic refinement of the model is highly effective and errors in the initial interpretation can often be corrected. If the resolution of the X-ray data is low (∼3 Å), however, atomic refinement is less effective and it may be very difficult to correct any errors in this initial interpretation (Kleywegt & Jones, 1997 ▶). Although manual model building using a very good electron-density map can require less than a day for 100 or more residues, when the electron-density map is less clear the process can be much slower. It has been recognized for some time that automated procedures for model building would speed up the macromolecular structure determination process considerably and several procedures for doing this have been developed. Most of these procedures are based on the connectivity of the polypeptide chain or on the presence of regular structure (helices and β-strands, common motifs) in the chain. Greer (1985 ▶) devised a rapid procedure (‘bones’) for tracing the path of the polypeptide chain using the connectivity of regions of high electron density in the map. This procedure was extended by Swanson (1994 ▶) to allow threshold-independent tracing of connected regions in a map. Feigenbaum et al. (1977 ▶) used artificial intelligence methods to identify features in electron-density maps. Jones & Thirup (1986 ▶) and Jones et al. (1991 ▶) fitted electron density with fragments from known protein structures. Oldfield (2002 ▶) described a method for automated model building that began by identifying helices and strands and then extending these segments one amino acid at a time to trace a polypeptide chain. Cowtan (1998 ▶, 2001 ▶) and later Terwilliger (2001 ▶) used FFT-based approaches to identify the locations of helices, β-strands and other structure in an electron-density map by template matching. Holton et al. (2000 ▶) used machine-learning techniques to identify side chains in a map. McRee (1999 ▶) has described a semi-automated method for building main-chain and side-chain models in a map, beginning with the identification of Cα positions and fitting fragments from a main-chain library and continuing with using a rotamer library to fit side chains. Pavelcik et al. (2002 ▶) described an alternative and very rapid method for template matching of arbitrary fragments of structure to a map. Levitt (2001 ▶) uses a stepwise approach to model building, beginning with the ‘bones’ of Greer (1985 ▶) to identify helices and strands and extending them one amino acid at a time using ψ, ϕ angles from tables of allowed values. The most widely used automated model-building procedure in current use, ARP/wARP, has been described by Lamzin & Wilson (1993 ▶), Perrakis et al. (1999 ▶) and Morris et al. (2002 ▶). This procedure is very different from all those described above because it is based on an interpretation of the electron-density map in terms of individual atoms, iteratively followed by atomic refinement and an interpretation of the atomic coordinates in terms of a polypeptide chain. The requirement for atomicity limits the application of the method to electron-density maps at a resolution of about 2.3 Å, but for data at this resolution or better the method is exceptionally powerful for automatic model building and atomic refinement. Here, we describe a procedure for automated model building that is related to those described by Oldfield (2002 ▶), McRee (1999 ▶) and Levitt (2001 ▶), but which uses alternative approaches to carry out each of the constituent steps. The method of Cowtan (1998 ▶) is used as a sensitive method for identifying the locations of helices and β-strands. Correlations of template density and map density rather than density at atomic coordinates are used for refinement of the position and orientation of fragments. A fragment-placement method based on tripeptides from refined protein structure and related to the method of Jones & Thirup (1986 ▶) is used to extend segments of structure. Chain connectivity and the correct chain direction are determined by requiring that independently built segments must overlap at two or more consecutive Cα positions before they are merged into a single segment. METHODS As in previous methods for main-chain model building at moderate resolution (Oldfield, 2002 ▶; Levitt, 2001 ▶), our procedure is carried out in hierarchical steps. Firstly, helices and β-strands are located and fitted, with multiple interpretations of each of these secondary structures typically kept. Each helix or strand is then extended in an iterative fashion with libraries of tripeptides from refined protein structures. The collection of (overlapping) partially extended fragments are then assembled into a polypeptide chain by requiring that two or more consecutive Cα positions overlap for two segments to be merged, by requiring that there be no atomic overlaps and by beginning with the best-fitting segments. Each of these steps and the generation of templates and fragment libraries is described below. In all steps, space-group symmetry is used to identify positions that are equivalent in the unit cell and the distance between two points is considered to be the smallest distance between one of the points and any point symmetry-related to the other. Helical and β-strand templates 2.1. An averaged helical template similar to that described in Terwilliger (2001 ▶) was used to identify helical segments in a map. This template consists of the average electron density calculated from a collection of α-helical segments six amino acids in length (from phycoerythrin; PDB code 1lia; Chang et al., 1996 ▶; Berman et al., 2000 ▶), all superimposed on a standard α-helical segment (from myoglobin; PDB code 1a6m; Vojtechovsky et al., 1999 ▶; Berman et al., 2000 ▶). The template included all points within 4 Å of a main-chain or Cβ atom in the standard segment. The template was calculated at a resolution of 3 Å. An averaged β-strand template was constructed in the same fashion, except that the segments used in the template were four amino acids long. Fragment libraries 2.2. Four fragment libraries were constructed. One consisted of 17 α-helical segments from six to 24 residues long in the protein phycoerythrin. Each segment of more than six residues was superimposed on the standard helical segment in three positions: one with the N-terminal six residues of the segment superimposed on the standard segment, one with the C-terminal six residues superimposed and one with the middle six residues superimposed. In this way, a short helical segment that is identified can potentially be extended in either direction. A second library consisted of 17 β-strand segments from four to nine amino acids long from chain A of carboxypeptidase A (PDB code 1bav; Massova et al., 1996 ▶; Berman et al., 2000 ▶), superimposed on the standard β-strand fragment in the same way as for the helical segments. The third and fourth libraries consisted

Read the complete text