PROGRESS IN AUTOMATIC BIOMACROMOLECULAR MODEL BUILDING

F. Pavelcik

Department of Chemical Drugs, Faculty of Pharmacy, University of Veterinary and Pharmaceutical Sciences Brno. 61242 Brno, Department of Inorganic Chemistry, PRIF UK, 84215 Bratislava

Automatic model building is an open challenge in protein crystallography. Three principally different approaches have been developed so far in various protein laboratories: (i) ARP method (based on an interpretation of the electron density in terms of oxygen globs iteratively followed by atomic refinement). (ii) Skeltonization. (iii) Positioning of rigid protein fragments in the electron density by graphical or FFT methods.

Accurate automatic protein model building, based on an expansion of the electron density in spherical harmonics and spherical Bessel functions, phased rotational and translational function and on flexible fragment concept, has been described in a series of papers [1-3] by author. The protein model is built with an accuracy of about 0.2 Ǻ at resolutions of 1.2-1.9 Ǻ. Partial results (e.g. 90% of polyalanine structure) can be obtained at resolutions 2.0-2.3 Ǻ. The recent advances in the method development and computer program will be presented.

Phased rotational, translational and conformation function.

The phased rotational and translational function has been generalized from six-dimensional to multidimensional function. Conformational search was added. Instead of using different fragments for each conformer only one fragment is used and required conformation is calculated on fly. Only the best conformer is stored at particular position and refined.

Loop building.

Hydrophilic residues on the protein surface are usually part of the loop structure. Residues forming loops are sometimes disordered or having large thermal movements. The electron density at loop position is weak, difficult to interpret and in many cases confused with hydrogen-bonded water molecules. If sequence is correctly assigned to the main-chain, number of missing residues in known ad the loop can be constructed. In the method proposed, the both ends of two unconnected chains are extended randomly with AlphA0 fragments. Three factors are then optimized: overlap (mean distance) of peptide groups, electron density fit and Ramachandran energy. Ramachandran energy (2-D periodic function) is expressed as a set of Fourier coefficients. Fourier coefficients are calculated by Fourier summation from pseudo-atomic representation. The minima on Ramachandran map are modeled by different atoms and the curvature of the minimum by a temperature factor.

Least-square refinement of the group model.

Standard nonlinear least-squares were applied to protein refinement. Each group (fragment) is treated as a generalized atom. The number of parameter to be refined is reduced approximately by a factor of four compared to atomic refinement. The method is more-or-less equivalent to a rigid body refinement. The only difference is that in the group model the groups are overlapped.

[1] F. Pavelcik., J. Zelinka and Z. Otwinowski, Acta Cryst. D58 (2002) 275-283

[2] F. Pavelcik, Acta Cryst. A59 (2003) 487-494

[3] F. Pavelcik, Acta Cryst. D60 (2004) 1535-1544