Asilomar Conference home page
Rapid Homology Modeling with Molecular Mechanics and Dynamics.
Six of the comparative modeling targets (HPR, NDK, EDN, P450, CRABP and HALOFER) were predicted with the procedure we are developing for the fast and semi-automatic generation of homology models. The principle new features of this procedure are the use of an all-atom potential with no cutoff on the long range forces, the use of iterative distance geometry, and the use of different algorithms for optimization. The variations which were tested include the "genetic" algorithm and 4-dimensional embedding as well as combining molecular dynamics with conjugate gradients optimization. The molecular mechanics and dynamics program AMMP was used (Harrison 1993).First, the amino acid sequence of the target is aligned with sequences of related proteins of known structure using GCG. Then the sequences and the alignments are examined in order to reposition insertions and deletions preferentially at the protein surface and between elements of regular secondary structure. Incorrect placement of insertions and deletions is a potential source of major error. Ideally the structure with the highest number of identical residues and the fewest insertions or deletions is used as the starting model. Also, it should be a well-refined structure at relatively high resolution. The starting coordinates are altered to represent the target sequence. The peptide backbone and the atoms which are identical in the side chains are kept and the new atoms are generated by AMMP. The identical atoms are constrained to their original positions and the new atoms are generated with iterative distance geometry. This procedure results in compact structures, but not necessarily in chemically optimal structures. The structure is then minimized, first allowing only the new atoms to move, and then all of the atoms. The UFF all atom potentials with AMBER charges were used, and no cutoff radius was applied. (AMMP is able to run without cutoffs in times comparable to conventional programs which use cutoffs). We also include crystallographic waters when present and any common ligands because water molecules are often structurally important and functionally conserved in ligand binding sites. The most important parts of the model are the ligand binding sites, because their properties usually determine the biological activity of the protein. The final minimization of all atoms is usually done with a combination of conjugate gradients and short runs of molecular dynamics. The procedure can be repeated to test alternate positions of loops or specific residues. Alternative optimization strategies such as the "genetic" algorithm or 4-Dimensional embedding were alsoexplored. The conserved regions can also be restrained to be near the starting positions during minimization. Manual adjustment of loops or side chains can be used if necessary.
The minimization procedure has been shown to give good agreement
between calculated protein-ligand interaction energies and observed
binding energies for complexes with similar solubility and entropy
changes. With five crystal structures of HIV protease inhibitor
complexes, minimization produced RMS differences of .48 to .66 A on all
C-alpha atoms and .67 to .92 A for all atoms. These values are close to the
amount of difference observed between two crystal forms of the same
protein or structures refined in two different labs (.45 to .8
Angstroms for all atoms). This protein structure prediction experiment
will allow us to evaluate this strategy and improve the modeling
procedure.
Asilomar Conference home page