|
The process of reconstructing an all-atom peptide, or protein, model from a subset of the atomic coordinates is the subject of a considerable body of literature. Numerous applications for this process exist, including enhancement of low-resolution models from crystallography into usable models or conversion of the coarse structures typical of ab initio folding computations or comparative protein modelling techniques into all-atom models.
It is common to split the process of reconstructing all-atom structures into two separate tasks. The first task is to predict the full backbone, and the second is to build side chains onto this. This partitioning of the problem is considered reasonable because it is observed that the backbone geometry of well defined secondary structure elements is mostly invariant to the identity of its amino acid residues.
Previously published backbone reconstruction methods either generate de novo backbone conformations or utilise fragment libraries derived from the PDB to locate possible structures which do not violate the provided Cα trace. These conformations are evaluated using either energetic, homology-based or geometric criteria. BB applies a dead-end elimination procedure, a knowledge-based scoring function and a library of 3-residue fragments to construct all backbone atoms. In a previous study, BB was found to produce structures with an average all-backbone atom RMSD of 0.41 Angstroms from the Cα coordinates alone.
Features include:
- Easy to use.
- Free downloads.
- Accurate predictions.
Downloads:
- RPM version 0.9-1 for Redhat 8.0 on x86 PCs.
- PDF, PostScript installation and usage manual.
- PDF preprint of Journal of Computational Chemistry paper. (currently withheld)
- PDF preprint of Bioinformatics paper. (currently withheld)
- tar.gz Test data set used in Journal of Computational Chemistry paper. (currently withheld)
- Other binary downloads will be made available soon. Email me if you have any particular requests.
Important note about the forcefield:
The forcefield data included in the above RPM was not generated with a very optimal training set for general purpose backbone reconstruction. This data is instead biased for generating reliable core backbone structures at the expense of loop backbone structures, as is desirable for my protein threading software. This results in an overall RMSD that is about 25% worse than I'd expect with a more general forcefield. I will regenerate a general forcefield in the near future, but don't have time to do this at present. I'll put such an updated forcefield on this website as soon as I get time to produce it.
| 
Introduction