Application of Coarse-Grained Force Fields to Predict Protein Structures
Given the profound importance of energy functions in molecular modelling, the design of appropriate protein force fields has attracted considerable attention and much research, using a variety of techniques and algorithms. While significant progress has been made in the last few years, we still do not have satisfactory solutions. Our research programme is working towards finding a satisfactory solution of a coarse-grained force field and applying it to predict protein structures. In 2004 another Critical Assessment of protein Structure Prediction (CASP) test will be held, in which participants are asked to predict structures of proteins before the experimentally determined structures is released. With an expected 50 proteins to predict in this Olympics of protein structure prediction, we expect to invest many hundreds of person-hours and several thousands of CPU hours on HPC platforms in the CASP test. The reward from this unique blind test is an objective evaluation of our methods and a direct performance comparison with other competitors' methods.
|
Principal Investigator Thomas HuberDepartment of Mathematics University of Queensland |
Project d84, f86 |
|
Co-Investigators Ian LenaneSarah Thomas Department of Mathematics University of Queensland |
RFCD Codes 249901 |
Significant Achievements, Anticipated Outcomes and Future Work
By further fine tuning the underlying force field terms, further performance improvements could be achieved in our protein structure prediction package Wurst. Specifically, we developed a conceptional new approach to predict residue contacts in protein structures from sequence information alone. While other methods predict if a pair of residues is in contact, our approach uses windows of residues and recognises contact patterns using a trained neural network. Currently we are working to extend this approach to predict interchain orientation, which is then used to compute protein-protein interactions.
Computational Techniques Used
Home-built code is used for the numerical optimisation which produces the coarse-grained force field. In the heart of the code lie the functions for scoring protein sequences with candidate structures and optimisation algorithms such as gradient based (quasi-Newton) function optimiser and Monte Carlo/simulated annealing. The code to train a neural net for contact pattern recognition is based on the well optimised SNNS package, while the code to query the trained neural net are stand alone routines that can easily be used by themselves or in Wurst. All our code is written in portable ANSI C and is well optimised for scalar platforms.
Publications, Awards and External Funding
External Funding and Awards
None.
Publications
Torda, A.E., Procter J. and Huber, T.,
Wurst: Protein threading mixing statistical scoring function,
sequence profiles and optimised substitution matrices,
Nucl. Acid Res. (2004) in press.
Hamilton, N., Burrage, K., Ragan M.A. and Huber, T.,
Protein contact prediction using patterns of correlation,
Proteins (2004) in press.