dTASSER

dTASSER is the implementation of the PROSPECTOR/TASSER methodology for protein structure prediction from J. Skolnick into the Berkeley Open Infrastructure for Network Computing (BOINC) hosted at PREDICTOR@HOME in C. L. Brooks, III lab.

Despite considerable efforts in recent years, predicting the native structure of a protein from its amino acid sequence is still one of the most important and challenging problems in structural biology. Based on recent studies, it appears that a combination of theoretical methods and new computational technologies are necessary to further advance this area of research. PREDICTOR@HOME is a world-community experiment and effort to use distributed world-wide-web volunteer resources to assemble a supercomputer aimed at testing and evaluating new algorithms for protein structure prediction. Additionally, new structure prediction algorithms have emerged in the last few years: one of the most recent is TASSER (Threading / ASSembly / Refinement).

In the first stage of the project the resources and computing power of a distributed system such as PREDICTOR@HOME are initially used to critically test TASSER on a number of biological systems, including both soluble and transmembrane proteins. Once validated, our goal in the near future is to use these tools in other projects such as the structural prediction of viral capsids or aiding to the efforts in trying to cure diseases like tuberculosis.

How it works

J. Skolnick's methodology is a hierarchical approach that consists of template identification (PROSPECTOR), followed by tertiary structure assembly by rearranging continuous template fragments (TASSER). Assembly occurs using replica exchange hyperbolic Monte Carlo (RXHMC) sampling under the guide of an optimized, reduced force field that includes knowledge-based statistical potentials and spatial restraints extracted from threading alignments. Models are selected from the RXHMC trajectories in the low-temperature replicas using a clustering algorithm (sPICKER).

The pipeline

PROSPECTOR

[PROTEINS 56:502 (2004)]

PROtein Structure Predictor Employing Combined Threading to Optimize Results: Third generation of threading algorithms.

For a given Target:

• Thread the sequence through a library of representative templates
• Threading is an iterative sequence-structure alignment approach
• The scoring function consists of:
  • Sequence profiles from FASTA / BLOSUM_62
  • Secondary structure propensities from PSIPRED
  • Consensus contact predictions from previous alignments
• Finally, the templates are categorized as easy or hard on the basis of Z-score significance and alignment consistency

TASSER

[PNAS 101:7594 (2004)]

Threading ASSEmbly Refinement: Direct descendant of TOUCHSTONE; ab initio protein structure prediction. [PNAS 98:10125 (2001), Biophys. J. 85:1145 (2003)]

CAS force field: Cα Atoms and Side chain centers of mass (SG) are used to represent the protein. This force field uses the following information:

• Predicted secondary structure propensities
• Consensus predicted SG contacts
• Protein specific SG-pair potentials
• Statistical propensities: local Cα correlations, hydrogen bonding and hydrophobic burial interactions

So, for a given threading Template:

1. Build initial full-length model
2. Submit model to replica exchange Monte Carlo sampling.
3. The conformational movement updates consist on:
   • Off-lattice: rigid fragment translation / rotation
   • On-lattice : bond movements
4. Save trajectories for further analysis (clustering)

sPICKER

[J. Comput. Chem. 25:865 (2004)]

Structure PICKER: Direct descendant of SCAR [J. Comp. Chem. 22:339 (2001)]

• Two step clustering algorithm: 1) each replica, 2) their centroids

Identifies near-native folds by clustering structures:

• One step; shrunken representative set of conformations
• The (combined) models of the highest structure density are significantly closer to native than the lowest energy structure [PROTEINS S7:91 (2005)]

Performance

• Success rate of best models is strongly coupled to template quality
• In a significant number of cases, TASSER can improve the models over the best template alignments

[CASP7 Meeting]

Can the sampling be improved?

Several results suggest that it can be done if:

• Running longer simulations; this brings the top models closer to the native structure
• Running TASSER with different random seeds; this helps to explore more throughly the conformational space

but...

TASSER extended sampling takes more than 2 weeks on a single processor for a 400 residue protein!

Combining protein structure prediction and parallel distributed computing: dTASSER

02/2007 Mauricio Carrillo-Tripp @ The Scripps Research Institute