Abinitio Protein Structure Prediction

Vamsi Krishna Kosaraju, Xue Feng

(Dept of Chemistry)

Introduction

The word “abinitio” means “from the beginning”. The goal of any abinitio protein structure
prediction protocol is to predict the native fold from amino acid sequence alone. The problem
of finding a native structure for a given protein sequence has remained unsolved for a long time.
This is with the exception of comparative modelling where the knowledge of homologous
structure is used to predict the structure of new protein or threading methods where
secondary structure (fold) information is used to model the new protein.

In last 10 years, very significant developments have been reported in the area of abinitio
protein structure prediction in terms of search methodology required for searching
astronomical conformational space (i.e. Monte Carlo simulations or Molecular dynamic
simulations) and the scoring functions used to identify the correct native fold. In spite of
progress, we still face several roadblocks. For example, there is no single method that fully
complies with the definition of “abinitio” - most of the methods use partly structural knowledge
from the protein databank. These methods currently require huge computational resources,
therefore the length of the protein structure to be predicted is limited to 100 - 150 residues.

Using abinitio protocol, we made an attempt to predict the structure of N-terminal fragment of
p35. This protein is known to be an activator of Cyclin - Dependent Kinase 5 and analogous to
other activators (cyclin) of the CDK family. CDK5-p35 complex has been identified to play a
significant role in several neuronal functions and deregulation. It has been reported that
fragmentation of p35 triggers the collapse of neurons which eventually lead to formation of
plaques (major hallmark of Alzheimer’s disease). The C-terminal portion of p35 along with
CDK5 has been solved by X-ray Crystallography and N-terminal portion (147 residues – p10)
remains to be solved which would unravel the mystery of Structural activation.
 Start

Fold Assignment

failed

Target – Template Alignment

Abinitio
Modeling

Comparative Modeling

Model Assessment and

validation

End

Typical Protein Structure Prediction Process

 Start

Primary Amino acid Sequence

Conformational Search

Prediction of native Fold

Model Assessment

End

Simple Flow of Abinitio Prediction

 CDK5

P25

p10

Rosetta Methodology:

Rosetta is one of the best implemented algorithms that come close to the definition of abinitio
structure prediction. Rosetta prediction begins with the identification of fragments (3mers and
9mers) from structure databases (I-sites) that have consistency with local sequence preferences.
All fragments are assembled into models with global properties. Finally, scoring function is
employed to identify the best model from decoy population.

Cluster Computing:

Rosetta structure prediction demands huge computational resources to search through gigantic
number of protein conformations. The AbintioRelax algorithm implemented in the rosetta
package is originally responsible for structure prediction. This process is operated in two steps:
The first step is coarse-grained fragment search through conformational space using knowledge
based scoring function to identify native-like structure. The second step is all atom refinement
which is a computational intensive process. We used atlas4 and atlas5 clusters from the
Computer Centre to run multiple of serial jobs to generate 200,000 decoys. The top 10% decoys
based on the Rosetta score were selected and further used for clustering to identify the best
global conformation.

Results and Analysis:

Our preliminary results show that the final model that we selected is one of the most highly
conserved structure among the decoys, identified through clustering analysis where the RMSD
among the top decoys was plotted against the Rosetta score. Since there is no additional
experimental data available to validate the p10 structure, we used structure match algorithm to
identify the possible matches from protein databank (PDB). Surprisingly, the p10 model
showed minor similarities with cyclins. Additional evidence from in vitro experimental studies is
also consistent with structural model i.e. phosphorylation sites. We also employed other
structure prediction algorithms to predict the same structure. Results are encouraging and we
will be taking the structure to the next level.

Structure Model of p10 predicted from abinitio simulations

Conclusion:

The model of N-terminal fragment of p35 (p10) that we built using Rosetta abinitio protocol
appears to be promising from preliminary analysis.