previous section previous page next page next page

Online Lectures on Bioinformatics


Protein Structure

Prediction Methods

Knowing a protein's 3-dimensional structure helps us to understand its functionality and provides means for planning experiments and drug design. Experimental methods given by X-ray crystallography and NMR spectroscopy to determine protein structure are essential. The Brookhaven Protein Data Bank (PDB) is the repository for those structures. Files including atom coordinates which are suited for visualization by graphical molecule viewers like rasmol can be obtained at this site. PDB is also searchable with a sequence as a query, e.g. with the BLAST service located at NCBI with a polypeptide as a query. Nevertheless, experimental methods are technically very difficult and expensive and the gap in the number between sequenced proteins and known structures increases. Thus, model building of proteins is of great importance.

When a protein first is unfolded in vitro and then released again, it folds back to the same three-dimensional structure it had before. Thus, various prediction methods are based on the assumption, that the three-dimensional protein structure is determined by its primary structure.

Structure prediction methods are coarsely divided into three categories [5]:

  1. Comparative modelling
    If the sequence to model has a homologue in the PDB which it is very similar to, the homologue may be used as target and a structural model is built on the basis of this template.

  2. Fold recognition
    In absence of a significantly similar sequence with known structure, various methods put together in the term "Fold Recognition", exist to find a suited template for model building. In opposite to Comparative Modelling, these methods try to determine how well a known structure fits the sequence to model.

  3. Ab initio prediction
    In contrast to the above methods, the goal of ab initio prediction is to build a model for a given sequence without using a template by minimizing knowledge based energy functions (Baker et al., Univ. of Washington, Seattle) and lattice models (Samudrala et al., Levitt Lab, Stanford).

The most reliable results are obtained, if the structure of a very close related sequence is known and comparative modelling can be applied. If not, Secondary Structure Prediction provides a key for further analysis. Methods summarized in the next section are used to predict secondary structures. This information may be used for aligning secondary structures and detecting a suitable template for fold recognition methods. If latter methods fail to work, ab initio prediction needs to be chosen.

Comments are very welcome.