|
|
|
||||||||||||||
 |
 |
|
|
|
||||
|
Sequence Space by Chris Dodge
Introduction Direct visualisation and exploration of biological datasets is one of the many challenges facing bioinformatics, however mathematical techniques can be used to process such datasets and present them in forms more amenable to standard visualization practices. A disadvantage of such techniques is that obtaining an intuitive feel for the data after such processing is difficult and can require an understanding of the mathematics used. In this project we have taken a method used to cluster aligned protein sequences and combined views on the resulting data with more conventional protein data viewers. As all of the viewers interact with one another, it is hoped that an understanding of the processed data can be obtained without the need to fully understand the mathematics, but rather by seeing the correspondence between conventional and mathematical views. We have taken a method developed by Casari, Sander and Valencia [1] that positions protein sequences from a multiple alignment as points in a high dimensional, generalized "sequence space". An ordination method called Principle Component Analysis is then used to reduce the number of dimensions so that we can view the data normally. Ordination, as described by Higgins [2], "is used to take a set of objects, initially arranged in a high-dimensional space and represent these in a small number of dimensions, while preserving the inter-object distances as much as possible". There are many ways in which we could define sequence space, and the advantage of the method used here if that it is based on protein sequence alone. As the input data is a set of aligned sequences, it is reasonable to assume that they will be clustered somewhere in sequence space. Using ordination to reduce the dimensionality of the data while preserving inter-object distance should yield clusters in a subspace viewable in two or three dimensions. A further advantage of this method is that we can project each residue into the subspace, and so directly determine which residues are responsible for the protein family under analysis and also the clusters (subfamilies) within the family. Additionally, the conservation of residues within the whole family and the subfamilies can be clearly identified. Conserved residues are often of functional significance, so this technique can be used to predict functional residues including those which may modulate family function (i.e. be responsible for a subfamily) without prior biological knowledge of the protein family. Implementation There are two main parts to the code:
This view shows two dimensions of the resulting N dimensions following ordination of sequence space. Usually, sequence space is reduced to six dimensions, which can be selected and viewed here. Each point on the plot represents one protein of the original multiple alignment.
This view shows three dimensions of the resulting N dimensions following ordination of sequence space. The viewer allows free or continuous rotation of the 3D space, which gives an improved view of the subfamily clustering. Each cross represents one protein from the original multiple alignment.
Each point on this plot represents a residue that has been projected into the same subspace as the 2D protein view. A correspondence between the 2D protein and 2D residue space can be seen in that the protein points and residue points are more concentrated along three directions from the origin (blue cross). Thus we can determine which residues are responsible for the clusters seem in the protein view. Additionally, the further from the origin a residue is the more conserved it is.
This view shows the residue data in three dimensions.
This simply shows each protein in the original data set with associated data. Selected proteins are highlighted in blue. Click on figure to see the enlargement.
In the image below, highlighted proteins can be seen that were made by selecting a small cluster in the 2D protein view, and highlighted residues from the corresponding area in the 2D residue view. A good correspondence is shown between the selected proteins and those proteins in which the selected residues appear. If we had selected the most conserved residues in this cluster (i.e. those at the extremity of the 2D residue view) then we could perhaps make the prediction that they are somehow involved in the specific function of that cluster.
Click on the image to see the enlargement.
To Georg Casari, Chris Sander and Alfonso Valencia for the original work on the algorithm used here, details of which are found in [1]. Again to Chris Sander for suggesting and overseeing the work done during this project. To Dirk Walther for kindly allowing me to use his Java 3D protein structure viewer (Webmol). Article by: Chris Dodge Resources and further information http://www.ebi.ac.uk/ http://industry.ebi.ac.uk/SeqSpace/ Chris Dodge- formerly EMBL-EBI - now Synomics. chris@dodgies.demon.co.uk Industry Programme http://industry.ebi.ac.uk/ Stanford University Genomic Resources web server http://genome-www.stanford.edu/ http://genome-www.stanford.edu/structure/webmol/ Dirk Walther walther@cmpharm.ucsf.edu Literature:
External sites are not endorsed by EMBL-EBI |
||||||
|
||||||
|
|
|
Direct questions or comments to Bioinformer Editor. This page last modified Friday, 16 July, 1999. (c) 1997-1999 EMBL-EBI. All Rights Reserved. |