Fast protein structure-function prediction with new MNSP descriptors

Janez Konc, Milan Hodoscek, Dusanka Janezic

National Institute of Chemistry, Hajdrihova 19, SI-1000 Ljubljana, Slovenia

We designed a new algorithm for predicting a protein's function from its structure. Large surface depressions in a protein's surface often coincide with a binding pocket of an endogenous ligand. Functional groups of amino-acids, which protrude into such clefts, define interactions between an enzyme and a ligand. By comparing the clefts of two proteins we can measure similarity of their biochemical functions. The algorithm first calculates a derivative of smooth protein surface. Depressions are then discovered by the following procedure. We take a set of evenly dispersed surface points and for each one we construct a vector originating in this point, perpendicular to the surface and orientated away from the protein. If a vector intersects with the surface at some other point, its origin point is in a concave part of the protein. By assigning each such point a property of an underlying aminoacid, we describe possible interactions of the protein surface with ligand. From these basic elements we recursively construct MNSP (Multilevel Neighborhood of Surface Properites) descriptors. Each surface point gets its own descriptor, which consists of the property of this point and the properties of points closer than some cutoff distance. Once we have created two sets of MNSP descriptors, one set for each protein, we are able to compare them with the following equation: number of descriptors that are common to both proteins are divided by the total number of descriptors. Because the algorithm uses descriptors that contain only implicit information about the 3D orientation, it is much faster compared to standard methods that use 3D alignment. Built-in efficient algorithm for finding clefts makes it much easier to use because no pre-proccessing of the protein is needed. For its simplicity and fast calculation it may be used for cross comparison of an entire protein database, giving as a result a simple value of similarity.

Email: konc@cmm.ki.si

Address: Čufarjeva 1, 1000 Ljubljana