ITEE Innovation Expo 2008 : Project Details

Geometric hashing for the discovery of protein binding surfaces

Protein-Protein interactions play a fundamental role in many biological processes, hence the ability to determine whether two given proteins will bind is of great importance in many fields such as drug design and understanding disease. The function of a protein is closely related to the structure of its interaction site. With this in mind, this project developed a kernel function able to separate pairs of proteins into binding or non-binding classifications based on their structure.

A protein is made of a string of amino acids, each forming a link in the underlying backbone. Amino side chains are atoms that extend from the backbone and allow for atomic interactions. A distribution of the distances between all the interaction points provides a numeric description of the protein binding site, retaining the structural information of the site invariant of position or rotation. A protein-protein interaction is hence described by two histograms; hashes of two binding sites classified as binding.

Support Vector Machines learn where to split samples of data in a given space to separate the classes, in this case a pair of histograms is classed as binding or non-binding. Real data however is rarely linearly separable by class, instead a kernel function measures a difference between two given samples and is designed to separate the classes, allowing an SVM to learn to classify new unknown samples.