Machine Learning and Friends Lunch

From Sequence to Structure: A Holistic Approach

Abstract

Proteins perform a variety of functions inside all living things. To determine the role a protein plays inside the body, its three-dimensional shape has to be known. This knowledge also helps to understand diseases and to design drugs to cure these diseases. The problem of computationally determining the shape (or structure) of a protein is largely unsolved, despite several decades of research in this area. If this problem were solved, our understanding of biology and medicine would change quite radically.

For the purposes of our research we view protein structure prediction as search for the global minimum in a high dimensional energy function. This is no easy task, numerous local minima and the huge configuration space contribute to make it difficult. In this talk, I will introduce model-based search which is a new high dimensional search algorithm based upon active learning. Model-based search integrates domain knowledge and information generated during search into a model. The model is used to focus computation. Our results indicate that model-based search is significantly better at finding lower energy minima than the best current technique. We expect model-based search to have equally positive results when used on other problem instances of high dimensional extremum finding.

Back to ML Lunch home