Computational Intelligence, Volume 22, Number 3/4, 2006 VISIO-SPATIAL CASE-BASED REASONING: A CASE STUDY IN PREDICTION OF PROTEIN STRUCTURE JIM DAVIES Institute of Cognitive Science, Carleton University, Ottawa, Canada J ANICE GLASGOW AND TONY KUO School of Computing, Queen’s University, Kingston, Canada We show that visio-spatial representations and reasoning can be used as a similarity metric for case-based protein structure prediction. Our system retrieves pairs of α-helices based on contact map similarity, then transfers and adapts the structure information to an unknown helix pair. We show that similar protein contact maps predict a similar three-dimensional protein structure. The success of this method provides support for the notion that changing representations can enable similarity metrics in case-based reasoning. Key words: case-based reasoning, protein structure, analogy, bioinformatics, computational biology. 1. INTRODUCTION It is well known that the right representation greatly facilitates reasoning (Amarel 1968) and there is a growing recognition of the need for intelligent architectures to accommodate a diversity of representations (McCarthy et al. 2002). The guiding theory of our research is that changing representations allows reasoners to see similarities in one representation type that might be difficult to detect in another. For example, teleological representations of a human face and the front of a car may have very little semantic overlap. In this research, we focus on visio-spatial representations. In our example, representing the headlights and eyes as circles, and the grill and mouth as a centrally located hole allows connections to be drawn between these components. As people often have visio-spatial experiences when solving problems (Casakin and Goldschmidt 1999; Farah 1988; Monaghan and Clement 1999; Shepard and Cooper 1988), an important step in establishing our above theory is to show that visio-spatial representations can be used to solve a variety of problems. In this article, we provide support for this notion in the domain of protein structure prediction. We describe the problem, and then how we use visio-spatial reasoning on images to solve it. 1.1. Protein Structure Prediction A primary goal of molecular biology is to understand the biological processes of macro- molecules in terms of their physical properties and chemical structure. Because knowing the structure of macromolecules is crucial to understanding their functions, and all life cru- cially depends on protein function (Hunter 2004), an important part of molecular biology is understanding the three-dimensional (3D) structure of proteins. Proteins are composed of one or more chains of amino acid residues. The description of which residues appear and in what order is the protein’s “primary structure.” According to the laws of chemistry, the chains twist, fold, and bond at different points, forming a complex 3D shape. Subchains form regular “secondary structures,” the two main types being α-helices and β -strands. The 3D structure of a chain is its “tertiary structure,” and the overall protein Address correspondence to Jim Davies, Institute of Cognitive Science, Carleton University, Ottawa, Ontario, K1S 5B6, Canada; jim@jimdavies.org. C  2006 Blackwell Publishing, 350 Main Street, Malden, MA 02148, USA, and 9600 Garsington Road, Oxford OX4 2DQ, UK.