Neural associative memories for molecular databases

We describe recent and on going research into the use of neural associative memories in a molecular database system. The aim of the work is to allow rapid access to a large database of 3D images of molecules using possibly incomplete information provided by the user. Two approaches are being pursued. Both are based upon the use of a class of neural networks known as correlation matrix memories (CMMs) which possess certain computationally attractive features. The first approach uses CMMs to match featural descriptions derived from 2D bond diagram molecular models. The second is aimed at matching structural descriptions derived from complex and flexible 3D molecular models. We focus on the first approach but briefly outline the second method. We describe the architecture of the CMMs, and compare them to conventional database systems. The results of a theoretical study into CMM matching performance are presented, and used to illustrate the potential advantages of our strategy