Neural networks applied to protein structure

The Technical University of Denmark, Lyngby, Denmark.

Abstract

Neural networks are computational strategies inspired by the brain, which have remarkable cognitive properties. They are trained rather than programmed and this process requires access to a sufficiently large body of reliable data. Applied to proteins, they thus need data which link amino-acid sequence to secondary structure motifs and three-dimensional atomic coordinates. Given these, and a suitable training algorithm, the hope is that the trained network will be able to predict secondary and tertiary structures when it is presented with an arbitrary input sequence. This approach has proved to be capable of predictive reliability around 70% for secondary structures and it has also done rather well with the tertiary structures of some proteins, provided that there is a high degree of homology between the training and test proteins. The fundamentals of the neural network approach will be described, and an attempt will be made to evaluate its future potential as more data become available.