Natural versus Random Proteins: Nouvel Neural Network Approach Based on Time Series Analysis

A bstract We study the set of about 35000 primary structures of natural proteins of length more than 360 residues and the same size set generated via partial or total randomization. Associated to every sequence composed of 20 amino acids, a time series is formed from hydropathy values of the first 360 residues. To measure the absolute deviations of hydropathy index on different time scales, the 24-dimensional vector of total log-amplitudes is introduced. We describe then a configuration of the 1-hidden layer neural network which is trained to solve the binary classification problem of natural and random sequences. A satisfactory distinguishing accuracy random/natural of 88% is obtained.