A back-propagation neural network predicts absorption maxima of chimeric human red/green visual pigments

Abstract

The absorption spectra of human red and green visual pigments have peak wavelengths, lambda max, that differ by 31 nm, yet the opsins differ in only 15 amino acids. Mutagenesis studies have demonstrated that seven of the 15 amino acids determine the spectral shift. We trained neural networks to predict the lambda max of any red/green chimeric protein. Seven mutants were excluded from the original training set. The trained networks were able to predict the lambda max for the excluded mutants. As an additional test, five new chimeric pigments were constructed and lambda max determined. The neural networks correctly predicted the lambda max of all five mutants. The use of neural networks is a novel approach to the problem of wavelength modulation in visual pigments.