Performance of a neural-network-based determination of amino acid class and secondary structure from 1H-15N NMR data

A neural network which can determine both amino acid class and secondary structure using NMR data from 15N-labeled proteins is described. We have included nitrogen chemical shifts, 3JHNH alpha coupling constants, alpha-proton chemical shifts, and side-chain proton chemical shifts as input to a three-layer feed-forward network. The network was trained with 456 spin systems from several proteins containing various types of secondary structure, and tested on human ubiquitin, which has no sequence homology with any of the proteins in the training set. A very limited set of data, representative of those from a TOCSY-HSQC and HNHA experiment, was used. Nevertheless, in 60% of the spin systems the correct amino acid class was among the top two choices given by the network, while in 96% of the spin systems the secondary structure was correctly identified. The performance of this network clearly shows the potential of the neural network algorithm in the automation of NMR spectral analysis.