NEURAL NETWORKS FOR THE PEAK-PICKING OF NUCLEAR-MAGNETIC-RESONANCE SPECTRA

Peak-picking is the lowest-level task of the interpretation of two-dimensional, and multidimensional Nuclear Magnetic Resonance (NMR) spectra in general, for protein structure determination. It consists of individuating peaks on two-dimensional frequency spectra, for further elaboration. The performances of several feedforward artificial neural networks trained with back propagation with temperature on the task of peak-picking are compared. The best one averages less than an approximate 5% error on well-defined spectral regions. The performances of the network are comparable with those of a human expert; the consequences of this fact on the possibility of improving further the performance of the network are discussed.