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

Peak-picking is the lowest-level task of the interpretation of two-dim ensional, and multidimensional Nuclear Magnetic Resonance (NMR) spectr a in general, for protein structure determination. It consists of indi viduating peaks on two-dimensional frequency spectra, for further elab oration. The performances of several feedforward artificial neural net works trained with back propagation with temperature on the task of pe ak-picking are compared. The best one averages less than an approximat e 5% error on well-defined spectral regions. The performances of the n etwork are comparable with those of a human expert; the consequences o f this fact on the possibility of improving further the performance of the network are discussed.