A common sense approach to peak picking in two-, three-, and four-dimensional spectra using automatic computer analysis of contour diagrams

Abstract

The analysis of multidimensional NMR spectra has been a challenging problem since the earliest two-dimensional experiments were reported as stacked plots (I). The first step in analysis involves obtaining a list of chemical-shift coordinates for the cross peaks. Initially, simple programs were used to generate contour plots of twodimensional NMR data, which were analyzed manually. As the utility and versatility of multidimensional NMR spectroscopy grew, several automated methods of peak picking have been developed. This Communication describes a new peak-picking algorithm which is based on contour diagrams and designed for the automated interpretation of higher dimensional 3D and 4D spectra. The oldest and most robust method of analysis is the manual interpretation of 2D contour plots. The strength of manual peak picking results from the relative ease with which the human eye can discriminate real peaks from artifacts and noise. As the proteins studied have become larger, the number of spectra to be analyzed and the number of cross peaks within a spectrum have increased dramatically, with the result that significantly more time and energy are required for the tedious manual peakpicking step. Interactive graphics software, which dynamically maintains a list of peak positions, has to some extent helped with this time-consuming step, particularly with regard to bookkeeping. Although more time will be saved with the automation of peak picking, manual inspection of spectra with an interactive graphics program will always be necessary to verify and edit automated results. Approaches to automated peak picking can be divided into three types: (a) thresholdbased methods; (b) multiplet-symmetry-based methods; and (c) peak-shape-based methods. The simplest automated peak-picking algorithm is based primarily on the intensity of local extrema exceeding a threshold value (2). Uninteresting regions of the spectrum, such as tl noise ridges, are defined to avoid selecting peaks along these artifacts. Since some real peaks have very low intensities, the threshold must be set close to the noise level, which unfortunately results in a very large number of local extrema being picked due to the noise. Thus, by itself the threshold method fails by selecting too many or too few peaks, but is ideally suited as a filter for more sophisticated methods.