Automated Analysis for NMR Spectra Using Artificial Intelligence |
The aim of this work is to develop an expert system for automated analysis of protein nuclear magnetic resonance (NMR) spectra. NMR spectroscopy provides an important technology for determining three-dimensional (3D) structures of biological molecules. In order to interpret these data, the spectroscopist must correlate each peak in the spectrum with a specific atom in the protein sequence. Once these "resonance assignments" are available, the intensities, splittings, and other spectral features of the NMR peaks can be interpreted to provide structural information and a 3D image of the protein's structure. We have developed computer software capable of fully automated analysis of resonance assignments from NMR spectra. This expert system, AUTOASSIGN, utilizes concepts from artificial intelligence to rapidly complete the analysis of resonance assignments. AUTOASSIGN is an expert system for automated analysis of resonance assignments in protein NMR spectra. Proteins themselves are polymers of amino acids; there are twenty naturally occurring types of amino acids which are put together in different sequences to make different proteins. These chains then fold into unique 3D structures, which are the targets of structure determination and rational drug design. Different NMR experiments generate spectra which contain different kinds of information. Our laboratory is expert both in developing new NMR experiments which provide structural information for proteins, and in applying this technology to determine 3D structures of proteins. Typically, these spectra correlate NMR peaks in 2D plots or in 3D or 4D volumes. Such 3D or 4D data are difficult to analyze by humans, and can be more systematically analyzed by computer.
AUTOASSIGN integrates an object-oriented knowledge base of amino acid structures with methods of constraint propagation from artificial intelligence. The input to AUTOASSIGN is a set of 3D and 4D NMR spectra and the amino acid sequence of the protein (known from molecular biology methods). The output is a list assigning each peak in these spectra to a specific atom in the protein's structure. This mapping (or assignment process) allows further interpretation of the NMR data, which are used to generate 3D structures of the protein using other software.
The program has been developed and tested on literally hundreds of simulated NMR data sets, and on real data sets for five different proteins. Each real NMR data set requires 1 - 2 weeks of NMR data collection. The current version of AUTOASSIGN is highly robust and achieves complete and error-free analysis of real data for proteins containing 50 to 160 amino acids in 10 - 30 min on a Sun Sparc2 workstation. The next step of the work involves automation of the second stage of strucuture determination, namely the generation of distance constraints from multidimensional NMR spectra. These data can then be used to compute 3D structures of proteins using existing computer programs. In this way the complete process of structure determination can be automated.
