Software Approaches for Determination of 3-Dimensional Molecular Structures from Multi-Dimensional NMR

Abstract

Two and higher dimensional NMR spectroscopies offer extraordinary power for detailed structure elucidation of proteins, nucleic acids and other important biomolecules. The methodology of elucidating biopolymer structures at atomic resolution from NMR spectroscopic data incorporates primarily NOESY experiments, but also may add spin-spin coupling constants and other information measured from COSY and other NMR experimentation. There are several important challenges that must be overcome in order for this methodology to be generally applicable to a broad range of biomolecules. One of the most important long-term goals of this research arena is to be able to determine structures with a confidence level sufficient to allow utilization of the structural information without confirmatory experimentation such as single X-ray structures.

In order to achieve this long-term goal, a number of issues must be dealt with: 1) primary 2D (nD) data reduction must incorporate techniques to allow accurate determination of sufficient NOESY cross-peak volumes; 2) computational schemes must be developed which not only determine refined molecular structures from the experimental information, but which also reflect confidence levels in the determined structures based on intelligent error analysis through all procedures; 3) corrections for these calculations must include, at a minimum, correction for dynamics variations, correction and recovery for missed spectral assignments, and wide sampling of possible molecular geometries.

Development of automated and assisted multi-dimensional NMR spectral assignment techniques is critical for many of these studies, where hundreds or even thousands of cross peaks may be significant for analysis. Techniques incorporating automated NOESY walks, pattern recognition for identification of specific sites, and other techniques will have to be used together for optimal spectral assignment. Of course 3- and higher dimensional spectroscopy will also assist in this area.

At Syracuse University, one of the primary goals realized at this time, is optimal preparation of the data for analysis. Use of non-linear processing techniques based on the maximum likelihood method (MLM) and specialized protocols increases the number of cross peaks that can be used for 3D structure determination. Experiments and computation underway indicates that these non-linear techniques have broad applicability and that, across a range of spectral conditions, they are robust and quantitative (or where dynamic range is too high, corrections may be possible to quantitate the smallest peaks). Preliminary results on synthetic and mixed data show superior quantification of cross-peak volumes over a range of peaks sizes exceeding 50:1.

A second area of investigation at Syracuse University involves utilization of parallel and distributed computing methods. These are initially being applied to two applications: 1) 3D NMR data processing and 2) using a genetic algorithm for NMR molecular modeling.

The basic idea is to utilize, in parallel, workstation and other computers coexistent on local and wide-area computer networks. In cases where specialized computing hardware such as MIMD parallel computers (examples: Alliant FX/80, Hypercubes, etc.) or SIMD architectures (example: Connection Machine) are available, an additional opportunity is present to dissect a computational application and allocate appropriate portions to that specialized hardware. This type of distribution of processing tasks is included in the work underway. Thus, on a computational network such as the one existing at Syracuse University which incorporates a large number of Sun work stations, IBM RISC System 6000’s, and a large configuration Connection Machine 2, as well as an Alliant FX/80, more than an order of magnitude speedup in realization of compute and I/O applications such as 3D NMR data processing and matrix manipulation found in aspects of NMR molecular modeling.