September 25, 2005, ACD/Labs SMASH 2005 Seminar, Verona, Italy

Use of Genetic Algorithms with NOESY and ROESY Data to Predict the Stereochemistry and 3D Configurations of Small Molecules

Yegor D. Smurnyy, Mikhail E. Elyashberg, Kirill A. Blinov, Brent A. Lefebvre

Abstract

Presented here is a genetic algorithm-based method to predict relative stereochemistry and 3D configurations of small molecules (100-1000 Da) based on NOESY and ROESY data. This tool is ideally suited to natural products and can help decrease our dependence on expensive X-Ray Crystallography.

The solution can be found in a global optimization of a given target function. To construct this function, the set of NOE signals are treated as additional distance restraints and used in a molecular mechanics (CharMM parameterization) calculation. After optimization, the resulting root mean square deviation (RMSD) between d_real and d_NOE is calculated. Here d_real represents a distance obtained from molecular mechanics, and d_NOE is estimated from NOESY or ROESY peak intensities.

The target function (RMSD, in this case) can be minimized in numerous ways. Experience shows that for small molecules (100-300 Da, 3-6 stereocenters) a straightforward approach over all stereoisomers (8-64 variants) is acceptable. For more complicated cases, implementation of an alternative approach is necessary. In this case a genetic algorithm is proposed based on a stochastic approach, which mimics natural evolution. A possible solution is presented by a vector in which key features of a solution (stereocenter configurations) are encoded. A pool (10-30 items) of such vectors is then a subject for crossover and random mutation. The crossover consists of splitting two "ancestors" into parts and assembling "offspring" from these parts. Mutation represents random alternating of a position in a solution. The goal is to reach as diverse regions as possible of the solution space to avoid being trapped in a local minimum. After these procedures, the offspring with the worst RMSD values are eliminated to maintain the pool size and the evolution continues.

Experience shows that this simple scheme can be applied to molecules with 8-12 stereo-centers. However, it fails for more complex structures; as the solutions pool degenerates. A number of special procedures have been created for these complex systems which allowed the correct outcome to be achieved for many natural products, even when they were as complicated as brevetoxin. Upon numerous tests, the correct answer is found after only 1000-1500 Generations, (contrast that to 8 million possible stereoisomers!).

Figure 1: Brevetoxin 2, a natural product of 900 Da and 23 stereocenters; resulting in 8 Million stereoisomers.

Details of the special procedures that were necessary to make the most complicated systems solvable will be provided. Examples demonstrating the algorithm's efficiency for solving the stereochemistry and 3D configurations of structures are presented here.

Download the poster in Adobe Acrobat format (226 Kb PDF file).