Prediction of three-dimensional molecular structures using information from infrared spectra

Abstract

A method has been developed for obtaining the three-dimensional (3D) structure of a molecule directly from the information in an infrared spectrum. By training a counterpropagation network with coded 3D structures and IR spectra it is possible to simulate the structure code of a query spectrum. To obtain an initial structure, the simulated code is compared with structure codes generated from a database of over 100,000 molecules. The molecule with the most similar structure code defines the so-called initial model for the 3D structure. By means of stochastic methods, this initial structure is modified in various ways to gain the best similarity with the structure code from the neural network. Thus, information from infrared spectra can be used to generate the corresponding 3D structure. The scopes and limitations of the method will be discussed with different examples of compounds using experimental IR spectra from various applications. The intention of this method is to assist in the interpretation of IR spectra and in structure elucidation in routine analysis.