| Abstrakt: |
In this work we propose the use of a neural network to predict the crystal mode of packing of small organic molecules from just their 3-D molecular structure. A sample of 31 molecules, of quite different chemical characters and known crystal structures, has been employed. These molecules, encoded by the 1-D Fourier transform of their 3-D point charge distributions, are used as input in a Kohonen neural network. Although no molecular packing information is given, the resulting similarity output maps self-classify the molecules after the type of H-bonding pattern they present, or according to their observed molecular packing mode, when no H-bonds exist. The corresponding crystal packings of these molecules were encoded similarly by the Fourier transform of a finite cluster of molecules sought from their known crystal structures. The self-classification of these encoded packings, on the Kohonen map, presents good correlation with the classification found for the isolated molecules, and both correlate well also with the visually observed types of packings in the crystal structure. Thus, it seems that the Fourier transform of an isolated molecule includes enough packing information to allow its classification into packing modes. Finally, the same neural network, trained with part of the set of 31 molecules, supervised with their crystal packing, is used to predict the encoded packing of molecules not included in the training, in order to classify them into a mode of packing. |
