Molecular flexibility in crystal structure prediction

• Main Author: Kazantsev, Andrey
• Other Authors: Pantelides, Costas ; Adjiman, Claire
• Published: Imperial College London 2011
• Subjects: 539.6
• Online Access: http://ethos.bl.uk/OrderDetails.do?uin=uk.bl.ethos.542489

The packing of molecules in solids greatly affects the properties of the bulk materials. This is particularly important for the pharmaceutical industry, where the discovery of crystal forms at a late stage of process development can have disastrous consequences. As a result, the importance of polymorphism in crystal structures of organic molecules has been recognised for many years. This thesis presents computational developments that can complement experimental form screening of molecules for which conformational flexibility is significant. Current methods for crystal structure prediction are limited by the extent of molecular flexibility that can be practically handled due to the prohibitive computational cost associated with quantum mechanical calculations integrated in most of the successful approaches. In order to reduce the number of quantum mechanical evaluations, local approximate models can be defined for the estimation of the intramolecular energy, molecular geometry and the conformationally dependent intermolecular electrostatic model. A novel algorithm, CrystalOptimizer, for the accurate local minimisation of the lattice energy of crystals involving flexible organic molecules is presented. The main novelty of the algorithm is the use of dynamically constructed and updated local approximate models which essentially make available the full accuracy of quantum mechanical models at each and every iteration of the minimisation algorithm, requiring only a small number of explicit quantum mechanical calculations. This has made possible the accurate treatment of molecules involving a relatively large numbers of atoms with significant flexibility in torsional and bond angles and even bond lengths. The performance of the algorithm is critically assessed and demonstrated on a set of single and multi-component crystals. An extension of an existing algorithm for the identification of low energy crystal structures of flexible molecules, CrystalPredictor, is also described. In the proposed modification, the intramolecular energy and the molecular conformation are modelled using local approximate models. This provides a more realistic model for the effects of the flexible degrees of freedom on the molecular geometry and lattice energy. The use of deterministic low-discrepancy sequences ensures an extensive and uniform coverage of the multivariable search space. A parallelised implementation of the algorithm allows minimisations from several hundreds of thousands of initial guesses to be carried out in reasonable time. A further computational benefit is derived by the storage of the information used to construct the local approximate models in databases, which can be re-used in subsequent re-minimisation of structures with more accurate models for the lattice energy. The usefulness of these modifications is demonstrated on the ROY molecule, for which the structures of all experimentally known polymorphs are identified by the algorithm. By combining the above algorithms, a comprehensive multi-stage methodology for ab initio determination of the crystal structure of a given molecule based solely on its atomic connectivity is presented. The application of the methodology to two large and flexible molecules of pharmaceutical interest is also demonstrated.