Date Published: December 01, 2016
Publisher: International Union of Crystallography
Author(s): Isaac Sugden, Claire S. Adjiman, Constantinos C. Pantelides.
This article describes an important improvement in the CrystalPredictor II code: adaptive Local Approximate Models (LAMs). This improvement allows the most efficient use of computational effort to cover a flexible molecule’s conformational space, and is illustrated with a crystal structure prediction (CSP) investigation into the sixth blind test molecule 26.
The primary aim of crystal structure prediction (CSP) techniques is to produce a ranked list of all the potential crystal structures for a molecule or set of molecules. Because of the significant effect that crystal structure has on solid-state properties, such as colour, solubility and hygroscopicity, such a ranked list offers a wealth of information and many opportunities to improve the development of new crystalline materials (Price et al., 2016 ▸; Neumann et al., 2015 ▸). In the case of the pharmaceutical industry, the appearance of a new or unexpected form or polymorph can have major legal and economic ramifications, particularly if solubility/bioavailability are affected, as illustrated by the cases of the appearance of Ritonavir form II (Chemburkar et al., 2000 ▸) and the Zantac litigation (Seddon, 1999 ▸). Furthermore, the ability to tune a molecule’s solid-state properties through predictive approaches would be very useful to industries that rely on crystalline materials. Therefore, significant benefits are offered by the possibility of predicting a molecule’s crystal structure(s), especially when this is possible via ab initio techniques that rely only on molecular structure information.
The recent blind test on crystal structure prediction methods, organized by the Cambridge Crystallographic Data Centre, sought to evaluate the capabilities of current computational methods in predicting the crystal structures of organic molecules. Five targets were chosen, representing challenges to the crystal structure prediction community. The two versions of CrystalPredictor were deployed by two of the participating groups, in combination with CrystalOptimizer, to identify Z′ = 1 structures. This approach resulted in the identification of the known experimental structures within the predicted energy landscapes in most cases. However, in the case of molecule (XXVI), shown in Fig. 1 ▸, the multiple flexible torsion angles present particular difficulties, which are discussed here and motivate the development of an improved version of CrystalPredictor II.
This section presents an adaptive algorithm that automatically positions LAMs at points in the search domain of the independent degrees of freedom, where necessary to ensure the required degree of accuracy. Firstly, the revised algorithm for generating new LAMs is summarized in §3.1, with examples of its implementation given in §§3.2 and 3.3.
The proposed algorithm is now applied to molecule (XXVI) from the sixth blind test (cf. §2). As shown in Fig. 1 ▸, the molecule has 7 flexible degrees of freedom, several of which have broad ranges of flexibility.
The 2016 blind test (Reilly et al., 2016 ▸) revealed that achieving an appropriate balance between computational cost and accuracy in the global search for crystal structures remains a challenge for large molecules. The algorithm presented in this paper addresses this issue by introducing the adaptive placement of LAMs within the CrystalPredictor II algorithm, an improvement on the uniform grid scheme which had proved too computationally demanding to apply to molecule (XXVI). A higher density of LAM points is automatically achieved in chemically interesting areas of conformational space, thereby resulting in a more efficient use of expensive ab initio calculations. This, in turn, allows the CrystalPredictor II algorithm to handle larger molecules and to explore larger areas of conformational space, through an effective global search methodology. The successful application of this new approach to molecule (XXVI) realises one of the aims of the blind tests, namely to drive innovation in CSP by providing unique and challenging molecular systems.