Peter T. A. Galek

Successful prediction of a model pharmaceutical in the fifth blind test of crystal structure prediction

The range of target structures in the fifth international blind test of crystal structure prediction was extended to include a highly flexible molecule, (benzyl-(4-(4-methyl-5-(p-tolylsulfonyl)-1,3-thiazol-2-yl)phenyl)carbamate, as a challenge representative of modern pharmaceuticals. Two of the groups participating in the blind test independently predicted the correct structure. The methods they used are described and contrasted, and the implications of the capability to tackle molecules of this complexity are discussed.

Knowledge-based model of hydrogen-bonding propensity in organic crystals

A new method is presented to predict which donors and acceptors form hydrogen bonds in a crystal structure, based on the statistical analysis of hydrogen bonds in the Cambridge Structural Database (CSD). The method is named the logit hydrogen-bonding propensity (LHP) model. The approach has a potential application in identifying both likely and unusual hydrogen bonding, which can help to rationalize stable and metastable crystalline forms, of relevance to drug development in the pharmaceutical industry. Whilst polymorph prediction techniques are widely used, the LHP model is knowledge-based and is not restricted by the computational issues of polymorph prediction, and as such may form a valuable precursor to polymorph screening. Model construction applies logistic regression, using training data obtained with a new survey method based on the CSD system. The survey categorizes the hydrogen bonds and extracts model parameter values using descriptive structural and chemical properties from three-dimensional organic crystal structures. LHP predictions from a fitted model are made using two-dimensional observables alone. In the initial cases analysed, the model is highly accurate, achieving approximately 90% correct classification of both observed hydrogen bonds and non-interacting donor-acceptor pairs. Extensive statistical validation shows the LHP model to be robust across a range of small-molecule organic crystal structures.

Knowledge-based H-bond prediction to aid experimental polymorph screening

With an ever increasing regulatory and financial emphasis on solid form screening in the pharmaceutical industry, a knowledge-based method has been developed to assess crystal stability based on hydrogen bonding. An application is illustrated for the polymorphic drug ritonavir (Norvir™). The method quickly suggests a real threat of polymorphism in this compound by quantifying the likelihood of competing H-bonds, and strongly supports the relative stability of form II over form I. For the first time, H-bond geometry data are also reported following structure redeterminations deposited recently in the Cambridge Structural Database. The methods speed and versatility are emphasized, facilitating future application in assisting solid form selection of a diverse range of compounds.

Quantifying Homo‐ and Heteromolecular Hydrogen Bonds as a Guide for Adduct Formation

An investigation into the predictability of molecular adduct formation is presented by using the approach of hydrogen bond propensity. Along with the predictions, crystallisation reactions (1a-1j) were carried out between the anti-malarial drug pyrimethamine (1) and the acids oxalic (a), malonic (b), acetylenedicarboxylic (c), adipic (d), pimelic (e), suberic (f), azelaic acids (g), as well as hexachlorobenzene (h), 1,4-diiodobenzene (i), and 1,4-diiodotetrafluorobenzene (j); seven (1a to 1g) of these successfully formed salts. Five of these seven salts were found to be either hydrated or solvated. Hydrogen bond propensity calculations predict that hydrogen bonds between 1 and acids a-g are more likely to form rather than the H bonds involved in self-association, providing a rationale for the observation of the seven new salts. In contrast, propensity of hydrogen bonds between 1 and h-j is much smaller as compared to other bonds predicted for self-association/solvate formation, in agreement with the observed unsuccessful reactions.

Knowledge-based hydrogen bond prediction and the synthesis of salts and cocrystals of the anti-malarial drug pyrimethamine with various drug and GRAS molecules

We have previously reported on hydrogen bond propensity calculations for the potential formation of adducts between pyrimethamine and dicarboxylic acids. Here we extend the range of potential synthon interactions using a variety of potential coformers. Specifically calculations were performed to predict the possibility of the formation of molecular adducts, 1a–1h, between the anti-malarial drug pyrimethamine (1) and (a) carbamazepine, (b) theophylline, (c) aspirin, (d) α-ketoglutaric acid, (e) saccharin, (f) p-coumaric acid, (g) succinimide and (h) L-isoleucine. The bonds of highest propensity were predicted between 1 and coformers (b–h), indicating a high probability of formation of adducts between 1 and b–h. In contrast the bonds of highest propensity were between reactants and the solvent for the adduct 1a, indicating either a high probability of the reactants crystallizing as solvates or incorporation of solvent into the adduct lattice. Experimental results agreed with the propensity calculations with the formation of a solvated cocrystal (1a·CH3OH). The successful application of hydrogen bond propensity calculations to the prediction of likely outcomes of these cocrystallization reactions suggests that this may be a useful tool in designing more targeted screening experiments.

Knowledge-based approaches to co-crystal design

The use of knowledge-based methods has been intimately connected with the field of co-crystal design since the seminal papers of Etter and Desiraju in the 1990s. Here we explain and exemplify how rational co-crystal design has been carried out in the past using crystal structure knowledge as well as presenting emerging methodologies for knowledge-based co-former selection.

Evaluation of molecular crystal structures using Full Interaction Maps

The specific crystalline form of a compound has a significant impact on its solid state properties. A key requirement for chemists developing crystalline materials is therefore to understand and evaluate the crystal form under investigation. We show here how the visualisation of molecular interaction maps within the context of a crystal structure can be used to evaluate the stability of polymorphic structures, assess multiple types of non-covalent interactions and provide a platform for crystal morphology analysis. Examples of three industrially-relevant compounds – sulfathiazole, anastrozole and cipamfylline – illustrate this well. A qualitative agreement with experimental stability data is observed for the five sulfathiazole crystal forms. The anastrozole crystal structure is demonstrated to optimise interactions to the strongest acceptor sites even though there are no conventional hydrogen-bond donors in the structure. Finally, the fastest growing plane of the needle-like morphology of cipamfylline is shown to have more H-bond donor and acceptor interactions per surface area than the slower growing planes.

One in half a million: a solid form informatics study of a pharmaceutical crystal structure

We introduce a knowledge-based approach to the evaluation, analysis and prediction of the properties of a crystal form; described inclusively as Solid Form Informatics. This approach is exemplified using the recently published crystal structure of the drug lamotrigine in the context of the Cambridge Structural Database (CSD). Analysis at the molecular, intermolecular and supramolecular level is carried out using the range of software available in the CSD System alongside new research applications. This work provides a template for the thorough analysis of any crystal structure and paves the way toward a fully automated structural analysis for the drug formulation scientist, with the aim to better provide answers to the fundamental questions raised during the drug development process. To conclude, the knowledge gained about the structure is applied to predict the potential for a co-crystal formulation of the drug and to automatically select optimal co-crystal formers. The crystal structure of lamotrigine was the half-millionth structure to enter the CSD. We demonstrate how the 499,999 structures that preceded it are central to the analyses presented.

Universal prediction of intramolecular hydrogen bonds in organic crystals

A complete exploration of intramolecular hydrogen bonds (IHBs) has been undertaken using a combination of statistical analyses of the Cambridge Structural Database and computation of ab initio interaction energies for prototypical hydrogen-bonded fragments. Notable correlations have been revealed between computed energies, hydrogen-bond geometries, donor and acceptor chemistry, and frequencies of occurrence. Significantly, we find that 95% of all observed IHBs correspond to the five-, six- or seven-membered rings. Our method to predict a propensity for hydrogen-bond occurrence in a crystal has been adapted for such IHBs, applying topological and chemical descriptors derived from our findings. In contrast to intermolecular hydrogen bonding, it is found that IHBs can be predicted across the complete chemical landscape from a single optimized probability model, which is presented. Predictivity of 85% has been obtained for generic organic structures, which can exceed 90% for discrete classes of IHB.

Truly prospective prediction: inter- and intramolecular hydrogen bonding

Accurately predicting which H-bonds might form in an organic crystal structure is demonstrated in a prospective, chronological setting using our recently developed method. The extent of correct classification for present and absent H-bonds is assessed whereby training and trial data are separated on the basis of age; the choice of cut-off is termed the model date. An encouragingly high predictivity is maintained for H-bonds in target structures published over the last 7–8 years, indicating promising future application toward novel structures. Predictions are computed using probability models trained for chosen target compounds using existing crystal structures in the public domain. An extension to the prediction of intramolecular H-bonds is also applied and is seen to be significant when considering the combination of possible H-bonds for target structures. Two target systems are selected for illustration: an amino-chloride salt hydrate and an amido-carboxylic acid.