Chem Soc Rev
`
`REVIEW ARTICLE
`
`View Article Online
`View Journal
`
`Facts and fictions about polymorphism†
`
`Cite this: DOI: 10.1039/c5cs00227c
`
`Aurora J. Cruz-Cabeza,*a Susan M. Reutzel-Edensb and Joel Bernsteincd
`
`We present new facts about polymorphism based on (i) crystallographic data from the Cambridge
`Structural Database (CSD, a database built over 50 years of community effort), (ii) 229 solid form screens
`conducted at Hoffmann-La Roche and Eli Lilly and Company over the course of 8+ and 15+ years
`respectively and (iii) a dataset of 446 polymorphic crystals with energies and properties computed with
`modern DFT-d methods. We found that molecular flexibility or size has no correlation with the ability of a
`compound to be polymorphic. Chiral molecules, however, were found to be less prone to polymorphism
`than their achiral counterparts and compounds able to hydrogen bond exhibit only a slightly higher
`propensity to polymorphism than those which do not. Whilst the energy difference between polymorphs
`is usually less than 1 kcal mol1, conformational polymorphs are capable of differing by larger values (up
`to 2.5 kcal mol1 in our dataset). As overall statistics, we found that one in three compounds in the CSD
`are polymorphic whilst at least one in two compounds from the Roche and Lilly set display polymorphism
`with a higher estimate of up to three in four when compounds are screened intensively. Whilst the
`statistics provide some guidance of expectations, each compound constitutes a new challenge and
`prediction and realization of targeted polymorphism still remains a holy grail of materials sciences.
`
`Received 13th March 2015
`
`DOI: 10.1039/c5cs00227c
`
`www.rsc.org/chemsocrev
`
`1. Introduction
`
`‘‘Polymorphism has been mainly studied in its phenomenological
`aspects, while its structural and energetic aspects have been
`alluded to in diverse fields of research, but in spite of a large body
`of data, have never been considered in a systematic way. . . The
`control of crystal polymorphism has practical advantages in many
`branches of the chemical industry, in fact, all those which deal with
`the organic solid state.’’1
`Since 1991, the phenomenon of polymorphism – the ability of a
`compound to crystallize in more than one crystal structure – has
`been the subject of growing interest (Fig. 1). A literature search on
`the topic renders over eleven thousand scientific publications in
`WebOfScience2 and over six thousand patent documents world-
`wide from 1966–2013 in Espacenet3 (Fig. 1). Although some key
`contributions to the subject were made in the late 60s and 70s, a
`significant communal interest in the subject did not occur until
`the early 90s (Fig. 1). The phenomenon of polymorphism, however,
`
`a Roche Pharma Research and Early Development, Therapeutic Modalities,
`Roche Innovation Center Basel, Basel, Switzerland. E-mail: aurorajosecruz@gmail.com
`b Small Molecule Design & Development, Eli Lilly and Company, Lilly Corporate Center,
`Indianapolis, IN 46285, USA
`c Faculty of Natural Sciences, New York University Abu Dhabi, P.O. Box 129188,
`Abu Dhabi, United Arab Emirates
`d Department of Chemistry, Ben-Gurion University of the Negev, P.O. Box 653,
`Beer Sheva 84105, Israel
`† In memory of our friend and mentor, the late Frank H. Allen, and in honor of
`the 50 years anniversary of the Cambridge Crystallographic Data Centre.
`
`This journal is © The Royal Society of Chemistry 2015
`
`was already recognized almost 200 years ago,4 and it has a
`somewhat turbulent history.
`The first example of a polymorphic organic compound was
`benzamide, identified and studied by Liebig and Wohler in
`1832.5 Although the crystal structure of the stable form was
`determined as early as 1959,6 a labile form was discovered in
`20057 whilst the original metastable form resisted solution
`until 2007.8,9
`The century following the Wohler–Liebig discovery witnessed
`considerable activity in the study of polymorphism.10–16 For
`instance, the first issue of Zeitschrift fur Kristallographie, founded
`in 1877 by the legendary P. von Groth,12 contained a paper by his
`student, Otto Lehmann, with a diagram of a hot stage microscope
`and a ‘‘time versus temperature’’ curve, clearly indicating the four
`polymorphs of ammonium nitrate.17 Although the subject was not
`a molecular crystal, the study is a classic example of the recording
`of thermal events associated with transitions between polymorphic
`crystal forms.
`Interest in structural polymorphism declined during the
`early decades of the development of structural crystallography
`and, although many instances of polymorphism had been
`documented based on thermal11 and optical14 data, in most
`cases their structural characterization awaited developments in
`rapid single crystal structure determination. That gap has been
`closed to a significant extent, especially for ‘‘classic’’ (i.e. iconic)
`molecules (e.g. benzene, benzamide, etc.). It is possible –
`indeed, not unlikely – that for at least some of those molecules
`the polymorphic landscape has not been fully mapped since
`
`Chem. Soc. Rev.
`
`Merck 2006
`Argentum v. Merck
`IPR2018-00423
`
`Published on 24 September 2015. Downloaded by New York University on 24/09/2015 13:18:15.
`
`1
`
`

`

`Review Article
`
`View Article Online
`
`Chem Soc Rev
`
`Fig. 1 Number of publications, citations to those and patents related to polymorphism. Landmark contributions are indicated and commented further
`in the text. Inner graph corresponds to the citations history of the McCrone & Haleblian, J. Pharm. Sci., 1969 review (669 citations in Google Scholar, five-
`year bins).
`
`sufficient ‘‘time and money’’18 have not been invested into
`exploring their polymorphs.
`Polymorphism, after all, has not always been a sought
`phenomenon and it has been overlooked in the past, especially
`in the early days of structural crystallography. There are a
`number of reasons for this relative neglect. First, for many
`years carrying out a crystal structure determination was a major
`
`task; hence, the time and effort involved in solving the crystal
`structure of another crystal form of the same molecule often was
`not justified. Furthermore, prior to the early 1970’s determination
`of non-centrosymmetric crystal structures, structures with Z0 4 1,
`structures with disorder or even those that did not grow into
`‘‘good’’ single crystals, presented major challenges for the
`crystallographer and were often abandoned. This is most likely
`
`Aurora J. Cruz-Cabeza is a Postdoctoral Fellow at F. Hoffmann-La
`Roche Ltd in Basel (Switzerland). After her BSc degree in chemistry
`from the University of Jae´n (2002), she earned a Masters degree in
`Heterogeneous Catalysis from the University of Co´rdoba (2004) and a
`PhD in Physical Chemistry from the University of Cambridge (2008).
`Aurora has worked as a researcher in several pharmaceutical
`companies (Pfizer and Roche), the University of Amsterdam and the
`Cambridge Crystallographic Data Centre.
`Susan M. Reutzel-Edens is a Senior Research Advisor in Small
`Molecule Design & Development at Eli Lilly and Company. She
`obtained her BS degree in chemistry from Winona State
`University (1987), then earned her PhD in organic chemistry at
`the University of Minnesota (1991). Susan brought her experience
`in hydrogen-bond directed co-crystallization and interest
`in
`crystal polymorphism to Eli Lilly, where she developed Lilly’s
`solid form design program and for two decades led a team of cross-functional scientists charged with finding commercially-viable
`crystalline forms for small-molecule drug products.
`Joel Bernstein obtained a BA degree at Cornell University and a PhD in physical chemistry at Yale University. Following
`postdoctoral stints in X-ray crystallography at UCLA and the Weizmann Institute of Science, he joined the faculty of Ben-Gurion
`University where he was the incumbent of the Carol and Barry Kaye Professorship of Applied Science until 2010 and is now Professor
`Emeritus. Currently, Joel is a professor at the newly founded New York University Abu Dhabi. He has published over 180 research
`and review articles and is the sole author of a book entitled ‘‘Polymorphism in Molecular Crystals’’ (Oxford University Press).
`
`From left to right: Aurora J. Cruz-Cabeza,
`Susan M. Reutzel-Edens and Joel Bernstein
`
`Chem. Soc. Rev.
`
`This journal is © The Royal Society of Chemistry 2015
`
`Published on 24 September 2015. Downloaded by New York University on 24/09/2015 13:18:15.
`
`2
`
`

`

`Chem Soc Rev
`
`the reason that the structures of the four polymorphs of the
`relatively simple molecule benzidine (with Z0 = 4.5, 3.0. 1.5 and
`4.5 respectively) remained unreported until well
`into the
`21st century.19 Another important factor for the decline in
`interest in polymorphism was the advance of other analytical
`methods that readily provided increasingly precise and reproducible
`data for characterizing and defining compounds and materials.
`During that period numerical data became the mode for
`defining materials. That is still very much the case, but the
`relative ease and decreasing cost in time and money of carrying
`out crystal structure determinations, combined with the facility
`of publishing digital color photos of crystals and crystal
`structures in the chemical literature has considerably aided
`in nurturing the renaissance of interest and activity in poly-
`morphism and its manifestations.
`In spite of the lack of widespread activity in polymorphism
`research during the middle decades of the 20th century, there
`were two active groups that made important contributions in
`that period. One was the group led by Ludwig Kofler at the
`University of
`Innsbruck (followed successively by Marie
`Kuhnert-Brandstatter and Artur Burger) and Walter McCrone,
`originally at Cornell, and later as an independent consultant.
`Both published books in the 1950’s with major emphasis on the
`polymorphism of organic materials.20,21 A 1980 translation of the
`Kofler book by Walter C. McCrone is available from McCrone
`Associates, Inc. McCrone’s 1965 chapter on polymorphism18
`remains one of the classic papers on the subject together with
`his seminal 1969 Journal of Pharmaceutical Sciences review
`with Haleblian,22 the first specific review publication relating
`polymorphism to the pharmaceutical industry, see Fig. 1. The
`citation history of that publication is illustrative of the evolving
`interest in polymorphism catalyzed to a great extent by the
`pharmaceutical industry. From the inner graph in Fig. 1 it can
`be seen that following an initial rise in citations of the 1969
`McCrone/Haleblian paper during the 1970’s, a pattern normally
`expected for a review, interest apparently waned until a renewal
`marked by a fairly steep rise in the number of citations starting
`around 1995. This rapid rise in interest in the McCrone/Haleblian
`paper is likely related to the high profile 1991 patent litigation on
`ranitidine hydrochloride (Zantacs)23,24 which at the time was the
`world’s largest selling drug ($3.45 billion year1 – nearly twice as
`much as the next largest selling drug) and dealt to a large extent
`with various issues surrounding the polymorphism of the active
`ingredient. In support of this contention there was a parallel
`increase in scientific publications dealing with polymorphism,
`echoed by an increase in the number of patents issued containing
`‘‘polymorph’’ in the title or abstract after 1991 (Fig. 1). Interestingly,
`there is also a spike in the number of publications in 2002,
`following the 1998–1999 saga of the removal and subsequent
`relaunch of Abbott’s reformulated drug ritonavir due to the
`formation of an undesirable new polymorph.25,26
`As interest in polymorphism has increased, many aspects of
`the subject have been addressed either directly or in passing.
`Some representative (but by no means comprehensive) scientific
`contributions towards our understanding of polymorphism
`(Fig. 1) after the Haleblian and McCrone review include:
`
`View Article Online
`
`Review Article
`
`(i) reports on conformational (1978),27 disappearing (1995)28
`and concomitant polymorphs (1999),29
`(ii) contributions towards our fundamental understanding
`of the thermodynamics (1979)30,31 and phase transitions (1991)32
`in polymorphs,
`(iii) the structural and energetic study of polymorphs under
`room temperature conditions by Gavezzotti and Filippini (1995),1
`(iv) studies on polymorphism in the context of crystal
`engineering,33–35
`(v) numerous studies on polymorphism in the context of
`pharmaceutical materials7,8,12–15,36,37 including studies of landmark
`polymorphic drug systems such as ritonavir25,26 and aspirin,38–41
`(vi) applications of Hirshfeld surfaces42 and computational
`chemistry43 to the study of polymorphism, and
`(vii)
`the development of new methods for surveying
`the crystal forms landscape, among them automation for
`high-throughput crystallization,37 the liquid assisted grinding
`technique44 or crystallizations in the presence of polymers.45
`In spite of an impressive array of contributions across a
`broad spectrum of their chemical and physical aspects, poly-
`morphic systems are in many ways still enigmatic, echoing the
`1937 observation by Buerger and Bloom ‘‘with the accumula-
`tion of data, there is developing a gradual realization of the
`generality of polymorphic behavior, but to many chemists
`polymorphism is still a strange and unusual phenomenon.’’46
`This contribution presents a systematic study of polymorphism
`from diverse sources. The first of these is based on the Cambridge
`Structural Database (CSD). We analyze the data and attempt to
`correct for certain biases in order to extract meaningful statistics.
`We also compute energetics for 215 polymorphic families with
`modern DFT-d techniques. These structure-based statistics are
`then compared to experimental polymorph screening statistics
`from 229 studies conducted at F. Hoffmann-La Roche Ltd (here-
`after Roche) and Eli Lilly and Company (hereafter Lilly) over more
`than nine and fifteen years respectively. In this article, we address
`several fundamental aspects of the phenomenon and we question
`previous assertions promoted in the literature, many based on
`chemical intuition rather than scientific evidence. These lead to
`the correction of some common misconceptions that have been
`perpetuated in the polymorphism literature and suggest that the
`facts about polymorphism lie beyond chemical intuition and
`predictability.
`
`2. Datasets and methods
`2.1 Datasets derived from the Cambridge Structural Database
`(CSD)
`the polymorphic datasets. Crystallo-
`2.1.1 Retrieval of
`graphic data were retrieved from the CSD vs. 5.33 (Nov 2011)
`using Conquest.47 The structure searches were restricted to
`organic molecules containing only the most common atomic
`elements (C, H/D, N, O, S and halogens). Crystal structures with
`all atomic coordinates determined (with the exception of
`hydrogen atoms) were retrieved and no polymeric structures
`were allowed. Only crystal structures containing the keyword
`
`This journal is © The Royal Society of Chemistry 2015
`
`Chem. Soc. Rev.
`
`Published on 24 September 2015. Downloaded by New York University on 24/09/2015 13:18:15.
`
`3
`
`

`

`Review Article
`
`‘‘polymorph’’ (i.e. the compound was described in the literature
`as being polymorphic) were kept.
`In the CSD, a REFCODE consists of a six-letter code followed
`by two numbers. A REFCODE family (the 6 letter code) contains
`all determined crystal structures for a given compound (including
`polymorphic crystals and structure redeterminations). For the
`initial statistics of the CSD, we worked with REFCODE families
`(the six letter code). A REFCODE family corresponds to a unique
`composition (e.g. a unique compound or a unique mixture of
`compounds in a particular stoichiometry).
`Three different polymorphic datasets were constructed:
` Polymorphic dataset of neutral single components (POL):
`REFCODE families of single component crystal structures –
`2048 polymorphic families.
` Polymorphic dataset of neutral multicomponents (MULTI-
`POL): REFCODE families containing at least 2 or more compo-
`nents, all in neutral form – 303 polymorphic families.
` Polymorphic dataset of salts (SALTS-POL): REFCODE
`families containing at least two ionised components – 347
`polymorphic families.
`2.1.2 Retrieval of the monomorphic datasets. In building
`the monomorphic datasets, the same search criteria used for
`the polymorphic searches were applied to the entire CSD vs.
`5.33 (Nov 2011). REFCODE families belonging to the poly-
`morphic sets were then removed.
`Three different monomorphic datasets were constructed:
` Monomorphic dataset of neutral single components
`(MONO): REFCODE families of single component crystal struc-
`tures – 105 601 monomorphic families.
` Monomorphic dataset of neutral multicomponents
`(MULTI-MONO): REFCODE families containing at least 2 or more
`components, all in neutral form – 21 622 monomorphic families.
` Monomorphic dataset of salts (SALTS-MONO): REFCODE
`families containing at least two ionised components – 16 285
`monomorphic families.
`2.1.3 Molecular geometries and descriptors. Molecular
`geometries were retrieved from the crystal structures and exported
`as molecular files using Conquest.47 OpenBabel was used for
`molecular format conversions and the addition of hydrogen
`atoms48 to molecules with unresolved hydrogen atom positions.
`Molecular descriptors were calculated using the ChemAxon
`cheminformatics plugin.49 Properties such as number of atoms,
`molecular weight (Mw), number of asymmetric centers, number of
`aliphatic/aromatic rings or number of hydrogen bond donors and
`acceptors were calculated. In addition, we defined and calculated a
`descriptor referred to as DOFlex (or molecular degrees of flexibility)
`as the sum of: (1) the number of acyclic rotatable bonds, (2) the
`number of groups attached to triple bonds and (3) the number of
`aliphatic rings which could potentially also change their geometry.
`A compound was defined to be drug-like if it satisfied the Lipinski
`rule-of-five criteria: Mw r 500, log P r 5, H-bond donors r 5 and
`H-bond acceptors r 10.50
`2.1.4 Polymorphic subset for optimization. The subset of
`polymorphic molecules and crystals taken for calculations was
`constructed by searching the best R-factor list of the CSD vs.
`5.33 (Nov 2011) using the same criteria as for the POL subset.
`
`View Article Online
`
`Chem Soc Rev
`
`Only different polymorphic crystals are kept in the best R-factor
`list, hence there are no redeterminations. Only REFCODE
`families with more than one REFCODE were kept.
`Since the calculation of lattice energies with accurate methods
`requires a considerable amount of computational time, we applied
`further filtering criteria in order to obtain a manageable subset.
`(1) Only structures with an R factor o 5%.
`(2) Only structures with resolved hydrogen atom positions.
`(3) Only polymorphic families containing structures with
`less than 192 atoms per unit cell.
`(4) Only ambient pressure polymorphic forms.
`The subset used for optimization (POLcalc) consisted of 289
`polymorphic molecules and 596 crystal structures.
`2.1.5 Calculation of lattice energies. We used periodic
`density functional
`theory with van der Waals corrections
`(DFT-d) for geometry relaxations of the polymorphic structures
`in the POLcalc subset. The PBE functional51 was used with PAW
`pseudopotentials52,53 and the Grimme’s van der Waals corrections
`(d2)54 as implemented in the VASP code (version 5.3.3).55–58
`A kinetic energy cut-off of 520 eV was used. The Brillouin zone
`was sampled using the Monkhorst–Pack approximation59 on a
`grid of k-points separated by approximately 0.07 Å (the minimum
`k-point sampling used was 2  2  2 k-points). All atoms and
`unit cell parameters were allowed to optimize and structural
`relaxations were halted when the calculated force on every atom
`was less than 0.003 eV Å1.
`Energies obtained from DFT-d codes are normally given per
`unit cell. We normalized the energies to the number of mole-
`cules in the unit cell so that energies can be compared per
`molecule across the polymorphs. We will refer to the calculated
`energies per mol as EDFT-d.
`2.1.6 Optimised subset of polymorphic structures (POLDFT-d).
`After attempting the optimization of the 596 crystal structures,
`some additional filtering was applied. Polymorphic families with
`structures that did not converge in the optimization procedure were
`removed. The converged crystal structures were compared with
`the experimental structures (used as input in the optimization
`procedure) using the COMPACK algorithm60 with a 20 molecule
`cluster and the standard settings. Some of the optimized struc-
`tures deviated considerably from the experimentally determined
`ones. This could be due to errors in the experimental structures.
`In fact, previous studies have used DFT-d calculations to assess
`the correctness of experimental crystal structures.61 We removed
`polymorphic families containing optimized structures that
`deviated considerably from the experimental X-ray structures.
`These included structures not matching 20 out of 20 molecules
`in the COMPACK comparison or structures matching 20 mole-
`cules but having an rmsd20[r] 4 0.45 Å.
`After the above-mentioned filtering, 215 polymorphic families
`containing 446 crystal structures remained. We will refer to this
`subset as POLDFT-d and use it for further calculations and data
`analysis.
`
`2.2 Datasets from Roche & Lilly
`As evidenced very much by the historical record in Fig. 1 and
`discussed earlier, much of the progress in understanding the
`
`Chem. Soc. Rev.
`
`This journal is © The Royal Society of Chemistry 2015
`
`Published on 24 September 2015. Downloaded by New York University on 24/09/2015 13:18:15.
`
`4
`
`

`

`Chem Soc Rev
`
`chemistry of polymorphism, its manifestations and ramifications
`has been driven by practical demands and considerations. The
`rapidly increasing volume of literature on this subject contains
`many examples of individual studies on polymorphic systems.62
`However, since every compound represents totally unknown
`territory in terms of the crystal landscape, there is perhaps no
`better means for demonstrating the variety and vagaries of
`polymorphism behavior than the cumulative record of two
`groups of experienced practitioners. Thus we have compiled
`solid form statistics from 229 solid form screens conducted by
`Roche and Lilly comprising screens of 145 structurally diverse
`parent compounds (72 Roche & 73 Lilly) and 84 different salts
`(Lilly). The screenings were generally conducted early in drug
`product development to support commercial form selection and
`ranged in scope from limited to comprehensive. As might be
`expected in industrial settings, screens were most often limited
`by design, time or material supply, though material quality
`(purity) may arguably have also been a factor. Some screenings
`would have been stopped because of project termination whilst
`other compounds would have been screened for polymorphs
`several times at different stages of development. All of the
`compounds were screened by conventional (manual + semi-
`automated) methods, with screen designs tailored to the solu-
`bility properties of the starting materials, when appropriate.
`Small subsets of the Roche and Lilly datasets were also subjected
`to high-throughput methods to pilot the use of automation for
`polymorph screening. If the high-throughput method yielded a
`new XRD-pattern, follow up experiments would be repeated in a
`manual way.
`In constructing the Roche and Lilly datasets, crystalline
`forms were counted only when sufficient physical and chemical
`data were acquired to support their existence. However, the
`criteria for establishing a new crystal form were slightly different
`at the two locations. Whereas a new solid form is designated at
`Roche only if it can be obtained at least twice and is characterized
`by various analytical techniques, at Lilly, a single occurrence of a
`new form might be sufficient, provided the supporting data are
`unequivocal (e.g. a crystal structure from a single crystal isolated
`from a batch of material). Importantly, amorphous forms and
`
`View Article Online
`
`Review Article
`
`unconfirmed crystalline forms, of which there were many, have
`not been included in the survey for either the Roche or the Lilly
`datasets. As such, this tabulation of solid form diversity among
`typical pharmaceuticals must be considered conservative.
`
`3. Choosing representative
`monomorphic structures for the CSD
`datasets
`
`To obtain meaningful statistics of polymorphism in the CSD, it
`is important to define suitable monomorphic datasets as a
`basis for comparison. The fact that a given compound only has
`one unique crystal structure recorded in the CSD says very little
`about its tendency to exhibit polymorphism. Thus as a general
`caveat, it must be remembered that any statistical analysis
`based on the CSD relates only to reported crystal structures.
`Many compounds in the CSD have been studied only once
`crystallographically and often crystal structure determination
`has been a means for molecular structure validation.
`One might initially expect that structures in the polymorphic
`and monomorphic datasets would span a similar range of
`molecular size. However, if we plot the normalized distributions
`of the structures in the POL and MONO datasets as a function of
`molecular size, i.e., the number of atoms (Fig. 2a), we observe
`large and apparently significant differences. The maximum of
`the distribution for the POL dataset is located around 30 atoms
`whilst that of the MONO dataset appears at approximately
`40 atoms. Thus if the complete MONO dataset of the CSD is
`used for statistical analysis, one might conclude that smaller
`molecules are far more likely to be polymorphic than larger
`molecules and that the occurrence of polymorphism in the CSD
`is 1.9%. A much more likely explanation is that our polymorphic
`datasets are somewhat biased for smaller molecules. In other
`words, on average, smaller molecules serve as model compounds
`for studies concerning polymorphism whilst larger molecules
`(which are generally less likely to be commercially-available and
`are harder to synthesize) are more often studied by X-ray single
`crystal diffraction determination only once.
`
`Published on 24 September 2015. Downloaded by New York University on 24/09/2015 13:18:15.
`
`Fig. 2 Normalized distributions for the POL (orange) and MONO (black) datasets (a) and the POL and MONO subset with redeterminations (b) as a
`function of the number with atoms.
`
`This journal is © The Royal Society of Chemistry 2015
`
`Chem. Soc. Rev.
`
`5
`
`

`

`Review Article
`
`Table 1 Number of compounds in the original and refined datasets
`
`Monomorphic
`dataset
`
`MONO
`SALTS-MONO
`MULTI-MONO
`
`Unique
`compositions
`
`Unique compositions
`with redeterminations
`
`105 601
`16 285
`21 622
`
`3731 (3.5%)
`639 (3.9%)
`805 (3.7%)
`
`One way of testing for that bias is to consider just those
`compounds that have been more intensively studied (for a variety
`of reasons). If instead of using the complete MONO dataset we use
`a subset of the molecules for which there exists at least one crystal
`structure redetermination (3731 molecules), we can be assured
`that the majority of those compounds have been crystallized at
`least twice resulting in the same crystal structure.63 This modest
`adjustment has a dramatic effect, causing the POL and MONO
`molecular size distributions to become very similar (Fig. 2b).
`We applied this criterion (only monomorphic families with at
`least one crystal structure redetermination) to our three MONO
`datasets, Table 1. The monomorphic subsets with redeterminations
`tend to be between 3.5–3.9% of the original monomorphic data-
`sets. This means, only between 5–6% of the unique compositions
`in the CSD (monomorphic and polymorphic) are studied and
`determined with X-ray diffraction more than once.
`For the statistics reported below, we will only use the
`monomorphic subsets with redeterminations and we will refer
`to them as MONOre, SALTS-MONOre and MULTI-MONOre.
`We refer to ‘‘monomorphic’’ compounds as those crystallising
`with the same composition in only one crystal structure. For
`example if compound A crystallizes by itself in two different
`polymorphs, A will be counted as polymorphic. If compound A
`also crystallizes with water in a single hydrate crystal structure, then
`‘‘A + H2O’’ is accounted for as a monomorphic multicomponent
`crystal. Our use of monomorphic and polymorphic here is exclusive.
`We refer, therefore, to polymorphism of unique compound(s) in
`unique compositions and stoichiometries.
`
`4. Statistical facts about
`polymorphism
`4.1 Statistics on polymorphism occurrence (CSD, Roche & Lilly)
`The true occurrence of polymorphism is very difficult to determine
`and depends to a large extent on the choice of the data sample.
`This is demonstrated for three attempts summarized in Table 2.
`During the period 1948–1961 McCrone regularly reported the
`results of crystal growing experiments64 with approximately
`25% of the organic compounds exhibiting polymorphism.
`A more recent survey of the pharmaceutical compounds in
`
`Table 2 Occurrence of polymorphism according to various sources
`
`View Article Online
`
`Chem Soc Rev
`
`the European Pharmacopeia yielded 42% polymorphism.65
`A summary of the 245 compounds studied by SSCI with
`the specific goal of screening for crystal forms yielded 48%
`exhibiting polymorphism.66
`For the purpose of this study we have compiled three sets of
`statistics on polymorphism: (i) from 8035 unique crystal com-
`positions from CSD (ii) from 72 solid form screens performed at
`Roche and (iii) from 157 solid form screens performed at Lilly.
`On the one hand, the CSD dataset contains a very large amount
`of information – much larger than any single group compile –
`but the degree of form screening for the reported compounds
`may vary enormously. We expect unbiased data on polymorphism
`for all kinds of crystal compositions. On the other hand, the
`Roche and Lilly sets of compounds are much smaller but have
`been intensively screened specifically for multiple crystal forms.
`The Roche dataset contains screening efforts for identifying
`polymorphs of 72 parent compounds over the course of 8+ years.
`The Lilly dataset contains screening efforts for identifying poly-
`morphs of 73 parent compounds and 84 different salts over the
`course of 15+ years. It is important to note that, whilst solvates
`and hydrates were found during these polymorph screenings, the
`screenings were targeted to the original parent compound or salts
`polymorphism but less usually on the hydrates and rarely that of
`solvates. We expect, therefore, that the industrial datasets will be
`more representative of polymorphism of neat parent compounds
`and salts.
`The statistics on polymorphism for these three datasets
`together with their 95% confidence intervals are presented in
`Table 3. The occurrence of polymorphism is given separately
`for different types of crystal compositions: (i) single component
`crystals, (ii) cocrystals, solvates and hydrates and (iii) salts and
`hydrates of salts.
`The occurrence of polymorphism in single component
`compounds was found to be 37  1% in the CSD compared
`to 53  12% in the Roche and 66  11% in the Lilly datasets.
`The difference in polymorph occurrences for single component
`crystals is due in part to the inherent nature of the data. Whilst
`the CSD dataset contains compounds crystallised at least twice
`as single crystals suitable for study by X-ray diffraction, the
`Roche and Lilly datasets contain data from diverse types of
`polymorph screenings where many crystallization conditions
`were explored. Moreover, characterization of the forms comprising
`the Roche and Lilly datasets extended beyond single crystal X-ray
`diffraction, in some cases, increasing the probability of identifying
`polymorphs. For example, because high temperature polymorphs,
`particularly those produced by reversible phase transitions
`on increasing temperature, are generally easier to characterize
`(at high temperatures) using other techniques, they should be
`
`Source
`Microscopy studies by McCrone (1948–1961)64
`From European Pharmacopeia (1964–2004)65
`From SSCI polymorph screens of organic compounds (1991–2007)66
`
`Data type
`
`Compounds
`
`Polymorphism
`occurrence (%)
`
`Organic compounds
`Single component organic compounds
`Organic compounds
`
`140
`598
`245
`
`25
`42
`48
`
`Chem. Soc. Rev.
`
`This journal is © The Royal Society of Chemistry 2015
`
`Published on 24 September 2015. Downloaded by New York University on 24/09/2015 13:18:15.
`
`6
`
`

`

`Chem Soc Rev
`
`View Article Online
`
`Review Article
`
`Table 3 Occurrence of polymorphism (Pol. occ.) in the CSD, Roche and Lilly datasets
`
`Data type
`
`Single-component
`
`Neutral multicomp.
`Cocrystals
`Solvatesc
`Hydrates
`
`Salts
`Unhydrated salts
`Hydrated salts
`
`CSD
`U.C.a
`
`5941
`
`403
`318
`387
`
`820
`166
`
`Pol. occ. (%)
`
`37
`
`36
`25
`20
`
`37
`28
`
`C.I.b (%)
`
`(36, 38)
`
`(31, 41)
`(21, 30)
`(16, 24)
`
`(33, 40)
`(21, 35)
`
`Roche
`U.C.a
`
`Pol. occ. (%)
`
`68