research papers
A method to estimate statistical errors of properties derived from chargedensity modelling
^{a}Institut Galien Paris Sud, UMR CNRS 8612, Université Paris Sud, Faculté de Pharmacie, Université ParisSaclay, 5 rue JeanBaptiste Clément, ChâtenayMalabry, 92296, France, ^{b}Laboratoire Structures, Propriétés et Modélisation des Solides (SPMS) UMR CNRS 8580, Ecole CentraleSupélec, 3 rue JoliotCurie, GifsurYvette Cedex, 91192, France, ^{c}CRM2, UMR CNRS 7036, Institut Jean Barriol, Université de Lorraine, Vandoeuvre les Nancy Cedex, France, and ^{d}The Barcelona Institute of Science and Technology, Institute of Chemical Research of Catalonia (ICIQ), Avinguda Països Catalans 16, Tarragona, 43007, Spain
^{*}Correspondence email: christian.jelsch@univlorraine.fr
Estimating uncertainties of property values derived from a chargedensity model is not straightforward. A methodology, based on calculation of sample standard deviations (SSD) of properties using randomly deviating chargedensity models, is proposed with the MoPro software. The parameter shifts applied in the deviating models are generated in order to respect the variance–covariance matrix issued from the leastsquares This `SSD methodology' procedure can be applied to estimate uncertainties of any property related to a chargedensity model obtained by leastsquares fitting. This includes topological properties such as coordinates, electron density, Laplacian and ellipticity at critical points and charges integrated over atomic basins. Errors on electrostatic potentials and interaction energies are also available now through this procedure. The method is exemplified with the charge density of compound (E)5phenylpent1enylboronic acid, refined at 0.45 Å resolution. The procedure is implemented in the freely available MoPro program dedicated to chargedensity and modelling.
Keywords: Monte Carlo methods; electron density; uncertainty; topology; intermolecular interactions.
CCDC reference: 1829445
1. Introduction
Errors on electrondensityderived properties, such as topological characteristics or electrostatic potential, are generally poorly addressed in the relevant literature. To the best of our knowledge, no available computer software designed for chargedensity analysis on the basis of multipolar modelling computes properly analytical standard deviations on electrondensityderived properties. For instance, in the XD2006 program (Volkov et al., 2006), there is a feature that allows one to compute estimated uncertainties of the electron density ρ(r), of the Laplacian ∇^{2}ρ and of values using the variance–covariance matrix, but it only accounts for the contributions of some of the parameters used in the Hansen & Coppens (1978) model, i.e. monopole and multipole populations. It implies that the propagation of errors due to the contributions of the atomic coordinates and of the contraction/expansion coefficients κ and κ′ is not taken into account. This could lead, consequently, to an overall underestimation of standard deviations on electrondensityderived properties.
Estimating uncertainties on properties derived from a chargedistribution model is yet essential to avoid any false or overinterpretation of these properties. When several experimental Xray diffraction data sets collected during distinct and independent measurements are available for the same compound, it becomes possible to study the reproducibility of the refined chargedensity model and to estimate uncertainties of derived properties through the determination of their sample standard deviations (SSDs). Such an approach was followed in a few studies, but often with questionable statistical significance given the sometimes very sparse sampling used [down to two models (Dittrich et al., 2002; Grabowsky et al., 2008), a larger sample (up to four data sets) but varying experimental temperatures or setups (Messerschmidt et al., 2005; Förster et al., 2006)].
Closely related but still different compounds (such as peptide bond properties in different amino acids) were also investigated (Flaig et al., 1999). In an article dedicated to the transferability of atomic parameters in alanylXalaninetype tripeptides, Grabowsky et al. (2008) computed the global average of the standard deviations (noted experimental reproducibility indices ) obtained in those studies, for various electrondensityderived properties of the QTAIM (quantum theory of atoms in molecules; Bader, 1990; Bader et al., 1987) framework. For instance, they obtained, this way, average experimental errors (ρ) = 0.07 e Å^{−3} and (∇^{2}ρ) = 3.3 e Å^{−5} associated, respectively, with electron density and with Laplacian values at the bond critical points.
The most comprehensive and statistically sound reproducibility study on a wide range of electrondensityderived parameters was undertaken by Kamiński et al. (2014). They used 13 independently collected highresolution Xray diffraction data sets of αoxalic acid dihydrate. From these data, obtained using similar experimental setups, they derived 13 oxalic acid chargedensity models which were refined following identical strategies. This approach allowed them to analyse the normality of the error distribution in experimental data and in residual electron densities using the Shapiro–Wilk statistical test and, more importantly, to obtain very informative results in terms of dispersion of structural/chargedensity model parameters and of chargedensityderived property values. They have shown, for instance, that among the multipole model parameters, the valence populations present large reproducibility deviations, reaching up to 40% of the corresponding atomic net charge. Conversely, multipole populations were characterized by moderate dispersions. Thus high reproducibility was achieved among the refined models. The multipole populations expected to be close to zero due to atom were indeed statistically negligible. In the same way, concerning chargedensityderived properties, Kamiński et al. (2014) were able to evidence a significantly smaller dispersion of electrondensity values on weak intermolecular (hydrogen bonds) critical points [10^{−3} < < 3 × 10^{−2} e Å^{−3}] compared with covalent bonds [3 × 10^{−2} < < 6 × 10^{−2} e Å^{−3}] (CP = BCP = bond critical point) and, in any case, lower than the (ρ) value of 0.07 e Å^{−3} obtained by Grabowsky et al. (2008). The methodology proposed by Kamiński et al. (2014) provides standard deviations on any properties derived from the chargedensity model, as well as possible rules of thumb for property uncertainties in any chargedensity model of comparable quality. However, this approach is very resource and timeconsuming as it implies the collection of a statistically significant number of diffraction data sets at subatomic resolution. The uncertainties obtained may also not account totally for all systematic errors present in the data measurements.
Krause et al. (2017) recently presented a method based on R_{free} calculations. Sample standard deviations computed on the relevant models refined on subsets of the measured reflections (for example, 20 subsets of 95% reflections) can yield a rough estimate of the standard deviation on topological properties of the electron density. However, the R_{free} method has two drawbacks. Firstly, when strong reflections are omitted (put in the test set), the results of these refinements versus the remaining data are significantly influenced. This effect does not have much impact on the of protein structures (which have poor R factors and a large number of reflections) but is crucial for the of quantitative electron densities.
Secondly, the estimated uncertainty on a derived property obtained using this method depends on the number N of complementary R_{free} refinements performed. The discrepancy between the refined models decreases with N, as the number of free reflections omitted in the validation sets decreases proportionally to 1/N.
Here, we present a method allowing the estimation of uncertainties on properties derived from a chargedensity model. This method consists of a statistical Monte Carlo random sampling procedure, based on the variance–covariance matrix obtained after the convergence of the leastsquares refinement.
The leastsquares method is widely used for the structural and chargedensity ).
of crystal structures. The optimization procedure that uses the matrix of normal equations has a great power of convergence. The inversion of the full normal matrix also provides the variance–covariance matrix of the refined parameters and permits one to determine the precision of the refined structure model (Hamilton, 1964The current study addresses the uncertainty on properties related to the precision of measurements. The accuracy of properties which is related to systematic errors in measurements is however a different issue.
In the present paper, the methodology for estimation of uncertainties is illustrated with the chargedensity analysis of an organic compound: (E)5phenylpent1enylboronic acid (hereafter noted BOH2, Fig. 1). The unique electronic and physicochemical properties of boronic acid make this kind of compound very useful as a pharmaceutical agent. are strong Lewis acids. They can be used as enzyme inhibitors in Suzuki crosscoupling reactions, Diels–Alder reactions, carboxylic acid activation or selective reduction of among many other uses (Yang et al., 2003). In recent years, have also been reported as interesting building blocks in covalent organic frameworks (Côté et al., 2007; Spitler & Dichtel, 2010; Ding et al., 2011). To the best of our knowledge, this article is the first experimental chargedensity study of a boronic acid compound.
2. Experiment
2.1. Crystallization
For the current experiment, crystals were grown by slow evaporation of an ethanol/water solution of the compound BOH2 in a few days at room temperature. A single, colourless crystal of dimensions 0.34 × 0.18 × 0.10 mm was selected for the diffraction experiment. The compound crystallized in the centrosymmetric Pbca. More data on the orthorhombic crystal of BOH2 are given in Table 1.
‡At sinθ_{max}/λ = 1.22 Å^{−1}. 
2.2. Data collection
A singlecrystal highresolution and highly redundant Xray data collection of the BOH2 compound was performed on a Rigaku MicroMaxHF rotatinganode diffractometer equipped with a Pilatus 200K hybrid pixel detector using Mo Kα radiation (λ = 0.71073 Å). The crystal was mounted on a Kapton micromount. The data collection was carried out at 90 (1) K under a stream of nitrogen using the Oxford 700 Plus Cryosystems gasflow apparatus.
The diffraction data were collected using ω scans of 0.5° intervals with the CrystalClearSM Expert 2.1b29 software (Rigaku, 2013) up to a resolution of 0.41 Å (sinθ/λ < 1.22 Å^{−1}). The exposure times were 5 and 40 s per frame for low and highresolution data, respectively. Data reduction and absorption correction were performed using the CrysAlisPro 1.171.38.37f package (Rigaku Oxford Diffraction, 2015); the internal R(I) factor was 3.06% for all reflections (Table 1).
2.3. Structure solution and refinement
The structure of the BOH2 compound has already been determined (Gelbrich et al., 2000). In our study, the structure of BOH2 was solved using the SIR2014 software (Burla et al., 2015). In particular all the H atoms were located in the difference Fourier map. An initial independent atom model (IAM) was undertaken using the SHELXL2014 software (Sheldrick, 2015).
2.4. Multipolar refinement
The chargedensity model was refined against diffraction intensities using the program MoPro (Guillot et al., 2001; Jelsch et al., 2005). The program is based on the multipolar scattering factor formalism of Hansen & Coppens (1978) and allows the definition of restraints on stereochemistry, thermal motion and chargedensity parameters. Data resolution was truncated at 0.45 Å as the very high resolution reflections showed decreasing values of 〈F_{o}^{2}〉/〈F_{c}^{2}〉 well below unity, as verified with the XDRK software (Zhurov et al., 2008). For the same reason, an I/σ_{I} > 0.35 cutoff was applied. The evolution of 〈F_{o}^{2}〉/〈F_{c}^{2}〉 as a function of reciprocal resolution s is shown in the supporting information.
The multipole expansion was done at the octupolar level for B, C and O atoms and the dipole level for H atoms. The core and valence spherical scattering factors were calculated using the wavefunctions for isolated atoms from Su & Coppens (1998) and the coefficients were taken from Kissel et al. (1995).
The MoPro program has numerous functionalities with respect to constraints, restraints and similarity applying to the stereochemistry and charge density. For the H atoms, the values of anisotropic U_{ij} parameters were fixed to those obtained from the SHADE3 server (Madsen & Hoser, 2014). The H—X distances of H atoms were restrained to the values obtained from neutron diffraction studies (Allen & Bruno, 2010) with a restraint sigma σ_{rest} of 0.01 Å. Distance X—H similarity restraints were also applied to chemically equivalent groups (σ_{rest} = 0.01 Å).
The chargedensity model was subsequently refined against diffraction intensities. The electrondensity maps, local topological properties and intermolecular electrostatic energies were computed using the VMoPro module of the MoPro suite (Guillot et al., 2001; Jelsch et al., 2005), while the molecular view with thermal ellipsoids and the isosurface representations were produced with MoProViewer (Guillot et al., 2014).
Automatic restraints of chemical equivalence and ) were applied to the electrondensity parameters such as contraction/expansion κ and κ′, valence and multipole populations P_{val} and P_{lm}. The optimal weight σ_{opt} of the restraints applying to the chargedensity parameters (atom equivalence and local symmetry) was set to 0.2, as determined by minimizing the global R_{free} factor (Brünger, 1992; Zarychta et al., 2011). The parameters κ and κ′ of H atoms were restrained to be similar (σ_{rest} = 0.02).
(Domagała & Jelsch, 2008The molecular parameters including scale factor, xyz, U_{ij}, P_{val}, P_{lm}, κ and κ′ were refined together with the block diagonal option and finally using the full normal matrix until convergence, yielding wR^{2}(I) = 3.6%. The crystallographic details of the are given in Table 1.
The topological charges were integrated on atomic basins using the program BADER (Tang et al., 2009). A parallelepiped embedding the BOH2 molecule extracted from the was defined with a margin of 3 Å around the atomic nuclei. For each deviating model, the total electron density of the molecule inside this parallelepiped was computed using the program VMoPro, with a grid step of 0.05 Å along each direction and then saved as a Gaussian cube file. Then, the program BADER was used for atomic basin definition and charge integration. The sum of the integrated electron charges was smaller than the total number of electrons in the molecule with an average lack of 0.47 e (SSD = 0.0028 e) for a total number of 102 electrons. The unattributed was evenly redistributed on the 29 integrated atomic basin charges.
3. Methodology
3.1. Leastsquares and uncertainties
The leastsquares MoPro (Guillot et al., 2001; Jelsch et al., 2005), software dedicated to chargedensity A multipolar chargedensity model defined according to the formalism of Hansen & Coppens (1978) can be refined for crystal structures when ultra high resolution Xray diffraction data have been measured. For macromolecular structures, the transferability principle can be used to define a multipolar electrondensity model. When the is performed against the reflection intensities, the minimized function E is defined as
is implemented inwhere and are the calculated and observed reflection intensities, respectively, and is the vector of the model parameters being considered in the corresponding W_{H} represents a weight for each reflection H. This weight can be taken as the squared inverse estimated error of the measured intensity.
stage. The factorThe structurefactor amplitude is obtained by summation over all atoms a in the and all symmetry operators s of the as follows:
where, for each atom a, f_{a} is the atom scattering factor, β is the dimensionless thermal tensor and X_{a} is the atom coordinates.
The leastsquares n parameters in the model and after linearization of the calculated reflection intensities around the current vector , the minimization of E is performed by solving the matrix system of normal equations:
is performed iteratively. At each cycle ofwhere is the n^{2} symmetric normal matrix, is the unknown shift vector to apply to the n variables refined. V is a vector of dimension n with elements like
The normal matrix element A_{ij} concerning the refined parameters x_{i} and x_{j} is obtained from the summation of the products of the intensity (or structurefactor) derivatives over the reflections H:
The normal matrix elements [equation (5)], through the calculated intensities, incorporate implicitly the contribution of symmetryrelated atoms in the as can be seen in equation (2). At the convergence, the variance–covariance matrix of the model parameters is obtained from B, the inverse of matrix A (Hamilton, 1964). The ith diagonal term of the matrix B provides an estimated standard deviation (e.s.d.), noted σ(x_{i}), of the parameter x_{i}. If the weighting scheme W_{H} used in the leastsquare function is not properly scaled, all e.s.d.'s have to be multiplied by the goodness of fit (GOF ≠ 1):
The C_{ij} between the parameters x_{i} and x_{j} in the is obtained by the equation
3.2. Generation of randomly deviating chargedensity models
The procedure is started from the converged chargedensity model at X^{min}. The values of the parameter vector X are assumed to be distributed according to a multidimensional Gaussian probability density function with mean and variance–covariance matrix .
If there were no correlations between parameters, the matrices A and B would be diagonal and the shifts dx_{i} to apply to each parameter to obtain a deviating model would be the e.s.d.'s multiplied by a random number:
In other words,
where R is a vector of random and independent real numbers normally distributed with a zero average and a unitary variance.
In real situations, the variance–covariance matrix B is symmetric positivedefinite but not diagonal as parameters show some correlations. The deviating parameter vector X values are generated using the following practical procedure.
Since the normal matrix A is symmetric definitepositive, the matrix A is orthogonally diagonalized at the value where E is minimal leading to the expression
where Q is an orthogonal matrix and D is diagonal which contains the strictly positive eigenvalues of A.
Therefore, its inverse matrix B can be written as
and the matrix S, a square root matrix of B, can be obtained as
The deviating parameter vector X is obtained by applying
Each element of the vector dX is a linear combination of the R elements and thus the vector X follows a multivariate Gaussian distribution. The mean vector of this distribution is equal to :
By propagation of uncertainties, the variance–covariance matrix of this multivariate Gaussian distribution is defined, using expression (9), as
The events of the normal distribution of the vector R are generated using a random Gaussian number generator with zero mean and unitary sigma. The software MoPro generates random Gaussian numbers using the `ratio of uniform deviates' method introduced by Kinderman & Monahan (1977) and augmented with quadratic bounding curves by Leva (1992). To avoid rare events which would lead to meaningless deviating chargedensity models, the algorithm is modified to generate random numbers following a truncated Gaussian function. This modification consists of reducing the infinity support of the Gaussian probability density function to a [−4; 4] interval and of normalizing the resulting function in order to obtain a unitary variance.
Following the Monte Carlo procedure described in this section, several deviating chargedensity models are generated using equation (12). The studied properties are computed on all these models and the SSDs are deduced from the sample values. The method is applied in the current study to the BOH2 molecule.
The number of deviating models required depends on the expected precision of SSDs. For any property P, assuming it follows a normal distribution with , if a sample of N events, , is taken from its distribution, the SSD, estimator of , can be defined as
The quantity follows a χ probability distribution with N  1 (for more details, see Appendix A). This implies that the expected relative standard deviation of the estimator SSD can be approximated as follows:
where is the expected uncertainty value and the standard deviation of the estimator SSD. We can select a number of events large enough to have an expected relative standard deviation value [equation (16)] smaller than a limit value e_{s} (see Fig. S1 in the supporting information). This information is relevant to estimate the number of deviating models necessary for a proper estimation of model property uncertainties. Expression (16) is however only strictly valid for a sample of random values from a normally distributed population. It is used in our study to estimate the uncertainty of the SSD for derived properties, assuming that their distributions are normal. For example, using N = 20 deviating models, for any derived property SSD, a relative standard deviation of 16% is expected. This precision is enough to estimate standard deviations of the considered properties with one significant digit.
The method is tested on the chargedensity model of the BOH2 molecule, by generating a series of 20 randomly deviating models from which various derived properties are calculated, along with their SSDs.
For some examples of derived properties, a larger sample of 500 models has been used to obtain population histograms and to check the nature of population distributions. These histograms (Laplacian and ellipticity at the bond Fig. S2. It appears almost all properties have unimodal and Gaussianlike population distributions. The histogram of wR^{2}(I) factors is also shown; the value for the refined model is 3.616%, while for the perturbed structures the R values are always higher and the average wR^{2}(I) is 3.694% with a SSD of 0.007.
electrostatic energy) are provided in4. Results
4.1. Geometry properties: distances and angles
The good accordance between SSD and e.s.d. values has been verified for the parameters used to describe the structure and charge density. For example, Fig. S3 shows the agreement between e.s.d.'s issued from the leastsquare normal matrix inversion using equation (6) and the SSD values obtained from 20 deviating structures for the atomic fractional coordinates.
The SSD has been calculated for the bond distances and angles. For comparison, e.s.d.'s have also been retrieved from the error propagation method implemented in the MoPro software, and the relative differences between the SSD and e.s.d., SSD − e.s.d./e.s.d., are calculated to check the reliability of the error propagation method.
In the plots SSD versus e.s.d. for the interatomic distances and angles between nonH atoms (Fig. 2), the points are distributed along the y = x line. Moreover, the maximal value of is 25% for interatomic distances and 30% for interatomic angles, which implies a good agreement between SSD and e.s.d. if only one significant digit is expected.
4.2. Electron density
The statistical procedure used to estimate standard deviations can be extended to any molecular property, including the static electron density. The static deformation electron density in the (C11, B1, O2) plane is considered as an example in Fig. 3(a). The SSD map (Fig. 3b) in the (C11, B1, O2) plane shows significant features near atomic nuclei, which is expected as the electron density takes large values and varies drastically in their vicinity with the nuclei coordinate shifts. These features around nuclei are anisotropic, which can be related to the positive and negative multipolar deformation density in the map (Fig. 3a) due to the formation of covalent bonds or the electron lone pairs in the Oatom case.
The SSD(ρ) level is found to be below 0.015 e Å^{−3} on the covalent bonds between nonH atoms; for bonds involving H atoms SSD(ρ) is, in comparison, slightly higher, but still below 0.020 e Å^{−3}.
4.3. Topology of covalent bonds
For each deviating model, VMoPro and the results processed by statistical analysis for SSD estimation. For each the distances between bonded atomic nuclei and the corresponding BCP position are reported, with the topological properties, in Table 2.
topological analysis is performed using the software

The intramolecular bonds involving the B atom have the largest uncertainties on the Laplacian values. The B—O bonds in particular show positive ∇^{2}ρ_{CP} Laplacian values and the largest uncertainties among covalent bonds between nonH atoms, as boron is a very light element with respect to oxygen. The B—O bonds also have the most accurate distances X—CP and Y—CP with uncertainties below 10^{−3} Å (Table 2). Among the X—H bonds, the ones with O atoms have the largest (in magnitude) Laplacian values and SSDs; the relative uncertainties are however similar, around 2.4% for all X—H bonds (Table 2).
Uncertainties of electron densities ρ and Laplacian values ∇^{2}ρ at the CPs show, in the case of X—Y bonds (hereafter, X and Y stand for nonH atoms), average values around 0.010 e Å^{−3} and 0.42 e Å^{−5} while their maxima reach, respectively, 0.014 e Å^{−3} and 0.93 e Å^{−5}. In the case of X—H bonds, the average uncertainties of ρ and ∇^{2}ρ are quite comparable with the previous ones, with, respectively, 0.014 e Å^{−3} and 0.42 e Å^{−5} and maximal values of 0.031 e Å^{−3} and 0.98 e Å^{−5}. It must be noted that, in both cases, uncertainties are dramatically below the root mean square discrepancies reported by Grabowsky et al. (2008) in a study of the charge densities of The SSD(ρ_{CP}) values are in accordance with those found in the SSD map of the static deformation density (Fig. 3b). The ρ electron density and its relative SSD are shown in Fig. 4 along the B1—O1 bond path and the SSD(ρ) error is two orders of magnitude smaller than ρ. The errors on ρ on the B1—O1 bond are comparatively lower than those on the C—O bond of oxalic acid exemplified in the Kamiński et al. (2014) study. The mean error over density 〈SSD(ρ)/ρ〉 is 1% while for the Laplacian 〈SSD(∇^{2}ρ)/∇^{2}ρ〉 reaches 3%.
The SSD values of the (λ_{1}, λ_{2}, λ_{3}) Hessian matrix ∂^{2}ρ/∂x_{i}∂x_{j} eigenvalues at the bond CPs are shown in Table S1. Examples of population histograms for the Laplacian and ellipticity at the CP of the bond B1—O1 are shown in Figs. S2(a), S2(b). It is relevant to note that the ellipticity at an electrondensity CP, which is, by definition, positive, can have a drastically asymmetric statistical density distribution when its reference value derived from the converged model is small relative to its SSD (Fig. S2b).
The plot of SSD values of distances X⋯CP versus Y⋯CP for the X—Y covalent bonds (nonH atoms) is illustrated in Fig. 5. A remarkable equality between uncertainties in the distances X⋯CP and Y⋯CP can be observed, the SSDs being generally in the 2 × 10^{−4}–4 × 10^{−4} Å range. This result can be simply explained by the high accuracy of heavyatom nucleus positions relative to BCP positions, making the uncertainty of the CP position the predominant cause of error. This is confirmed by the lower order of magnitude of the X—Y distance uncertainties compared with the ones on distances X⋯CP and Y⋯CP (Fig. 5, Table 2).
The X⋯CP and H⋯CP distance SSDs involving X—H bonds are higher, mostly in the 4 × 10^{−4}–8 × 10^{−4} Å range (Fig. S4). The observed SSD values of X⋯CP and H⋯CP distances are, in this case, more dissimilar, but of the same order of magnitude as the d(X, H) SSD. It has to be recalled here that Hatom positions were restrained during the model (§2.4); therefore the d(X⋯H) values and their uncertainties obtained depend partly on the distance restraints used.
The knowledge of uncertainties is crucial to assess the pertinence of discussions on the property values. For instance, the histogram of ∊ ellipticities with SSDs on the C—C bond CPs allows one to compare the values visually (Fig. 6). With respect to SSD values, the formally double bond C10=C11 clearly has a higher ellipticity than all other bonds. Among the four formally single bonds, the differences between ∊ values are generally significant as the standard deviation between values (0.067) is 5.7 times larger than the average SSD uncertainty (0.012). The discrepancies among the aromatic bonds are less meaningful with a standard deviation between values of 0.020, which is only two times larger than the average SSD uncertainty (0.011).
4.4. Topology of intermolecular interactions
Intermolecular interactions play a key role in crystal engineering which is an important field in chemical crystallography; therefore estimation of errors on their properties is extremely timely. In the BOH2 crystal packing, 17 unique interatomic contacts shorter than 3 Å were identified between the reference molecule and its environment, involving eight distinct neighbour molecules (Table 3). The intermolecular (3,−1) CP search has been done using the software VMoPro on the 20 deviating models. All the O⋯H hydrogen bonds show nonambiguous bond paths between the two atoms. Two of the intermolecular contact CPs have unstable bond paths, in the sense that they lead to different linked atoms within the deviating models (Table 3). The first nonstable bond path involves the phenyl H4 atom of the molecule (−x + , y − ½, z) which is connected to the phenyl C atoms of the reference molecule, C1 in 13 deviating models and C6 in the seven others. The second ambiguous bond path involves another weak phenyl⋯phenyl interaction between the H3 atom of the reference molecule and either C4 (three in 20 cases) or H4 of the molecule (x − ½, y, −z + ½) (17 in 20 cases). The C⋯H contacts can be considered as very weak hydrogen bonds [respectively, ρ = 0.0364 (8) and ρ = 0.0432 (9) e Å^{−3}] with the phenyl moiety as acceptor. Moreover, two reported van der Waals contacts between H atoms, at d(H⋯H) > 2.7 Å, yield a CP and bond path detected only in some of the deviating models and are reported in italics in Table 3. Globally, the bond paths and CPs are found to be stable in the models perturbed at standard deviation in all the strongest interactions and most of the weaker ones.
For the properties at the intermolecular bond CPs which are systematically detected in all models, the uncertainties of ρ_{CP} and Laplacian ∇^{2}ρ_{CP} values are of the same magnitude as those shown by Kamiński et al. (2014) and do not exceed 6 and 4% in relative values, respectively. Similar uncertainty could also be observed in the intermolecular area from the static deformation density SSD map (Fig. 3b), where SSD values tend to be lower than the 0.005 e Å^{−3} contour level outside of the molecule.
The mean error over density at the intermolecular CPs 〈SSD(ρ)/ρ〉 is 3%. For the Laplacian, the mean 〈SSD(∇^{2}ρ)/∇^{2}ρ〉 is 3.7% on the two hydrogen bonds while it reaches only 1.3% on the weaker interactions. The relative errors are similar for ρ_{CP} on the covalent bonds and nonbonded interactions. Conversely, Laplacian values generally have a lower relative error on weak interactions compared with strong hydrogen bonds or covalent bonds.
The SSD of G_{CP} and V_{CP}, the kinetic and density (Espinosa et al., 1998), respectively, derived from ρ_{CP} and ∇^{2}ρ_{CP}, can also be computed. The E_{HB} = −V_{CP} of the two O⋯H—O hydrogen bonds present in the BOH2 crystal packing was estimated. For O1⋯H2O, E_{HB} = 37.9 (9) and for O2⋯H1O, E_{HB} = 41 (2) kJ mol^{−1}; the relative errors are therefore 2.3 and 5.6%, respectively.
4.5. Atomic charges
The atoms in molecules (AIM) topological analysis is extended to the integrated topological properties. The series of topological analysis results are used to estimate the uncertainties of atomic basin charges (Table 4). The integrated charge SSDs are found to be higher for C atoms (0.03 to 0.06 e) than for H, B and O atoms in the current structure (below 0.02 e in general). The average SSD value over all atomic charges is 0.024 e and the maximal SSD is obtained for the C5 atom of the phenyl moiety with 0.058 e. Such values are smaller but of the same order of magnitude as the typical uncertainties of atomic valence populations obtained from the variance–covariance matrix at the end of the multipolar (Table 4). The SSDs of atomic basin electronic charges Q_{topo} are plotted against the e.s.d. of atomic valence populations P_{val} (Fig. 7). The valence population e.s.d. and SSD values are in good agreement (Table 4, Fig. S9). For most of the atoms, the valence population e.s.d. and SSD values are larger than the SSD of the corresponding integrated atomic basin charge. However, as SSD values of topological charges are consistently around a few tenths of an electron while their values can vary by several orders of magnitude (between 3 × 10^{−3} e for H72 and 2.4 e for the B atom), the corresponding relative uncertainties of Q_{topo} atomic charges can reach high values, especially for atoms bearing low integrated charges. For some H atoms, the uncertainty is larger than their weak charge (H71, H72, H82, H91) (Table 4). For the O atoms which bear a negative Q_{topo} charge of about −1.3 e, the relative error reaches on the other hand only 1.4%. The strongly positive atomic charge Q_{topo} of the B atom bonded with these two O atoms leads to a low relative error of 0.6%.

The two definitions of charges Q_{topo} and P_{val} derived are generally in good agreement, except for the B and O atoms which show much larger Q_{topo} charges. When the charge integration is carried out on the promolecule with spherical neutral atoms (IAM), the B atom turns out to have Q_{topo} = +1.60 e charge, while for the O1 atom Q_{topo} is −1.04 e, values which are far from the zero charge of a neutral atom. Therefore, the raw topological charges are not always to be compared with the P_{val} derived charges when atoms with very dissimilar atomic numbers, such as B and O, form a Nonzero Q_{topo} charges were recently reported by Stachowicz et al. (2017) for a CaF_{2} crystal when using the IAM model.
4.6. Electrostatic potential
The 0.001 a.u. (a.u. = atomic units) electrondensity isosurface of the isolated molecule was chosen to map the molecular electrostatic potential Φ and its sample standard deviation (Fig. 8a). On this surface, the SSD of the electrostatic potential on the molecular surface lies between 5 × 10^{−3} and 2 × 10^{−2} e Å^{−1} and the average `signal over uncertainty' ratio reaches 4.8. As depicted in Fig. 8 and Fig. S5, there is no clear correlation between the electrostatic potential SSD on the isosurface and its absolute value on the electrondensity isosurface (correlation coefficient = 17%). Regions of highest can be seen nearby the H2, H3 and H4 phenyl ring H atoms and close to the B atom (blue patches on Fig. 8b). These locally large values can be explained by the fact that these H atoms present the largest thermal displacement parameters (2.9 < B_{eq} < 3.2 Å^{2}) in the BOH2 compound, leading to larger uncertainties on their positions. Similarly, the high values observed in the vicinity of the B atom can be explained by an e.s.d. on its valence population that is twice as large as those of their neighbour O atoms (Table 4), locally increasing the SSD of the molecular electrostatic potential. Molecular surface points which are mostly under the electrostatic influence of these atoms show consequently particularly large values. Nearly 90% of the considered surface points present values lying between 0.008 and 0.016 e Å^{−1}, distributed around the 0.012 e Å^{−1} average value and spanning the whole electrostatic potential values range (−0.16 to +0.32 e Å^{−1}). The threedimensional distribution of values is presented in Fig. S6 by the mean of three superimposed 0.04, 0.02 and 0.01 e Å^{−1} isosurfaces. As expected, the increases strongly in close vicinity to the atomic nuclei, where electrostatic potential variations become large due to the perturbed nuclei positions and valence populations in the 20 considered models contributing to the statistics. The volume of space located between the 0.01 and 0.02 e Å^{−1} isosurfaces of encompasses typical intermolecular interaction distances, i.e. regions where electrostatic potential is usually interpreted. The ratio is useful to estimate the electrostatic potential statistical significance on various regions of the electrondensity surface (Kamiński et al., 2014). This property, mapped on the electrondensity surface, is represented in Fig. 8(c). The electrostatic potential is therefore statistically very significant in regions of strong values, with reaching 16 in our case. Conversely, becomes lower than unity when the electrostatic potential falls below ∼0.02 e Å^{−1}, in absolute value, which can be interpreted as a broadening of the zero potential contour regions on the molecular surface, as represented in white in Fig. 8(c) using a significance criterion of . Regions located either side of this lowpotential stripe can then be considered as either electropositive or electronegative with a high degree of confidence.
4.7. Electrostatic energy
Eight unique dimers of molecules in contact have been identified in the crystal packing. Considering each dimer, the intermolecular electrostatic energy is computed for the 20 perturbed models using the EP/MM method (Volkov et al., 2004) as implemented in the software VMoPro, and the SSD is calculated as the uncertainty estimator (Table 5).

Absolute errors of intermolecular electrostatic interaction energies in dimers of BOH2 molecules appear consistently of a few kJ mol^{−1}. Therefore the SSD relative error is as low as 7% for the largest value E_{elec} = −62.2 kJ mol^{−1} but for the weakest interactions the SSD is larger than E_{elec} itself. Such large relative errors confirm clearly that weak electrostatic interaction energies of a few kJ mol^{−1} cannot be interpreted as either stabilizing or destabilizing. This is perfectly in line with the chemical accuracy in computational chemistry, generally considered to be around 5 kJ mol^{−1} (or 1 kcal mol^{−1}) (Perdew et al., 1999). For the energy summed over all dimers, the error reaches 19%. As the energy value results from an integration product between the electron density and the electrostatic potential, the relative errors of the two factors accumulate to yield a larger relative error.
Examples of Gaussianlike population histograms for electrostatic interaction energies are shown in the supporting information for symmetry operations (x − 1, y, z) and (−x, −y + 2, −z) (Fig. S2).
4.8. Parameters to take into account
In the method proposed, the generation of a series of deviating models is done by the calculation of the square root matrix S [equation (11)] which is obtained after diagonalization of the full normal matrix A whatever the derived property of interest. In practice, the procedure bears some similarity to a step [equation (3)], but the inverted normal matrix B is replaced by its square root S [equations (11) and (12)] and the vector V is replaced by random numbers R.
The SSDs of the dimers' electrostatic energy obtained from 20 deviating models generated starting with the reduced leastsquares normal matrix obtained excluding the contributions of the U_{ij} thermal displacement parameters are also shown in Table 5. Nearly all these SSDs are smaller compared with the standard procedure where the full normal matrix is used. The SSD of the total E_{elec} value is significantly reduced from 14.8 to 8.9 kJ mol^{−1} when thermal displacement parameters are excluded from the normal matrix. Although the U_{ij} parameters are not directly involved in the equation describing the static electron density and the electrostatic potential, they do have an impact on the magnitude of SSD values.
This is due to the properties of the inversion of the symmetric positivedefinite matrix. It is demonstrated in Appendix B that, when more parameters are refined, the diagonal elements of the inverted normal matrix B = A^{−1} take larger values. Consequently, when the number of refined parameters is increased, the e.s.d.'s of parameters become larger and the SSDs of derived properties also tend to increase. This is especially the case when there are significant correlations between parameters. Obtaining very high resolution in the diffraction data set tends actually to globally diminish the correlations between parameters (Jelsch et al., 2000) and helps in the deconvolution between thermal displacement and chargedensity parameters.
Some properties may involve only part of the parameters, such as electrostatic interaction energy between molecular fragments. If this property depends only on a few atom parameters, the `square root matrix S calculation' step [equation (11)] could in principle be performed considering the reduced normal matrix corresponding only to these specific atomic parameters. This will however lead to an underestimation of SSD values. It is therefore recommended that SSDs are obtained using a full normal matrix issued from all parameters. For this reason, thermal parameters should be taken into account in the normal matrix calculation when generating the perturbed structures, although they do not have a direct impact on the charge density.
5. Conclusion
At the convergence of a leastsquares crystallographic P in the perturbed models appears to be generally within one SSD from the final refined value; in the case of topological integrated charges and electrostatic energies, it was, for instance, found that .
against diffraction data, the e.s.d.'s of the parameters used to model the molecular structure and electron density can be directly retrieved. However, the uncertainties on derived molecular properties are not readily available. To estimate the errors of properties, series of models at `standard deviation' from the final refined model can be easily generated by using vectors of random numbers and a square root of the inverted normal matrix. The SSDs obtained for the properties derived from a sample of such deviating structures can be used as estimated values of their uncertainties. For instance, samples of 20 perturbed structures yield SSD values with an expected relative precision of 16%. The average value of propertiesIn the BOH2 structure, the SSD of the electron density at the X—Y bond CPs is in the range 0.01 to 0.04 e Å^{−3}, which represents 0.5 to 2% in relative value. The average SSD on the corresponding Laplacian values is 0.42 e Å^{−3} and the average relative error SSD(∇^{2}ρ)/∇^{2}ρ is 3%. The average uncertainty on the ellipticity ∊ on X—Y bond CPs is found to be around 0.01 and is usually not dependent on the ellipticity value ranging here from 0.002 to 0.33. For X—H bonds, the average SSD(∊) is 0.007, while the maximal value ∊ is 0.012. For interacting molecular dimers of the BOH2 molecule in the crystal, the error on the electrostatic energy is typically in the 2 to 4 kJ mol^{−1} range. Intermolecular topological bond paths were found to be stable and preserved in most of the 17 interactions, except for four weak contacts. The SSD of the electrostatic potential on the molecular surface lies between 5 × 10^{−3} and 2 × 10^{−2} e Å^{−1}. High absolute values of electrostatic potential, which are usually interpreted as electronegative or electropositive sites, are shown to be significant with high signalovernoise ratios.
The availability of estimated errors is important for the proper interpretation of experimental chargedensity results, for instance, in the comparison of properties among similar chemical groups in a molecule, or of independent molecules in the
Discrepancies found in the properties of chemically equivalent atoms or of covalent bonds are physically meaningful only if they are significantly larger than the estimated error.The comparison of closely related but different compounds such as topological properties in different et al. (1999) and Grabowsky et al. (2008) is also more pertinent when an estimation of errors is available.
as investigated by FlaigOne should also recall that the actual errors obtained by the SSD method give information about the precision but may not take into account the effects of systematic errors on model accuracy. The structural and chargedensity parameters may be driven away from their `true' values to compensate for the systematic errors, while the crystallographic R factors may not be significantly worsened.
APPENDIX A
For any parameter Y distributed following a normal distribution with , if a sample of N events, , is taken from its distribution, the sample standard deviation s, biased estimator of , can be defined as
The quantity follows a probability distribution with N  1 with a s^{2} sample variance. The standard deviation of a χ probability distribution function with k is defined as
Using this expression, the standard deviation of the sample variance s^{2} distribution can be derived,
and, by propagation of the uncertainty, the standard deviation of the sample standard deviation s distribution is approximated by
where is the mean of the sample standard deviation s distribution.
As an estimator of , the expected value of the sample standard deviation s is well known to be approximately equal to and thus the relative standard deviation on becomes
APPENDIX B
The goal of this appendix is to discuss the way, after the convergence of an m+n parameter model to calculate the e.s.d.'s of only m parameters. The standard way is to derive the e.s.d.'s from the diagonal terms of the inverse matrix of the full (m+n, m+n) normal matrix A. To reduce computational resources, one could deduce e.s.d.'s starting from a principal (m, m) submatrix U of the normal matrix A corresponding to the m parameters considered. This second method will lead to underestimated e.s.d.'s as explained below. This demonstration uses the properties of symmetric positivedefinite matrices.
In the simple case of a x_{1} and x_{2}), the normal matrix can be written as:
with only two parameters (where r is the ratio of the weighted scalar product between the two sets of intensity partial derivatives and [see equation (4)] and the product of their weighted norms. r can be considered as the cosine between two vectors and therefore −1 ≤ r ≤ 1.
The inverted normal matrix is then
One can note that −r is equal to the between the two parameters. The e.s.d.'s are increased, as they are divided by (1r^{2}) when the two parameters are refined together, illustrating the strong impact of a large parameter correlation.
In the general case, let us suppose the full normal matrix A is invertible, positivedefinite and decompose it into
with U and W principal (m,m) and (n,n) submatrices of A corresponding, respectively, to the m parameters of interest and the n extra ones, and V its offdiagonal (m,n) submatrix.
As principal submatrices of the positivedefinite matrix A, U and W are positivedefinite matrices. The inverse matrix of A, noted , is also positivedefinite and can be partitioned into four blocks as
where C and E are the positivedefinite principal (m,m) and (n,n) submatrices, respectively, and D is the offdiagonal (m,n) submatrix.
The product yields the identity matrix :
By identification, the following relations apply:
As mentioned previously, the submatrix U is invertible. Thus, the equations (27b) and (27c) imply, respectively, that the matrices D and C satisfy
Then, the combination of equations (28) and (29) yields
As a principal submatrix of the positive matrix A^{−1}, E is also a positivedefinite matrix and its Cholesky decomposition is written, with L a lower triangular matrix, as follows:
Then, the combination of equations (30) and (31) gives
Let us define the (m,n) matrix M as ; equation (32) becomes
The diagonal elements of the product matrix , noted T, are positive numbers:
Therefore, the first m diagonal element of is augmented compared with those of matrix :
Assuming GOF = 1, equation (35) implies the following inequality between the e.s.d.'s of the m parameters of interest derived from the full normal matrix A [e.s.d.^{A}_{ii} = (A_{ii}^{1})^{1/2}] and those derived from the reduced matrix U [e.s.d.^{U}_{ii} = (U_{ii}^{1})^{1/2}]:
In other words, deducing the e.s.d.'s on the m parameters from a reduced normal matrix U leads to an underestimation of uncertainties. This conclusion is valid whatever the subset of parameters considered. One can note that in the particular case in which there were no correlations between the m parameters and the remaining n others, the matrix V would be equal to zero. As a result, A would be a block diagonal matrix, M = 0 and A^{1}_{ii} = U^{1}_{ii}: the e.s.d.'s of parameters would not be underestimated if deduced from the reduced normal matrix U.
Supporting information
CCDC reference: 1829445
https://doi.org/10.1107/S2053273318004308/ae5043sup1.cif
contains datablock I. DOI:Structure factors: contains datablock I. DOI: https://doi.org/10.1107/S2053273318004308/ae5043Isup2.hkl
Supporting information file. DOI: https://doi.org/10.1107/S2053273318004308/ae5043Isup3.cml
Supporting figures and table. DOI: https://doi.org/10.1107/S2053273318004308/ae5043sup4.pdf
C_{11}H_{15}BO_{2}  F(000) = 816 
M_{r} = 190.04  D_{x} = 1.165 Mg m^{−}^{3} 
Orthorhombic, Pbca  Mo Kα radiation, λ = 0.71073 Å 
Hall symbol: p_2ac_2ab  Cell parameters from 50963 reflections 
a = 7.5200 (1) Å  θ = 2.6–60.0° 
b = 9.3837 (1) Å  µ = 0.08 mm^{−}^{1} 
c = 30.7120 (5) Å  T = 100 K 
V = 2167.21 (5) Å^{3}  Block, colourless 
Z = 8  0.34 × 0.18 × 0.10 mm 
Rigaku MicroMaxHF Pilatus 200K diffractometer  11803 independent reflections 
Radiation source: finefocus sealed tube  10496 reflections with > 2.0σ(I) 
Confocal Max  optic monochromatorR_{int} = 0.031 
Detector resolution: 5.8140 pixels mm^{1}  θ_{max} = 52.2°, θ_{min} = 2.7° 
Fullsphere data collection, phi and ω scans  h = 0→16 
Absorption correction: multiscan CrysAlisPro 1.171.38.37f (Rigaku Oxford Diffraction, 2015) Empirical absorption correction using spherical harmonics, implemented in SCALE3 ABSPACK scaling algorithm.  k = 0→20 
T_{min} = 0.472, T_{max} = 0.999  l = 0→68 
117942 measured reflections 
Refinement on F^{2}  Primary atom site location: structureinvariant direct methods 
Leastsquares matrix: full  Secondary atom site location: difference Fourier map 
R[F^{2} > 2σ(F^{2})] = 0.026  Hydrogen site location: difference Fourier map 
wR(F^{2}) = 0.039  All Hatom parameters refined 
S = 0.95  w = 1/[3.3*σ^{2}(F_{o}^{2})] 
14958 reflections  (Δ/σ)_{max} < 0.001 
574 parameters  Δρ_{max} = 0.32 e Å^{−}^{3} 
0 restraints  Δρ_{min} = −0.16 e Å^{−}^{3} 
Experimental. CrysAlisPro 1.171.38.37f (Rigaku Oxford Diffraction, 2015) Empirical absorption correction using spherical harmonics, implemented in SCALE3 ABSPACK scaling algorithm. 
Refinement. Refinement of F^{2} against reflections. The threshold expression of F^{2} > 2sigma(F^{2}) is used for calculating Rfactors(gt) and is not relevant to the choice of reflections for refinement. Rfactors based on F^{2} are statistically about twice as large as those based on F, and Rfactors based on ALL data will be even larger. 
x  y  z  U_{iso}*/U_{eq}  
B1  0.251627 (19)  0.997843 (15)  0.028837 (5)  0.011002 (12)  
C1  0.96386 (2)  0.89957 (2)  0.172477 (6)  0.015083 (13)  
C2  0.92648 (3)  0.97351 (3)  0.211051 (7)  0.019388 (17)  
C3  1.02734 (3)  1.09177 (3)  0.223317 (8)  0.02290 (2)  
C4  1.16784 (3)  1.13760 (3)  0.197292 (8)  0.022720 (19)  
C5  1.20751 (3)  1.06375 (3)  0.159143 (8)  0.021551 (17)  
C6  1.10594 (3)  0.94602 (2)  0.146808 (7)  0.018120 (16)  
C7  0.85128 (3)  0.77401 (2)  0.159105 (8)  0.018915 (15)  
C8  0.67225 (3)  0.81543 (2)  0.138779 (7)  0.016543 (14)  
C9  0.68873 (2)  0.88966 (2)  0.094495 (6)  0.015373 (13)  
C10  0.51313 (2)  0.893953 (19)  0.071361 (5)  0.014376 (13)  
C11  0.43514 (2)  1.006991 (19)  0.052461 (6)  0.012834 (13)  
H2  0.819 (3)  0.9396 (12)  0.2309 (6)  0.0371 (12)  
H3  0.9964 (12)  1.1450 (18)  0.2528 (10)  0.0408 (13)  
H4  1.242 (2)  1.230 (3)  0.2063 (3)  0.0396 (12)  
H5  1.313 (3)  1.0982 (12)  0.1390 (7)  0.0369 (13)  
H6  1.1366 (11)  0.8916 (18)  0.1174 (9)  0.0332 (11)  
H10  0.4449 (18)  0.796 (3)  0.07101 (15)  0.0307 (10)  
H11  0.5012 (16)  1.105 (2)  0.05400 (12)  0.0276 (8)  
H71  0.922 (2)  0.712 (2)  0.1362 (8)  0.0356 (10)  
H72  0.8237 (10)  0.711 (2)  0.1870 (9)  0.0378 (11)  
H81  0.598 (2)  0.720 (3)  0.13438 (19)  0.0355 (10)  
H82  0.602 (2)  0.881 (2)  0.1604 (7)  0.0327 (10)  
H91  0.7387 (16)  0.992 (3)  0.09795 (18)  0.0289 (10)  
H92  0.778 (3)  0.8308 (18)  0.0752 (6)  0.0320 (10)  
H1O  0.226 (2)  1.200 (3)  0.0175 (2)  0.0252 (11)  
H2O  0.058 (5)  0.8762 (5)  0.0102 (6)  0.0264 (12)  
O1  0.163964 (19)  1.115603 (16)  0.013417 (5)  0.012248 (10)  
O2  0.16948 (2)  0.868002 (16)  0.023395 (6)  0.014655 (11) 
U^{11}  U^{22}  U^{33}  U^{12}  U^{13}  U^{23}  
B1  0.00972 (5)  0.00844 (5)  0.01485 (6)  0.00044 (4)  −0.00137 (4)  −0.00029 (4) 
C1  0.01166 (5)  0.01836 (6)  0.01523 (6)  0.00001 (5)  −0.00236 (5)  0.00324 (5) 
C2  0.01600 (6)  0.02737 (8)  0.01479 (7)  −0.00194 (6)  0.00015 (5)  0.00141 (6) 
C3  0.02169 (8)  0.03025 (10)  0.01677 (8)  −0.00307 (8)  −0.00256 (7)  −0.00358 (8) 
C4  0.01826 (7)  0.02603 (9)  0.02388 (9)  −0.00496 (7)  −0.00482 (7)  −0.00090 (7) 
C5  0.01485 (7)  0.02505 (9)  0.02475 (9)  −0.00372 (6)  0.00160 (6)  0.00188 (7) 
C6  0.01488 (6)  0.02091 (7)  0.01856 (7)  −0.00037 (6)  0.00202 (5)  0.00102 (6) 
C7  0.01656 (6)  0.01716 (6)  0.02302 (8)  −0.00075 (5)  −0.00683 (6)  0.00576 (6) 
C8  0.01272 (6)  0.02000 (7)  0.01691 (7)  −0.00187 (5)  −0.00238 (5)  0.00355 (5) 
C9  0.01142 (5)  0.01929 (6)  0.01541 (6)  0.00064 (5)  −0.00238 (4)  0.00201 (5) 
C10  0.01289 (5)  0.01346 (5)  0.01678 (6)  0.00061 (4)  −0.00458 (4)  0.00010 (4) 
C11  0.01071 (5)  0.01216 (6)  0.01563 (6)  −0.00073 (4)  −0.00272 (5)  0.00079 (4) 
H2  0.030 (5)  0.053 (5)  0.028 (5)  −0.008 (2)  0.006 (2)  0.003 (2) 
H3  0.040 (5)  0.053 (6)  0.030 (6)  0.000 (2)  −0.002 (2)  −0.015 (3) 
H4  0.036 (5)  0.036 (5)  0.047 (5)  −0.011 (3)  −0.009 (2)  −0.006 (2) 
H5  0.025 (5)  0.049 (5)  0.036 (5)  −0.011 (2)  0.004 (3)  0.006 (2) 
H6  0.029 (5)  0.043 (5)  0.027 (5)  −0.0004 (18)  0.002 (2)  −0.007 (3) 
H10  0.029 (4)  0.019 (4)  0.044 (5)  −0.004 (2)  −0.0095 (17)  0.0027 (17) 
H11  0.025 (3)  0.021 (4)  0.037 (4)  −0.007 (2)  −0.0035 (14)  0.0028 (15) 
H71  0.030 (4)  0.037 (4)  0.040 (5)  0.004 (2)  −0.007 (2)  −0.007 (3) 
H72  0.035 (4)  0.041 (4)  0.037 (5)  −0.0087 (18)  −0.0088 (19)  0.017 (3) 
H81  0.034 (4)  0.027 (4)  0.046 (5)  −0.009 (2)  −0.0125 (18)  0.0084 (18) 
H82  0.030 (4)  0.039 (4)  0.029 (4)  0.004 (2)  0.002 (2)  0.002 (2) 
H91  0.030 (4)  0.025 (5)  0.032 (4)  −0.006 (2)  −0.0037 (15)  0.0017 (16) 
H92  0.028 (4)  0.038 (4)  0.031 (4)  0.011 (2)  −0.001 (2)  −0.005 (2) 
H1O  0.023 (4)  0.016 (5)  0.036 (5)  −0.004 (3)  −0.0049 (17)  0.0012 (18) 
H2O  0.021 (5)  0.017 (4)  0.040 (4)  0.0015 (16)  −0.010 (3)  0.0012 (15) 
O1  0.01025 (4)  0.00765 (4)  0.01884 (6)  0.00014 (4)  −0.00214 (4)  0.00036 (3) 
O2  0.01175 (5)  0.00762 (4)  0.02459 (7)  0.00039 (4)  −0.00509 (5)  0.00004 (4) 
B1—O1  1.3711 (2)  C7—C8  1.5340 (3) 
B1—O2  1.3762 (2)  C7—H71  1.06 (3) 
B1—C11  1.5614 (2)  C7—H72  1.06 (3) 
C1—C7  1.5079 (3)  C8—C9  1.5331 (3) 
C1—C6  1.3975 (3)  C8—H82  1.05 (3) 
C1—C2  1.4014 (3)  C8—H81  1.06 (3) 
C2—C3  1.3959 (4)  C9—C10  1.5001 (2) 
C2—H2  1.06 (3)  C9—H91  1.04 (3) 
C3—C4  1.3929 (4)  C9—H92  1.05 (3) 
C3—H3  1.06 (4)  C10—C11  1.3439 (2) 
C4—C5  1.3935 (4)  C10—H10  1.06 (3) 
C4—H4  1.07 (3)  C11—H11  1.04 (2) 
C5—C6  1.3955 (3)  H1O—O1  0.93 (4) 
C5—H5  1.06 (3)  H2O—O2  0.94 (4) 
C6—H6  1.06 (3)  
O1—B1—O2  117.113 (12)  C8—C7—H71  108.1 (6) 
O1—B1—C11  122.759 (12)  C8—C7—H72  107.3 (6) 
O2—B1—C11  120.118 (13)  H71—C7—H72  109.1 (9) 
C7—C1—C6  121.287 (16)  C9—C8—C7  113.902 (14) 
C7—C1—C2  120.296 (18)  C9—C8—H82  109.7 (7) 
C6—C1—C2  118.414 (17)  C9—C8—H81  108.2 (5) 
C1—C2—C3  120.846 (19)  C7—C8—H82  109.5 (7) 
C1—C2—H2  119.3 (9)  C7—C8—H81  107.6 (5) 
C3—C2—H2  119.8 (9)  H82—C8—H81  107.8 (8) 
C4—C3—C2  120.18 (2)  C8—C9—C10  111.174 (13) 
C4—C3—H3  121 (1)  C8—C9—H91  111.1 (6) 
C2—C3—H3  119.1 (9)  C8—C9—H92  108.2 (6) 
C3—C4—C5  119.43 (2)  C10—C9—H91  110.0 (6) 
C3—C4—H4  120.0 (8)  C10—C9—H92  108.0 (6) 
C5—C4—H4  120.6 (8)  H91—C9—H92  108.3 (8) 
C4—C5—C6  120.31 (2)  C9—C10—C11  127.586 (13) 
C4—C5—H5  120.0 (9)  C9—C10—H10  114.2 (6) 
C6—C5—H5  119.7 (9)  C11—C10—H10  118.2 (7) 
C1—C6—C5  120.812 (17)  C10—C11—B1  122.888 (13) 
C1—C6—H6  119.7 (9)  C10—C11—H11  117.8 (6) 
C5—C6—H6  119.5 (9)  B1—C11—H11  119.3 (7) 
C1—C7—C8  113.930 (14)  B1—O1—H1O  113.6 (9) 
C1—C7—H71  109.2 (7)  B1—O2—H2O  112.5 (9) 
C1—C7—H72  109.1 (6)  
B1—C11—C10—C9  179.41 (3)  C7—C8—C9—H92  47 (1) 
B1—C11—C10—H10  −2.0 (4)  C8—C9—C10—C11  131.14 (4) 
C1—C7—C8—C9  67.25 (3)  C8—C9—C10—H10  −47 (1) 
C1—C7—C8—H82  −56 (2)  C9—C8—C7—H71  −54 (2) 
C1—C7—C8—H81  −172.9 (4)  C9—C8—C7—H72  −171.8 (4) 
C1—C6—C5—C4  −0.39 (3)  C9—C10—C11—H11  −0.9 (3) 
C1—C6—C5—H5  −179.1 (3)  C10—C9—C8—H82  −71 (2) 
C1—C2—C3—C4  −0.32 (3)  C10—C9—C8—H81  46 (1) 
C1—C2—C3—H3  −179.6 (3)  C10—C11—B1—O1  173.27 (3) 
C2—C1—C7—C8  78.15 (3)  C10—C11—B1—O2  −5.59 (3) 
C2—C1—C7—H71  −161.0 (8)  C11—C10—C9—H91  7.7 (4) 
C2—C1—C7—H72  −42 (1)  C11—C10—C9—H92  −110 (2) 
C2—C1—C6—C5  −0.40 (3)  C11—B1—O1—H1O  2.1 (4) 
C2—C1—C6—H6  −179.9 (3)  C11—B1—O2—H2O  177.0 (4) 
C2—C3—C4—C5  −0.48 (3)  H2—C2—C3—H3  1.2 (5) 
C2—C3—C4—H4  178.1 (3)  H3—C3—C4—H4  −2.6 (5) 
C3—C4—C5—C6  0.83 (3)  H4—C4—C5—H5  0.9 (6) 
C3—C4—C5—H5  179.5 (3)  H5—C5—C6—H6  0.4 (5) 
C3—C2—C1—C7  −178.62 (3)  H10—C10—C9—H91  −170.9 (6) 
C3—C2—C1—C6  0.75 (3)  H10—C10—C9—H92  71 (2) 
C4—C3—C2—H2  −179.5 (3)  H10—C10—C11—H11  177.7 (5) 
C4—C5—C6—H6  179.1 (3)  H11—C11—B1—O1  −6.4 (4) 
C5—C4—C3—H3  178.8 (3)  H11—C11—B1—O2  174.7 (4) 
C5—C6—C1—C7  178.97 (3)  H71—C7—C8—H82  −177.5 (5) 
C6—C1—C7—C8  −101.20 (3)  H71—C7—C8—H81  66 (2) 
C6—C1—C7—H71  19.7 (8)  H72—C7—C8—H82  65 (2) 
C6—C1—C7—H72  139 (1)  H72—C7—C8—H81  −52 (1) 
C6—C1—C2—H2  180.0 (3)  H81—C8—C9—H91  169.0 (6) 
C6—C5—C4—H4  −177.8 (3)  H81—C8—C9—H92  −72 (2) 
C7—C1—C6—H6  −0.5 (3)  H82—C8—C9—H91  52 (1) 
C7—C1—C2—H2  0.6 (3)  H82—C8—C9—H92  170.3 (6) 
C7—C8—C9—C10  165.61 (3)  H1O—O1—B1—O2  −179.0 (4) 
C7—C8—C9—H91  −71 (2)  H2O—O2—B1—O1  −2.0 (4) 
D—H···A  D—H  H···A  D···A  D—H···A 
C10—H10···O1^{i}  1.06 (3)  2.58 (2)  3.4297 (2)  137 (1) 
O1—H1O···O2^{ii}  0.93 (4)  1.77 (4)  2.6967 (2)  176 (1) 
O2—H2O···O1^{iii}  0.94 (4)  1.82 (4)  2.7549 (2)  177 (1) 
Symmetry codes: (i) −x+1/2, y−1/2, z; (ii) −x+1/2, y+1/2, z; (iii) −x, −y+2, −z. 
Acknowledgements
We would like to thank Dr Alejandro DíazMoscoso for the synthesis of the compound.
References
Allen, F. H. & Bruno, I. J. (2010). Acta Cryst. B66, 380–386. Web of Science CSD CrossRef CAS IUCr Journals Google Scholar
Bader, R. F. W. (1990). Atoms in Molecules: a Quantum Theory, 1st ed. International Series of Monographs on Chemistry 22. Oxford: Clarendon Press. Google Scholar
Bader, R. F. W., Carroll, M. T., Cheeseman, J. R. & Chang, C. (1987). J. Am. Chem. Soc. 109, 7968–7979. CrossRef CAS Web of Science Google Scholar
Brünger, A. T. (1992). Nature, 355, 472–475. PubMed Web of Science Google Scholar
Burla, M. C., Caliandro, R., Carrozzini, B., Cascarano, G. L., Cuocci, C., Giacovazzo, C., Mallamo, M., Mazzone, A. & Polidori, G. (2015). J. Appl. Cryst. 48, 306–309. Web of Science CrossRef CAS IUCr Journals Google Scholar
Côté, A. P., ElKaderi, H. M., Furukawa, H., Hunt, J. R. & Yaghi, O. M. (2007). J. Am. Chem. Soc. 129, 12914–12915. Web of Science PubMed Google Scholar
Ding, X. S., Guo, J., Feng, X. A., Honsho, Y., Guo, J. D., Seki, S., Maitarad, P., Saeki, A., Nagase, S. & Jiang, D. L. (2011). Angew. Chem. Int. Ed. 50, 1289–1293. CrossRef CAS Google Scholar
Dittrich, B., Koritsánszky, T., Grosche, M., Scherer, W., Flaig, R., Wagner, A., Krane, H. G., Kessler, H., Riemer, C., Schreurs, A. M. M. & Luger, P. (2002). Acta Cryst. B58, 721–727. Web of Science CSD CrossRef CAS IUCr Journals Google Scholar
Domagała, S. & Jelsch, C. (2008). J. Appl. Cryst. 41, 1140–1149. Web of Science CrossRef IUCr Journals Google Scholar
Espinosa, E., Molins, E. & Lecomte, C. (1998). Chem. Phys. Lett. 285, 170–173. Web of Science CrossRef CAS Google Scholar
Flaig, R., Koritsánszky, T., Janczak, J., Krane, H.G., Morgenroth, W. & Luger, P. (1999). Angew. Chem. Int. Ed. 38, 1397–1400. Web of Science CrossRef CAS Google Scholar
Förster, D., Wagner, A., Hübschle, C. B., Paulmann, C. & Luger, P. (2006). Z. Naturforsch. B, 62, 696–704. Google Scholar
Gelbrich, T., Sampson, D. & Hursthouse, M. B. (2000). University of Southampton, Report Archive. Google Scholar
Grabowsky, S., Pfeuffer, T., Morgenroth, W., Paulmann, C., Schirmeister, T. & Luger, P. (2008). Org. Biomol. Chem. 6, 2295–2307. Web of Science CSD CrossRef PubMed CAS Google Scholar
Guillot, B., Enrique, E., Huder, L. & Jelsch, C. (2014). Acta Cryst. A70, C279. CrossRef IUCr Journals Google Scholar
Guillot, B., Viry, L., Guillot, R., Lecomte, C. & Jelsch, C. (2001). J. Appl. Cryst. 34, 214–223. Web of Science CrossRef CAS IUCr Journals Google Scholar
Hamilton, W. C. (1964). Statistics in Physical Science. New York, USA: The Ronald Press Company. Google Scholar
Hansen, N. K. & Coppens, P. (1978). Acta Cryst. A34, 909–921. CrossRef CAS IUCr Journals Web of Science Google Scholar
Jelsch, C., Guillot, B., Lagoutte, A. & Lecomte, C. (2005). J. Appl. Cryst. 38, 38–54. Web of Science CrossRef IUCr Journals Google Scholar
Jelsch, C., Teeter, M. M., Lamzin, V., PichonPesme, V., Blessing, R. H. & Lecomte, C. (2000). Proc. Natl Acad. Sci. USA, 97, 3171–3176. Web of Science CrossRef PubMed CAS Google Scholar
Kamiński, R., Domagała, S., Jarzembska, K. N., Hoser, A. A., SanjuanSzklarz, W. F., Gutmann, M. J., Makal, A., Malińska, M., Bąk, J. M. & Woźniak, K. (2014). Acta Cryst. A70, 72–91. Web of Science CSD CrossRef IUCr Journals Google Scholar
Kinderman, A. J. & Monahan, J. F. (1977). ACM Trans. Math. Software (TOMS), 3, 257–260. CrossRef Google Scholar
Kissel, C., Speranza, F. & Milicevic, V. (1995). J. Geophys. Res. 100(B8), 14999–15007. CrossRef Google Scholar
Krause, L., Niepötter, B., Schürmann, C. J., Stalke, D. & HerbstIrmer, R. (2017). IUCrJ, 4, 420–430. Web of Science CSD CrossRef CAS PubMed IUCr Journals Google Scholar
Leva, J. L. (1992). ACM Trans. Math. Software, 18, 449–453. CrossRef Google Scholar
Madsen, A. Ø. & Hoser, A. A. (2014). J. Appl. Cryst. 47, 2100–2104. Web of Science CrossRef CAS IUCr Journals Google Scholar
Messerschmidt, M., Scheins, S. & Luger, P. (2005). Acta Cryst. B61, 115–121. Web of Science CSD CrossRef CAS IUCr Journals Google Scholar
Perdew, J. P., Kurth, S., Zupan, A. & Blaha, P. (1999). Phys. Rev. Lett. 82, 2544–2547. CrossRef CAS Google Scholar
Rigaku (2013). CrystalClear SM Expert 2.1 b29. Rigaku Corporation, Tokyo, Japan. Google Scholar
Rigaku Oxford Diffraction (2015). CrysAlisPro 1.171.38.37f. Rigaku Oxford Diffraction, Yarnton, England. Google Scholar
Sheldrick, G. M. (2015). Acta Cryst. C71, 3–8. Web of Science CrossRef IUCr Journals Google Scholar
Spitler, E. L. & Dichtel, W. R. (2010). Nat. Chem. 2, 672–677. CrossRef CAS Google Scholar
Stachowicz, M., Malinska, M., Parafiniuk, J. & Woźniak, K. (2017). Acta Cryst. B73, 643–653. CSD CrossRef IUCr Journals Google Scholar
Su, Z. & Coppens, P. (1998). Acta Cryst. A54, 646–652. Web of Science CrossRef CAS IUCr Journals Google Scholar
Tang, W., Sanville, E. & Henkelman, G. (2009). J. Phys. Condens. Matter, 21, 084204. Web of Science CrossRef PubMed Google Scholar
Volkov, A., Koritsanszky, T. & Coppens, P. (2004). Chem. Phys. Lett. 391, 170–175. Web of Science CrossRef CAS Google Scholar
Volkov, A., Macchi, P., Farrugia, L. J., Gatti, C., Mallinson, P. R., Richter, T. & Koritsanszky, T. (2006). XD2006. Revision 5.34. University of New York at Buffalo, New York, USA. Google Scholar
Yang, W. Q., Gao, X. M. & Wang, B. H. (2003). Med. Res. Rev. 23, 346–368. CrossRef CAS Google Scholar
Zarychta, B., Zaleski, J., Kyzioł, J., Daszkiewicz, Z. & Jelsch, C. (2011). Acta Cryst. B67, 250–262. Web of Science CSD CrossRef CAS IUCr Journals Google Scholar
Zhurov, V. V., Zhurova, E. A. & Pinkerton, A. A. (2008). J. Appl. Cryst. 41, 340–349. Web of Science CrossRef CAS IUCr Journals Google Scholar
This is an openaccess article distributed under the terms of the Creative Commons Attribution (CCBY) Licence, which permits unrestricted use, distribution, and reproduction in any medium, provided the original authors and source are cited.