Aspherical scattering factors for SHELXL - model, implementation and application.

A new aspherical scattering factor formalism has been implemented in the crystallographic least-squares refinement program SHELXL. The formalism relies on Gaussian functions and can optionally complement the independent atom model to take into account the deformation of electron-density distribution due to chemical bonding and lone pairs. Asphericity contributions were derived from the electron density obtained from quantum-chemical density functional theory computations of suitable model compounds that contain particular chemical environments, as defined by the invariom formalism. Thanks to a new algorithm, invariom assignment for refinement in SHELXL is automated. A suitable parameterization for each chemical environment within the new model was achieved by metaheuristics. Figures of merit, precision and accuracy of crystallographic least-squares refinements improve significantly upon using the new model. open access.

Introduction

According to a recent search in the Cambridge Structural Database (CSD) (Groom et al., 2016 ▸) SHELXL (Sheldrick, 2015 ▸) is by far the most widely used computer application for crystallographic least-squares refinement (Cruickshank, 1970 ▸; Rollett, 1970 ▸) in small-molecule crystallography. In macromolecular crystallography (Sheldrick & Schneider, 1997 ▸) it is mostly used for structures with atomic resolution (better than 1.2 Å) (Sheldrick, 1990 ▸), especially when standard deviations of the refined bond distances and angles are being sought (Dauter, 2003 ▸) and when atomic displacement parameters (ADPs) (Trueblood et al., 1996 ▸) can be refined. SHELXL has seen continuous development and improvements since about 1970, but the first official release was 1976 (Sheldrick, 2008 ▸). The implementation of new features often followed research by other scientists after algorithms and concepts reached a certain level of maturity and showed their value. This is exemplified by the choice of refinement against intensities and not structure factors as a default (Hirshfeld & Rabinovich, 1973 ▸; Arnberg et al., 1979 ▸). In other cases genuinely new ideas and features were implemented in the program, as for instance in the new RIGU restraints (Thorn et al., 2012 ▸). In these code modifications and extensions, the emphasis has always been on robustness and reliability, user friendliness, conceptual elegance (≃ simplicity), compatibility with previous versions and speed. We think it is this philosophy, as well as some early design decisions, including the choice of the programming language Fortran, that led to the success of SHELXL.

SHELXL implements the independent atom model (IAM) (Doyle & Turner, 1968 ▸). In this simple and elegant model each element in the periodic table has a single corresponding uncharged scattering factor, even when atoms are involved in chemical bonding and thus are not strictly identical. The analytical representation of the IAM scattering factor, which is the Fourier transform of an atomic electron density , is usually incorporated as an overlay of four Gaussian functions,with element-specific coefficients and exponents and optionally a constant c, as tabulated in the International Tables for Crystallography, Volume C (Prince, 2004 ▸),1 The main reason for using Gaussian functions for the IAM is the execution speed of analytical Fourier transform. However, their properties are not entirely satisfactory (Murshudov, 2016 ▸): firstly, form factors should be proportional to at high resolution, but Gaussian functions cannot mimic this proportionality; secondly, these functions are not ideally suited to model atomic charges to go beyond the IAM, since charge transfer mainly affects diffraction at low resolution.

Since its founding study (Coppens, 1967 ▸) charge-density research2 has shown that the experimental electron-density distribution (EDD) deviates from its IAM representation in bonding and lone pair regions (Koritsánszky & Coppens, 2001 ▸), and this was predicted as early as 1915 (Debye, 1915 ▸). Today, with major improvements in detector technology and routine low-temperature measurements, such deformation density3 becomes visible ever more frequently in the residual electron density . We believe that with this information becoming more routinely available, modern least-squares refinement programs should be able to model it. When this is achieved, structures become more precise as well as more accurate (Coppens et al., 1969 ▸; Brock et al., 1991 ▸; Jelsch et al., 1998 ▸; Dittrich et al., 2007 ▸). This article is about a new aspherical scattering factor model that was implemented in SHELXL.

Methods and models

Scattering factor models in charge-density research and quantum crystallography

Developments to improve scattering factors started in the early 1960s.4 The most established way to model aspherical EDD was conceived by Robert F. Stewart and is often called the rigid pseudoatom or multipole model (Stewart, 1976 ▸). An alternative earlier model with different angular functions but similar capabilities was developed by Hirshfeld (1971 ▸). Stewart’s approach has been modified by Hansen & Coppens (1978 ▸). Their multipole model (from now on called the SHC model) includes refinable radial screening parameters and has been widely used to ‘measure’ experimental EDD by least-squares refinement of aspherical population parameters normalized for electron density. The SHC model dominated charge-density research for decades, and contributed to answering numerous research questions at the interface between chemistry and physics (Koritsánszky & Coppens, 2001 ▸). Several computer programs have implemented the SHC model: MOLLY (Hansen, 1978 ▸), VALRAY (Stewart et al., 1998 ▸), MoPro (Jelsch et al., 2005 ▸), XD2006 (Volkov et al., 2006 ▸), Jana2006 (Petříček et al., 2014 ▸) and WinXPRO (Stash & Tsirelson, 2014 ▸).

More recently, it has become increasingly obvious that quantum-chemical computations in the framework of density functional theory (DFT) are powerful enough to faithfully reproduce molecular experimental EDD for light-atom structures.5 Therefore, the atom-centered SHC multipole model has more recently also been used to represent transferable EDD obtained from DFT computations. These efforts led to the construction of the invariom6 and other scattering factor databases.7 An alternative to tabulating scattering factors is Hirshfeld atom refinement (Jayatilaka & Dittrich, 2008 ▸; Capelli et al., 2014 ▸), where a quantum-chemical mol­ecular EDD is partitioned to provide tailor-made aspherical atomic scattering factors for the molecule of interest. Here the philosophy is different than in classical charge-density research: the experimental results are atomic positions and ADPs,8 but not valence EDD anymore, since only positional and displacement parameters are refined. Both approaches lead to a considerably better agreement of and , and more accurate and precise structures can be measured. This becomes most apparent for H atoms and their bond distances (Stewart et al., 1975 ▸; Dittrich et al., 2005 ▸; Woińska et al., 2014 ▸; Dittrich et al., 2017 ▸; Malaspina, Edwards et al., 2017 ▸; Malaspina, White et al., 2017 ▸).

Work by Dadda et al. (2012 ▸) and Nassour et al. (2014 ▸) deserves special emphasis, since some of the concepts used here are similar to their ‘virtual-atom’ refinement. Differences are technical in nature, but nevertheless important for everyday use. While these authors maintain a separation of core and valence electron density in analogy to the SHC model, we retain the IAM and model additional deformation density. Both approaches share the motivation to better model conventional data sets or those measured for macromolecular crystals with limited scattering power.

The bond-oriented deformation density (BODD) model in perspective

A new Gaussian-based scattering factor model is being introduced here. It is fully integrated into a well established refinement program and is therefore more widely applicable than previous scattering factor databases and Hirshfeld atom refinement alike. For instance, the BODD model can be used to model twinned (Herbst-Irmer & Sheldrick, 1998 ▸) as well as disordered (Müller et al., 2006 ▸) structures. The electron-density model described next is not intended to refine an experimental EDD, but was solely designed to model experimental data better on the basis of suitably partitioned (Dittrich et al., 2004 ▸) theoretical computations of model compounds.

One important characteristic of the new model – in contrast to the SHC multipole model with its monopole populations and the ‘virtual-atom’ model – is that the scattering factors themselves do not carry a charge. Rather, the model constitutes a re-distribution of an atomic EDD into bonds and lone pairs for each individual atom. This is an important feature, because it allows the simultaneous combination of aspherical and conventional IAM scattering factors, e.g. to model compounds that also contain heavier elements. The validity and particular usefulness of this combination have been exemplified for coordination compounds (Wandtke et al., 2017 ▸), albeit with a different refinement program and relying on the SHC model: the main challenge in the refinement of these metal-containing compounds with database parameters is that an exponentially increasing number of chemical environments would need to be tabulated for the central atoms.9 Coordination compounds are also amenable to Hirshfeld atom refinement. However, we then observe an exponential increase in computation time for compounds with electron-rich atoms due to the number of basis functions required. Both earlier approaches are hence unsuited for everyday use.

The EDD not described adequately by the IAM is on the one hand the EDD in bonding regions and on the other hand the EDD associated with lone pairs. Therefore, the concept of BODD is divided into two parts, one for modeling the bonding electron density (BEDE) and one for modeling the lone pair electron density (LONE). Both can be evoked by new commands in SHELXL.

Bonding electron density: BEDE

The idea behind BEDE is based on density deformation functions similar to dipoles. They add electron density (ED) in the direction of the bond, and in order to keep the overall electron count correct, subtract ED at the atomic positions. If, for example, a covalent bond between two C atoms is described, the deformation will look as presented in Fig. 1 ▸. The correction is usually applied to a bond from both directions.

Figure 1

Left: basic concept of BEDE instruction for a standard covalent bond seen from the bonding direction going from atom1 to atom2. Right: generating asphericity from Gaussian functions while avoiding atomic charges.

In the BEDE model Gaussian functions [see equation (1)] are positioned on the bond and at the atomic position, resulting in two Gaussians each with position r, amplitude A and spread B. This amounts to six parameters per bonding direction of an atom. Because EDD should not simply be added, the amplitude A at the atomic position of atom1 is the negative of the amplitude at the bond, reducing the number of parameters by one. The position of the Gaussian function is fixed with respect to the atomic coordinates, which leads to four parameters per bond per atom.

The instructions for SHELXL have the following syntax:

BEDE atom1 atom2 r A B1 B2 where: atom1 is the first atom involved in the bond, r is the distance between the Gaussian function and atom1 along the bond to atom2, and A is the amplitude for the Gaussian function on the bond. Likewise A is also the negative amplitude of the Gaussian function at atom1’s position, as visualized in the right part of Fig. 1 ▸. B1 is the spread of the Gaussian function with +A and B2 the spread of the Gaussian function with −A. Both B values are multiplied by the displacement parameters of atom1. Reference values for A, B1 and B2 are obtained by refining them as SHELXL free variables against intensity data generated by Fourier transformation of theoretical electron density. For structure refinement against experimental data they can be held fixed, but it turns out to be advantageous to refine up to three global scale factors for them.

Lone pair electron density: LONE

The task of finding the direction for a function to model EDD in lone pairs is not as straightforward. Here a procedure analogous to deducing the orientation of H atoms from the other bonds connected to an atom is relied upon. SHELXL uses the number m to classify what kind of geometrical arrangement an atom is in. An overview of all possible positions for lone pairs and their H-atom analogs (where applicable) is presented in Fig. 2 ▸. The syntax thus includes m instead of a second atom, and in some cases an additional angle:

Figure 2

Concept and the different options m for the LONE instruction. The H-atom treatment corresponding to the same m is also shown, including angle expectations according to the VSEPR model.

LONE m atom A B1 B2 r [angle]

The angle applies only to m = 2, 3, 7 and 9 and is in those cases fixed like it is for the H atoms. As known from the VSEPR theory (Gillespie, 1963 ▸, 1970 ▸) the angle may deviate from that of the ideal geometry due to valence shell repulsion.10 Similar to BEDE B2 is the coefficient in the exponent of the subtracted Gaussian function. In those cases where more than one Gaussian function with amplitude A is created, the Gaussian at the named atom will have an amplitude which is a multiple of A, so that the number of electrons stays balanced. For example, for m = 2 the subtraction at the atom will be of 2A. A special case is m = 12, which corresponds to a disordered methyl group with two half-occupied positions rotated from one another by 60°, thereby placing 12 half lone pairs in a circle, where the amplitude of the Gaussian function at the atom will be −6A with A being the height of each of the 12 Gaussian functions. For m = 6 and 7 there is no hydrogen analog, because these LONE instructions are meant to be applied to π-bonds. While m = 6 places Gaussian functions above and below the atom named in the direction perpendicular to the plane defined by two bonds (preferably to non-H atoms), m = 7 does the same for terminal atoms where in order to define a plane the second bond is connected to the neighboring atom. m = 9 is similar to m = 7 but places the two Gaussian functions in the plane of the atoms instead of perpendicular to it. The case of m = 15 can be used in cases of lone pair placement for atoms with four or five bonds and one lone pair, for example SF4 or BrF5. For the modeling of π-bonds a combination of several BEDE and LONE instructions is suggested, and there is a ‘LONE 6’ command for this purpose.11 In the case of a carbonyl bond there are no other bonds originating from the O atom, so the subtraction of ED comes from the direction of the C atom.

In summary, BEDE and LONE provide an additional deformation model on top of the IAM that seamlessly extends its capabilities. This in turn means full downward compatibility with earlier refinements that can still be performed exactly as before. We think that this is a requirement for a widely applicable approach, since for some applications and research questions the IAM remains entirely sufficient and appropriate. Another characteristic is that the BODD approach, as suggested by its full name ‘bond-oriented deformation density’, does not require a local atomic coordinate system. Every Gaussian function used is attached or directed to an atom that thereby becomes aspherical. A disadvantage in this context is that BEDE and LONE do not reach the same sophistication in representing electron density as invarioms, transferable aspherical atom models (TAAMs) or Hirshfeld atom refinement; fine features of the EDD are not as well represented in the scattering factor model.12 The goal of the new model is not to replace earlier charge-density or current quantum-crystallography (Massa et al., 1995 ▸; Grabowsky et al., 2017 ▸; Tsirelson, 2018 ▸) models, but to open up a still rather narrow research area to the wider crystallographic community. BODD can complement more sophisticated approaches in applications where conceptual simplicity, easy implementation and execution speed are important. Speed is why we chose Gaussian functions over Slater functions, as the former permit fast analytical Fourier transform.

Usage

BEDE and LONE commands are fully compatible with all other features of SHELXL. The user first performs a conventional refinement. Subsequently the BEDE and LONE instruction commands are added. Currently this is done automatically by evoking the APEX3 software.13 The APEX3 graphical user interface now contains a Python plugin that assigns the asphericity parameters and allows manual modification by the user. Technically the BEDE and LONE parameters are contained in an external file, which is referred to in the INS file. After assignment the BODD refinement can be directly performed by a click of the mouse.

As pointed out earlier, BEDE and LONE are not meant for free refinement of asphericity parameters with experimental data. In contrast to the multipole model, which was designed for this purpose and where the functions were chosen to minimize parameter correlation, it is strictly not advised to freely refine BEDE and LONE parameters, as results cannot be expected to be chemically or physically meaningful. Rather, BEDE and LONE parameterization is done using DFT computations of suitable model compounds, for the choice of which we rely on the notation of the respective entries of the invariom database (Dittrich et al., 2013 ▸). The best set of individual BEDE and LONE values are contained in a database and were obtained by metaheuristics; Appendix A contains full details of the respective technical implementation. These parameters should therefore only be used as fixed additions to the IAM scattering factors. The philosophy is hence similar to invariom and Hirshfeld atom refinement.

Results

Effects of BODD modeling on common quality indicators

In this section the benefits of modeling aspherical EDD due to chemical bonds and lone pairs via the proposed method are investigated using a series of structure models taken from the literature. For this purpose model quality is assessed by the value of computed before and after applying the method. A drop of is expected after applying the BODD method. Since the absolute scale of the aspherical EDD parameters taken from the database is unknown in each particular structure, three scale factors were added in the refinements, multiplying all A, B1 and B2 parameters.

Table 1 ▸ summarizes the effects of modeling aspherical EDD for 15 structures of organic compounds.14 The table proves that the addition of aspherical EDD to the IAM structural model leads to a better fit to the data and most likely improved the physical significance of the structural models in all cases.

Table 1

R 1 values before and after applying the BODD model for several structure models

The six-character abbreviation is the IUCr publication code.

Compound name	IUCr code	Literature	R 1 (IAM)	R 1 (BODD)
Xylitol	sh5011	Madsen et al. (2004 ▸)	2.34	2.07
Morphine	lc5024	Scheins et al. (2005 ▸)	3.33	2.56
(E)-1-[4-(Hexyloxy)phenyl]-3-(2-hydroxy-phenyl)prop-2-en-1-one	hb6948	Fadzillah et al. (2012 ▸)	4.06	3.38
Chelidamic acid methanol solvate	eg3095	Tutughamiarso et al. (2012 ▸)	3.66	3.21
Baicalein nicotinamide (1/1)	fg3251	Sowa et al. (2012 ▸)	5.80	5.62
2-(1,4,7,10-Tetraazacyclododecan-1-yl)cyclohexan-1-ol (cycyclen)	dt3014	de Sousa et al. (2012 ▸)	4.50	4.29
2-(1H-Indol-3-yl)-2-oxoacetamide	yp3017	Sonar et al. (2012 ▸)	6.35	6.08
cis-2-(4-Fluorophenyl)-3a,4,5,6,7,7a-hexahydroisoindoline-1,3-dione	fg3250	Smith & Wermuth (2012 ▸)	3.51	3.22
5-Ethynyl-2′-deoxycytidine	fg3269	Seela et al. (2012 ▸)	3.02	2.12
N,N-Dibenzyl-N′-(furan-2-carbonyl)thiourea	fa3263	Pérez et al. (2012 ▸)	4.51	4.22
3-Phenylcoumarin	zj2091	Matos et al. (2012 ▸)	3.97	3.58
3,4,6-Tri-O-acetyl-1,2-O-[1-(exoethoxy)ethylidene]-β-D-manno-pyranose·0.11H2O	bi3042	Liu et al. (2012 ▸)	4.84	4.69
4-[(E)-(4-Ethoxyphenyl)iminomethyl]phenol	bt5991	Khalaji et al. (2012 ▸)	2.98	2.49
N,N′-Bis(2-methylphenyl)-2,2′-thiodibenzamide	fg3262	Helliwell et al. (2012 ▸)	3.55	3.23

Since three scale factors were added to the structure model, the data-to-parameter ratio is reduced. To show that improvements in the fit are significant, a series of five structures was investigated in more detail in order to test whether the drop in is caused by improving the model or by overfitting. This was done by computing for all five structures before and after applying the BODD model (Lübben & Grüne, 2015 ▸). quantifies the amount of bias in the structural model. If the amount of bias is reduced after applying the BODD model, the BODD model does in fact improve the structural model significantly, and this is indeed the case here.

The bias b of a structural model can be quantified via as follows: A positive value of b indicates a reduction of bias upon introducing the BODD model. A negative value indicates added bias.

Table 2 ▸ shows that application of the BODD model reduces the amount of bias, supporting the hypothesis that it improves the structural model by taking bond electron density into account.

Table 2

Refinement details of corresponding IAM and BODD refinements

Compound						b	Resolution (Å)
sh5011	2.34	2.07	0.27	2.77	2.44	0.03	0.41
lc5024	3.33	2.56	0.77	4.49	3.75	0.08	0.44
hb6948	4.06	3.38	0.68	5.31	4.54	0.04	0.73
eg3095	3.66	3.21	0.45	4.61	4.12	0.02	0.82

Statistical analysis shows that the improvement in atomic positions for the non-H atoms is rather small (results not shown). This is similar to invariom refinement and Hirshfeld atom refinement. More interestingly, it is possible to freely refine meaningful H-atom positions with BODD with high-quality data. There have been detailed studies on this matter using neutron data as reference since the first study of this kind with theoretical aspherical scattering factors (Dittrich et al., 2005 ▸) was performed, and similar results (albeit with slightly shorter average X⋯H distances than in Hirshfeld atom refinement or invariom refinement due to the use of only dipolar functions) can also be obtained with the BODD model. We decided not to report such results here, since our method should be generally applicable; when high-quality data are not available it is probably preferable to use AFIX constraints and elongate the X⋯H bond vector rather than to freely refine H-atom positions. We then suggest using an elongation factor of 1.14 for BODD. Moreover, the interested reader can easily perform such tests, since the model will be widely available.

While non-H-atom positions remain rather unaffected when including BODD contributions, including asphericity has a profound and systematic effect on the ADPs for both H and non-H atoms. Here inclusion of BEDE and LONE parameters leads to systematically larger H-atom ’s and smaller non-H-atom ADPs in better agreement with the reference invariom refinement, in analogy to earlier results for the multipole model (Jelsch et al., 1998 ▸; Dittrich et al., 2007 ▸). These findings are also in agreement with our earlier article (Lübben et al., 2014 ▸) on H-atom ADPs. Using a temperature-dependent multiplier (Madsen & Hoser, 2015 ▸) ratio for the H-atom constraints (or free refinement of ) will thus be useful for BODD refinements with data measured at cryogenic temperatures. Another observation concerns the overall scale factor (OSF). As reported before, when applying theoretical aspherical multipole scattering factors (Volkov et al., 2007 ▸), the OSF also usually becomes smaller (see Table 3 ▸). The situation is different for xylitol, where data were affected by extinction. The suggested weighting scheme values also become smaller overall; this also holds for the second value which up-weighs reflections similar to . These findings can be complemented by statistical analysis. Given a set of N values the mean value and its population standard deviation are defined by The population standard deviation or root-mean-square deviation (RMSD) gives an indication of the spread of the values around the mean. To assess the similarity of individual ADP values we provide the mean difference MD (rather than the mean value) and its standard deviation in Table 3 ▸. Subsequently the ADPs are also visually compared using Peanut plots (Hummel et al., 1990 ▸).

Table 3

Mean difference (MD) in Å and σ(MD) (both ×10−4) of corresponding IAM and BODD refinements using invariom refinement as a reference model with the number of values analyzed

Values of the overall scale factors (OSF) are also given.

Compound	MD INV-IAM(σ)	MD INV-BODD(σ)	No. of values	OSF IAM	OSF BODD
sh5011	−1.5 (3)	−1.2 (2)	60	1.21	1.21
lc5024	−2.3 (3)	−1.2 (2)	132	0.73	0.70
hb6948	−7.4 (17)	−0.2 (9)	144	1.00	1.01
eg3095	−7.6 (24)	−2.0 (12)	90	1.00	0.97

Next the residual density maps after BODD refinement were compared with the conventional IAM residual density maps. Figs. 3 ▸ ▸ ▸ to 6 ▸ show that the BODD model is indeed able to fit most of the aspherical EDD, since the features are significantly reduced in comparison with the IAM residuals.

Figure 3

Left: xylitol (sh5011) residual IAM density at ISO level 0.1 e Å−3. Center: BODD deformation density added to the IAM. Right: remaining residual density after adding the BODD deformations. IAM / = 0.33/−0.20; BODD / = 0.23/−0.22.

Figure 4

Left: morphine (lc5024) residual IAM density at ISO level 0.2 e Å−3. Center: BODD deformation density added to the IAM. Right: remaining residual density after adding the BODD deformations. IAM / = 0.59/−0.35; BODD / = 0.33/−0.29.

Figure 5

Left: (E)-1-[4-(hexyloxy)phenyl]-3-(2-hydroxy-phenyl)prop-2-en-1-one (hb6948) residual IAM density at ISO level 0.08 e Å−3. Center: BODD deformation density added to the IAM. Right: remaining residual density after adding the BODD deformations. IAM / = 0.22/−0.27; BODD / = 0.05/−0.05.

Figure 6

Left: chelidamic acid methanol solvate (eg3095) residual IAM density at ISO level 0.12 e Å−3. Center: BODD deformation density added to the IAM. Right: remaining residual density after adding the BODD deformations. IAM / = 0.38/−0.24; BODD / = 0.13/−0.06.

Studying the residual electron-density maps of compound eg3095 (Fig. 6 ▸) reveals that even in cases where the residual density of the IAM model is noisy, and hence apparently does not contain information on bonding electron density, applying the BODD model does not introduce errors (recognizable as new features). Although the reduction of bias by modeling aspherical EDD is smaller for this example, an improvement is still seen. This shows that the BODD model is robust and can safely be used even in the absence of obvious bonding residual density features.

Figs. 3 ▸ ▸ ▸ to 6 ▸ also show the deformation density ( ), demonstrating that the modeled density is in agreement with a chemist’s expectations. All figures show electron-density maps which were generated with the program SHELXLE (Hübschle et al., 2011 ▸).

While statistical analysis (Table 3 ▸) quantitatively shows that ADPs from BODD and invariom refinement are more similar than both are to the IAM ADPs, this can be more easily conceived visually through Peanut plots. Both BODD and invariom model ADPs are systematically smaller than those from the IAM as indicated by the red (smaller) RMSD surfaces in Figs. 7 ▸ ▸ ▸ to 10 ▸. For xylitol, extinction again affects the ADPs, but overall the results are fully consistent with earlier invariom results (Dittrich et al., 2007 ▸).

Figure 7

Left: xylitol (sh5011) Peanut plot of IAM minus BODD RMSD with a magnification factor of 10. Center: similar IAM minus invariom Peanut plot for comparison. Right: BODD minus invariom Peanut plot.

Figure 8

Left: morphine (lc5024) Peanut plot of IAM minus BODD RMSD with a magnification factor of 10. Center: similar IAM minus invariom Peanut plot for comparison. Right: BODD minus invariom Peanut plot.

Figure 9

Left: (E)-1-[4-(hexyloxy)phenyl]-3-(2-hydroxy-phenyl)prop-2-en-1-one (hb6948) Peanut plot of IAM minus BODD RMSD with a magnification factor of 10. Center: similar IAM minus invariom Peanut plot for comparison. Right: BODD minus invariom Peanut plot.

Figure 10

Left: chelidamic acid methanol solvate (eg3095) Peanut plot of IAM minus BODD RMSD with a magnification factor of 10. Center: similar IAM minus invariom Peanut plot for comparison. Right: BODD minus invariom Peanut plot. Only the main molecule is shown.

Discussion

All aspherical atom approaches, the new BODD model, SHC databases and Hirshfeld atom refinement alike, have strengths and weaknesses. Least-squares refinement with the SHC multipole model is nearly as fast as the BODD implementation in SHELXL, and a rather accurate atomic scattering factor can be stored by using up to 27 parameters, in addition to information on a local atomic coordinate system and an atom type. However, the possibility of having a wrong local coordinate system in cases where an unambiguous definition with SHC scattering factor databases is impossible sometimes requires manual checking and intervention. This problem is solved in BODD, where the local coordinate system is directly deduced from neighboring atoms without user intervention. The relative orientation of BODD with respect to the unit-cell axes is thus hidden from the user.

In comparison with database approaches, Hirshfeld atom refinement can be slightly more accurate – depending on the basis set used – than database scattering factors and is more rigorously defined in terms of the physical–chemical model. It can also include most of the crystal field effect.15 Combining the variational principle for single-point energy minimization with least-squares refinement via experimental intensities to convergence in turn, Hirshfeld atom refinement is clearly attractive, but also very time-consuming, as it entails considerable computing efforts which increase to the power of four with the number of basis functions. Moreover, Hirshfeld atom refinement cannot handle polymeric structures, and although it can, in principle, be used for disordered structures, the computational effort then increases further. In contrast, disorder can be handled in the BODD model and also in invariom refinement (Dittrich et al., 2016 ▸). The most attractive point for Hirshfeld atom refinement is that it can be combined with X-ray wavefunction fitting (Jayatilaka, 1998 ▸), and this combination is called X-ray wavefunction refinement (Grabowsky et al., 2012 ▸). Both form part of a new research area called quantum crystallography, where recent efforts clearly have the potential to rejuvenate charge-density research (Genoni et al., 2018 ▸) in general. The BODD model has a different focus: user-friendliness and speed, and we again emphasize that the BODD model is not intended to replace detailed studies of the electron density, which require or benefit from using more sophisticated approaches. The BODD model describes the deformation density intuitively using the familiar concepts of bonding electrons and lone pairs rather than the less intuitive multipoles.

Conclusion

The BODD model to implement aspherical scattering factors in SHELXL provides deformation density in addition to the IAM, is fast, backwards compatible and a user-friendly alternative to database approaches based on the SHC multipole model or to Hirshfeld atom refinement. The pre-computed values of the new BODD database can be directly used via the APEX3 interface and no further computations are required. Scattering factor assignment is nearly fully automatic; when an ideal scattering factor cannot be found in the database suitable alternatives are identified. The model permits the combination of conventional IAM and aspherical scattering contributions, e.g. for refinement of a metal complex where no aspherical scattering factor for the metal atom is needed or available.

Table 4

Definition of node colors for each node level

Level	Color	Definition
0	r	0
	g	Sum of bond lengths to bonding partners
	b	0
	a	1
1	r	Distance to root atom
	g	Sum of bond lengths to bonding partners
	b	1
	a	1
2	r	Distance to root atom
	g	0
	b	2
	a

Table 5

Row content of the environmental archetype data table

Column	Definition
Environmental graph	The environmental graph is serialized and stored
Inner hash	Inner hash attribute of the graph object
Invariom name	Invariom name generated for the root atom