Neutron crystallographic refinement with REFMAC5 from the CCP4 suite

Hydrogen (H) atoms are abundant in macromolecules and often play critical roles in enzyme catalysis, ligand-recognition processes and protein–protein interactions. However, their direct visualization by diffraction techniques is challenging. Macromolecular X-ray crystallography affords the localization of only the most ordered H atoms at (sub-)atomic resolution (around 1.2 Å or higher). However, many H atoms of biochemical significance remain un­detectable by this method. In contrast, neutron diffraction methods enable the visualization of most H atoms, typically in the form of deuterium (²H) atoms, at much more common resolution values (better than 2.5 Å). Thus, neutron crystallography, although technically demanding, is often the method of choice when direct information on protonation states is sought. REFMAC5 from the Collaborative Computational Project No. 4 (CCP4) is a program for the refinement of macromolecular models against X-ray crystallographic and cryo-EM data. This contribution describes its extension to include the refinement of structural models obtained from neutron crystallographic data. Stereochemical restraints with accurate bond distances between H atoms and their parent atom nuclei are now part of the CCP4 Monomer Library, the source of prior chemical information used in the refinement. One new feature for neutron data analysis in REFMAC5 is refinement of the protium/deuterium (¹H/²H) fraction. This parameter describes the relative ¹H/²H contribution to neutron scattering for hydrogen isotopes. The newly developed REFMAC5 algorithms were tested by performing the (re-)refinement of several entries available in the PDB and of one novel structure (FutA) using either (i) neutron data only or (ii) neutron data supplemented by external restraints to a reference X-ray crystallographic structure. Re-refinement with REFMAC5 afforded models characterized by R-factor values that are consistent with, and in some cases better than, the originally deposited values. The use of external reference structure restraints during refinement has been observed to be a valuable strategy, especially for structures at medium–low resolution.

1. Introduction

Knowledge of protonation states and hydrogen (H) atom positions in macromolecules can be critical in helping to formulate functional hypotheses and, generally, in providing a more complete characterization of the biological processes under investigation. H atoms are responsible for the reversible protonation of active site residues involved in enzymatic reactions (Ahmed et al., 2007; Fisher et al., 2012; Wan et al., 2015). They are also necessary for the formation of hydrogen bonds that stabilize macromolecular structures, contributing to the establishment of biological interfaces (Engler et al., 2003; Niimura et al., 2004; Oksanen et al., 2017). Additionally, as H atoms are often involved in determining specificities in protein–ligand recognition processes, their identification and localization may help in the development and design of new therapeutics (Combs et al., 2020; Kovalevsky et al., 2020; Kneller et al., 2022).

The positions of many H atoms in macromolecules can be estimated using the coordinates of their parent atoms (those to which they are covalently bound) and known geometric properties (Sheldrick & Schneider, 1997). This is the case, for example, for amide H atoms in the protein backbone, for those bound to C^α atoms, for those attached to aromatic C atoms etc. However, many H atoms of biochemical interest, for example those on the side chains of histidines, protonated aspartates and glutamates, or those associated with multiple favourable positions (the hydroxyl groups of the amino acids serine, threonine and tyrosine), cannot be located on the basis of simple geometric considerations, but need to be determined experimentally (Fisher et al., 2009; Gardberg et al., 2010).

Although H atoms represent a large fraction of the total atomic content of macromolecules (∼50% and ∼35% of protein and nucleic acid atoms, respectively) their experimental visualization is not straightforward. In X-ray macromolecular crystallography they contribute little to the total scattering, thus even at (sub-)atomic resolution (<1.2 Å) only a fraction of all H atoms are typically observed in electron density maps (Howard et al., 2004; Petrova & Podjarny, 2004). For instance, in the case of the 0.85 Å resolution room-temperature X-ray structure of crambin, less than 50% of all H atoms could be identified (Chen et al., 2012). These tend to be the most ordered ones, which are seldom interesting from a functional viewpoint (Fig. 1a). At comparable resolution, H atoms can be expected to be more visible in cryogenic-sample electron microscopy (cryo-EM) maps than in electron density maps due to the nature of the electrostatic potential (Clabbers & Abrahams, 2018; Maki-Yonekura et al., 2023). Yamashita et al. (2021) analysed H atom density from X-ray crystallo­graphic and cryo-EM single-particle analysis (SPA) data for apoferritin structures deposited in the PDB (Berman et al., 2000) and EMDB (Lawson et al., 2016), highlighting that even at 2.0 Å resolution it is possible to see some H atoms in cryo-EM maps. For extremely well-behaved samples, the recent `resolution revolution' in cryo-EM SPA has allowed atomic resolution to be achieved (Nakane et al., 2020; Yip et al., 2020). In the structure of apoferritin at 1.2 Å resolution, most H atoms (approximately 70%) are easily discernible (Fig. 1b). However, a recent microcrystal electron diffraction (microED) experiment on triclinic lysozyme reported at subatomic resolution only allowed the identification of 35% of H atoms (Clabbers et al., 2022).

Figure 1 Map examples. (a) Electron density maps for Tyr12 (PDB entry 3kyu, 1.1 Å resolution) show positive peaks for all aromatic H atoms; however, the H atom on the hydroxyl group is not visible. (b) Cryo-EM maps for Tyr32 (PDB entry 7a4m, 1.22 Å resolution) show positive difference peaks for all H atoms. (c) Neutron scattering length density maps for Tyr146 (PDB entry 1cq2, 2.0 Å resolution) show positive difference peaks for all H atoms (in the form of ^2H), including the H atom on the hydroxyl group. Electron and neutron scattering length density maps were calculated using REFMAC5 (Murshudov et al., 2011), contoured at the +1.0σ (2mFo − DFc in grey) and +3.0σ (mFo− DFc in green) levels. Cryo-EM weighted and sharpened Fo (grey) and omit (Fo − Fc, green) maps were calculated using Servalcat (Yamashita et al., 2021) and contoured at the +1.5σ and +3.0σ levels, respectively. Molecular-graphics representations were produced with Coot 1.0 (Emsley et al., 2010).

Neutron macromolecular crystallography is a powerful technique that allows the direct visualization of H atoms at more conventional resolutions (Blakeley & Podjarny, 2018). In contrast to X-rays, which interact with atomic electron clouds, neutrons are scattered by nuclei (Fermi & Marshall, 1947). Atoms that are abundant in macromolecules typically possess positive neutron scattering lengths (0.665 × 10⁻¹², 0.936 × 10⁻¹² and 0.581 × 10⁻¹² cm for C, N and O, respectively) that contribute favourably to the signal-to-noise (S/N) ratio of Bragg peaks. Although the scattering length of the common protium isotope (¹H; note that in this article we use the conventional ¹H and ²H notation to indicate protium and deuterium isotopes, respectively, whilst we use H when referring to hydrogen atoms in general) is small and negative (−0.374 × 10⁻¹² cm), its replacement with the heavier deuterium isotope ²H (scattering length 0.667 × 10⁻¹² cm) makes them readily visible in neutron diffraction maps at 2.0–2.5 Å resolution or better (Fig. 1c). Another important advantage of neutron diffraction for structure determination is the absence of global and specific radiation-induced damage, which can be a serious limitation when using X-ray or electron sources (Baker & Rubinstein, 2010; Garman, 2010).

Crystallographic refinement is one of the final steps in the process of solving a macromolecular structure by diffraction methods (Tronrud, 2004). Various protocols are applied to maximize the agreement between the diffraction data and model parameters, which typically include atomic coordinates, atomic displacement parameters (ADPs) and occupancy values (Shabalin et al., 2018). Refinement of macromolecular models using neutron diffraction data can currently be carried out using packages initially developed for X-ray crystallo­graphic refinement and modified to include neutron scattering lengths and the ability to deal with the refinement of individual H atom positions. They include the nCNS patch (Adams et al., 2009), which is an extension of the Crystallography and NMR System (CNS) package (Brünger et al., 1998), and SHELXL2013 (Gruene et al., 2014). SHELXL2013 is the most recent version of the SHELXL refinement program originally developed for small molecules and later adapted to macromolecules (Sheldrick, 2015). Another widely used package for neutron refinement is phenix.refine (Afonine et al., 2012), which is distributed as a part of the Phenix suite (Liebschner et al., 2019). This program also includes the option of performing joint neutron/X-ray refinement, a concept first introduced in the field of small-molecule crystallography (Coppens et al., 1981) and later applied to macromolecules with its nCNS implementation. Although effective joint neutron/X-ray refinement ideally requires the two data sets to be collected from the same crystal under the same conditions, it has the great advantage of increasing the available experimental data, thus compensating for the increased number of parameters arising from the explicit addition of H atoms to the model.

Here, we describe an extension of the crystallographic refinement package REFMAC5 (Murshudov et al., 2011) from the CCP4 suite (Agirre et al., 2023) for the refinement of macromolecular models using neutron crystallographic data. Our implementation introduces a new parameter, dubbed the `deuterium fraction', representing the ¹H/²H fraction that is refined during the optimization procedure. It also allows the effective use of stereochemical restraints from high-resolution reference structures, if available. We have tested REFMAC5 (version 5.8.0415) for the refinement of neutron models using ¹H/²H fraction parameters for selected or all H atoms together with restraints to a high-resolution known X-ray reference structure. Our evaluation involved the re-refinement of 97 PDB entries and one novel structure (FutA). The results of the refinement process are discussed in this study.

2. Methodology and results

2.1. Reassessment of X—H restraint distances for macromolecular refinement

Macromolecular crystallographic refinement takes advantage of prior chemical knowledge. Information on `ideal' bond lengths, bond angles and other chemical properties are incorporated into the target function and used in restrained refinement as subsidiary conditions to improve the model parameters (Waser, 1963; Diamond, 1971; Jack & Levitt, 1978; Konnert & Hendrickson, 1980). Much of the available prior chemical knowledge used in macromolecular crystallographic refinement derives from high-resolution small-molecule X-ray diffraction experiments and the corresponding structures deposited in databases such as the Cambridge Structural Database (CSD; Groom et al., 2016) and the Crystallography Open Database (COD; Gražulis et al., 2012). The values of X—H (where X is a non-H `parent' atom) bond lengths derived from X-ray diffraction experiments reflect the relative positions of the atomic electron clouds. However, the distances between H nuclei and their parent atoms are longer than those between the electron clouds. This is because the valence electron density for H atoms is shifted towards their parent atoms (Coppens, 1997). Thus, to properly model and refine macromolecular models against neutron diffraction data, bond-distance information should take this into account.

In addition to X-ray crystallographic structures, the CSD also contains a limited set of small-molecule structures determined by neutron crystallography. Neutron entries in the CSD have almost doubled in recent years, from 1213 in 2009 to 2362 (1452 organic and 910 metal–organic compounds) in 2021. An analysis of X—H bond lengths using the 2009 CSD neutron database was reported by Allen & Bruno (2010) that reassessed information derived from the limited earlier data of the late 1980s and early 1990s (Allen et al., 1987, 1992; Orpen et al., 1989). We took advantage of the recent enrichment in neutron structures in the CSD and re-evaluated X—H bond-length values. We employed the same approach as Allen & Bruno (2010) by selecting nonpolymeric organic compounds without disorder and with R factors ≤ 0.075 (647 entries). Entries derived from powder diffraction data were excluded. Neutron entries were retrieved using ConQuest (Bruno et al., 2002) and mean, median and standard deviation values for the X—H bond-length distributions were estimated using Mercury (Macrae et al., 2020). O—H and N—H bond lengths were estimated by removing groups involved in very short hydrogen bonds, as reported by Allen & Bruno (2010).

In an orthogonal approach, we also derived X—H nuclear distances from quantum-mechanics (QM) calculations. Initially, the stereochemical restraints generator AceDRG (Long et al., 2017) was employed to provide initial coordinates for 2652 molecules constituted of twenty or fewer atoms selected from DrugBank (Wishart et al., 2018). The cutoff value on atom numbers was chosen to ensure computational efficiency while providing a pool size comparable to that of CSD entries. For geometry optimization, density functional theory (DFT) calculations were carried out with the self-consistent field wavefunction of restricted Hartree–Fock type as implemented in GAMESS-US (Schmidt et al., 1993). The hybrid generalized gradient approximation functional, B3LYP, was used with the (6-311++G**) basis set that includes both polarization and diffuse functions. The solvent effect was calculated using the polarizable continuum model with water as solvent. More than 70% of the calculations ran successfully, producing optimized coordinates for 1874 out of 2652 molecules. We did not perform a detailed analysis of calculations that ended prematurely.

Table 1 summarizes nuclear bond distances for the most common X—H bond classes. It also provides the values as reported by Allen & Bruno (2010) for reference. Overall, the recent nuclear bond distances derived from the CSD in 2021 are fully consistent with those previously derived in 2009. Nuclear distances obtained from theoretical calculations are also consistent with the experimentally derived values. The AceDRG data table has been updated to use the median (m) and standard deviation (σ) values for all X—H nuclear distances from the CSD 2021 data (Fig. 2a).

Table 1
Nuclear X—H bond lengths

X represents (C, N, O, S) atoms covalently bound to H. For C, different types of hybridization are given, with C(ar) indicating aromaticity. The first set of values (CSD-2009) is that of Allen & Bruno (2010) obtained from an analysis of curated neutron diffraction structures available within the CSD in 2009. The values in the CSD-2021 column are updated values from our analysis using curated CSD entries as of 2021. The values in the QM column are the result of QM calculations carried out as described in the text. The letters μ, σ, m and n represent the mean, standard deviation, median and number of observations, respectively, for each X—H class. For QM, the standard deviation is not applicable. All bond lengths are in Å.

	CSD-2009				CSD-2021				QM
	μ	σ	m	n	μ	σ	m	n	μ	σ	m	n
C(sp^1)—H	1.042	0.022	1.044	5	1.042	0.022	1.044	5	1.063	—	1.063	9
C(sp^2)—H	1.082	0.013	1.084	109	1.083	0.015	1.085	163	1.087	—	1.085	538
C(sp^3)—H	1.089	0.010	1.091	1118	1.087	0.010	1.092	1397	1.093	—	1.093	12985
C(ar)—H	1.083	0.017	1.085	721	1.084	0.018	1.085	1251	1.083	—	1.083	3906
C(sp^2)—N—H2	1.013	0.010	1.012	141	1.014	0.012	1.013	177	1.010	—	1.009	1055
C(sp^3)—N—H2	1.002	0.010	1.002	4	1.002	0.052	1.018	68	1.020	—	1.019	1172
C(sp^3)—O—H	0.970	0.012	0.971	169	0.969	0.018	0.972	186	0.966	—	0.964	1229
S—H	1.338	—	1.338	1	1.338	—	1.338	1	1.345	—	1.345	83

Figure 2 Example of a dictionary mmCIF file from the updated version of AceDRG. (a) 3D representation of the adenosine triphosphate (ATP) monomer. N, C, O and P atoms are shown in blue, teal, red and orange, respectively. X—H bonds are represented by grey sticks with their nuclear and X-ray diffraction-derived bond lengths (in Å) highlighted in orange and light blue, respectively. (b) Extract from the monomer description of the ATP component dictionary. The category _chem_comp_bond describes the bonded atoms, bond types and the ideal values of bond lengths and uncertainties associated with them. In this example, we show the ideal X—H bond lengths and standard deviations for nucleus positions (_chem_comp_bond.value_dist_nucleus and chem_comp_bond.value_dist_nucleus_esd; orange) and electron positions (_chem_comp_bond.value_dist and _chem_comp_bond.value_dist_esd; light blue).

2.2. Inclusion of X—H nuclear distances in the CCP4 Monomer Library (CCP4-ML)

The CCP4-ML, also referred to as the REFMAC5 dictionary (Vagin et al., 2004), currently contains close to 35 300 entries for all standard and most nonstandard amino acids, nucleotides, saccharides and various ligands. Each entry, identified as a monomer, possesses a unique code and provides stereochemical information about the constituent atoms, bond distances, bond angles and torsion angles as well as stereochemical centres and planes. Statistics for these geometric parameters have been generated by AceDRG using data from the COD. In addition, the CCP4-ML also contains more than 100 descriptors that specify covalent linkages between monomers and associated chemical modifications. The latter define all of the chemical and geometric changes that occur to monomers following chemical reactions (for example removal of one of the O atoms in peptide-link formation). Covalent links refer to covalent interactions between monomers (for example, peptide links, sugar-peptide links, DNA/RNA links; Nicholls, Joosten et al., 2021; Nicholls, Wojdyr et al., 2021).

The CCP4-ML has recently been updated to contain X—H nuclear distances (orange in Fig. 2b) as _chem_comp_bond.value_dist_nucleus and _chem_comp_bond.value_dist_nucleus_esd in addition to the distances between electron clouds (light blue in Fig. 2b) (Nicholls, Wojdyr et al., 2021). X—H nuclear distances can now also be used to refine models from electron-derived experiments (cryo-EM SPA and microED), as both H atom `positions' (electron and nucleus) contribute to the scattering.

2.3. CCP4 implementation of neutron macromolecular crystallographic refinement

2.3.1. `Deuterium fraction' parametrization

Neutron crystallographic experiments on macromolecules are typically carried out on ¹H/²H-exchanged crystals to maximize the S/N ratio (Kossiakoff, 1984). This can be performed by replacing exchangeable ¹H atoms with ²H by soaking macromolecular crystals in deuterated media (Niimura & Podjarny, 2011). Alternatively, perdeuteration, which replaces all H atoms with ²H, can be carried out at the protein-production stage by overexpressing the protein(s) of interest in Escherichia coli or yeast strains in heavy water-based medium supplied with a perdeuterated carbon source such as glycerol. Protein perdeuteration is a more effective method of improving the S/N ratio as it dramatically lowers the incoherent background while enhancing the coherent scattering signal (Shu et al., 2000; Fisher et al., 2014). In addition, it avoids map-cancellation issues due to the negative scattering length of protium (Blakeley & Podjarny, 2018; Logan, 2020). Currently, most neutron entries in the PDB (157 out of 213) reflect experiments carried out on partially deuterated samples, as ¹H/²H exchange is simpler and less expensive than perdeuteration. However, the establishment of dedicated deuteration facilities and advanced experimental protocols have made perdeuteration more accessible to users (Meilleur et al., 2009; Budayova-Spano et al., 2020; Pierce et al., 2020).

In the refinement procedure implemented in REFMAC5, we have introduced a new quantity that represents the deuterium fraction for individual H atoms. This method is similar to the `deuterium saturation' implemented in SHELXL (Gruene et al., 2014). In this parametrization, protium ¹H and deuterium ²H isotopes at each H position are not considered as separate entities. Instead, H atoms are represented by a unique set of coordinates that are associated with their isotope fraction, which is optimized during the minimization of the target function. The scattering factor for the ¹H/²H mixture is calculated using

where f_i(s) is the total contribution of protium and deuterium isotopes to the scattering factor of the ith H atom, s is the Fourier space vector, m_i is the deuterium fraction parameter, which is an adjustable parameter, and b_H and b_D are the neutron scattering lengths of the ¹H and ²H isotopes, respectively. Neutron scattering lengths are tabulated in the CCP4 atomsf_neutron library, retrieved from https://www.ncnr.nist.gov/resources/n-lengths/list.html (Sears, 1992). The refined output model in mmCIF format contains only H atoms (no ¹H/²H or ²H sites) and a new _atom_site.ccp4_deuterium_fraction column representing the value of the deuterium fraction for each of the H atoms in the model. Users have the option to refine deuterium fraction parameters for either only polar or all H atoms. This method simplifies the model output as there is no ¹H/²H duplication for the same set of coordinates, for example, when alternative conformations are introduced into the structure (Figs. 3a and 3b). The presence of only `generalized' H atoms with their corresponding deuterium fraction parameter also reduces the risk of bookkeeping errors. In the deuterium fraction representation, all ²H atoms are converted to H atoms and their presence is indicated by their corresponding deuterium fractions (Figs. 3c and 3d). We note that this new item can only be added to mmCIF files, which is now the model deposition standard. For PDB files that have fixed-column format, ¹H and ²H are present at each H position and the deuterium fraction is indicated in the occupancy column.

Figure 3 Comparison between the traditional representation of partially ^1H/^2H-exchanged structures and perdeuterated structures and the deuterium fraction representation in mmCIF files. (a) Traditional exchangeable ^1H/^2H sites representation extracted from the mmCIF file of PDB entry 1vcx. The ^1H atom bonded to the main-chain N atom of Ile40 is partially exchanged with ^2H. ^1H and ^2H isotopes have separate atom rows in the atom table with alternative locations A and B (green). The sum of their total occupancy (green) is set to 1.0 (the occupancy values of the ^1H and ^2H atoms are 0.07 and 0.93, respectively). The H atoms bonded to CA and CB of Ile40 are not exchanged during the partial deuteration procedure; their occupancy value is equal to 1.0. (b) Deuterium fraction representation created by REFMAC5. A new column has been created that specifies the fraction of the deuterium substitution (where 100% is fully deuterated) for the exchanged H atoms. ^2H atoms are not present in the atom table, only ^1H atoms with the corresponding deuterium fraction parameters (red). The ^1H atom bonded to the main-chain N atom of Ile40 has a deuterium fraction value of 0.92, while the ^1H atoms bonded to CA and CB of Ile40 are not exchanged, hence the deuterium fraction for these H atoms is zero. (c) Traditional perdeuterated sites representation extracted from the mmCIF file of PDB entry 3rz6. Here, all of the H atoms of Ile7 have been substituted with ^2H atoms. There are no ^1H atoms in the traditional perdeuterated structures. The occupancy of the ^2H atoms is set to 1.0 by convention. (d) Deuterium fraction representation for perdeuterated structures. All ^2H atoms are converted to ^1H atoms, the corresponding deuterium fractions are refined and values close to 1.0 are obtained.

2.3.2. Reference structure restraints

Neutron macromolecular crystallographic data often suffer from limited completeness and high resolution is not always achievable. Therefore, a useful strategy to increase the data-to-parameter ratio in refinement is that of joint neutron/X-ray refinement, provided that an isomorphous X-ray data set is available. This approach, which was originally implemented in nCNS and is available within phenix.refine in the Phenix suite, has been employed for the refinement of several macromolecular structures (Liebschner et al., 2018).

Neutron diffraction data sets are often of poorer quality compared with X-ray data. The low flux of available neutron beams requires either large crystals or very long exposure times for smaller crystals to obtain measurable diffraction data. Consequently, neutron data sets often have low completeness due to the limited data-collection time available on neutron crystallographic instruments. Additionally, the incoherent scattering of H atoms can lead to low S/N ratios.

Combining two sources of information, X-ray and neutron, can potentially mitigate some of the challenges when refining models against neutron data alone. The current joint refinement method uses a combined target function to optimize a single atomic model simultaneously against two data sets (X-ray and neutron; Afonine et al., 2010; Liebschner et al., 2020). To satisfy the refinement of the target function, the X-ray and neutron crystals should be isomorphous and ideally the data should be collected under the same conditions. This, however, cannot always be accomplished.

Any differences in the underlying structures of macromolecules analysed using different experimental methods can cause problems and require special consideration. This has been observed, for example, in the joint refinement of macromolecular models against X-ray and NMR data (Kovalevskiy et al., 2018). Joint refinement can be useful in identifying discrepancies between structures obtained under different experimental conditions. However, if attempting to achieve a single model, it is important to ensure that any approach involving the co-utilization of data from different experimental sources does not suffer from excessive bias due to fundamental structural differences. Therefore, there is a preference to avoid joint refinement in cases where other strategies to stabilize neutron refinement exist.

One such strategy is to utilize structural information from homologous X-ray models via the use of external restraints. Such restraints have been useful in the refinement of low-resolution X-ray (Headd et al., 2012; Nicholls et al., 2012; Smart et al., 2012; Schröder et al., 2014; Sheldrick, 2015; van Beusekom et al., 2018) and cryo-EM (Afonine et al., 2018; Nicholls et al., 2018) structures. This approach is robust to structural differences between the target and reference models by employing an anharmonic penalty function, which avoids pulling the model into conformations that are not supported by the data.

The purpose of external restraints is twofold. Firstly, to inject prior structural information: the target (neutron) model is pulled towards the conformation adopted by the reference (X-ray) structure, which helps to improve the model stereochemistry/geometry. Secondly, to increase the effective data-to-parameter ratio, thus stabilizing refinement and helping to avoid overfitting. The importance of the latter should not be underappreciated, especially given that neutron data are typically limited and noisy. This approach can be applied if a high-resolution model related to the target structure to be refined is available. Fortunately, when performing neutron crystallographic studies of macromolecules, the corresponding high-resolution X-ray models are invariably determined first and thus are generally available. Given that X-ray models provide significantly more accurate coordinates for all non-H atoms than their neutron counterparts, their use as a source of prior structural information appears to be a reasonable approach towards improving neutron refinement.

The CCP4 program ProSMART (Nicholls et al., 2014) generates such external restraints by distilling the local structure of a known reference model. Here, we used Pro­SMART to identify matching atoms by aligning the target model and an X-ray reference model before generating interatomic distance restraints between proximal non-H atoms within a given distance threshold (default 4.2 Å), which should be long enough to capture information about secondary structure whilst being short enough to allow differences in global conformation. The resulting external restraints were subsequently used by REFMAC5 during refinement of the target neutron model.

2.4. Performance analysis by re-refinement of PDB entries

To test our current implementation, we re-refined 97 of the available neutron PDB entries (45.5% of the total) using REFMAC5. Of these, 55 are structures that were originally refined against neutron data only and 42 are entries deposited following a joint neutron/X-ray refinement protocol. We selected our test pool based on the availability of experimental data (including complete cross-validation sets) and a wide resolution range (upper limit 1.05–2.75 Å).

For each entry, coordinate files (in PDB and mmCIF format) and crystallographic data (mmCIF format) were downloaded from the PDB. Each mmCIF reflection file was then converted into MTZ format, which serves as the standard format used by CCP4 programs (Agirre et al., 2023). For the `neutron-only' entries, the CCP4 program CIF2MTZ was utilized to convert mmCIF to MTZ format. For entries refined using a joint X-ray/neutron protocol, their mmCIF reflection files should contain two distinct data blocks: one for X-ray diffraction and one for neutron diffraction. However, a few entries have been erroneously deposited with a single data set. Since the refinement process within REFMAC5 was only performed against neutron reflections, those were extracted and converted to MTZ format using GEMMI (Wojdyr, 2022). In cases where only intensities were available, they were converted into amplitudes using the Servalcat `fw' function (Yamashita et al., 2021), which implements the French–Wilson procedure (French & Wilson, 1978).

To compare refinement statistics with those reported in the PDB, all ¹H and ²H atoms present in the models were retained without regeneration. REFMAC5 is able to read ¹H/²H sites and ²H atoms using the Servalcat REFMAC5 controller (`refmacat'), which uses GEMMI for restraint generation (Yamashita et al., 2023). ²H atoms are converted to H atoms with deuterium fraction parameters by GEMMI, and their distances are adjusted using nuclear values from the CCP4-ML. In cases such as PDB entry 5ksc, where the original model does not contain any ¹H (or ²H) atoms except for water molecules, GEMMI was employed to add them at riding positions.

If H atoms are generated, it is necessary to initialize their deuterium fraction prior to refinement. Users can choose to initialize all H atoms or only polar H atoms. For perdeuterated structures, in which all H atoms are replaced by ²H atoms, the initialization process sets the deuterium fraction parameter to 1 for all H atoms. In the case of ¹H/²H-exchanged structures, the deuterium fraction is only set to 1 for H atoms exchanged with ²H. Subsequently, the refinement process is performed to optimize the deuterium fraction. Initialization was not used for the refinement of most of the entries containing ¹H/²H or ²H sites, while it was necessary for a few entries, such as PDB entries 1c57, 1cq2, 5ksc and 1xqn, where only ¹H atoms were present in the models.

Our standard refinement protocol consisted of five cycles of restrained positional and individual ADP refinement using the data in the published resolution range. Three cycles of deuterium fraction refinement were performed after each cycle of individual atomic refinement. For perdeuterated samples we allowed refinement of the deuterium fraction for all H atoms, whilst for ¹H/²H-exchanged samples only polar H atoms had this parameter included in the optimization. H atom positions have been refined individually with all available restraints (bond lengths, angles, planarity and torsion angles) to ensure proper geometry. We found that this procedure allows deuterium fraction parameters to converge as the models had previously been refined by the original depositors.

2.4.1. Re-refinement of PDB entries originally refined against neutron data only

Using the protocol described earlier, we used REFMAC5 to re-refine 55 PDB entries that were originally refined using neutron data only. Entries were chosen over a wide resolution range from medium–low resolution (2.7 Å, PDB entry 2efa) to subatomic resolution (1.05 Å, PDB entry 4ar3). R-factor statistics for all 55 re-refined models are given in Table 2.

Table 2
Re-refinement of selected neutron-only models from the PDB

Comparison of published R-factor statistics and those obtained by re-refinement using REFMAC5.

PDB information			REFMAC5 refinement statistics
PDB code	Published resolution range (low–high) (Å)	Published R values (work/free) (%)	Initial R values (work/free) (%)	Final R values (work/free) (%)	Data completeness (%)	No. of reflections
1c57	15.79–2.40	27.0/30.1	29.7/33.0	19.9/25.4	87.38	8129
1cq2	6.00–2.00	16.0/25.0	18.6/25.7	14.9/24.7	91.07	7528
1iu6	10.00–1.60	20.1/22.8	21.1/23.4	18.4/22.9	87.16	5775
1v9g	25.05–1.80	22.2/29.4	26.9/33.5	22.9/33.7	85.31	1949
1vcx	27.60–1.50	18.6/21.7	19.0/21.7	17.8/21.8	81.94	6620
1wq2	20.00–2.40	22.9/28.9	28.6/32.1	21.5/28.7	92.29	6232
1xqn	32.82–2.50	26.6/32.0	30.0/31.5	23.0/29.4	74.99	6088
2dxm	8.00–2.10	19.7/26.0	20.2/26.6	18.8/26.4	64.12	20178
2efa	80.00–2.70	21.6/29.1	24.3/27.6	22.3/28.0	95.66	2154
2gve	10.00–2.20	26.8/31.9	26.3/30.3	23.4/29.8	93.73	22133
2wyx	19.41–2.10	22.3/25.8	25.8/28.7	22.5/27.7	86.85	15033
2xqz	53.19–2.10	22.5/25.9	25.9/29.1	19.0/26.5	77.26	13374
2yz4	33.64–2.20	27.9/31.2	28.1/31.3	22.0/27.6	66.06	8044
2zoi	70.00–1.50	19.2/21.9	19.9/22.4	18.5/21.5	89.64	14526
2zpp	20.00–2.50	22.1/26.0	22.8/26.8	20.5/25.9	98.61	2612
2zwb	20.00–1.80	22.3/24.7	22.8/25.0	18.0/23.1	95.93	9862
3a1r	30.84–1.70	19.5/23.8	18.5/22.9	16.1/21.8	81.89	10300
3fhp	41.43–2.00	16.8/24.7	18.6/23.4	16.8/22.2	81.46	4615
3kmf	20.00–2.00	25.0/30.0	28.9/29.1	27.9/31.0	86.35	31611
3q3l	36.42–2.50	22.1/26.8	23.0/26.8	22.9/26.6	73.25	27157
3ryg	27.11–1.75	18.1/20.0	22.8/23.8	20.4/25.5	92.21	4884
3rz6	21.77–1.75	20.8/23.8	25.2/25.6	21.3/25.1	79.55	4215
3rzt	27.21–1.75	20.2/24.9	25.4/29.3	21.9/29.2	75.60	3989
3ss2	24.33–1.75	21.1/24.2	23.9/26.2	20.1/26.0	77.19	4080
3u2j	12.10–2.00	23.2/27.2	23.5/27.0	22.4/26.5	86.54	14521
4ar3	15.71–1.05	19.9/23.7	19.7/23.0	18.8/22.4	88.73	21580
4ar4	27.46–1.38	18.6/22.6	17.9/22.1	15.5/21.0	91.56	9958
4bd1	9.97–2.00	21.9/25.7	22.4/25.0	14.6/17.6	87.72	18290
4c3q	10.00–2.20	19.2/24.0	21.3/25.1	17.3/24.2	88.19	13256
4fc1	10.00–1.10	21.1/25.3	22.8/25.5	21.5/24.1	76.85	10549
4g0c	20.00–2.00	26.7/28.3	27.2/26.0	24.3/28.1	84.40	13742
4k9f	27.14–1.75	19.9/24.1	21.6/26.0	17.8/25.7	74.37	3671
4q49	20.00–1.80	18.7/21.5	18.6/20.9	15.5/20.4	89.45	18803
4y0j	20.00–2.00	26.3/29.1	29.2/30.6	26.6/31.8	80.89	13095
4zz4	51.19–1.80	19.7/22.1	19.5/20.6	17.7/23.4	71.33	7524
5a90	38.91–1.70	19.2/22.7	21.1/24.1	18.9/23.2	88.61	28772
5gx9	33.45–1.49	15.8/20.0	17.3/21.0	17.0/20.9	93.50	14375
5ksc	10.00–2.10	24.0/28.0	29.6/31.4	23.5/30.9	70.78	12424
5mnx	22.17–1.42	16.6/20.6	18.5/21.9	18.3/21.8	90.57	36197
5mny	22.09–1.43	16.4/19.3	18.1/20.5	18.2/20.2	93.83	36397
5mnz	21.32–1.45	16.9/20.1	18.1/21.1	18.0/21.0	90.20	33722
5mo1	19.68–1.49	17.5/21.6	18.6/22.3	18.9/22.5	93.38	32293
5mo2	22.14–1.50	16.1/20.0	17.4/20.8	17.2/20.5	88.77	30027
5ty5	14.78–2.30	23.9/25.2	27.2/27.0	24.4/29.8	73.88	34761
5vg1	12.00–2.10	18.7/26.5	21.4/27.5	19.8/27.0	75.47	13304
5vnq	16.70–2.20	24.2/28.0	27.5/30.1	23.7/29.0	71.29	7734
5zo0	22.86–1.65	18.6/22.9	19.7/23.2	18.2/22.4	86.21	21422
6c78	14.76–1.75	18.9/21.7	23.9/25.6	19.8/24.9	85.31	25773
6gtj	32.66–1.80	23.2/27.6	24.7/28.6	19.7/27.6	78.58	21858
6h1m	21.76–2.15	21.8/24.9	22.3/25.3	18.5/22.5	91.42	12707
6l26	24.76–1.44	16.8/20.6	18.9/22.1	16.9/21.0	88.73	33763
7jor	17.23–2.05	24.9/28.8	25.4/29.5	23.2/29.5	79.44	12714
7kks	14.64–2.20	25.7/28.2	28.1/30.7	22.4/29.0	98.34	23326
7kkw	14.65–2.30	24.9/30.2	27.9/31.8	22.5/31.6	98.38	20628
7vei	17.70–2.00	17.0/21.5	16.8/21.5	14.2/21.3	98.30	7555

For some entries (for example PDB entries 1wq2, 3rz6, 4c3q, 4fc1, 5a90, 5gx9 and 7kkw in Table 2), we observe that the initial R_work and R_free values are higher than those reported in the PDB. In the case of PDB entry 1wq2, the PDB header reports values of 22.9% and 28.9% for R_work and R_free, respectively, while the paper indicates values of 28.2% and 30.1% (Chatake et al., 2003). The latter values are similar to the initial R factors from REFMAC5 (28.6% and 32.1% for R_work and R_free, respectively). Following refinement, the R_work and R_free values from REFMAC5 become comparable to the deposited values, suggesting convergence of the refinement procedure (Table 2).

For several structures, including the low-resolution PDB entries 1c57, 1wq2, 1xqn, 2gve and 2yz4, the medium-resolution PDB entries 3fhp, 3u2j, 4bd1 and 6h1m and the high-resolution PDB entries 2zoi, 2zwb, 3a1r, 4ar3, 4ar4, 4fc1 and 4q49, the R_work and R_free values obtained from REFMAC5 are lower compared with the deposited values, often improving by ∼2–3 percentage points. However, for a few other entries the final R-factor values obtained from REFMAC5 are slightly higher. One explanation is that in this study the models have been re-refined without any additional refinement strategy that could significantly improve the refinement statistics. For example, the application of TLS refinement (Winn et al., 2001, 2003), as well as the use of anisotropic ADP refinement for high-resolution structures and jelly-body restraints, could potentially improve the refined model. One general point of consideration, however, is that the calculation of scaling factors used in the R-factor equation is different among refinement packages and this can lead to differences in R factors. Although overall R values are not the only metrics to consider when evaluating the quality of a structural model, which cannot be properly assessed without careful map analysis, the values obtained from this test set indicate that our implementation for neutron crystallographic refinement performs satisfactorily.

2.4.2. Re-refinement of PDB entries originally refined using a joint neutron/X-ray strategy

We also tested the refinement of 42 models previously obtained through joint neutron/X-ray refinement utilizing solely neutron data and incorporating the deuterium fraction parameterization. Table 3 presents all joint neutron/X-ray models featuring neutron data from lowest to highest resolution that were selected for re-refinement within REFMAC5. The table compares the R-factor statistics published for these selected entries with the R factors obtained through their re-refinement using REFMAC5.

Table 3
Re-refinement of selected joint neutron/X-ray models from the PDB using neutron diffraction data only

Comparison of published R-factor statistics and those obtained by re-refinement using REFMAC5.

PDB information			REFMAC5 refinement statistics
PDB code	Published resolution range (low–high) (Å)	Published R values (work/free) (%)	Initial R values (work/free) (%)	Final R values (work/free) (%)	Data completeness (%)	No. of reflections
2r24	40.11–2.19	25.7/29.1	25.9/29.7	22.1/30.0	72.76	10892
3r98	53.54–2.40	20.7/25.1	20.8/26.1	18.4/25.7	74.99	11610
3r99	53.54–2.40	20.7/25.0	20.8/26.1	18.5/25.6	74.99	11610
3vxf	44.87–2.75	18.3/23.4	17.7/23.6	17.0/24.3	73.72	7670
3x2o	18.83–1.50	22.8/25.1	23.6/25.6	21.1/24.0	93.49	23109
3x2p	19.68–1.52	21.8/26.0	22.8/26.6	22.0/25.5	91.46	25101
4cvi	39.84–2.41	17.6/24.3	17.6/23.8	14.8/22.9	73.74	11426
4dvo	20.00–2.00	19.0/21.4	19.8/22.1	17.9/22.6	91.62	28663
4gpg	37.61–1.98	19.5/25.9	20.0/25.9	20.0/26.0	69.97	10446
4ny6	26.77–1.85	17.6/22.4	19.6/22.3	15.2/21.1	89.97	4654
4pdj	32.40–1.99	23.0/27.1	23.6/25.9	20.4/26.1	78.41	8315
4pvm	36.38–2.00	20.9/27.1	21.4/27.1	19.0/27.4	76.96	12062
4pvn	52.28–2.30	20.9/26.2	21.4/26.8	18.6/27.0	98.62	10831
4qcd	21.19–1.93	16.7/22.7	17.8/22.5	17.5/22.6	79.33	16540
4qdw	20.00–1.80	16.6/17.9	19.2/18.2	16.2/17.5	72.86	31459
4s2g	20.00–2.00	16.4/18.2	17.5/18.4	14.4/18.8	93.51	13125
4xpv	20.00–2.00	26.4/30.4	27.3/30.0	24.8/29.9	80.58	11251
5a93	15.03–2.20	21.7/23.6	23.0/22.8	15.9/23.7	71.94	11004
5cg5	53.57–2.40	18.6/22.9	19.3/22.9	17.5/22.8	98.45	17458
5cg6	22.12–2.40	26.0/28.7	26.0/25.3	24.2/27.0	98.25	17245
5jpr	36.71–2.20	23.6/31.0	23.6/31.0	20.2/31.2	68.13	8761
5mon	22.13–1.42	17.0/18.1	20.7/21.4	18.2/20.9	90.56	36196
5moo	22.09–1.43	17.0/18.5	19.9/20.9	18.0/20.3	93.83	36397
5mop	21.33–1.45	17.2/18.4	19.4/20.6	17.7/20.7	90.20	33722
5moq	25.43–1.50	15.0/16.7	17.1/18.5	15.3/18.5	89.39	30123
5mor	19.67–1.49	19.6/20.7	22.8/23.8	18.9/22.0	93.38	32293
5mos	22.15–1.50	16.6/18.0	18.7/19.5	16.9/19.8	88.77	30027
5xpe	17.02–2.09	22.5/27.8	24.2/28.2	21.0/28.7	78.13	9595
5zn0	33.76–1.90	18.8/24.7	19.6/24.9	17.4/25.6	97.05	22817
6bbr	15.00–2.30	22.1/24.6	25.8/25.1	21.4/26.1	90.97	7336
6bbz	15.43–2.20	19.8/22.0	24.2/24.0	19.7/24.6	92.18	8108
6bq8	40.00–2.20	23.2/28.8	22.8/28.4	20.5/30.6	72.77	5686
6exy	28.20–1.70	15.0/18.7	18.6/19.5	15.8/19.5	95.37	14401
6u0c	15.00–2.10	21.8/25.0	23.2/27.1	19.0/26.3	84.10	10060
6u0e	14.51–1.89	21.7/25.4	26.1/28.0	21.8/26.9	91.19	14921
6u0f	13.79–2.00	21.9/24.1	25.2/26.5	20.6/25.9	86.60	12087
7a0l	40.00–2.10	21.5/23.0	23.0/23.2	17.0/26.0	78.62	18131
7d6g	26.33–2.10	17.7/21.9	20.3/22.1	16.3/22.4	71.99	6598
7jun	14.96–2.50	20.1/25.3	20.3/25.2	17.9/26.5	83.14	7617
7tx3	14.12–1.89	22.6/27.5	23.0/27.9	21.8/28.5	93.70	22968
7tx4	12.75–2.35	17.7/25.9	17.9/25.9	16.4/26.1	85.83	5371
7tx5	32.87–2.30	18.7/25.9	19.4/26.5	18.1/27.1	75.39	5882

The R_work and R_free values obtained from REFMAC5 [Table 3; Final R values (work/free) column] by refining joint models using only neutron data were found to be similar to those obtained from joint neutron/X-ray refinement [Table 3; Published R values (work/free) column]. For some entries, the R factors are slightly improved compared with the published values. It is widely acknowledged that refinement solely using neutron data may lead to overfitting due to the explicit refinement of H atom parameters. However, the gap observed between the R_work and R_free values obtained from REFMAC5 is not substantial (the mean ΔR is ∼6%). Thus, this strategy can be a viable alternative when joint refinement is not feasible.

2.4.3. Re-refinement using external restraints

To improve the quality of neutron atomic models, especially at low resolution, re-refinement was performed by incorporating X-ray reference structure restraints. A subset of models obtained by neutron refinement only and by joint neutron/X-ray refinement, featuring neutron data at low resolution and a few at high resolution, were selected for this analysis. For the `neutron-only' entries (PDB entries 1c57, 2efa, 2yz4 and 2zpp), the corresponding X-ray reference structures were chosen from the PDB based on their high structural similarity to the neutron refined structures. The `Find Similar Assemblies' option in the PDB uses Structure Similarity Search (Guzenko et al., 2020) to assess global 3D shape similarity, providing a Structure Match Score indicating the probability as a percentage that the structure match is similar to the query. The X-ray structures chosen reported the highest Structure Match Score. If a suitable X-ray reference model is not known, we recommend running a BLAST search (Altschul et al., 1997) over the whole PDB by inputting the FASTA sequence of the target neutron model.

For the joint neutron/X-ray structures selected a different protocol was applied. Firstly, these models were subjected to refinement against their corresponding X-ray data using REFMAC5, with a total of ten refinement cycles. The output model obtained from this refinement process was subsequently employed as a reference model.

ProSMART (Nicholls et al., 2014) takes as input the neutron target model and X-ray reference structure model in PDB or mmCIF format and generates interatomic distance restraints between proximal non-H atoms reported in a restraint file. The refinement was performed by simultaneously refining non-H atoms of the model by using restraints generated by ProSMART and by using the deuterium fraction parametrization for H atoms (twenty refinement cycles interleaved with three deuterium fraction refinements¹). PDB information for the neutron and X-ray models selected, as well as the published refinement statistics and those obtained by REFMAC5, are shown in Table 4.

Table 4
Selected neutron models and corresponding X-ray reference models for re-refinement within REFMAC5 using external restraints

Comparison of published R-factor statistics and those obtained by re-refinement using REFMAC5.

PDB code		Published resolution range (low–high) (Å)		Published R values (work/free) (%)		REFMAC5 R values (work/free) (%)
Neutron	X-ray	Neutron	X-ray	Neutron	X-ray	Neutron refinement with external restraints
1c57	1dq6	15.79–2.40	8.00–1.90	27.0/30.1	18.6/NA	20.2/24.9
2efa	1b2a	80.00–2.50	55.00–1.70	21.6/29.1	18.8/23.0	21.9/25.3
2yz4	1dq6	33.64–2.20	8.00–1.90	27.9/31.2	18.6/NA	21.7/27.3
2zpp	1b2g	20.00–2.50	10.00–1.80	22.1/26.0	20.0/22.6	20.8/25.4
3r98	3r98	53.54–2.40	43.89–2.10	20.7/25.1	16.6/20.3	17.8/25.2
3r99	3r99	53.54–2.40	43.89–2.10	20.7/25.0	16.6/20.3	17.9/25.3
3vxf	3vxf	44.87–2.75	29.22–1.60	18.3/23.4	16.1/18.4	16.8/23.5
3x2o	3x2o	18.83–1.50	28.23–1.00	22.8/25.1	13.5/15.3	21.2/23.9
3x2p	3x2p	19.68–1.52	37.75–0.99	21.8/26.0	13.4/14.2	22.0/25.7
4cvi	4cvi	39.84–2.41	14.80–2.10	17.6/24.3	13.4/17.6	14.0/22.8
4dvo	4dvo	20.00–2.00	29.90–1.55	19.0/21.4	NA/NA	18.4/22.2
4gpg	4gpg	37.61–1.98	50.00–1.89	19.5/25.9	14.6/20.3	20.9/25.7
4pvn	4pvn	52.28–2.30	43.13–1.95	20.9/26.2	15.6/18.5	18.7/26.8
5cg6	5cg6	22.12–2.40	44.31–1.70	26.0/28.7	19.6/21.1	24.0/27.8
5xpe	5xpe	17.02–2.09	46.50–1.64	22.5/27.8	15.5/18.5	20.5/29.7
5zn0	5zn0	33.76–1.90	36.01–1.10	18.8/24.7	18.6/21.2	16.7/25.5
6bq8	6bq8	40.00–2.20	10.00–2.00	23.2/28.8	19.9/24.5	19.7/29.2
6exy	6exy	28.20–1.70	31.93–1.10	15.0/18.7	12.3/13.7	16.0/19.8
6u0e	6u0e	14.51–1.89	29.01–2.10	21.7/25.4	18.4/23.5	21.4/26.7
7d6g	7d6g	26.33–2.10	40.00–1.65	17.7/21.9	15.7/18.6	14.6/22.1
7tx4	7tx4	12.75–2.35	61.05–1.90	17.7/25.9	16.6/22.4	15.2/26.0

The incorporation of external restraints has been observed to improve both the R_work and R_free values for low-resolution neutron structures. Specifically, the R factors are improved by ∼2–3 percentage points in certain cases (PDB entries 1c57, 2efa, 2yz4 and 4cvi; Table 4, Neutron refinement with external restraints) compared with both the published values and those obtained using deuterium fraction refinement only. Moreover, certain high-resolution structures (PDB entries 3x2o and 3x2p) also demonstrate improved R factors, which indicate that these restraints can improve the quality of neutron models regardless of the resolution.

2.5. Selected examples of neutron crystallographic refinement

2.5.1. Re-refinement of the neutron structure of chloride-free urate oxidase in complex with its inhibitor 8-aza­xanthine

Our re-refinement runs reported in previous sections mainly looked at global refinement statistics. As a selected example that involved a more detailed inspection of neutron maps, we carried out a re-refinement of the joint neutron/X-ray structure of perdeuterated urate oxidase (UOX) in complex with its 8-azaxanthine (8AZA) inhibitor (PDB entry 7a0l; McGregor et al., 2021).

In many organisms, the degradation of uric acid (UA) to 5-hydroxyisourate (5-HIU) is catalysed by cofactor-independent UOX (Kahn et al., 1997). In a two-step reaction, UA first reacts with O₂ to yield dehydroisourate (DHU) via a 5-peroxoisourate intermediate (Bui et al., 2014). This is then followed by a hydration step, in which DHU is hydroxylated to 5-HIU (Kahn, 1999; Wei et al., 2017). The joint structure of perdeuterated UOX in complex with its 8AZA inhibitor, relevant to the hydration step, has recently been determined using X-ray and neutron data at 1.33 and 2.10 Å resolution, respectively (McGregor et al., 2021). Joint refinement was carried out with phenix.refine (Afonine et al., 2010). It showed that the catalytic water molecule (W1) is present in the peroxo hole as neutral H₂O (D₂O), oriented at 45° with respect to the organic ligand. It is stabilized by Thr57 and Asn254 on different UOX protomers as well as by an O—H⋯π interaction with 8AZA. The active site Lys10–Thr57 dyad features a charged Lys10–NH₃⁺ side chain engaged in a strong hydrogen bond with Thr57^OG1, while the Thr57^OG1–HG1 bond is oriented toward the π system of the ligand, on average.

Re-refinement of the UOX:8AZA complex with REFMAC5 was performed against neutron data alone using deuterium fraction parameterization and external restraints. ¹H and ²H atoms on previously modelled residues and water molecules were maintained at their positions and were not regenerated. Deuterium fraction parameters were refined for all H atoms. H atom positions were refined individually. External restraints were generated using ProSMART by first re-refining the model against its X-ray data (ten cycles) and using the output model as a reference structure. Data-collection and refinement statistics are given in Supplementary Table S1.

8AZA is bound as a monoanion deprotonated at N3 and omit neutron maps confirm that W1 is neutral (Fig. 4a). This is supported by the presence of positive peaks for two ²H atoms whose deuterium fraction values refine to 0.77 (H1) and 0.84 (H2). The protonation state of the Lys10–Thr57 active site dyad has also been investigated. Omit neutron maps for Lys10 show that the residue is positively charged due to the presence of a `tri-lobe' density distribution around NZ (Fig. 4b). All H atoms bound to Lys10 refine with a high deuterium fraction parameter value (>0.80). The direction of the OG1–HG1 bond in Thr57 was not easily identified in the original work (McGregor et al., 2021). Here, omit maps reveal positive density for Thr57<sup>HG1</sup> at the 2.5σ level (Fig. 4b). We refined the orientation of the OG1–HG1 bond using the REFMAC5 `hydrogen refine rpolar' (rotatable polar) option, resulting in an optimal fit to the density. The deuterium fraction parameter for HG1 refined to 0.81. The orientation of the OG1–HG1 bond suggests the formation of another O—H⋯π interaction with N7 of 8AZA at 2.56 Å and a hydrogen bond is also formed between Lys10^HZ1 and Thr57^OG1 at a distance of 1.87 Å (Fig. 4b). Overall, our results are fully consistent with those from the previous study (McGregor et al., 2021), and mechanistic considerations can be found therein.

Figure 4 Neutron structure of the UOX:8AZA complex. (a) W1 is present as a neutral H2O molecule. A 2mFo − DFc neutron scattering length density map for 8AZA and the O atom of W1 is shown in grey at the +1.0σ level. An omit mFo − DFc neutron map indicates the presence of two deuterons (H1 and H2) as suggested by the elongated positive density (in green at the +3.0σ level) next to the O atom. (b) Representation of a portion of the active site highlighting the protonation of the Lys10–Thr57 dyad and the hydrogen-bonding network. H difference neutron density for Lys10^NZ and Thr57^OG1 is shown in green at the +3.0σ and +2.5σ levels, respectively. Hydrogen bonds are shown as grey dashed lines and their distances are shown in purple.

2.5.2. Refinement of the neutron structure of the Prochloro­coccus iron-binding protein FutA

Finally, we employed REFMAC5 for the refinement of a novel neutron structure. The marine cyanobacterium Prochlorococcus plays a significant role in global photosynthesis (Huston & Wolverton, 2009). However, its growth and productivity are constrained by the limited availability of iron. Prochloro­coccus encodes the FutA protein that can accommodate the binding of iron in either its ferric (Fe³⁺) or ferrous (Fe²⁺) state. The structure of FutA has recently been determined using a combination of structural biology techniques at room temperature, revealing the redox switch that allows the binding of both iron oxidation states (Bolton et al., 2023).

The X-ray structure of FutA, determined at a resolution of 1.7 Å, shows that the iron-binding site involves four tyrosine side chains (Tyr13, Tyr143, Tyr199 and Tyr200) and a solvent molecule, forming a trigonal bipyramidal coordination. The presence of Arg203 in the second coordination shell suggested the possibility of X-ray-induced photoreduction of the iron centre, leading to a ferrous (Fe²⁺) binding state. To investigate the protonation of active site residues surrounding the iron, the neutron structure of FutA was determined at 2.1 Å resolution using ¹H/²H-exchanged crystals, taking advantage of deuterium fraction refinement (Bolton et al., 2023). The final model is characterized by R_work and R_free values of 18.2% and 25.0%, respectively. Data-collection and refinement statistics are given in Supplementary Table S2. The neutron structure reveals that the side chain of Arg103 is protonated and thus carries a positive charge, with all of its exchangeable H atoms refining with deuterium fraction parameter values of >0.50 (Fig. 5). Neutron maps suggest that the iron-coordinating residues Tyr13, Tyr143, Tyr199 and Tyr200 exist as tyrosinates. The H-omit map for the water molecule (W1) confirms its presence as neutral H₂O, supported by the presence of positive peaks for two H atoms whose deuterium fraction values refine to 0.84 (H1) and 1.0 (H2) (Fig. 5). In contrast to the room-temperature X-ray structure, Arg203 is not involved in any interactions and does not contribute to the second coordination shell. Consequently, the iron-binding site is composed of four negatively charged tyrosinates, a positively charged arginine in the second shell and a neutral water (W1), suggesting that this coordination cages neutralized ferric iron. This was further confirmed by the serial femtosecond X-ray structure and electron paramagnetic resonance (EPR) measurements. Coordinates and structure factors of the neutron FutA structure have been deposited in the PDB as entry 8oen. This represents the first neutron structure to be refined using REFMAC5 and deposited within the PDB. Further mechanistic information on FutA is discussed in a separate publication (Bolton et al., 2023).

Figure 5 FutA (ferric state) determined by neutron diffraction at 2.1 Å resolution. The iron-binding site is formed by four tyrosines (Tyr13, Tyr143, Tyr199 and Tyr200), a solvent molecule (W1) and Arg103 in the second coordination shell. The positive density (green mesh, mFo − DFc omit map at the +3.0σ level) indicates that these atoms have undergone ^1H/^2H exchange and suggests that Arg103 is positively charged whilst W1 is neutral. The side chain of Arg203 is not oriented towards the binding site and does not engage in polar interactions. N, C and O atoms are shown in blue, dark green and red, respectively. Iron is shown in gold and H atoms are in grey.

3. Conclusions and availability

Neutron crystallography offers a unique advantage in the determination of H atom positions, enabling the investigation of many biological processes. Despite its great potential in structural biology, the number of biological structures deposited in the PDB to date (25 July 2023) using neutron-only data (or joint neutron/X-ray) is extremely small (213) compared with those solved by X-ray crystallography (176 935), nuclear magnetic resonance (NMR; 14 034) and electron microscopy (EM; 16 239). This is due to technical limitations such as low neutron beam flux, long data-collection times and limited access to neutron beamlines. Nonetheless, recent advances in instrumentation, experimental protocols and computational tools have significantly advanced the field. As a result, the number of neutron structures deposited in the PDB has significantly increased in the last few years. The period 2015–2022 alone has seen the deposition of more than half of the total neutron structures (130 out of 213) and this is likely to further accelerate in the coming years.

The CCP4 suite now provides tools for the refinement of macromolecular models using neutron diffraction data. Recent developments include the extension of the CCP4 Monomer Library by incorporating H atom nucleus distances. These restraints are required to ensure the correct H atom positions in neutron crystallographic refinement. Moreover, the inclusion of H nucleus positions holds potential for the further refinement of H atoms of cryo-EM models, as both H atom positions (electron and nucleus) contribute to the scattering. New features and refinement strategies have been implemented in REFMAC5 for the refinement of neutron models: specifically, the introduction of the deuterium fraction parameter for H atoms. One of the benefits of this approach is that it generates models containing only H atoms, without any ¹H/²H or ²H sites. For each H atom, the models incorporate a deuterium fraction parameter that indicates the level of deuteration in the sample. This results in clearer and more easily interpretable models, minimizing the bookkeeping errors that may arise when alternative conformations are present in the models. Re-refinement of neutron structures using REFMAC5 has yielded R-factor values that are in line with the originally deposited values, including those obtained previously through joint neutron/X-ray techniques. Additionally, for certain neutron entries the refinement process has led to improvements in model quality. Another valuable strategy is the use of external reference structure restraints during the refinement of models obtained by neutron diffraction. Incorporating restraints from X-ray reference structures has demonstrated an enhancement in the accuracy and reliability of neutron models, particularly in low-resolution cases.

The ability to perform neutron crystallographic refinement using REFMAC5 will be available in CCP4i2 (Potterton et al., 2018) and CCP4 Cloud (Krissinel et al., 2022) in an upcoming version of CCP4 that uses Refmacat instead of REFMAC5. This option can be enabled by selecting the appropriate diffraction experiment type (X-ray, Electron or Neutron) in the `Advanced' tab of the Refinement task interface, in which case the appropriate form factors and relevant default behaviours are used during model refinement. In Neutron mode, the graphical interface provides the ability to choose whether to refine all, only polar or only rotatable polar H atom positions, to use H atom torsion-angle restraints, to refine the ¹H/²H fraction for all H atoms (for perdeuterated crystals) or just polar H atoms (for ¹H/²H exchanged samples), and the choice of whether to reinitialize ¹H/²H fractions prior to refinement. Relevant keywords and documentation for neutron crystallographic refinement will be available in the documentation section of the CCP4 website (https://www.ccp4.ac.uk/).

Footnotes

¹As a rule, when external restraints or jelly-body refinement are used, more refinement cycles are needed to achieve convergence.