research papers
Improving the accuracy and resolution of neutron crystallographic data by threedimensional profile fitting of Bragg peaks in reciprocal space
^{a}Neutron Scattering Division, Neutron Sciences Directorate, Oak Ridge National Laboratory, 1 Bethel Valley Road, Oak Ridge, TN 37831, USA, ^{b}Computer Science and Mathematics Division, Computing and Computational Sciences Directorate, Oak Ridge National Laboratory, 1 Bethel Valley Road, Oak Ridge, TN 37831, USA, ^{c}Institut für Biologie, HumboldtUniversität zu Berlin, Philippstrasse 13, Leonor Michaelis Haus, 10115 Berlin, Germany, ^{d}Department of Chemistry, University of Alabama in Huntsville, 301 Sparkman Drive, Huntsville, AL 35899, USA, and ^{e}Neutron Technologies Division, Neutron Sciences Directorate, Oak Ridge National Laboratory, 1 Bethel Valley Road, Oak Ridge, TN 37831, USA
^{*}Correspondence email: sullivanbt@ornl.gov
Neutron crystallography is a powerful technique for directly visualizing the locations of H atoms in biological macromolecules. This information has provided key new insights into enzyme mechanisms, ligand binding and hydration. However, despite the importance of this information, the application of neutron crystallography in biology has been limited by the relatively low _{1/2}) at high resolutions, decreased R_{free} factors, extended resolutions and improved nuclear density maps. Importantly, additional features are revealed in nuclear density maps that may provide additional scientific information. These results suggest that threedimensional profile fitting will help to extend the capabilities of neutron macromolecular crystallography.
of available neutron beams and the large incoherent neutron scattering from hydrogen, both of which contribute to weak diffraction data with relatively low signaltobackground ratios. A method has been developed to fit weak data based on threedimensional profile fitting of Bragg peaks in by an Ikeda–Carpenter function with a bivariate Gaussian. When applied to data collected from three different proteins, threedimensional profile fitting yields intensities with higher correlation coefficients (CCKeywords: neutron crystallography; integration; profile fitting.
1. Introduction
Neutron crystallography can provide structural, chemical and functional information on biological macromolecules that is difficult or impossible to obtain using other techniques (Blakeley et al., 2008). One of its main advantages is the ability to directly visualize hydrogen (H) or deuterium (D) atoms at modest resolutions of around 2.0–2.5 Å (Bacik et al., 2017; Kwon et al., 2016; Casadei et al., 2014; Coates et al., 2008; Wan et al., 2015; Chen & Unkefer, 2017). Despite its potential to elucidate the molecular mechanisms behind a wealth of phenomena (Langan et al., 2018; Schaffner et al., 2017), the application of neutron crystallography remains limited by the relatively weak intensity of available neutron beams and the high neutron scattering background arising from by hydrogen within the sample (O'Dell et al., 2016). While more powerful beamlines and advances in sample preparation have helped to address these challenges, there are also opportunities to develop more advanced computational tools to improve the accuracies of the measured neutron crystallographic data and of the resulting refined structures. Previously, we have developed new computational tools for joint Xray and neutron that result in more accurate structures (Afonine et al., 2010). In this work, we focus on a new computational tool to increase the accuracy of the neutron crystallographic data.
One existing approach to integrating neutron Bragg peaks is to use peakminusbackground integration methods. These integration schemes sum events from a predefined volume centered at the peak and subtract the local background, which is determined by summing events from a separate, nearby volume with appropriate geometric scaling. While these schemes have proven to be successful, they face several critical disadvantages. Firstly, they may not appropriately account for the asymmetric peak shape at pulsed neutron sources. Neutron Bragg peaks from instruments with pulsed, moderated sources have a long tail on the high timeofflight (TOF) end which is difficult to distinguish from background, resulting in either a decreased signaltonoise ratio (with a generous peakvolume definition) or artificially decreased intensities (when this tail is considered to be background). In addition, they demand very precise knowledge of the location of each peak. For large unitcell experiments in particular, being only a few pixels off can decrease the integrated intensity by factors of up to 50% with aggressive integration schemes. Using peakminusbackground integration, peaks that fall on or near detector edges may not be integrated accurately. In the case of a standard data set collected on the MaNDi beamline at the Spallation Neutron Source (SNS; Coates et al., 2015), integration errors arising from peaks near detector edges may affect as many as one fifth of the peaks. Finally, as scientifically pertinent problems continue to demand higher resolution and the analysis of larger unit cells (Azadmanesh et al., 2017), it becomes more difficult to quantify peak intensity as peaks become closer to each other and eventually overlap.
To address these issues, profile fitting has historically been employed. While analytical consideration of singlecrystal Bragg peak intensities was first given serious consideration in 1962, Diamond was the first to demonstrate increased crystallographic data quality as a result of profile fitting (Alexander & Smith, 1962; Diamond, 1969). A decade later, profile fitting was extended to large unit cells using the `oscillation method' (Rossmann, 1985; Harrison et al., 1985) and has since been developed further (Pavese & Artioli, 1996; Leslie, 2006; Kabsch, 2006). While these techniques are appropriate for monochromatic Xray and neutron Bragg peaks, planned user programs at pulsed neutron sources such as the European Spallation Source (ESS), Lund, Sweden and the Second Target Station at SNS, Oak Ridge, USA will enable the widespread use of TOF techniques. To maximize the effectiveness of experiments at current and future pulsed neutron sources, it is imperative to have algorithms that exploit the information provided by TOF profiling.
Crystallography beamlines at modern pulsed neutron sources use timeresolved area detectors to record diffracted neutrons. Recently, there have been a handful of proposals to fit TOF profiles to integrate peaks. Yano et al. (2016) demonstrated that profile fitting provides improved model structures from protein data. To carry out their profile fitting, the authors fitted the observed profiles to a Gaussian profile convolved with two backtoback exponentials that phenomenologically describe the profiles. This is similar to the functional form proposed by Gutmann (2017), who noted that it describes the peak asymmetry arising from the tail well. The first report to examine fitting in (Schultz et al., 2014) demonstrated decreased R factors using peaks integrated along the TOF profile compared with peakminusbackground integration. A complete description of the peak, however, must be threedimensional to account for the two detector spatial dimensions and the TOF. Equivalently, these three dimensions can be expressed in A preliminary report (Tomoyori & Tamada, 2016) suggested that threedimensional profile fitting will be beneficial to data quality, but examined only a handful of peaks in detector space.
Here, we present an algorithm for integrating Bragg peaks by threedimensional profile fitting in R values for protein data sets yet, of particular interest, produces a significantly increased CC_{1/2} at high resolutions (Karplus & Diederichs, 2012). To assess the accuracy of each integration method, we carry out refinements of models from Xray data against peaks from each integration method. In each case examined, profile fitting yields R_{free} factors demonstrating an increased accuracy from profile fitting. The first data set, perdeuterated E166Q βlactamase mutant, shows a decrease in R_{free} of 2.3% at 1.89 Å resolution. The second data set, H/Dexchanged PsbO (an extrinsic subunit of photosystem II), shows a decrease in R_{free} of 2.3% at 2.2 Å resolution. The third data set, perdeuterated Pseudomonas aeruginosa peptidyltRNA hydrolase 1 (PaPth1), shows a decrease in R_{free} of 2.7% from initial at 2.60 Å resolution. The increased resolution in data sets such as that of PaPth1 makes it possible to better visualize important features such as water molecules. Finally, the resulting nuclear density maps from each integration method are compared. Reflective of their decreased R_{free} values, nuclear density maps refined against profilefitted intensities show better agreement with the atomic model. Given these results, it is clear that threedimensional profile fitting has the potential to advance the capabilities of neutron crystallography.
The primary objective of this work is to improve data quality through more accurate integration of weak peaks and peaks that are partially recorded at the edge of detectors. However, we expect that threedimensional profile fitting will also benefit the deconvolution of any overlapping peaks. After describing the algorithm in detail, we compare its performance with standard spherical integration schemes using three complete representative data sets collected on the MaNDi beamline. Two data sets are perdeuterated and one is H/Dexchanged, demonstrating the effectiveness of this technique for both types of samples. It is shown that profile fitting yields comparable merging2. Methods
2.1. Data collection
For initial testing, strong peaks from a scolecite data set recorded on the TOPAZ beamline at SNS, Oak Ridge, USA (Jogl et al., 2011) were used. Protein data that contained many considerably weaker peaks were collected on the MaNDi beamline (Coates et al., 2015). The protein datacollection strategy was optimized using the CrystalPlan package (Zikovsky et al., 2011) and the numbers of orientations recorded are presented in Tables 1, 2 and 3. Crystallization of the E166Q βlactamase mutant was carried out as described in Tomanicek et al. (2010), while PsbO was crystallized as described in Bommer et al. (2017). Crystallization of PaPth1 was achieved as described in McFeeters et al. (2016).



2.2. Moderator characterization by Monte Carlo simulations
Neutron emission from the decoupled poisoned hydrogen moderator as viewed by the TOPAZ and MaNDi beamlines was simulated using MCSTAS (Nielsen & Lefmann, 2000) as described in Gallmeier (2010). Briefly, Monte Carlo simulations of the moderator output were fitted to the Ikeda–Carpenter (IC) function (Ikeda & Carpenter, 1985),
where ψ_{IC} is the intensity of neutrons from the moderator, α and β are energydependent constants, R is the energydependent ratio of slow to from the moderator and t′ = t − t_{0} > 0. This fit was performed for 141 logarithmically spaced energies ranging from 1 × 10^{−5} to 100 eV and the value of each parameter at each energy was fitted to a fourthorder Padé approximant. These values were used as an initial guess for fitting TOF profiles using the IC function.
2.3. Data reduction for profile fitting and strong peaks
The integration scheme was tested using the Mantid framework (Arnold et al., 2014), which allows the quick conversion of recorded event data to First, an orientation matrix (UB matrix) is determined from several hundred bright peaks in Given the UB matrix, the locations of all observable peaks were predicted using the PredictPeaks algorithm in Mantid. For the samples and resolutions presented in this work, the peaks did not overlap, as verified by ensuring that all integrated peaks were separated by at least the outer radius of the background used for spherical integration (§2.6). The procedure for each predicted peak is illustrated in the yellow box in Fig. 1.
For each peak with index h = (h, k, l), a histogram of recorded events from (h − η, k − η, l − η) to (h + η, k + η, l + η) is generated in η is a parameter that determines how large a volume in is considered for background removal. In practice, this parameter can be varied in the range ∼0.2–0.5 with little effect on the resulting intensities. For the current work η = 0.25 was used. From this histogram, the background must be differentiated from the peak signal. To determine the appropriate background threshold, a nearestneighbors smoothed histogram is generated. The threshold above which voxels (threedimensional `pixels' in reciprocal space) will be included in the peak will be determined from this smoothed histogram. Given that the energy of each peak is known and that emission from the moderator has been characterized by Monte Carlo simulations (§2.2), the expected TOF profile of each peak is known and only needs to be scaled for the number of neutrons. Thus, to determine the background threshold, it is sufficient to fit this expected profile to the resulting TOF profile at each background level until a satisfactory profile is found (χ^{2} ≃ 1). To achieve this, the TOF profile is generated by creating a histogram of events binned by TOF (TOF ∝ Lsin(θ)/q), effectively summing the remaining two directions. This profile is fitted to the Ikeda–Carpenter function, ψ_{IC}, convolved with a Gaussian and a tophat function to account for detector broadening and finite protonpulse duration, respectively. This is illustrated in Figs. 2(a) and 3(c), which show the TOF profile both before (blue) and after (orange) background subtraction. The background level is taken as the intensity with which the TOF profile is best described by the predicted TOF profile. These voxels (for example the slice shown in Fig. 3b) are used to construct the threedimensional model of the peak.
To generate the full threedimensional profile, it is natural to consider the reciprocalspace histogram in spherical coordinates q(q_{x}, q_{y}, q_{z}) → q(q_{r}, q_{φaz}, q_{2θ}) as 1/q_{r} ≃ TOF and q_{φaz} and q_{2θ} are described by a bivariate Gaussian distribution ψ_{BVG}, where φ_{az} denotes the azimuthal coordinate (in the xy plane) and 2θ is the standard scattering angle coordinate (angle from the z axis). The angular distribution is fitted to a twodimensional histogram in φ_{az} and 2θ, effectively summing q_{r} (Figs. 2b and 3f). ψ_{IC} and ψ_{BVG} at this point are effectively independent probability distributions. Incorporating a scale factor, A, and a constant background term, B, the resulting threedimensional model, ψ, is given by their product: ψ = A(ψ_{IC} × ψ_{BVG}) + B, where A and B are determined by a leastsquares fit to the threedimensional event histogram in Generating the model in which scales linearly with q to provide an undistorted view of the threedimensional peak profile, allows discretization at the level of instrument resolution rather than by generating thick slices, minimizing quantification error. A threedimensional rendering of a peak and its model are shown in Figs. 3(g) and 3(h), while twodimensional slices are shown in Figs. 2(c), 2(d) and 4(b). For completeness, it should be noted that this threedimensional model is generated from a (2 + 1)dimensional fit to simplify the leastsquares optimization from a computational point of view. In practice, no difference was found between these fits and full threedimensional profile fits.
2.4. Profile fitting for weak peaks and peaks on detector edges
While the procedure described in §2.3 works well for strong peaks, it is expected that profile fitting will most benefit the integration of weak peaks where the background and peak are nearly indistinguishable. An example of such a peak is shown in Fig. 4(a). While the TOF direction can still be fitted using the moderator characterization (Fig. 4c), there are too few counts to create a fittable angular histogram (Fig. 4d). To circumvent this, and given that the profile of ψ_{BVG} changes slowly with φ_{az} and 2θ, the angular distribution ψ_{BVG} is assumed to be the same as a nearest neighbor in (q_{φaz}, q_{2θ}) from a library of strong peaks. For the work presented here, profiles were applied if the peak had fewer than 250 events (as determined by spherical integration). The strongpeak library was constructed from peaks containing more than 500 events (as determined by spherical integration) for each data set. The parameters defining peak shape for the strongpeaks libraries for E166Q βlactamase and PsbO are shown in Fig. 5.
Since peaks near the detector edges may not be fully recorded, the profiles of the strong peaks can also be used to recover their intensity. In the present work, profiles were applied to edge peaks if the peak location was predicted to be 15 or fewer detector pixels from a detector edge. The merging statistics of peaks near the edge (between 1 and 15 pixels) are shown in Table 4.

2.5. Calculation of I and σ(I)
Reliable N_{obs}, it is clear that for each voxel N_{peak} = N_{obs} − N_{bg}, where N_{peak} and N_{bg} are the number of diffracted neutrons in the peak and background, respectively. The peak intensity I is then defined as I = . Following the same reasoning as Pflugrath (1999), the variance, σ^{2}(I), of this intensity is just the sum of the associated variances. Assuming Poisson statistics (), this can be expressed as
depends on accurate integration and error determination. Defining the observed number of neutrons in each voxel in aswith the final term being the variance of the fit. At this point, quantification of peak intensity depends on how the volume of the peak is defined (i.e. which voxels are summed over) and how the background is determined. For the present work, the intensity is determined by summing the model intensities of voxels that are above 5% of the maximum value of N_{model}. The background is assumed to be constant throughout the volume of the peak and is assumed to be the average number of neutrons in the (h − η, k − η, l − η) to (h + η, k + η, l + η) volume that is not considered a peak and is accessible with the detector coverage of the instrument.
2.6. Spherical integration of peak intensities
For comparison with traditional integration, the same peak sets were analyzed using the standard integration and via the IntegratePeaksMD algorithm in Mantid. Peaks from E166Q βlactamase and PsbO were integrated with a radius of 0.021 Å^{−1} and the background shell was taken from 0.022 to 0.026 Å^{−1}, while PaPth1 was integrated with a radius of 0.018 Å^{−1} with a background shell from 0.019 to 0.022 Å^{−1}.
protocol at MaNDi in parallel with profilefitted peaks. The only difference between the two data sets is how they were integrated. Spherical integration was performed2.7. Analysis of integrated intensities and details
After integration, protein peak intensities were scaled using LAUENORM from the LAUEGEN package (Campbell, 1995) and the merging statistics presented in Tables 1, 2 and 3 and Supplementary Tables S1, S2 and S3 were calculated using PHENIX (Adams et al., 2010). For threedimensional profile intensity data, data were rejected if χ^{2} of either the TOF, BVG or threedimensional scaling fit was too large (χ^{2} > 50). Peaks with I/σ(I) < 1.0 from either profile fitting or spherical integration were rejected. Peaks were also removed if the peak center was one detector pixel from the edge. The statistics presented in these tables are discussed in Karplus & Diederichs (2012). To generate initial models for a Protein Data Bank (PDB) entry for the same protein generated from Xray crystallography was used as a starting point. This model was aligned with the data using via Phaser (phenix.phaser). At this point, H or D atoms were added using phenix.ready_set. This model was refined using phenix.refine (Afonine et al., 2012) against data sets integrated using each integration method. The peak data for including the selection of the working and testing data sets, are the same except for the intensities and uncertainties resulting from the integration method. For each protein, models were refined from the same initial model for nine iterations using phenix.refine. For E166Q βlactamase, atomic positions, atomic B factors and occupancies were refined. Because was performed at above 2 Å for PsbO and PaPth1, individual atomic positions were not refined, although rigidbody was allowed. Overall R factors from refinements are shown in Tables 1, 2 and 3, while Fig. 6 and Supplementary Tables S6, S7 and S8 show CC and R from the refinements for each resolution shell.
To directly compare strong and weak peaks, merging statistics for E166Q βlactamase and PsbO are presented in Tables S4 and S5. For these tables, peaks were separated by being either above or below the median I/σ(I) for each data set for each integration method. Merging statistics were calculated in PHENIX exactly as was performed for the whole data set. The same comparison is not presented for PaPth1 as the low number of peaks (<15 000 in the final data set) makes it difficult to directly compare the split peak sets.
3. Results
3.1. Results for the E166Q βlactamase mutant
A summary of merging statistics and βlactamase mutant against peaks from each integration method is presented in Table 1. Shellbyshell merging statistics are given in Supplementary Table S1, while Supplementary Table S4 shows the same statistics for weak and strong peaks separately. The most drastic difference in merging statistics is in Pearson's CC_{1/2}, at high resolutions (Supplementary Table S1). I/σ(I) is higher at low resolution and approaches I/σ(I) = 1 more quickly at high resolution.
from refining the initial model of the E166QAtomic positions were refined during the E166Q βlactamase The models refined against profilefitted and spherically integrated data differed by an r.m.s.d. of 0.09 Å. Shellbyshell are shown in Fig. 6 and Supplementary Table S6. Overall, against the known model yields increased CC and decreased R values, particularly in the medium and highresolution shells. Individual residues have several structural differences as a result of profile fitting. One such residue is highlighted in Fig. 7.
3.2. Results for PsbO
A summary of merging and initial , while Supplementary Tables S2 and S5 show shellbyshell merging statistics. As with the E166Q βlactamase mutant, threedimensional profile fitting resulted in comparable overall merging R values and increased CC_{1/2}, especially at high resolutions. The overall I/σ(I) values are again higher at low resolution and approach unity more quickly for profilefitted peaks than spherically integrated peaks. Shellbyshell are shown in Fig. 6 and Supplementary Table S7, which show increased CC values and decreased R values in the medium and highresolution shells.
for PsbO is presented in Table 23.3. Results for PaPth1
A summary of merging and , shellbyshell merging statistics are shown in Supplementary Table S3 and shellbyshell are presented in Fig. 6 and Supplementary Table S8.
for PaPth1 is presented in Table 33.4. Effect on nuclear density
Better integration is expected to yield improved nuclear densities. Selected residues are shown in Fig. 7. One potential advantage of improved integration is the ability to resolve the location of additional atoms in aminoacid side chains, as illustrated by Ser86 in perdeuterated E166Q βlactamase (Fig. 7, Supplementary Fig. S1). Density maps from profilefitted intensities clearly resolve the OG atom (the top O atom in the images) and the bound D atom while maps derived from spherical integration are missing density for these atoms. Additionally, higher quality density maps allow atomic positions to be determined with higher certainty. Asn55 from the H/Dexchanged PsbO data set is shown in Fig. 7 and Supplementary Fig. S2. From inspection, it is clear that profile fitting results in better nuclear densities around the (top) ND2 atom and the bound DD21 and DD22 atoms. In addition, Phe28 from perdeuterated PaPth1 is shown in Fig. 7 and Supplementary Fig. S3. It is clear from inspection that the map from profilefitted intensities better matches the perdeuterated phenyl ring. Clearer definition in features such as this is expected to enable the discovery of new structural details.
4. Discussion
We have presented full threedimensional profile fitting of entire neutron crystallographic data sets for the first time. In contrast to other recent profile fitting performed in detector space (Tomoyori & Tamada, 2016; Yano et al., 2016; Gutmann, 2017), this integration is performed in As has been argued previously (Schultz et al., 2014), there are several convenient features of integrating in Most notably, the peak shapes are straightforward to model. In particular, it is straightforward to isolate peaks at high resolutions. In these peaks maintain separation, and even with a as large as that of PsbO (∼200 Å) there are no obvious effects of peak overlap. The background can be straightforwardly assessed over a large volume of by considering (h − η, k − η, l − η) to (h + η, k + η, l + η), which aids the quantitation of highresolution peaks over integration in detector space.
For these data sets, an overall increase in the average I/σ(I) was observed. Increases of approximately 25%, 40% and 15% were found for the E166Q βlactamase mutant, PsbO and PaPth1, respectively. This difference is likely to be related to the background level of each data set. Profile fitting significantly reduces the amount of nonpeak volume integrated and so it is expected that increases in signaltonoise will be seen in samples with higher background. It has been speculated (Tomoyori & Tamada, 2016) that there should be an increase of around 10% in signaltonoise resulting from profile fitting, while noting that applying learned peak shapes to weak peaks may increase this further. This is fairly consistent with our reported I/σ(I) values. Of particular interest, these data sets exhibit increased I/σ(I) at low resolution and decreased I/σ(I) at high resolution. This is likely to be an artifact of a high I/σ(I) resulting from the spherical integration method. Experience has shown that I/σ(I) does not fall to unity at high resolutions when using the spherical integration method, and while I/σ(I) does not fall to 1.0 using profile fitting, it more quickly approaches the unity limit.
It is also interesting to consider the merging statistics. As a complete data set, profile fitting leads to comparable merging R values for all three data sets presented. At higher resolutions, though, the merging R values for profilefitted peaks are slightly higher than those from spherically integrated intensities (Supplementary Tables S1, S2 and S3). These figures demonstrate that profilefitted intensities have a higher spread at high resolution, though not necessarily that the intensities are less accurate. To assess accuracy, we refined models from Xray data against peak sets which vary only in the integration method. Models refine better against profilefitted intensities, demonstrating that the technique produces more accurate intensities. The Pearson's CC_{1/2} has been argued to be the most reliable indicator of the quality of a data set (Evans, 2011; Diederichs & Karplus, 2013). For all three data sets, substantially higher CC_{1/2} values are observed at higher resolution. This increased consistency is, of course, a consequence of the relative insensitivity of profilefitted intensities to noise. In light of this, it is unsurprising that models refine better against profilefitted data.
To further verify that profile fitting has the largest effect in more accurately integrating highresolution data, shellbyshell and Supplementary Tables S6, S7 and S8. The CC_{1/2} and R values show that data–model agreement predominantly increases at medium and high resolutions. Taken together, these results strongly suggest that profile fitting more accurately integrates peaks for model by accurately integrating highresolution/weak peaks. The increase in CC_{1/2} is especially noticeable when comparing strong peaks with weak peaks. Supplementary Tables S4 and S5 compare peak sets which have been split into high and low I/σ(I). When considering the E166Q βlactamase data set (Supplementary Table S4), highresolution peaks have a CC_{1/2} above 0.19 in the outermost shells for profilefitted peaks, while spherically integrated peaks quickly fall to CC_{1/2} < 0.1. PsbO, which overall has a higher I/σ(I), shows similar results (Supplementary Table S5).
are presented in Fig. 6For weak peaks, ψ_{BVG} profiles in the nonTOF directions (φ_{az}, 2θ) were determined from a library of strong peaks. The notion of applying profiles from a library of strong peaks dates back to the 1980s in neutron crystallography (Sjölin & Wlodawer, 1981; Wilkinson et al., 1988) and has since proven to be beneficial in solving several protein structures. Of the Xray structures deposited in the PDB, peak integration for macromolecular crystallography has been dominated by XDS, MOSFLM, HKL and d*TREK (Kabsch, 2010; Leslie, 2006; Otwinowski & Minor, 1997; Pflugrath, 1999). More recently, DIALS has been released to facilitate the development of new algorithms and to process data from increasingly highthroughput crystallography facilities (Winter et al., 2018). While all of these packages use profile fitting to fit weak or incomplete peaks, MOSFLM and HKL integrate threedimensional peaks by summing a series of twodimensional images, a technique termed twodimensional integration. XDS, d*TREK and DIALS, on the other hand, integrate a full threedimensional model of the peak described as a threedimensional Gaussian. The integration scheme described in this work is most similar to threedimensional integration, except that the third dimension arises from TOF (rather than φslicing) and the functional form in the third dimension is an Ikeda–Carpenter function. The parameters defining peak shape from profile fitting are presented in Fig. 5, which shows the parameters for peaks with 0.4 mrad of the σ_{az} value of each data set. It is clear that the peak size decreases along the scattering direction with increasing scattering angle. In addition, the peak orientation, defined by the covariance ρ in clearly depends on the azimuthal angle. It is also clear that the peak profile changes appreciably for different samples. While using the profile of the nearest neighbors yielded more accurate intensities, the observed trends suggest that peaks can be modeled using the resolution function of the instrument and sample parameters which may further increase accuracy. It is also conceivable that a machinelearningbased approach could be developed to more accurately predict peak profiles for weak peaks.
In addition to more accurately integrating weak peaks, profile fitting offers the opportunity to recover data near the edge of detectors. As an example, merging statistics of pixels near the edge for the E166Q βlactamase data set are shown in Table 4. Of particular interest, the CC_{1/2} for the profilefitted data resembles CC_{1/2} for the entire data set, while spherically integrated peaks have a CC_{1/2} that quickly falls to 0. In traditional integration workflows, these intensities would typically be discarded or included despite poor quantification. While all of the data sets analysed so far were recorded using SNS detectors (Riedel et al., 2015), the capability to recover edge intensities also has the potential to benefit the integration of data recorded on positionsensitive tube detectors, which have considerably more gaps in detector coverage.
This algorithm has been implemented in the Mantid (Arnold et al., 2014) software package as the IntegratePeaksProfileFitting algorithm.
Supporting information
Supplementary Tables and Figures. DOI: https://doi.org//10.1107/S2059798318013347/mn5117sup1.pdf
Funding information
This work was funded through grant R01GM071939 from the National Institutes of Health. The neutron scattering measurements were carried out at the Spallation Neutron Source, which is sponsored by the Division of Scientific User Facilities, Office of Basic Energy Sciences, US Department of Energy under contract No. DEAC0500OR22725 with UTBattelle LLC. This research was also sponsored by the Applied Mathematics Division of ASCR, DOE, in particular under the ACUMEN project. Production and crystallization of PsbO was funded by Deutsche Forschungsgemeinschaft (DFG) grant SFB1078, TP A5. This work used samples grown at Oak Ridge National Laboratory's Center for Structural and Molecular Biology (CSMB), which is funded by the Office of Biological Environment Research in the Department of Energy's Office of Science.
References
Adams, P. D., Afonine, P. V., Bunkóczi, G., Chen, V. B., Davis, I. W., Echols, N., Headd, J. J., Hung, L.W., Kapral, G. J., GrosseKunstleve, R. W., McCoy, A. J., Moriarty, N. W., Oeffner, R., Read, R. J., Richardson, D. C., Richardson, J. S., Terwilliger, T. C. & Zwart, P. H. (2010). Acta Cryst. D66, 213–221. Web of Science CrossRef CAS IUCr Journals Google Scholar
Afonine, P. V., GrosseKunstleve, R. W., Echols, N., Headd, J. J., Moriarty, N. W., Mustyakimov, M., Terwilliger, T. C., Urzhumtsev, A., Zwart, P. H. & Adams, P. D. (2012). Acta Cryst. D68, 352–367. Web of Science CrossRef CAS IUCr Journals Google Scholar
Afonine, P. V., Mustyakimov, M., GrosseKunstleve, R. W., Moriarty, N. W., Langan, P. & Adams, P. D. (2010). Acta Cryst. D66, 1153–1163. Web of Science CrossRef CAS IUCr Journals Google Scholar
Alexander, L. E. & Smith, G. S. (1962). Acta Cryst. 15, 983–1004. CrossRef IUCr Journals Web of Science Google Scholar
Arnold, O., Bilheux, J.C., Borreguero, J., Buts, A., Campbell, S. I., Chapon, L., Doucet, M., Draper, N., Ferraz Leal, R., Gigg, M., Lynch, V. E., Markvardsen, A., Mikkelson, D. J., Mikkelson, R. L., Miller, R., Palmen, K., Parker, P., Passos, G., Perring, T. G., Peterson, P. F., Ren, S., Reuter, M. A., Savici, A. T., Taylor, J. W., Taylor, R. J., Tolchenov, R., Zhou, W. & Zikovsky, J. (2014). Nucl. Instrum. Methods Phys. Res. A, 764, 156–166. Web of Science CrossRef Google Scholar
Azadmanesh, J., Trickel, S. R., Weiss, K. L., Coates, L. & Borgstahl, G. E. O. (2017). Acta Cryst. F73, 235–240. Web of Science CrossRef IUCr Journals Google Scholar
Bacik, J.P., Mekasha, S., Forsberg, Z., Kovalevsky, A. Y., VaajeKolstad, G., Eijsink, V. G. H., Nix, J. C., Coates, L., Cuneo, M. J., Unkefer, C. J. & Chen, J. C.H. (2017). Biochemistry, 56, 2529–2532. Web of Science CrossRef PubMed Google Scholar
Blakeley, M. P., Ruiz, F., Cachau, R., Hazemann, I., Meilleur, F., Mitschler, A., Ginell, S., Afonine, P., Ventura, O. N., CousidoSiah, A., Haertlein, M., Joachimiak, A., Myles, D. & Podjarny, A. (2008). Proc. Natl Acad. Sci. USA, 105, 1844–1848. Web of Science CrossRef PubMed CAS Google Scholar
Bommer, M., Coates, L., Dau, H., Zouni, A. & Dobbek, H. (2017). Acta Cryst. F73, 525–531. Web of Science CrossRef IUCr Journals Google Scholar
Campbell, J. W. (1995). J. Appl. Cryst. 28, 228–236. CrossRef CAS Web of Science IUCr Journals Google Scholar
Casadei, C. M., Gumiero, A., Metcalfe, C. L., Murphy, E. J., Basran, J., Concilio, M. G., Teixeira, S. C. M., Schrader, T. E., Fielding, A. J., Ostermann, A., Blakeley, M. P., Raven, E. L. & Moody, P. C. E. (2014). Science, 345, 193–197. Web of Science CrossRef CAS PubMed Google Scholar
Chen, J. C.H. & Unkefer, C. J. (2017). IUCrJ, 4, 72–86. Web of Science CrossRef CAS PubMed IUCr Journals Google Scholar
Coates, L., Cuneo, M. J., Frost, M. J., He, J., Weiss, K. L., Tomanicek, S. J., McFeeters, H., Vandavasi, V. G., Langan, P. & Iverson, E. B. (2015). J. Appl. Cryst. 48, 1302–1306. Web of Science CrossRef CAS IUCr Journals Google Scholar
Coates, L., Tuan, H.F., Tomanicek, S., Kovalevsky, A., Mustyakimov, M., Erskine, P. & Cooper, J. (2008). J. Am. Chem. Soc. 130, 7235–7237. Web of Science CrossRef PubMed CAS Google Scholar
Diamond, R. (1969). Acta Cryst. A25, 43–55. CrossRef CAS IUCr Journals Web of Science Google Scholar
Diederichs, K. & Karplus, P. A. (2013). Acta Cryst. D69, 1215–1222. Web of Science CrossRef CAS IUCr Journals Google Scholar
Evans, P. R. (2011). Acta Cryst. D67, 282–292. Web of Science CrossRef CAS IUCr Journals Google Scholar
Gallmeier, F. X. (2010). SNS Source Descriptions for Use with MCSTAS. Tech. Rep. SNS106100200TR0195R00. Neutron Sciences Directorate, Oak Ridge National Laboratory. Google Scholar
Gutmann, M. J. (2017). Nucl. Instrum. Methods Phys. Res. A, 848, 170–173. Web of Science CrossRef Google Scholar
Harrison, S. C., Winkler, F. K., Schutt, C. E. & Durbin, R. M. (1985). Methods Enzymol. 114, 211–237. CrossRef CAS PubMed Google Scholar
Ikeda, S. & Carpenter, J. M. (1985). Nucl. Instrum. Methods Phys. Res. A, 239, 536–544. CrossRef Web of Science Google Scholar
Jogl, G., Wang, X., Mason, S. A., Kovalevsky, A., Mustyakimov, M., Fisher, Z., Hoffman, C., Kratky, C. & Langan, P. (2011). Acta Cryst. D67, 584–591. Web of Science CSD CrossRef CAS IUCr Journals Google Scholar
Kabsch, W. (2006). International Tables for Crystallography, Vol. F, edited by M. G. Rossmann & E. Arnold, pp. 218–225. Chester: International Union of Crystallography. Google Scholar
Kabsch, W. (2010). Acta Cryst. D66, 125–132. Web of Science CrossRef CAS IUCr Journals Google Scholar
Karplus, P. A. & Diederichs, K. (2012). Science, 336, 1030–1033. Web of Science CrossRef CAS PubMed Google Scholar
Kwon, H., Basran, J., Casadei, C. M., Fielding, A. J., Schrader, T. E., Ostermann, A., Devos, J. M., Aller, P., Blakeley, M. P., Moody, P. C. E. & Raven, E. L. (2016). Nature Commun. 7, 13445. Web of Science CrossRef Google Scholar
Langan, P. S., Vandavasi, V. G., Cooper, S. J., Weiss, K. L., Ginell, S. L., Parks, J. M. & Coates, L. (2018). ACS Catal. 8, 2428–2437. CrossRef Google Scholar
Leslie, A. G. W. (2006). Acta Cryst. D62, 48–57. Web of Science CrossRef CAS IUCr Journals Google Scholar
McFeeters, H., Vandavasi, V. G., Weiss, K. L., Coates, L. & McFeeters, R. L. (2016). Acta Cryst. F72, 220–223. Web of Science CrossRef IUCr Journals Google Scholar
Nielsen, K. & Lefmann, K. (2000). Physica B, 283, 426–432. Web of Science CrossRef CAS Google Scholar
O'Dell, W. B., Bodenheimer, A. M. & Meilleur, F. (2016). Arch. Biochem. Biophys. 602, 48–60. Web of Science CAS PubMed Google Scholar
Otwinowski, Z. & Minor, W. (1997). Methods Enzymol. 276, 307–326. CrossRef CAS PubMed Web of Science Google Scholar
Pavese, A. & Artioli, G. (1996). Acta Cryst. A52, 890–897. CrossRef CAS Web of Science IUCr Journals Google Scholar
Pflugrath, J. W. (1999). Acta Cryst. D55, 1718–1725. Web of Science CrossRef CAS IUCr Journals Google Scholar
Riedel, R., Donahue, C., Visscher, T. & Montcalm, C. (2015). Nucl. Instrum. Methods Phys. Res. A, 794, 224–233. Web of Science CrossRef Google Scholar
Rossmann, M. G. (1985). Methods Enzymol. 114, 237–280. CrossRef CAS PubMed Google Scholar
Schaffner, I., Mlynek, G., Flego, N., Pühringer, D., LibisellerEgger, J., Coates, L., Hofbauer, S., Bellei, M., Furtmüller, P. G., Battistuzzi, G., Smulevich, G., DjinovićCarugo, K. & Obinger, C. (2017). ACS Catal. 7, 7962–7976. Web of Science CrossRef PubMed Google Scholar
Schultz, A. J., Jørgensen, M. R. V., Wang, X., Mikkelson, R. L., Mikkelson, D. J., Lynch, V. E., Peterson, P. F., Green, M. L. & Hoffmann, C. M. (2014). J. Appl. Cryst. 47, 915–921. Web of Science CrossRef CAS IUCr Journals Google Scholar
Sjölin, L. & Wlodawer, A. (1981). Acta Cryst. A37, 594–604. CrossRef IUCr Journals Web of Science Google Scholar
Tomanicek, S. J., Blakeley, M. P., Cooper, J., Chen, Y., Afonine, P. V. & Coates, L. (2010). J. Mol. Biol. 396, 1070–1080. Web of Science CrossRef CAS PubMed Google Scholar
Tomoyori, K. & Tamada, T. (2016). J. Phys. Conf. Ser. 762, 012040. CrossRef Google Scholar
Wan, Q., Parks, J. M., Hanson, B. L., Fisher, S. Z., Ostermann, A., Schrader, T. E., Graham, D. E., Coates, L., Langan, P. & Kovalevsky, A. (2015). Proc. Natl Acad. Sci. USA, 112, 12384–12389. Web of Science CrossRef CAS PubMed Google Scholar
Wilkinson, C., Khamis, H. W., Stansfield, R. F. D. & McIntyre, G. J. (1988). J. Appl. Cryst. 21, 471–478. CrossRef Web of Science IUCr Journals Google Scholar
Winter, G., Waterman, D. G., Parkhurst, J. M., Brewster, A. S., Gildea, R. J., Gerstel, M., FuentesMontero, L., Vollmar, M., MichelsClark, T., Young, I. D., Sauter, N. K. & Evans, G. (2018). Acta Cryst. D74, 85–97. Web of Science CrossRef IUCr Journals Google Scholar
Yano, N., Yamada, T., Hosoya, T., Ohhara, T., Tanaka, I. & Kusaka, K. (2016). Sci. Rep. 6, 36628. Web of Science CrossRef PubMed Google Scholar
Zikovsky, J., Peterson, P. F., Wang, X. P., Frost, M. & Hoffmann, C. (2011). J. Appl. Cryst. 44, 418–423. Web of Science CrossRef CAS IUCr Journals Google Scholar
This is an openaccess article distributed under the terms of the Creative Commons Attribution (CCBY) Licence, which permits unrestricted use, distribution, and reproduction in any medium, provided the original authors and source are cited.