research papers
Dynamical scattering in iceembedded proteins in conventional and scanning transmission electron microscopy
^{a}Department of Chemistry and Centre for NanoScience, LudwigMaximiliansUniversity Munich, Butenandtstrasse 11, 81377 Munich, Germany, ^{b}Ernst RuskaCentre for Microscopy and Spectroscopy with Electrons (ERC1), Physics of Nanoscale Systems, Forschungszentrum Jülich, 52425 Jülich, Germany, ^{c}Ernst RuskaCentre for Microscopy and Spectroscopy with Electrons (ERC3), Structural Biology, Forschungszentrum Jülich, 52425 Jülich, Germany, and ^{d}Department Biology, HeinrichHeineUniversität Düsseldorf, 40225 Düsseldorf, Germany
^{*}Correspondence email: k.muellercaspary@cup.lmu.de
of biological macromolecules using cryogenic is based on applying the phase object (PO) assumption and the weak phase object (WPO) approximation to reconstruct the 3D potential density of the molecule. To enhance the understanding of image formation of protein complexes embedded in glasslike ice in a transmission electron microscope, this study addresses multiple scattering in tobacco mosaic virus (TMV) specimens. This includes the propagation inside the molecule while also accounting for the effect of structural noise. The atoms in biological macromolecules are light but are distributed over several nanometres. Commonly, PO and WPO approximations are used in most simulations and reconstruction models. Therefore, dynamical multislice simulations of TMV specimens embedded in glasslike ice were performed based on fully atomistic moleculardynamics simulations. In the first part, the impact of multiple scattering is studied using different numbers of slices. In the second part, different sample thicknesses of the iceembedded TMV are considered in terms of additional ice layers. It is found that singleslice models yield full frequency transfer up to a resolution of 2.5 Å, followed by attenuation up to 1.4 Å. Three slices are sufficient to reach an information transfer up to 1.0 Å. In the third part, ptychographic reconstructions based on (STEM) and singleslice models are compared with conventional TEM simulations. The ptychographic reconstructions do not need the deliberate introduction of aberrations, are capable of postacquisition aberration correction and promise benefits for information transfer, especially at resolutions beyond 1.8 Å.
Keywords: amorphous ice; cryogenic electron microscopy; dynamical scattering; integrative structural biology; molecular dynamics; image simulations.
1. Introduction
Understanding structural details down to the atomic scale is essential to decipher the chemical interactions of protein complexes and ultimately their function. In recent years, conventional ; Yip et al., 2020; Nakane et al., 2020; Zhang et al., 2020; Kato et al., 2019) by employing TEM imaging under parallel illumination conditions with the specimen embedded in glasslike amorphous ice. The direct structural interpretation of individual molecular images, however, remains challenging, primarily due to three aspects. First, only the projection of a rather extended 3D object along the electronbeam direction is observable. Second, CTEM phase contrast emerges from a complex entanglement of the specimen and electronoptical parameters. Third, dynamical scattering and propagation occur within the specimen, hampering the direct inversion of the recorded data to obtain the structure of interest when multiple scattering becomes significant. The first aspect is resolved in contemporary singleparticle cryoCTEM by recording large numbers of particles in different orientations, followed by their 3D alignment and reconstruction (Penczek et al., 1992). Similarly, correcting schemes are employed to account for the contrast transfer function (CTF) governing the second aspect, so as to reverse the distortion effects of the native spatial frequencies of the object in the imaging process, caused mainly by spherical aberration and an intentionally applied defocus to transfer phase to amplitude contrast (Crowther & Böttcher, 1996). Importantly, most of these correction schemes are inherently based on the assumption that the interaction of the electrons with the specimen can be treated within the (weak) phase object [(W)PO] approximation, which implies that dynamical scattering is negligible (Wade, 1992). Indeed, the dominant presence of light elements such as hydrogen, oxygen, carbon, nitrogen and sulfur often justifies the projection assumption in both forward simulations and object retrieval, despite the complexity of biological specimens with 10^{4} to 10^{6} atoms, although limitations of this simplification have been discussed (Hall et al., 2011). Moreover, different continuum and atomistic concepts have been used to include structural noise arising from the glasslike ice embedding (Himes & Grigorieff, 2021; Vulović et al., 2013; Shang & Sigworth, 2012) in forward simulations.
at cryogenic temperatures (cryoCTEM) has reached nearatomic to atomic resolution for proteins in their natural state using singleparticle (Herzik, 2020A detailed and comprehensive treatment of the effect of multiple electron scattering in vitrified proteins via multislice simulations is required to shed light on the ultimate resolution limits, accuracy and precision of contemporary singleparticle cryoCTEM. These simulations will also form the foundation for the applicability of novel acquisition and reconstruction schemes that employ cryoscanning TEM (cryoSTEM). In particular, the feasibility of differential phase contrast (Dekkers & de Lang, 1974) cryoSTEM utilizing segmented ring detectors acquired under infocus conditions has recently been demonstrated for biological specimens leading to nearatomic resolution (Lazić et al., 2022). When employing momentumresolved STEM techniques that use ultrafast cameras to collect a 4D dataset, the obtainable information can still be further extended. In analogy to cryoCTEM simulations, the interpretability of centre of mass (COM) images (Müller et al., 2014), as well as highly doseefficient direct ptychographic phaseretrieval schemes such as single sideband (SSB) (Rodenburg et al., 1993; Yang et al., 2015; O'Leary et al., 2021) and Wigner distribution deconvolution (WDD) (Rodenburg & Bates, 1992) ptychography, is equally governed by the validity of the weak phase approximation (for SSB) or by a multiplicative object (for WDD).
A coarse figure of merit for the breakdown of the projection assumption is provided by the Fresnel number (Vulović et al., 2014), which depends on the targeted resolution, the acceleration voltage and sample thickness. The Fresnel number is well known in light optics and helps to separate the validity of the Fresnel, Fraunhofer and ray optics concepts, respectively. Gureyev et al. (2020) presented a multislice study with biological molecules to compare the contributions of multiple scattering and free space propagation. For small molecules such as the amino acid aspartate (141 Da) and lysozyme (14.3 kDa), the conclusion was drawn that propagation is the dominant factor. However, the chosen proteins were rather small and had been simulated in a vacuum, reducing the scattering power and the likelihood of dynamical scattering as well as neglecting the impact of glasslike ice surrounding the protein.
Baxter et al. (2009) emphasized the importance of including structural noise, e.g. owing to the ice background, for the simulation of realistic images. To account for improved modelling of the surrounding ice, Himes & Grigorieff (2021) used coarse grained water pseudomolecules represented by randomly seeded single isotropic scattering centres of the water molecules with a density of lowdensity amorphous ice. To account for the hydration radius of the protein, they calculated hydration shells in the periphery of the protein derived by the protocol of Shang & Sigworth (2012). Vulović et al. (2013) employed (MD) simulations to model the glasslike ice embedding using empirical potentials.
In the present study, we treated both the protein and the ice fully atomistically. MD simulations were conducted to generate the atomic structure of the ice up to a thickness of 95.5 nm, followed by a hydration procedure to embed the protein. Subsequently, to obtain the electron wavefunction at the exit face of the specimen, dynamical multislice simulations were performed. The main focus of this study is to investigate the effect of dynamical scattering and propagation inside vitrified proteins using the tobacco mosaic virus (TMV) as a well known test sample in cryoEM (Sachse et al., 2007; Fromm et al., 2015; Weis et al., 2019). In particular, the deviations of treating the scattering problem as a singlescattering approach are quantified for cryoCTEM, as well as infocus cryoSTEM techniques such as integrated COM (iCOM) and ptychographic imaging.
2. Methods
2.1. Complete atomic model of specimen
In order to generate a simulation of iceembedded TMV, the most recent atomic model of TMV was retrieved from the Protein Data Bank [PDB code 6sag (Weis et al., 2019)] and helically expanded 147 times to generate an atomic model of three complete helical turns along the y axis. MD simulations were performed to model the glasslike ice background (Thompson et al., 2022) using the foursite rigid water model TIP4P/2005 (Abascal & Vega, 2005), as schematically shown in Fig. 1. The fourth site M is massless and contains the charge associated with the oxygen atom. Starting from a water crystal with a size of 251.1 × 251.1 × 15.5 Å with periodic boundary conditions, the MD simulations were initiated by applying a velocity distribution corresponding to a temperature of 280 K. Subsequently, isochoric and isobaric relaxations were performed for 0.2 and 0.1 ns, respectively, using time steps of 2 fs. This relaxation was followed by a rapid cooling at a rate of −1 K ns^{−1} for 203 ns (Hong & Kim, 2019; Bock & Grubmüller, 2022) and completed by an isobaric relaxation for 0.1 ns.
To combine the atomic model of TMV and the MDbased glasslike ice embedding we accounted for the hydration of the protein periphery by calculating shells with r_{3} = 3.0 Å (Shang & Sigworth, 2012; Himes & Grigorieff, 2021). The resulting probabilities were converted to binary masks using a binomial distribution representing the inner and outer boundary of the hydration shell covering the protein structure. To use the same model for multislice simulations with different slice thicknesses, and to avoid artificially strong scattering at the interface between the slices, we deleted atoms at the boundary when the distance to the next neighbour was smaller than the minimal distance found in a single slice. Fig. 1 shows a schematic setup of the simulation geometry of TMV embedded in glasslike ice. For the first series of simulations, we placed TMV with the minimum possible total specimen thickness of 195 Å. Initially, water molecules are thus only present beside the TMV, not on its top or bottom. Subsequently, Δz_{ice} is increased while the slice thickness Δz = 15 Å is kept constant so as to study the impact of top and bottom ice layers in vitrified specimens on the structural fidelity of the cryoCTEM images.
2.2. Single and multislice simulations
The dynamical interaction between the specimen and the electron wave, including multiple scattering and propagation, was modelled by the multislice algorithm following the theories of Cowley & Moodie (1957, 1959a,b). We briefly recall the central aspects of the multislice theory here to introduce single scattering, then dynamical scattering in general, and to define the terminology. The reader is referred to the comprehensive literature for detailed derivations (e.g., De Graef, 2003; Kirkland, 1998).
The multislice method provides a numerical solution to the relativistic Schrödinger equation by dividing the 3D electrostatic potential of the specimen into slices of thickness Δz. The potentials within the slices are projected along the electronbeam direction and placed in the slice centres, at which the interaction is taken into account. Between the slice centres, the electron wave is propagated along the distance Δz through potentialfree space. For Δz → 0, the multislice algorithm is an accurate solution to the relativistic scattering problem. In this work, isolatedatom scattering factors for electrons as tabulated by Lobato & Van Dyck (2014) were used.
2.2.1. Single scattering
Assuming a specimen thickness Δz and an incident wavefunction ψ_{in}(r), single scattering is formally described by
with being the Fresnel propagator for propagating a distance Δz/2, σ being the relativistic interaction constant, V_{1,Δz} being the projected potential of the first slice, r being a position vector perpendicular to the optical axis and ⊗ being the convolution. ψ_{in} is the incoming plane wave in CTEM and a focused probe in STEM. In practice, the convolution theorem is used, which translates the convolution in real space to a multiplication in Fourier space. There, the Fresnel propagator reads
with the spatial frequency vector k and the Fourier transform . Equation (2) is identical to the effect of defocus, i.e. a phase factor applied in the diffraction plane. The first propagation in CTEM leaves the plane wave unchanged, such that . The exponential term is often termed `phase grating'. Equation (1) consists of imprinting a phase modulation on ψ_{in}, such that it is termed a PO model. In any singleslice model, the interaction is fully described by the multiplication of the incident electron wave with an object transmission function (OTF).
If, additionally, the phase ϕ(r) = σV_{1,Δz}(r) is small, Taylor expansion of the phase grating yields 1 + iσV_{1,Δz}(r), known as the WPO model. Only in this case, the image intensity can be given analytically in closed form. The transfer of spatial frequencies is best studied in Fourier space and the Fourier transformed image intensity I of a WPO is
where A(k) is a potentially present objective aperture, k_{S} and k_{C} are spatial and temporal coherence envelopes, χ(k) is the aberration function, and δ(k) is Dirac's delta function. From this equation, the established Scherzer conditions for the defocus and the aperture can be derived to maximize contrast transfer within the Scherzer passband for spatial frequencies below the point resolution limit of the microscope.
2.2.2. Dynamical scattering
In the following, we refer to the term `dynamical scattering' as the consecutive action of phase grating and Fresnel propagation. To simulate dynamical scattering, the multislice scheme yields for the wavefunction ψ^{(n)} after interacting with the nth slice:
The exit wave results from a final propagation by Δz/2. According to the iterative application of equation (4), multiple scattering and propagation cannot be separated anymore for a given exit wave ψ^{exit}(r).
The small lattice vibrations at liquid nitrogen temperature were accounted for by multiplying the atomic scattering amplitudes with an isotropic Debye–Waller factor according to a Debye parameter of B = 0.3 Å^{2}. For those simulations where the surrounding glasslike ice was to be approximated by its mean inner potential (MIP) (e.g. Himes & Grigorieff, 2021), we used an artificially high Debye–Waller factor of 1809 Å^{2} for the ice regions, being five times the distance of the bond length between oxygen and hydrogen in the TIP4P/2005 water model. As a result, the potential of the ice becomes practically constant within the embedding, but without introducing abrupt potential steps where the hydration shell of the protein was applied.
2.3. Conventional image simulations
The CTEM simulations were performed using the exit waves formed under plane wave illumination. Aberrations, as well as partial spatial (k_{S}) and temporal coherence (k_{C}), were taken into account within the framework of linear imaging theory by multiplication of the diffraction pattern wavefunction with a phase defined by the aberration function χ and by partial coherence envelopes [k_{S}k_{C}exp(iχ)]. Additionally, an objective aperture with a radius of 20 mrad was applied if not stated otherwise. An inverse Fourier transform of equation (4) yields the image wavefunction, whose squared modulus corresponds to the recorded image. The simulation parameters can be found in Table 1. Dose considerations as well as detection performance are not included in the current CTEM image simulations.

2.4. simulations
For STEM, multislice simulations with a focused probe and 24 slices were performed, and the diffraction patterns for each scan pixel have been stored, resembling a recording with a pixelated STEM detector. This 4D dataset was used to calculate the centre of mass (Müller et al., 2014), from which the iCOM signal was obtained by an integration in Fourier space (Section S3 of the supporting information) (Lazić et al., 2016), the SSB reconstruction (Rodenburg et al., 1993) and the WDD (Rodenburg & Bates, 1992; McCallum & Rodenburg, 1992). Both iCOM and WDD and SSB yield the projected electrostatic potential of the specimen in PO and WPO approximation, respectively. WDD will only be discussed very briefly because, for lowdose applications, SSB is advantageous compared with WDD since no deconvolution is needed (O'Leary et al., 2021). Importantly, these methods allow one to reconstruct frequencies up to twice the radius of the probeforming aperture (Rodenburg et al., 1993). Therefore, we used 10 mrad for the probeforming aperture, being half as large as the CTEM counterpart. Because SSB and iCOM show the highest contrast in the absence of aberrations, the defocus was chosen as C_{1} = −97.5 Å, being the centre of the specimen, and no other aberrations were used. It turned out that 24 slices were by far sufficient and were chosen as the maximum number of slices in the STEM simulations.
For iCOM, diffracted intensities up to 1.4 times the bright field disc were used. To deconvolve the probe and σV, an aberrationfree probe was assumed, which only has consequences for the comparison in real space because it is a radially symmetric rescaling in Fourier space and ignored in the Fourier ring correlation (FRC) curve introduced below. For SSB, the complex value of each pixel was calculated from the average of the doubleoverlap regions in the 4D dataset, Fourier transformed with respect to the scan position, resulting in a constant transfer function up to a maximum spatial frequency equal to twice the aperture radius (O'Leary et al., 2021).
2.5. Comparison in real and Fourier space
We used two concepts to compare the reconstructions of the proteins for different slicings and imaging methods. An assessment of the impact of dynamical scattering in CTEM is best performed on the basis of the complex specimen exit wavefunctions simulated with different numbers of slices. This poses the challenge of introducing an error metric for a given complex exit wave regarding modulus and phase with respect to the ground truth; here a simulation with 120 slices. To this end, the complex normalized difference is used for comparison in real space and FRC is used in
The complex difference for two matrices is calculated bywhere X′ = X · exp(iϕ), with ϕ ∈ (0, 2π), which contains a constant phase offset such that is minimal in the glasslike ice region, with being the Frobenius norm of X. The division and the calculation of the absolute value are performed pointwise. The complex difference is invariant against uniform rescaling by a real factor a > 0, d(X, Y) = d(aX, aY).
The FRC curve measures the normalized crosscorrelation between resolution rings of two matrices (X and Y) in Fourier space. It is given by (van Heel et al., 1982; Saxton & Baumeister, 1982; van Heel & Schatz, 2005)
where denotes the complex conjugate of Y and Δk denotes the width of the resolution shell, which was chosen as the smallest present spatial frequency determined by the size of the field of view. A consequence of the normalization is that, except for the sign, the FRC is invariant against uniform rescaling with an isotropic function f(k) = f(k) in Fourier space.
3. Results
In order to understand the effects of dynamical scattering, we use a simulation with 120 slices as the ground truth. When only single scattering is considered, two quantities are of interest. Firstly, the OTF itself, whose phase or imaginary part are directly proportional to the projected potential of the specimen in PO and WPO approximation, respectively. These OTFs represent the electron wavefunction in the slice centre. However, multislice simulations yield the electron wavefunction at the exit face of the specimen. Secondly, we thus Fresnel propagated the PO and WPO results from the slice centre to the exit face and termed the results as PO_{exit} and WPO_{exit}, respectively. Initially, the impact of dynamical scattering is mainly studied at the level of exit wavefunctions, then at the basis of the image formed by the optical system, and different ice thicknesses are considered. Finally, a comparison between the performance of CTEM and recent phasecontrast STEM techniques is elaborated.
3.1. Dynamical scattering effects on exit waves under parallel illumination
The comparison of the phases of the OTF as well as the phases of the exit waves using single and 120slice interaction models are shown in Fig. 2. Figs. 2(a)–2(e) show the phases of the wavefunctions using different interaction models and Figs. 2(f)–2(i) depict the corresponding moduli. Interestingly, the phases of the PO and WPO approximation in Figs. 2(a) and 2(b), respectively, resemble the 120slice simulation best, indicating that a propagation to the exit face of the specimen leads to less accurate phase distributions of PO_{exit} and WPO_{exit} in Figs. 2(c) and 2(d), respectively. Nevertheless, all phase results show the glasslike ice on the left side and the typical ladderlike structure of TMV at the right side limited by the central TMV symmetry axis. A comparison of the moduli with the ground truth in Fig. 2(i) reveals systematic differences for both WPO approaches in Figs. 2(f) and 2(h), and a similar pattern is shown for PO_{exit} in Fig. 2(g). The of the modulus of the 120slice simulation ranges from 0.34 to 1.65 and is significant, suggesting that one should not conclude here that TMV is a PO. Clearly, the spatial information about the virus assembly is dominantly contained in the phase. In essence, this first overview means that none of the OTF models, defocused to the exit plane or not, yield suitable visual agreement with the multislice ground truth for phase and amplitude simultaneously. In that respect, PO_{exit} exhibits qualitative agreement in the phase and modulus for low and medium spatial frequencies, whereas high frequencies arise from Fresnel propagation by half the specimen thickness.
The obtained simulations raise two questions for further investigation. First, how many slices are actually needed to take into account dynamical scattering effects reliably? Second, do the observed differences between the single and multislice models have any impact on the results of singleparticle cryoCTEM?
First, the complex normalized realspace differences of exit waves for several models with respect to the 120slice simulation are depicted colourcoded in Fig. 3. Large errors for the moduli are found for singleslice OTF approaches according to Figs. 3(a)–3(d), decreasing by a factor of 2–4 for 2 slices and by a factor of 6–14 for 20 slices in Figs. 3(e) and 3(f), respectively. Interestingly, the PO approaches in Fig. 3(b) do not show distinct colour features distinguishing ice and TMV. Recalling that the difference to the ground truth is plotted here, this means that the PO approach yields a phase of a wave in the specimen centre that is in very good agreement with a full dynamical simulation, at least by visual inspection. This notion is also supported by the similarity of the PO phase with that of the 120slice simulation in Fig. 2. Larger and more systematic phase differences for the PO_{exit} and WPO_{exit} simulations in Figs. 3(c) and 3(d), i.e. the Fresnel propagation of the OTF from the slice centre to its exit face, are observed. At least for the phase, a closer match to the ground truth is found at the slice centre as opposed to the exit wave. In conclusion, the phase error decreases with the number of slices and becomes negligible at 20 slices. As the described properties represent realspace data of Figs. 2 and 3, they cannot be easily interpreted with respect to attainable resolution. Therefore, we turn to resolutionbased correlations included in FRC curves.
A comparison of the exit waves in Fourier space is provided in Fig. 4 for singleslice OTFbased models and multislice simulations with increasing number of slices (2, 3, 6 and 20). Considering an exit wave simulated with 120 slices as the ground truth again, with an increasing number of slices the FRCs gradually approach 1 across the full spatial frequency band up to 1 Å resolution. In fact, already three slices are sufficient to attain an FRC threshold of 0.4 at 1 Å^{−1}. Taking only two slices into account yields an FRC of 0 at 1 Å^{−1}, and still 0.25 at k = 0.9 Å^{−1}, corresponding to distances of 1.1 Å in the specimen. As the singleslice simulation is typically used for cryoCTEM, it warrants more careful interpretation. None of the (W)PO FRCs show a significant correct spatial frequency transfer beyond 0.7 Å^{−1}. From the FRC comparison considering the complexvalued waves, the propagation to the exit face of the specimen is favoured over the central slices, as the WPO and PO curves in Fig. 4 have nulls at smaller spatial frequencies than their PO_{exit} and WPO_{exit} counterparts. Among the exitwave approaches, the WPO_{exit} is the slightly better model. An FRC analysis separated into modulus and phase shows the same trend and is shown in Fig. S1 of the supporting information.
One might argue critically that the similarity of the wavefunction to the OTFs is the ultimate criterion, since the retrieval of the structure in terms of the single projection might be seen as the ultimate goal. In that case, the FRC of all simulations could have been performed with respect to the PO as ground truth, after propagating back to the centre of the specimen. Even in practice, one could move the objectivelens focus to this location; and indeed, the oscillatory behaviour of the FRC(PO) in Fig. 4 originates from the fact that two functions in different planes along the optical axis are compared here, introducing effects of the CTFs with different foci. In any case, one can revert this argumentation to the equivalent statement that if the PO approach was sufficient in the sense that the 120slice simulation yielded the same wavefunction in the slice centre after Fresnel propagation back to this plane then the forward propagation of the PO to the exit face of the specimen, PO_{exit} in Fig. 4, would yield the same result as the 120slice simulation. This is obviously not the case for higher spatial frequencies. Therefore, we conclude that a fewslice approach with slice thicknesses below ∼6 nm is needed to describe the interaction between specimen and incident electrons reliably.
The singleslice projection into an OTF can be considered physically relevant only in a limited manner, in contrast to the wavefunction created by dynamical scattering, as obtained by multislice simulation with a converged number of slices. This wavefunction must be considered to reflect the wave present in the object plane of the objective lens in experiments and therefore needs to be chosen as the ground truth. A significant deviation between OTFs and multislice simulation suggests using reconstruction techniques that either account for multiple scattering, introduce prior knowledge or do both.
3.2. Dynamical scattering in CTEM imaging
To simulate the experimentally observable CTEM image intensities, the WPO_{exit}, PO_{exit} and 120slice exit waves from Section 3.1 were subjected to the imaging process with a defocus of −500 nm and a spherical aberration of 1.5 mm. The results are compiled in Fig. 5, where the left column represents the image intensities as simulated. The right column depicts the result after correcting the sign of the CTF, also known as phase flipping, to mitigate contrast inversions, assuming equation (3) to be valid.
All three simulations show the outline of the helical assembly of TMV including features of the simulated glasslike ice matching the characteristic appearance of experimental cryoEM images. However, the WPO_{exit} case does not show internal features, in contrast to the PO_{exit} and 120slice cases, which are both very similar and agree in fine structure. The main difference between the PO_{exit} and 120slice cases is seen in the absolute contrast and the contrast between TMV and the glasslike ice, both being smaller in magnitude for 120 slices. The same effect is observed in the phaseflipped images on the right.
According to Fig. 2, the structural information of TMV is dominantly encoded in the phase of the exit wave, meaning that the recorded image should ideally be similar to Fig. 2(e). Therefore, Fig. 6(a) depicts the FRCs between this signal and Figs. 5(a)–5(c). On the other hand, the image intensity is the experimentally observable quantity, such that FRCs taking the image corresponding to the 120slice simulation as reference are shown in Fig. 6(b). As expected due to the focus setting, an oscillating CTF can be seen in both comparisons, while the singleslice models (W)PO_{exit} exhibit an additional beating of the envelope in Fig. 6(a). Interestingly, the null of the envelope at 0.55 Å^{−1} is in good agreement with the dashed black curve representing the Fouriertransformed Fresnel propagator for half the specimen thickness (Δz_{TMV}). From 0.7 Å^{−1}, the singleslice FRCs transit to noise. For two slices, the null of the envelope shifts to ∼0.7 Å^{−1}, in agreement with Fresnel propagation across 0.25Δz_{TMV}. Consequently, gradually replacing the dynamical interaction by freespace propagation reduces the consistency between the true exitwave phase and the CTEM intensity predicted by the respective models in terms of the envelope of the oscillating CTF.
Concerning Fig. 6(b), a slightly better agreement between all models is observed on the level of image intensities, which is due to the fact that the reference was subjected to an imaging process with the same envelopes and aberrations. Nevertheless, the (W)PO_{exit} cases yield suitable agreement only up to a spatial frequency of 0.7 Å^{−1}. To represent spatial frequencies up to 0.9 Å^{−1} correctly, at least two slices are needed, and 20 slices yield already perfect frequency transfer. Treating image formation using singleslice models is, therefore, only accurate up to a certain spatial frequency, corresponding to periodicities of ∼1.4 Å in the present case of TMV.
3.3. Atomistic and continuum modelling of ice embedding
The exit wave of the MIP approximation obtained by a 120slice simulation is compared with its fully atomistic counterpart from MDsimulated ice in Fig. 7. The modulus and phase of the MIP approximation show no contrast in the pure ice regions in Figs. 7(a) and 7(b), opposite to the amorphous structure in MD simulations, see e.g. Figs. 2(e) and 2(i). The complex realspace difference in Fig. 7(c) shows only structural noise and no contrast related to TMV. Structural noise due to the ice deteriorates the contrast between the protein and the embedding significantly, which may impact particle recognition. To model this transition realistically, MD simulations are thus necessary. Fig. 7(d) shows the FRCs of the fully atomistic result with (i) TMV in a vacuum, (ii) TMV with ice approximated by its MIP and (iii) solely amorphous ice from MD with empty space replacing TMV. A common feature of the computed FRCs is that values above 0.7 are rarely taken in the spatial frequency range above 0.2 Å^{−1}, suggesting that a continuum model does not match a fully atomistic treatment over a large range of the spatial frequency spectrum. Furthermore, the FRCs show that using a vacuum or the MIP for amorphous ice yield the same results because the effect of a constant potential offset leads to a rather small global modification of the electron wavelength in the specimen.
3.4. Influence of ice thickness in cryoCTEM imaging
Fig. 8 shows both the specimen exitwave phases and the CTEM image intensities for different specimen thicknesses. In particular, the ideal case with no ice embedding on top or below the TMV resulting in 19.5 nm thickness is depicted in Figs. 8(a) and 8(c), whereas Figs. 8(b) and 8(d) show the maximal case of 94.5 nm thickness with two additional ice layers of Δz_{ice} = 37.5 nm thickness encapsulating the TMV. Both results were obtained by fully atomistic MD. In both cases, the contrast is dramatically reduced for the thicker ice, concerning the internal structure of the TMV, as well as the contrast between the TMV and the adjacent ice embedding.
To mimic the averaging process for 2D class averages in singleparticle cryoEM, a total of 500 simulated CTEM images were added for thin, intermediate and thick specimens (34.5, 64.5 and 94.5 nm, respectively). Contrast inversions were accounted for by phase flipping according to the respective CTFs, again assuming equation (3) to be valid. In analogy to Fig. 6(a), FRCs with the exitwave phase for the minimal specimen thickness of 19.5 nm from Fig. 8(a) have been calculated and plotted in Fig. 8(e). The same curves can be observed when the averaged CTEM simulations are compared with the averaged CTEM simulation with no additional ice on top and bottom, as seen in Fig. S4. As expected from the visual impression in Figs. 8(c) and 8(d), increasing ice thickness leads to higher noise, and decreasing FRCs over the whole range of spatial frequencies are observed. In particular, FRCs drop linearly up to a frequency of 0.4 Å^{−1} and show a plateau with some modulations above. Interestingly, a threefold increase of the ice layer from 34.5 to 94.5 nm only leads to an FRC drop from ∼0.75 to 0.5. Nearly 40 nm of ice on both top and bottom face causes a marginal further FRC decrease by 0.1. To conclude, image quality deteriorates when specimen thickness increases, but more quickly as soon as ice layers of a few nanometres are present, whereas the additional effect imposed by growing ice with tens of nanometres thickness is noticeably smaller.
3.5. Comparison of cryoCTEM and phasecontrast cryoSTEM
The above analysis has shown several limitations of singleslice models. Nevertheless, it was also found that the (W)PO approaches perform accurately up to a resolution of 1.4 Å. These findings suggest to explore contemporary 4D STEM techniques employing the projection approximation, typically applied under infocus conditions and revealing a monotonous transfer. Fig. 9 compiles the results of STEM iCOM imaging and SSB ptychography of the 19.5 nm thick TMV specimen, together with a −500 nm defocused CTEM case for which we assumed an aberrationcorrected microscope without spherical aberration. Due to the inherent double resolution of the STEM techniques, the objective aperture in CTEM was chosen as twice as large to allow for a fair comparison. Figs. 9(a)–9(c) show that the STEM techniques yield striking contrast without aberrations and defocus. The power spectra in Figs. 9(d)–9(f) all contain the primary and secondary layer lines of TMV, despite the modulation by the Thon rings in the CTEM case in Fig. 9(d). Let us assume that, due to the inherent singleslice character of the STEM techniques considered here, one aims at measuring the complete projected structure of TMV according to the Coulomb potential. Therefore, Fig. 9(g) depicts the FRCs with the projected potential according to the PO approximation. The smooth and quasiidentical frequency transfers of iCOM, SSB and WDD reveal high structural fidelity of the reconstructed signals across the whole frequency range, with an FRC of 0.5 at 0.75 Å^{−1} and 0.25 at 1 Å^{−1}. The defocusrelated oscillation of the CTF for CTEM, however, is bound by an envelope with a null at 0.7 Å^{−1}, suggesting that a larger number of CTEM images with different foci over STEM images is needed to retrieve, in particular, high spatial frequencies faithfully. In this context, phasecontrast STEM appears promising, although a robust and comprehensive comparison with cryoCTEM will need to account for dose efficiency and noisetransfer characteristics in a future study.
4. Discussion
From a very general viewpoint, dynamical scattering impacts the direct interpretability of TEM data twofold. Firstly, the electron wave exiting the specimen differs significantly from the projected structure. Secondly, substantial multiple scattering violates the WPO approximation and thereby the availability of an analytical expression for the recorded image intensity, which is the prerequisite for a direct CTF correction scheme. With the structural fidelity of the exit wave and image intensity being the pivotal question in quantitative CTEM, the impact of dynamical scattering has naturally been addressed in cryoCTEM literature before (Vulović et al., 2014; Gureyev et al., 2020; Himes & Grigorieff, 2021). The outcome, however, has been largely focused on including the principal theory and simulation, but to a smaller extent on the impact on the resulting information transfer in singleparticle in comparison with the stateoftheart WPO approximation. In that respect, the present study sheds light on the quantitative analysis of how many slices are needed to obtain a certain accuracy in describing interaction and imaging, on the necessity of the MDbased atomistic modelling of the ice embedding, on the impact of various ice thicknesses on the specimen in the electronbeam direction, and on the perspective of recent cryoSTEM techniques for protein structure determination.
For TMV, a large protein assembly of high molecular weight, one can clearly draw the conclusion that singleslice approximations are an inaccurate way to describe the scattering in the specimen in comparison with multislice simulations, as was demonstrated at the level of specimen exit waves. However, we also found that this interpretation concerns mostly high spatial frequencies, since the WPO models show partly acceptable FRCs with a full multislice result up to 0.6 Å^{−1}. Beyond, at least 2–3slice interaction models are required to achieve 1 Å^{−1}.
In accordance with the exitwave analysis, CTEM images simulated with singleslice approaches differ substantially at high spatial frequencies from the image formed by the wavefunction from full dynamical scattering imaged with the same optical parameters (Fig. 6). In other words, interpreting highresolution features within recorded images in terms of singleslice models can be inaccurate, which includes the commonly applied WPObased CTF correction. At first sight, the observed relationship points to an inherent limit of the achievable spatial resolution independent of the quality of the imaging system used. Interestingly, to date, TMV structures have been experimentally resolved up to 1.9 Å (Weis et al., 2019), worse than for example apoferritin at 1.2 Å (Nakane et al., 2020; Yip et al., 2020). Given the wider TMV diameter of 18 nm and the high MIP in comparison with 13 nm apoferritin, multiple scattering will affect TMV rods more strongly than smaller and hollow apoferritin cages.
In addition to using at least a fewslice model for forward simulations and image interpretation, a clear recommendation can be made as to the use of continuum models and the impact of ice layers on the specimen surfaces. In fact, we here show that structural noise can be simulated atomistically, e.g. using amorphous ice including protein hydration obtained by MD. The obtained structural noise gives rise to a similar appearance in experimental images and deteriorates the TMV signal over a large range of frequencies (Fig. 7). Assuming a MIP offset as an embedding results in no improvement compared with considering vacuum. The current model does not include any beaminduced motion of the protein and water molecules during the exposure that have been characterized experimentally (Brilot et al., 2012; McMullan et al., 2015). This study does intentionally not deal with dose limitations or the recording process, both of which pose additional constraints to cryoEM image quality. The scattering physics described here is doseindependent, leaving dose and detection characteristics as a vital topic for future studies. Furthermore, lowangle scattering can be sensitive to electron redistributions due to chemical bonding (Madsen et al., 2021; Müller et al., 2010). This study exploited isolated atom scattering factors, which is sufficient to study the impact of dynamical scattering. Nevertheless, exploring the sensitivity of experimental cryo(S)TEM data to valencies and ionicities (Kirkland, 1998), e.g. via density functional theory simulations, would be highly interesting and for vitrified specimens.
It is commonly accepted that specimen thickness critically impacts the image quality in cryoEM and minimal ice embedding of the protein leads to the highest contrast images suitable for optimal highresolution image reconstructions (Nakane et al., 2020; Yip et al., 2020). In addition, our simulations show that image deterioration by additional ice layers is, however, nonlinearly dependent on ice thickness (Fig. 8). This observation for suggests that thick ice coverage reduces the FRC over a minimally embedded protein much less than expected after adding the first thin ice layers. However, contrast is also affected by (Angert et al., 1996), which needs to be considered for recordings without zeroenergyloss filter or correction of chromatic aberration. Regardless, the simulation results emphasize the requirement of experimentally controlling and minimizing ice thickness for obtaining optimal protein contrast.
As multislice approximations are currently not included in cryoEM et al., 1990). Such prior information may mitigate the violation of projection assumptions when generating atomic models. In cases when multiple scattering becomes relevant, it is possible to include multislice simulations of the 3D density in the iterative projection matching workflow, which could increase the precision orientation parameter determination for each experimental particle image. Moreover, by improving the simulations of CTEM images of atomic models through multislice approximations, it may ultimately become realisable to include the simulation based on the atomic protein model in the orientationdetermination iterative optimization until the atomic model matches the experimental images. The maximization of agreement between experimental diffraction data and the atomic protein model is commonly being performed in Xray protein crystallography (Brünger et al., 1998).
workflows, the computed 3D density maps will result in lowerresolution maps as well as other systematic density inaccuracies. When atomic models are built, chemical building blocks of amino acids are commonly placed in the obtained isosurface Coulomb potential (HendersonFormerly predicted (Rose, 1976) and recently implemented (Lazić et al., 2022) STEM phasecontrast imaging motivated the comparison of STEM techniques with cryoCTEM. Although iCOM, SSB and WDD ptychography involve the projection assumption, the FRC extends up to 1 Å^{−1} and is smooth across the whole frequency range. Note that the STEM FRCs still reach a value of 0.5 at 0.7 Å^{−1}, where the FRC between planewave PO and the multislice result is already 0 (Fig. 9). In general, the monotony of iCOM, SSB and WDD will be advantageous when small numbers of particles are studied, as the CTF correction in CTEM is an additional processing step enhancing noise that can only be overcome by averaging larger numbers of molecular views. While it was assumed here that all CTF oscillations can be corrected perfectly, a final comparison between CTEM and STEM needs to account for the detection stage, namely Poissonian noise and modulation transfer functions of the cameras. In this respect, SSB ptychography has performed highly efficient recently (O'Leary et al., 2021). Conceptually, 4D STEM possesses the advantage of allowing an enormous flexibility for data processing, with current developments going far beyond direct inversion schemes and singleslice models. In particular, ptychographic iterative engine algorithms perform very well with regards to spatial resolution (Maiden et al., 2009) and developments for inverse multislice ptychography (Bangun et al., 2022; Chen et al., 2021; Maiden et al., 2012) provide solutions beyond the projection assumption. Overcoming the (W)PO limitations to CTEM by 2–3slice phase retrieval in 4D STEM is thus a promising way to explore the opportunity of resolution and contrast enhancement in future cryoEM.
5. Data availability statement
The data that support the findings of this study are available from the corresponding author upon reasonable request.
Supporting information
Supporting information. DOI: https://doi.org/10.1107/S2052252523004505/rq5009sup1.pdf
Funding information
KMC acknowledges funding from the Deutsche Forschungsgemeinschaft under grant number EXC 2089/1390776260 (Germany's Excellence Strategy). KMC and MLL also acknowledge support from the Helmholtz Association (Germany) under contract VHNG1317.
References
Abascal, J. L. & Vega, C. (2005). J. Chem. Phys. 123, 234505. Web of Science CrossRef PubMed Google Scholar
Angert, I., Burmester, C., Dinges, C., Rose, H. & Schröder, R. (1996). Ultramicroscopy, 63, 181–192. CrossRef CAS Web of Science Google Scholar
Bangun, A., Melnyk, O., März, B., Diederichs, B., Clausen, A., Weber, D., Filbir, F. & MüllerCaspary, K. (2022). IEEE Trans. Comput. Imaging, 8, 996–1011. Web of Science CrossRef Google Scholar
Baxter, W. T., Grassucci, R. A., Gao, H. & Frank, J. (2009). J. Struct. Biol. 166, 126–132. Web of Science CrossRef PubMed CAS Google Scholar
Bock, L. V. & Grubmüller, H. (2022). Nat. Commun. 13, 1709. Google Scholar
Brilot, A. F., Chen, J. Z., Cheng, A., Pan, J., Harrison, S. C., Potter, C. S., Carragher, B., Henderson, R. & Grigorieff, N. (2012). J. Struct. Biol. 177, 630–637. Web of Science CrossRef CAS PubMed Google Scholar
Brünger, A. T., Adams, P. D., Clore, G. M., DeLano, W. L., Gros, P., GrosseKunstleve, R. W., Jiang, J.S., Kuszewski, J., Nilges, M., Pannu, N. S., Read, R. J., Rice, L. M., Simonson, T. & Warren, G. L. (1998). Acta Cryst. D54, 905–921. Web of Science CrossRef IUCr Journals Google Scholar
Chen, Z., Jiang, Y., Shao, Y.T., Holtz, M. E., Odstrčil, M., GuizarSicairos, M., Hanke, I., Ganschow, S., Schlom, D. G. & Muller, D. A. (2021). Science, 372, 826–831. Web of Science CrossRef CAS PubMed Google Scholar
Cowley, J. M. & Moodie, A. F. (1957). Acta Cryst. 10, 609–619. CrossRef IUCr Journals Web of Science Google Scholar
Cowley, J. M. & Moodie, A. F. (1959a). Acta Cryst. 12, 353–359. CrossRef CAS IUCr Journals Web of Science Google Scholar
Cowley, J. M. & Moodie, A. F. (1959b). Acta Cryst. 12, 360–367. CrossRef CAS IUCr Journals Web of Science Google Scholar
Crowther, R. & Böttcher, B. (1996). Structure, 4, 387–394. PubMed Web of Science Google Scholar
De Graef, M. (2003). Introduction to Conventional Transmission Electron Microscopy. Cambridge University Press. Google Scholar
Dekkers, N. H. & de Lang, H. (1974). Optik, 41, 452–456. Google Scholar
Fromm, S. A., Bharat, T. A., Jakobi, A. J., Hagen, W. J. & Sachse, C. (2015). J. Struct. Biol. 189, 87–97. Web of Science CrossRef PubMed Google Scholar
Gureyev, T., Quiney, H., Kozlov, A. & Allen, L. (2020). Ultramicroscopy, 218, 113094. Web of Science CrossRef PubMed Google Scholar
Hall, R., Nogales, E. & Glaeser, R. (2011). J. Struct. Biol. 174, 468–475. Web of Science CrossRef CAS PubMed Google Scholar
Heel, M. van, Keegstra, W., Schutter, W. & van Bruggen, E. (1982). Life Chem. Rep. 5, Suppl. 1, 69–73. Google Scholar
Heel, M. van & Schatz, M. (2005). J. Struct. Biol. 151, 250–262. Web of Science PubMed Google Scholar
Henderson, R., Baldwin, J., Ceska, T., Zemlin, F., Beckmann, E. & Downing, K. H. (1990). J. Mol. Biol. 213, 899–929. CrossRef CAS PubMed Web of Science Google Scholar
Herzik, M. A. Jr (2020). Nature, 587, 39–40. Web of Science CAS PubMed Google Scholar
Himes, B. & Grigorieff, N. (2021). IUCrJ, 8, 943–953. Web of Science CrossRef CAS PubMed IUCr Journals Google Scholar
Hong, S. & Kim, D. (2019). J. Phys. Condens. Matter, 31, 455403. Web of Science CrossRef PubMed Google Scholar
Kato, T., Makino, F., Nakane, T., Terahara, N., Kaneko, T., Shimizu, Y., Motoki, S., Ishikawa, I., Yonekura, K. & Namba, K. (2019). Microsc. Microanal. 25, 998–999. CrossRef PubMed Google Scholar
Kirkland, E. J. (1998). Advanced Computing in Electron Microscopy. Springer. Google Scholar
Lazić, I., Bosch, E. G. & Lazar, S. (2016). Ultramicroscopy, 160, 265–280. Web of Science PubMed Google Scholar
Lazić, I., Wirix, M., Leidl, M. L., de Haas, F., Mann, D., Beckers, M., Pechnikova, E. V., MüllerCaspary, K., Egoavil, R., Bosch, E. G. T. & Sachse, C. (2022). Nat. Methods, 19, 1126–1136. Web of Science PubMed Google Scholar
Lobato, I. & Van Dyck, D. (2014). Acta Cryst. A70, 636–649. Web of Science CrossRef IUCr Journals Google Scholar
Madsen, J., Pennycook, T. J. & Susi, T. (2021). Ultramicroscopy, 231, 113253. Web of Science CrossRef PubMed Google Scholar
Maiden, A., Humphry, M. & Rodenburg, J. (2012). J. Opt. Soc. Am. A, 29, 1606–1614. Web of Science CrossRef CAS Google Scholar
Maiden, A. & Rodenburg, J. (2009). Ultramicroscopy, 109, 1256–1262. Web of Science CrossRef PubMed CAS Google Scholar
McCallum, B. C. & Rodenburg, J. M. (1992). Ultramicroscopy, 45, 371–380. CrossRef Web of Science Google Scholar
McMullan, G., Vinothkumar, K. & Henderson, R. (2015). Ultramicroscopy, 158, 26–32. Web of Science CAS PubMed Google Scholar
Müller, K., Krause, F. F., Béché, A., Schowalter, M., Galioit, V., Löffler, S., Verbeeck, J., Zweck, J., Schattschneider, P. & Rosenauer, A. (2014). Nat. Commun. 5, 5653. Google Scholar
Müller, K., Schowalter, M., Rosenauer, A., Rubel, O. & Volz, K. (2010). Phys. Rev. B, 81, 075315. Google Scholar
Nakane, T., Kotecha, A., Sente, A., McMullan, G., Masiulis, S., Brown, P. M., Grigoras, I. T., Malinauskaite, L., Malinauskas, T., Miehling, J., Uchański, T., Yu, L., Karia, D., Pechnikova, E. V., de Jong, E., Keizer, J., Bischoff, M., McCormack, J., Tiemeijer, P., Hardwick, S. W., Chirgadze, D. Y., Murshudov, G., Aricescu, A. R. & Scheres, S. H. W. (2020). Nature, 587, 152–156. Web of Science CrossRef CAS PubMed Google Scholar
O'Leary, C. M., Martinez, G. T., Liberti, E., Humphry, M. J., Kirkland, A. I. & Nellist, P. D. (2021). Ultramicroscopy, 221, 113189. Web of Science PubMed Google Scholar
Penczek, P., Radermacher, M. & Frank, J. (1992). Ultramicroscopy, 40, 33–53. CrossRef PubMed CAS Web of Science Google Scholar
Rodenburg, J., McCallum, B. & Nellist, P. (1993). Ultramicroscopy, 48, 304–314. CrossRef Web of Science Google Scholar
Rodenburg, J. M. & Bates, R. H. T. (1992). Philos. Trans. Sect. A, 339, 521–553. Google Scholar
Rose, H. (1976). Ultramicroscopy, 2, 251–267. CrossRef Google Scholar
Sachse, C., Chen, J. Z., Coureux, P.D., Stroupe, M. E., Fändrich, M. & Grigorieff, N. (2007). J. Mol. Biol. 371, 812–835. Web of Science CrossRef PubMed CAS Google Scholar
Saxton, W. & Baumeister, W. (1982). J. Microsc. 127, 127–138. CrossRef CAS PubMed Web of Science Google Scholar
Shang, Z. & Sigworth, F. J. (2012). J. Struct. Biol. 180, 10–16. Web of Science CrossRef CAS PubMed Google Scholar
Thompson, A. P., Aktulga, H. M., Berger, R., Bolintineanu, D. S., Brown, W. M., Crozier, P. S., in 't Veld, P. J., Kohlmeyer, A., Moore, S. G., Nguyen, T. D., Shan, R., Stevens, M. J., Tranchida, J., Trott, C. & Plimpton, S. J. (2022). Comput. Phys. Commun. 271, 108171. Web of Science CrossRef Google Scholar
Vulović, M., Ravelli, R. B., van Vliet, L. J., Koster, A. J., Lazić, I., Lücken, U., Rullgård, H., Öktem, O. & Rieger, B. (2013). J. Struct. Biol. 183, 19–32. Web of Science PubMed Google Scholar
Vulović, M., Voortman, L. M., van Vliet, L. J. & Rieger, B. (2014). Ultramicroscopy, 136, 61–66. Web of Science PubMed Google Scholar
Wade, R. (1992). Ultramicroscopy, 46, 145–156. CrossRef CAS Web of Science Google Scholar
Weis, F., Beckers, M., von der Hocht, I. & Sachse, C. (2019). EMBO Rep. 20, e48451. Google Scholar
Yang, H., Pennycook, T. J. & Nellist, P. D. (2015). Ultramicroscopy, 151, 232–239. Web of Science CrossRef CAS PubMed Google Scholar
Yip, K. M., Fischer, N., Paknia, E., Chari, A. & Stark, H. (2020). Nature, 587, 157–161. Web of Science CrossRef CAS PubMed Google Scholar
Zhang, K., Pintilie, G. D., Li, S., Schmid, M. F. & Chiu, W. (2020). Cell Res. 30, 1136–1139. Web of Science CrossRef PubMed 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.