Removing Contamination-Induced Reconstruction Artifacts from Cryo-electron Tomograms

Abstract

Imaging of fully hydrated, vitrified biological samples by electron tomography yields structural information about cellular protein complexes in situ. Here we present a computational procedure that removes artifacts of three-dimensional reconstruction caused by contamination present in samples during imaging by electron microscopy. Applying the procedure to phantom data and electron tomograms of cellular samples significantly improved the resolution and the interpretability of tomograms. Artifacts caused by surface contamination associated with thinning by focused ion beam, as well as those arising from gold fiducial markers and from common, lower contrast contamination, could be removed. Our procedure is widely applicable and is especially suited for applications that strive to reach a higher resolution and involve the use of recently developed, state-of-the-art instrumentation.

Introduction

Cryo-electron tomography (cryo-ET) combines faithful sample preservation by vitrification with the three-dimensional (3D) imaging by electron microscopy (1). As of this writing, it is the only method that allows the comprehensive imaging of fully hydrated biological samples at a single nanometer scale. Consequently, it can provide structural information about molecular complexes and their constituents imaged in their physiological context. Thinning of vitrified samples by focused ion beam (FIB) milling can alleviate the fundamental limitation of transmission electron microscopy, i.e., sample thickness (2, 3). Currently, FIB milling is approaching a level where it can be applied routinely, thus allowing cryo-ET imaging of the interior of a wide variety of cell types. Together with recent developments in electron microscopy instrumentation, such as direct electron detector devices (DDD) and Volta phase plates for electron microscopy (EM), cryo-ET can deliver fine structural details of complexes imaged in their native, cellular environment (4, 5, 6, 7).

However, the combination of these high-end methods sometimes leads to unforeseen difficulties, such as the generation of surface contamination by the ion beam. Although advancements in sample preparation and handling may appear as the best strategy to solve such problems, this goal might not be attainable in many cases, because thinning vitrified cells by FIB is still a delicate method that requires careful, cell-type-dependent optimization. Consequently, developing image processing methods to alleviate the detrimental effects of these artifacts on the structures imaged might be more effective. Furthermore, because tomograms recorded with DDDs contain an improved high-frequency signal, even the fine artifacts induced by common cryo-ET contaminants may become more prominent and thus interfere with the high-resolution structural information. Therefore, computational image restoration approaches might be indispensable for postprocessing of cryo-tomograms obtained using state-of-the-art technology.

A number of computational stages involved in electron tomography are devoted to facilitate tomogram interpretation by applying some kind of signal restoration (7, 8). Denoising methods are commonly used as a postprocessing step to reduce the noise and improve the contrast (9, 10, 11, 12, 13). However, they are not well suited to remove the artifacts often present in tomograms that are due to the lack of projection views at high tilt angles or the limited angular sampling. These artifacts, especially exacerbated at the edges of highly dense features, are easily identifiable as long blurring rays in y slices (perpendicular to the tilt axis). A filter intended to downweight the available high tilt projections has been proposed to reduce the missing wedge artifacts at the expense of poorer resolution in the direction of the electron beam (14). Gold particles, typically added to the samples to facilitate tilt-series alignment, are particularly strong sources of streak artifacts. All existing methods that can remove the artifacts emanating from gold particles essentially substitute the gold particles in the tilt-series by some neutral density level estimated from the local neighborhood (15, 16, 17). This density substitution not only removes the projections of gold particles, but also other projection values along the same electron path. Therefore, these methods have very limited applicability in cases where the source of the artifacts is not confined to small areas.

Here we present a computational procedure that removes reconstruction artifacts arising from various types of contamination and restores the structural information in tomograms. The examples presented show applications to surface contamination associated with sample thinning by FIB, gold fiducial markers as well as other forms of surface contamination. The procedure is also evaluated on phantom data and on results of subtomogram averaging.

Materials and Methods

Sample preparation, vitrification, and thinning

Neuronal cultures

Primary cultures of neuronal cells were carried out according to the method of Kaech and Banker (18) with modified sample preparation for cryo-ET (19, 20). Gold Quantifoil grids (R1/4, and R2/4 200 mesh; Quantifoil Micro Tools, Jena, Germany) were additionally coated with a 20-nm carbon layer on the film side using a carbon evaporator (MED 020 Coating System; Bal-Tec, Balzers, Liechtenstein). Grids were sterilized with UV-light in the clean bench for 30 min and soaked into a solution of 1 mg/mL poly-L-lysine in 0.1 M borate buffer overnight and under light protection. Gold grids were repetitively washed in autoclaved distilled water, immersed in neuronal plating medium (MEM with Earl’s salts and L-glutamine, D-glucose 0.6%, FBS 5%), and stored in a CO₂ incubator. Rats of embryonic stages E18–E21 were sacrificed in accordance with the procedures accepted by the Max Planck Institute for Biochemistry (Martinsried, Germany) and hippocampi were dissected and isolated in HBSS-HEPES on ice. Tissues were cut in pieces of ∼1 mm and trypsinized at 37°C. Cells were isolated with a cell strainer and suspended in neuronal plating medium. Cells, 3.0 × 10⁵, were seeded in a 4-well-dish with a gold grid in each well. Cells were cultivated for 4 h before the medium was partially exchanged with preconditioned Neurobasal/B27 medium (Thermo Fisher Scientific, Waltham, MA) from glial preculture. On Day 3 of culture, Ara-C was added and every week medium was partially exchanged with fresh and preconditioned medium. Neurons were plunge-frozen after 15–19 days of culture. In some cases (see Fig. 4, d and e), the additional coating by carbon and conditioning with glial medium were not done. Immediately before plunging, concentrated BSA-coated, 10-nm fiducial markers (Aurion, Wageningen, The Netherlands) were diluted in culture medium and 3–4 μL was applied to the grid backside.

Figure 4.

Applications to samples that exhibit a common, weak granular surface contamination. The untilted image (a) and the tomogram (b) of neuronal cultures show an almost continuous granular layer. The artifacts in the y slice of the tomogram in (b) arise from the interference between the streaks of the fine surface contamination and those of gold particles. (Solid arrows) Pointers to streak artifacts. Images were acquired with a CCD camera. (c–e) Sparse contamination, images acquired with a DDD camera. (c) Untilted image of the tilt-series. (d and e, arrows) Pointers to the two y slices of the tomogram where streak artifacts originated by the contamination granules. In all cases, the procedure removed or attenuated the artifacts (initial tomograms are on the top and the restored on the bottom in b, d, and e). (Open arrows) Pointers to positions on the restored tomograms that correspond to the contamination. (Insets) Differences between the distorted and restored details (outlined by dashed lines). Scale bars = 100 nm.

Neurons were plunge-frozen on Quantifoil gold grids (Quantifoil Micro Tools) in a liquid ethane/propane mixture cooled by liquid nitrogen using a Vitrobot Mark IV (FEI, Hillsboro, OR) (see Figs. 1 and 3) or a manual plunger (see Fig. 4). The Vitrobot was set to 37°C, humidity 90%, blot time 10 s, blot force 10, and waiting time 5 s. Underneath the filter paper for blotting, Teflon (E.I. DuPont de Nemours, Wilmington, DE) sheets were used on both sides. Grids were blotted from excess ethane and stored in liquid nitrogen.

Figure 1.

Surface contamination and artifact removal. (a–c) Untilted image of the initial tilt-series, z slice, and y slice of the tomogram, respectively, showing surface contamination and the artifacts induced in the reconstruction. (d–f) Restored tomogram obtained by the artifact-removing procedure. The contamination was removed from the tilt-series, which caused the ripples to disappear from the tomogram. (Solid arrows) Pointers to contamination in the tilt-series (a) and tomogram (c). (Open arrows) Pointers to the corresponding positions in the restored versions (d and f). (Small lines on the left, b and c, e and f) Correspondence between the z- and y slices of the tomograms. Scale bar = 100 nm.

Figure 3.

Applications to samples affected by Pt-sputtering and large contamination. (a–c) A tomogram of HeLa cells containing Pt-sputtering-induced artifacts. (a) Tomogram before (left) and after the artifact-removing procedure (right). The z- and y slices are shown (top, bottom, respectively). (Dashed square) Area zoomed in (b) (arrow pointing to ripples). (c) Zoomed areas of another tomogram obtained under the same conditions, before and after the application of procedure. (d) A tomogram of neuronal synapse containing two medium/large contamination granules (denoted by arrowheads). Initial (left) and restored (right) tomograms are shown. (Top-left, bottom-left, and top-right panels) The z-, y-, and x slices, respectively. (Solid arrows) Pointers to ripples caused by the medium/large contamination granules. (Open arrows and arrowheads) Pointers to the corresponding positions in the restored tomograms. (Insets, with arrows) Zoomed views of the highlighted areas. (Small lines on the left, a and d) Correspondence between the x-, y-, and z slices of the tomograms. Scale bars = 100 nm.

Vitrified neuronal cultures shown in Figs. 1 and 3 were thinned using a DB FIB Quanta 3D FEG (FEI) equipped with a PP3000T cryo-system (Quorum Technologies, Guelph, Ontario, Canada) and an in-house-developed, open-nitrogen-circuit 360° rotatable cryo-stage. Platinum (Pt) was sputtered on the whole grid, for 60 s at 10 mA, in the PP3000T transfer system (Quora Technologies) before milling. For the standard wedge-milling routine the specimen was tilted such that the incident angle of the Ga⁺ ion beam, with respect to the sample surface, was 6°. Under this shallow angle, the milling was performed in three steps of sequentially decreasing ion beam current at 30 kV acceleration voltage. The first, rough cut was done with a 40-μm-wide rectangle pattern and a 0.5 nA beam current. Further thinning was then done with 0.3-nA and 0.1-nA beam currents. The final polishing step was performed at 0.05 nm. All milling steps were monitored by single-scan ion beam images using 100 ns pixel-dwell time, and by electron-beam images using a beam current of 6 pA at an acceleration voltage of 5 kV and 5 μs pixel-dwell time.

Synaptosomes

Cerebrocortical synaptosomes were prepared and vitrified as described previously in Dunkley et al. (21), Godino et al. (22), and Fernández-Busnadiego et al. (23) in accordance with the procedures accepted by the Max Planck Institute for Biochemistry. In brief, euthanized 6–8-wk-old male Wistar rats were decapitated, and the cortex was extracted and homogenized in homogenization buffer (HB; 0.32 M sucrose, 50 mM EDTA, 20 mM DTT, and one tablet of cOmplete Mini EDTA-free Protease Inhibitor Cocktail (10 mL, pH 7.4; Hoffmann-La Roche, Basel, Switzerland)) with up to seven strokes at 700 RPM in a Teflon (E. I. DuPont de Nemours) glass homogenizer. The homogenate was centrifuged for 2 min at 2000 g, and the pellet was resuspended in HB and centrifuged for another 2 min at 2000 g. Supernatants from both centrifugations were combined and centrifuged for 12 min at 9500 g. The pellet was resuspended in HB and loaded onto a three-step (3, 10, and 23%) Percoll (GE Healthcare, Little Chalfont, Buckinghamshire, UK) gradient in HB. The gradients were spun for 6 min at 25,000 g, and the material accumulated at the 10:23% interface was recovered and diluted to a final volume of 100 mL in HBM (HEPES-buffered medium; 140 mM NaCl, 5 mM KCl, 5 mM NaHCO₃, 1.2 mM Na₂HPO₄, 1 mM MgC_l2, 10 mM glucose, and 10 mM HEPES, pH 7.4). Percoll (GE Healthcare) was removed by an additional washing step with HBM by centrifugation for 10 min at 22,000 g, and the pellet was resuspended in HBM and immediately used in the experiments. All steps were performed at 4°C.

Synaptosomes were diluted to 0.7 mg/mL protein concentration determined by Bradford assay (Bio-Rad Laboratories, Hercules, CA) and preincubated for 30 min at 37°C. A 3-μL drop of 10-nm colloidal gold (Sigma-Aldrich) was deposited on plasma-cleaned, holey carbon copper EM grids (Quantifoil; Quantifoil Micro Tools) and allowed to dry. A 3 μL drop of synaptosomal suspension was placed onto the grid, allowed to equilibrate for 5 s, blotted with filter paper (Whatman Grade 1; GE Healthcare), and plunged into liquid ethane. Vitrified grids were kept in liquid nitrogen until imaging.

HeLa cells

HeLa cells were seeded on holey carbon-coated 200-mesh gold EM grids (Quantifoil Micro Tools) in Ibidi μ-slides (Ibidi, Munich, Germany) containing DMEM medium supplemented with 10% fetal bovine serum and cultured at 37°C with 10% CO₂. The EM grids were manually blotted from the reverse side and cells were plunge-frozen in a liquid ethane/propane mixture (2:1).

The same dual beam instrument (Quanta 3D FEG; FEI) as before was used to prepare thin electron transparent lamellae (2). To protect the milling front of the lamellae, a few hundred-nm-thick layer was formed on top of the whole grid by injection of gaseous organic platinum compound (Pt GIS; FEI) (modified from Hayles et al. (24)). A quantity of 15–30 μm-wide lamellae were prepared using a Ga⁺ ion beam at 30 kV at shallow angles between 8 and 14°. To eliminate the charging effect during acquisition, which is caused by the missing conductive carbon support film in the final lamella and affects the operation of the Volta phase plate, a few nanometers of thin metallic pure Pt layer was sputtered in an Argon atmosphere under cryo conditions in the PP3000T transfer system (Quorum Technologies) onto the lamella with the following parameters: 10 mA sputtering current, 500 V between stage and sputtering target, and 5–10 s of exposure at 4.5 × 10⁻² mbar.

EM and tomography

Images were recorded on Polara and Titan electron microscopes (FEI). Both microscopes were equipped with a field emission gun (operated at 300 kV), a computerized stage, and an imaging filter (operated in the zero-loss mode; Gatan, Pleasanton, CA). The tomogram shown on Fig. 4, d and e, was recorded on a MegaScan charge-coupled device (CCD) camera (Gatan), while all other shown tomograms were recorded on a K2 Summit DDD camera (Gatan). Pixel sizes at the specimen level were 0.82 nm for the CCD and 0.42 nm for DDD tomograms. Tilt series were binned two times (bin factor 4) so the final pixel sizes were 3.3 and 1.7 nm. A Volta phase plate was used for recording the tomograms shown in Figs. 1 and 3. The nominal underfocus was 0.5–1 μm with and 6–9 μm without the phase plate. Tilt series were typically recorded from −60° to +60° at 1.5–2° angular increment using Xplore3D (FEI) and SerialEM (25) for automated data collection. The tilt series were aligned using gold beads as fiducial markers and reconstructed with IMOD, except for those of HeLa cells that were aligned using fiducial-less alignment by patch tracking in IMOD (26).

Artifact-removing procedure

Three-dimensional reconstructions were computed with weighted back projection (WBP) using a standard ramp filter combined with a Hamming apodizing window, as implemented in tomo3d (27). All results shown here were obtained this way. WBP in IMOD and SIRT in tomo3d were also used to check the general applicability of the artifact-removing procedure introduced here.

The procedure itself was implemented in two ways, as a standalone executable and as a shell script that calls IMOD commands. Both versions were written for x86_64 architecture under LINUX and OSX, and are available upon demand or from http://sites.google.com/site/3demimageprocessing/masktomrec.

Results

Ion beam thinning of neuronal cultures induces surface contamination

Dissociated neuronal cultures were prepared and grown on EM grids following essentially the procedure we used previously in Lucić et al. (19). Cultures thinned by FIB consistently showed strong contamination in projection images (Fig. 1 a). Upon 3D reconstruction, it was clear that the contamination is localized on a sample surface (Fig. 1 c). This contamination was not seen when neurons were imaged by cryo-ET without thinning. Furthermore, the contamination also appeared on grids where platinum was not sputtered before milling. It is thus likely that its appearance is associated with thinning by FIB and that it is caused by the redeposition of milled material.

The contamination was predominantly located on the side of neuronal boutons that is away from the beam (on the lower-right side on both boutons visible on Fig. 1 a). Considering that the boutons extend in the z direction with respect to the neuronal processes that surround the boutons, it appears that the contamination is primarily located on the surface of the boutons that makes at most a shallow angle with the ion beam. We estimated from the tomogram that the surface delineated by the contamination is inclined at 0–25° with respect to the x,y plane (measured at the steepest direction). Considering that the ion beam was directed at the angle of 6° to nominal grid plane, it appears that the contamination is formed on the region of the bouton surface that is not directly exposed to the beam because it is sheltered by the central part of the bouton and also on the surface regions roughly parallel to the beam. This strengthens the argument that the ion beam causes contamination and suggests that cellular samples that have highly uneven (i.e., hilly) surface, such as neurons, are more prone to this type of contamination.

Surface contamination granules induced artifacts in the tomograms in the form of ripples in z slices, as seen in Fig. 1, a–c. They emanated as typical streaklike artifacts in y slices (very similar to those arising from gold particles) that at some point interfere with each other, thereby producing the high-frequency pattern observed in Fig. 1, b and c. The artifacts become stronger if lower angular sampling is used in tilt-series acquisition. In any case, the ripples can prevent small details from being discerned, and because they appear as fine filaments they might be misinterpreted and are likely to interfere with automated filament detection approaches, such as the one used on neuronal synapses (28). Therefore, these artifacts hinder further analysis and interpretation.

Procedure for removing surface contamination-induced reconstruction artifacts

We devised a procedure to remove the artifacts induced by the surface contamination and restore the distorted structural information. Our approach was to remove the contamination from the initial (contamination-containing) tilt-series because it causes artifacts in the tomogram, while keeping the useful information. The procedure starts by reconstructing the tomogram from the aligned initial tilt-series and shifting the grayscale values so that the mean is 0.

The next step is a generation of a mask enclosing the contamination in the tomogram. Obviously, the contamination has to be present in the tomogram. The mask can be generated manually, using simple tools from the available tomogram reconstruction and processing software packages. However, strategies to quickly design masks are needed for practical applicability of the procedure. Because the surface contamination comprises dense spots close to and overlapping each other, we devised a strategy based upon the grayscale density gradient. The gradient of the tomogram is calculated first, producing high values at the edges of the contamination spots. A wide Gaussian smoothing (e.g., σ = 4) then follows, so that high-gradient areas in close proximity (i.e., the contamination) merge, whereas other areas with isolated edges (e.g., membranes) are attenuated. Afterwards, the resulting Gaussian-smoothed gradient map is thresholded by density to yield the mask (see Fig. S1 in the Supporting Material). It is expected that in some situations false-positives or false-negatives arise. In those cases, complementary morphological operations (thresholding based on the size of connected components, dilation, etc.) usually help (29). Following this strategy, masks can be created in a matter of a few minutes.

After the initial tilt-series, the tomogram and the mask are generated, and the main part of the procedure is executed by iterating the following steps:

• 1)Apply the mask on this tomogram so that only the contamination remains and reproject the contamination (masked tomogram);
• 2)Subtract the reprojected contamination from this tilt-series to obtain a new tilt series; and
• 3)Reconstruct a new tomogram from the tilt-series generated in the previous step using the WBP and set the mean to 0.

A flowchart showing one iteration is presented in Fig. S2 a. As the procedure evolves, the contamination and as a result the reconstruction artifacts induced by the contamination are gradually removed, while the valuable projection information is preserved. In practice, the number of iterations is ∼5, as further processing provides only marginal improvements.

Fig. 1, d–f, shows the results of the artifact-removing procedure (five iterations), while the difference between the initial and the restored tomograms is shown in Fig. S3. The contamination has been removed from the tilt-series, but the remaining structural information is left unimpaired. For instance, a vesicle can be seen in Fig. 1 d while it is obscured behind the contamination in Fig. 1 a. The resulting tomogram now shows detailed information and the contamination is substituted by a fairly flat region.

We note that it is not necessary that the mask precisely traces the contamination. If needed, the contamination in the experimental tilt-series can be isolated by subtracting the initial tilt-series (Fig. 1 a) and the restored one (Fig. 1 d). We also tested this procedure when SIRT reconstruction was used instead of the WBP. The results were very similar to those where the WBP was used during the artifact-removing procedure and the SIRT was applied only at the last tilt series (data not shown), so we did not pursue this approach further.

Evaluation of the procedure

Several tests were conducted to evaluate the procedure using the tomogram and tilt-series (as described above). First, we focused on the number of iterations required to remove the artifacts. The standard deviation in the area of the tomogram where the contamination is located is expected to follow the progression of the procedure and show the amount of residual contamination. This calculation clearly demonstrates that most of the contamination removal takes place at the first three iterations (Fig. 2 a). Beyond the fifth iteration, the standard deviation does not change significantly. This number of iterations has been found empirically to be appropriate for all tomograms we tested.

Figure 2.

Evaluation of the artifact-removing procedure. (a) Standard deviation of the contamination region of the tomogram at different iterations. Most of the contamination is removed at the first 3–5 iterations. (b) Tests with a phantom composed of vertical lines with defined spatial frequency. Phantom distorted by artifacts (left) and the restored tomogram (right) were obtained in two different ways: the phantom was superposed in the whole (top) and on the contamination-only tomogram (bottom). The artifact-removing procedure recovered the phantom in both cases. (c) Subtomogram averaging test. (On the right) Gallery of ribosomes (z- and y slices) in (top) control, a tomogram of ribosomes distorted by contamination (middle), and after the application of the procedure (bottom). (On the left) Restored versus control (solid) and distorted versus control (shaded) FSC curves and the subtomogram averages (insets show two orthogonal planes of each volume) from restored (top), distorted (bottom), and control (middle, right) tomograms. Scale bars = 100 nm.

Next, we performed several tests with phantom data to check whether the procedure removes the contamination and the artifacts, and properly restores the original signal. Generation of test datasets is schematically illustrated in Fig. S2 b. In the Lines test 1 (Fig. S2 b), a phantom tomogram composed of straight vertical lines arranged at a defined spatial frequency (Lines phantom on Fig. S2 b) was reprojected and the density scale of the resulting tilt-series was set to match the experimental tilt-series (Fig. 1, a–c) without the surface contamination. This density-scaled phantom tilt-series was added to the experimental one and the result was subjected to standard WBP reconstruction. The WBP reconstruction showed the phantom severely distorted by the artifacts, whereas the artifact-removing procedure recovered the periodic line pattern (Figs. 2 b, top, and S4).

In the Lines test 2 (Fig. S2 b), only the experimental contamination (obtained by subtraction of the restored tilt series (Fig. 1 d) from the experimental one (Fig. 1 a)) was added to the phantom. The procedure almost completely restored the line pattern even though it was significantly distorted by the artifacts (Figs. 2 b, bottom, and S4). All these results demonstrate that the contamination induces reconstruction artifacts and that the procedure removes the artifacts and restores the signal.

To further explore the effects of the artifacts and assess the procedure, we performed subtomogram averaging of ribosomes (Ribosomes test in Fig. S2 b). A human ribosome structure was obtained from the EM Data Bank (EMD: 5592) (30) and scaled to the pixel size of the tomogram. A phantom tomogram was created by placing 240 ribosomes having random but known orientations at defined locations (Ribosomes phantom in Fig. S2 b). Those 240 ribosomes were spread throughout the tomogram so that they would be distorted at various levels by contamination artifacts (Figs. 2 c and S5). A phantom tilt-series was created by reprojecting and rescaling the phantom tomogram. For the subtomogram averaging, distorted, restored, and control tomograms were generated as follows: 1) Reprojected experimental contamination was added to the phantom tilt-series and a tomogram of ribosomes distorted by contamination artifacts was reconstructed from this combined series using the WBP. The resulting tomogram contained ribosomes distorted by contamination artifacts. 2) This tomogram was then subjected to the artifact-removing procedure to obtain the restored ribosomes tomogram. 3) For the control tomogram, the WBP reconstruction was applied to the phantom tilt-series. Fig. 2 c (right) displays a gallery of the artifacts-distorted ribosomes and the corresponding restored tomograms.

Ribosomes were extracted from the control, distorted, and restored tomograms (described in the previous paragraph) and aligned using their known positions and orientations. The subtomogram averages (insets in Fig. 2 c, left) were computed for all three cases by Fourier-space addition taking the missing wedge into consideration. The Fourier shell correlation (FSC) of both distorted and restored ribosome subtomogram averages were calculated in respect to the average from the control tomogram. The FSC curves clearly show the benefit of the procedure (Fig. 2 c, left). The FSC between the restored and the control tomograms (black FSC curve) showed a value close to 1 throughout the Fourier space and thus was much better than the FSC between the distorted ribosomes and the control (gray FSC curve). Artifacts are apparent on the distorted ribosome average, not only as the noisy surrounding background, but also on the ribosome structure itself. In contrast, the procedure yielded an average that is virtually the same as the control average.

Because subtomogram alignment was known, this subtomogram-averaging test shows the effects of the reconstruction artifacts induced by contamination as well as the benefits of the procedure. In a realistic subtomogram-averaging scenario that includes subtomogram alignment, the contamination-induced artifacts would likely contribute to improper alignment, resulting in a poorer average structure.

Formal development

It is instructive to derive the expression that governs the artifact-removing procedure. One iteration of the procedure that starts with tomogram and tilt series at the ith iteration (t_i and p_i, respectively) and results in tomogram at iteration i+1 (t_i+1) can be formally represented as

$$t_{i+1}=W\left(p_i-PM{t}_i\right),$$

where M is the masking operator that keeps the contamination and removes everything else, P is the reprojection, and W is the 3D reconstruction operator. It is understood that t₀ is the initial tomogram and p₀ is the tilt series used to reconstruct the initial tomogram. Using the fact that t_i = Wp_i, it follows that

$$t_{i+1}=t_i-WPM{t}_i.$$ (1)

The tomogram obtained after n iterations can be formally expressed as

$$t_n={\left(1-WPM\right)}^n{t}_0.$$ (2)

To understand the procedure better, we compared this expression with those of the similar iterative procedures that yielded suboptimal results. First, a procedure whereby an iteration consists of masking (removing) the contamination of a tomogram, followed by the reprojection of the masked tomogram and the reconstruction of the resulting tilt series, can be expressed as

$$t_{i+1}=WP{t}_i-WPM{t}_i.$$ (3)

This procedure removed the contamination, but caused the reconstruction artifacts to spread throughout the tomogram (Fig. S6). Comparing the formal expressions (Eqs. 1 and 3), we can conclude that the contamination removing term (WPMt_i) is the same, but the suboptimal procedure unnecessarily reprojects and reconstructs the tomogram (WPt_i), which leads to propagation of reconstruction artifacts.

Second, a single iteration consisting of reprojecting the contamination mask to get localization of the contamination, applying it to the tilt series so that the contamination is completely removed and reconstructing the resulting series, can be expressed as

$$t_1=t_1-W\left(PM{1}_3\right)p_0,$$ (4)

where 1₃ is an identity tomogram (all values are 1), so that M1₃ represents a 2D mask that is nonzero only at the location of the contamination and PM1₃ is the corresponding 2D mask that can be applied on a tilt series. It is obvious that removing all contamination from projections removes too much useful information. This was confirmed by applying this procedure (Fig. S7) and it can also be seen by comparing the second terms on the right-hand side of Eqs. 1 and 4.

It follows from Eq. 2 that the successful procedure should be equivalent to an alternative formulation whereby in each iteration the reprojected and reconstructed contamination is subtracted from the tomogram, instead of subtracting at the projection level as was stated in the procedure formulation. In other words, the reconstruction operator is expected to be linear. This was confirmed to be the case by the observation that results obtained using the alternative formulations were indistinguishable from those by the main procedure formulation (data not shown).

Therefore, the formal derivation of the artifact-removing procedure allowed us to reach a better understanding of the procedure and to explore alternative formulations.

Other applications

While the initial motivation to develop the procedure introduced here arose from ion-beam-induced surface contamination deposition on neuronal cells, the procedure itself is not limited to this preparation. Therefore, we applied this procedure to remove reconstruction artifacts arising from other types of contamination.

Structural investigations of various (nonneuronal) cell types can benefit from sputtering a small amount of platinum on FIB milled lamellas. While this approach increases the conductivity of the lamella, thus facilitating the use of the Volta phase plate for cryo-ET imaging (J. Mahamid, S. Pfeffer, M. Schaffer, E. Villa, R. Danev, L. Kuhn-Cuellar, F. Foerster, A.A. Hyman, J.M. Plitzko, and W. Baumeister, unpublished data), it also leaves a layer of fine granules on the sample surface. A layer of platinum sputtered on a lamella cut on HeLa cells can be clearly seen on tomographic slices (Fig. 3 a, left). This layer caused the formation of fine ripples throughout the tomogram, which might give a false impression of high-resolution information. It can be seen from y slices that these are reconstruction artifacts quite similar to those shown on Fig. 1. The application of our procedure successfully eliminated the artifacts (Fig. 3 a, right). This is better illustrated on the magnified insets on the same figure. Fig. 3 b is a zoom of the area boxed in Fig. 3 a, and shows a series of ripples (marked with an arrow) that disappeared after removing the artifacts. Fig. 3 c is another striking example of the benefits of the procedure. Here, several fine membranes are well defined (bottom panel) whereas they were difficult to perceive in the initial tomogram (top panel).

So far we were concerned with surface contamination associated with sample thinning by FIB. A larger contamination, such as that marked by arrowheads on Fig. 3 d, is commonly encountered in cryo-ET. Fig. 3 d (left), shows the initial tomogram (x,y,z slices). This contamination produced a series of ripples, as marked with arrows in the z- and x slices and enlarged in the insets. A spherical mask that covers the contamination was used. Figs. 3 d (right) and S3 show that the procedure successfully removed the large contamination and its associated artifacts. Although this tomogram also suffered from FIB-induced surface contamination, the artifacts arising from the two contamination types could be separated and they were both removed by the procedure. It is clear that the procedure cannot generate the signal that was lost because of the electron interaction with a large contamination, but can remove reconstruction artifacts arising from the contamination.

To test our procedure further, we applied it to two tomograms that had a weak but clearly detectable granular contamination, of the type that is commonly encountered in vitrified samples and is not related to FIB milling. In the first tomogram, the contamination formed almost continuous layers on both surfaces (Fig. 4, a and b). In our hands, its occurrence was not entirely predictable, but it could usually be avoided by ensuring that only the bottom of the liquid ethane container was in direct contact with liquid nitrogen, in effect increasing the temperature of liquid ethane. Even though the tomogram was recorded on a CCD camera and did not have a strong contrast, fine streak artifacts emanating from surface contamination and gold particles were visible in the initial tomogram and were significantly diminished in the restored tomogram. In the second tomogram, recorded on a DDD, only a few contamination granules were sufficient to produce streaks, which were to a large extent removed by the artifact-removing procedure (Fig. 4, c–e). Automatic mask generation with the procedures described above was challenging because of the low contrast of the contamination. In these cases, rough manual delineation of the z limits of the contamination granules, or even the mask itself, turned out to be a quick and convenient solution. Therefore, even though the contamination had an only slightly higher density than the biological structures of interest, the procedure effectively attenuated the streak artifacts caused by them.

Another obvious application of the artifact-removing procedure is the removal of gold particles from tomograms, as well as of the corresponding reconstruction artifacts. Streaks originating at gold particles were strongly attenuated by the procedure (Figs. 4 b and S8). Our technique is different from current approaches whereby a gold particle in a tilt-series is replaced by a value computed from the local neighborhood, thus ignoring information from other projections that overlap with the gold particles (15, 16, 17). Our technique uses these other values and, in this regard, the structural information in the tomogram is better preserved. Fig. S8 shows the benefits of our procedure compared to the technique used to erase gold particles in IMOD.

Discussion

New instrumentation advances for cryo-EM, which include the development of FIB and phase plates for EM and DDDs, are pushing the limits of structural investigations in situ (2, 3, 4, 5, 6). However, these technical advances also bring about challenges that need to be overcome to realize the full potential of cryo-ET. Here we present a computational procedure for removing 3D reconstruction artifacts caused by contamination present in vitrified biological samples. It is based on an iterative procedure involving reprojections of selected regions and 3D reconstructions and it was successful in removing, or at least greatly attenuating, artifacts arising from several different types of contamination, thus restoring the structural information.

Working with intact cells, we were confronted with two types of contamination associated with thinning of vitrified samples by FIB milling. The first type regularly occurs on vitrified neuronal cultures. It tends to be located on the region of sample surface that is sheltered from the ion beam by higher regions of the sample and possibly approximately parallel to it. It is likely that it is formed by the redeposition of material during milling and that a highly uneven (i.e., hilly) surface, created by numerous boutons present on the intertwined neuronal axons and dendrites, facilitates the generation of this contamination type on neuronal cells. The second type is due to postthinning platinum sputtering, which is beneficial for subsequent imaging using Volta phase plate for EM (J. Mahamid, S. Pfeffer, M. Schaffer, E. Villa, R. Danev, L. Kuhn-Cuellar, F. Förster, A. A. Hyman, J. M. Plitzko, and W.B., unpublished data) and is independent of the cell type. Both types appeared in our tomograms as a surface layer of moderately dense granular material. Three-dimensional reconstruction caused interference from multiple rays originating at individual granules, which resulted in fine ripples in tomograms at areas distant to the contamination (Figs. 1 and 3, a–c). It is clear that these artifacts would persist even if the tomogram size was adjusted so that the contamination was absent from tomograms. Because of their fine appearance, they can be mistaken for high-resolution information and obscure the interpretation of tomograms.

Other contamination types that we investigated commonly occur in cryo-tomograms. These include loosely scattered contaminants that are of only slightly higher density than the biological structures of interest. In this context, gold fiducial markers can be regarded as contaminants. The associated reconstruction artifacts consisted of very small ripples and rays, or streaks, which emanated from the contamination (Figs. 3 d and 4). Also, when phantom structures were superposed on experimental contamination, the reconstruction artifacts caused by the contamination severely distorted the superposed structures and reduced their resolution (Fig. 2).

Furthermore, we derived the formal expression for the procedure. This derivation helped us to reach a better understanding of the procedure and explore alternative formulations, which might be important for guiding future development.

The application of our procedure was successful in all cases. It removed ripples, removed or significantly attenuated streaks, and completely restored the superposed phantom structures and recovered the resolution. Therefore, tomographic studies that use CCDs, as well as those based on the advanced technology, such as DDDs and phase plates, may benefit from this procedure. The use of DDDs makes the artifacts more prominent and is often associated with the aim of reaching higher resolution, thus increasing the importance of the artifact-removing procedure for such studies.

Conclusion

We present here a computational approach for removing contamination-induced 3D reconstruction artifacts that is widely applicable and could significantly improve the quality and the resolution reached in cryo-tomograms, especially those obtained using the state-of-art technology.

Author Contributions

U.L., F.J.B.B., and M.K. prepared samples and recorded tomograms; U.L., M.S., and F.J.B.B. optimized and performed the FIB-milling procedure; W.B. provided essential materials and expertise; and J.-J. F. and V.L. designed and supervised research, developed and applied the computational procedure, and wrote the article.

Editor: Andreas Engel.