Actin was purified from rabbit skeletal muscle as described previously (Volkmann et al., 2005
). Purification of A. castellanii
, the bovine and budding yeast Arp2/3 complex, as well as the nucleation-promoting factors was described previously (Volkmann et al., 2001
). Samples for tomography were prepared as previously described for 2D analysis (Egile et al., 2005
). In brief, reactions were performed by mixing 2 μM Mg2+
-ATP–G-actin (10% pyrene labeled) with the Arp2/3 complex and nucleation-promoting factor (100 nM amoeba Arp2/3 complex with 200 nM Scar-VCA, 25 nM bovine Arp2/3 complex with 500 nM Scar-VCA, and 50 nM budding yeast Arp2/3 complex with 100 nM VCA-Bee1p). Actin polymerization was initiated by adding KMEI buffer (50 mM KCl, 2 mM MgCl2
, 1 mM EGTA, 0.2 mM DTT, 0.1 mM ATP, 0.02% azide, and 10 mM imidazole, pH 7.0). Polymerization was followed using a fluorescence spectrophotometer (MOS-250 spectrofluorimeter; Bio-Logic) equipped with BioKine 32 software (BioLogic) using 365 nm as the excitation wavelength and 407 nm as the emission wavelength. When the reaction reached the plateau, 2 μM phalloidin was added, and the samples were applied to electron microscopy grids as described in the next section.
Electron microscopy and data collection
For the negatively stained samples, the preparations were applied to glow discharged electron microscopy carbon-coated grids and partially blotted, and a 10-nm gold colloid solution (Ted Pella, Inc.) was applied and blotted. The sample was stained with 1 or 2% uranyl acetate and air dried. Tilt series of negatively stained samples were acquired at room temperature and 4-μm defocus at 31,000× (pixel size of 0.75 nm) from approximately −70 to 70° every 3° on a microscope (Tecnai TF30; FEI) operated at 300 keV using a 2,048 × 2,048-pixel CCD camera (Multiscan 894; Gatan), the serialEM package (Mastronarde, 2005
), and an advanced tomography holder (model 2020; Fischione Instruments) for data acquisition. For frozen-hydrated samples, protein solutions were applied to copper grids coated with holey carbon films (Quantifoil Micro Tools GmbH) and partially blotted, a 10-nm gold colloid solution was applied, and the sample was blotted again and plunged in liquid ethane. Tilt series of frozen-hydrated specimens were taken over holes and collected using a microscope (Tecnai F20; FEI) equipped with a 2,048 × 2,048-pixel slow scan CCD camera (TemCam-F224HD; Tietz Video and Image Processing Systems). Electron micrographs were recorded at a magnification of 26,500× (pixel size of 0.54 nm) from approximately −65 to 65° every 4° at 4-μm defocus using the TOM package (Nickell et al., 2005
) for data collection. The microscope was operated at liquid nitrogen temperature. Before data collection, conditions for all samples were refined using a microscope (Tecnai T12; FEI) at 120 keV.
We recorded a total of 18 tomographic tilt series of negatively stained and 27 tilt series of frozen-hydrated A. castellanii Arp2/3-mediated actin networks. For both the negatively stained and frozen-hydrated samples, several tomographic tilt series were repeated after a 90° sample rotation (double-tilt tomograms) to provide the best possible initial template models for the alignment (with minimal artifacts caused by missing data). We collected 10 tilt series of negatively stained filaments with yeast Arp2/3 complex and five tilt series with bovine Arp2/3 complex. To minimize beam damage, all data, including the negatively stained samples, were collected under strict low-dose conditions with a cumulative dose per tomogram not exceeding 100–300 e−/Å2.
Reconstructions were generated using IMOD (Kremer et al., 1996
). A minimum of 10 fiducial markers (10-nm gold beads) per reconstruction were used. 54, 161, and 279 branches were selected and boxed out from the tomograms of specimens with negatively stained bovine, yeast, and A. castellanii
Arp2/3 complex, respectively. Although the orientation of most branch junctions in these negative stain datasets are roughly perpendicular to the beam direction, different in-plane orientations as well as slight differences in the out-of-plane orientation increased data coverage in the averages substantially (Fig. S1 B). We selected 82 branches from the A. castellanii
cryotomograms. For tomograms of both negatively stained and frozen-hydrated specimen, we restricted the averaging and alignment of branches to a small neighborhood because the filaments are so flexible that they can bend close to the branch junctions.
Single-particle volume analysis
For each of the three negatively stained samples, we selected a well-defined double-tilt branch as a reference volume. For the frozen-hydrated samples, we used the average of the reconstruction of negatively stained A. castellanii
Arp2/3 complex branches as an initial reference. We aligned all branches to the respective references using a modified version of the density alignment algorithm in CoAn (Volkmann and Hanein, 1999
). The modification explicitly accounts for the missing data caused by tomographic data collection during the alignment step. The aligned branches were then averaged using the averaging module of CoAn, which contains a weighting function that accounts for the data coverage of the individual contributors. We applied a spherical Butterworth filter to the individual volumes to deemphasize features distal from the junction. The same filter was used for all boxed-out volumes. The boxed-out branches from cryotomograms were denoised using three rounds of iterative median filtering (van der Heide et al., 2007
) to improve alignment accuracy. The resulting alignment transforms were then applied to the original boxes that were then used for generating the average. For negatively stained samples, the 10% worst matching branches were excluded from the average; for frozen-hydrated samples, 25% were excluded. Then, all branches (including the ones removed from the average) were realigned to the average. For negatively stained samples, this procedure was repeated until there were no significant changes between subsequent averages (approximately five rounds). The refinement of the cryo data was restricted to two rounds because the resolution significantly worsened after the second round of refinement. To test for biasing toward the starting volume in the negatively stained samples, the complete process was repeated, choosing as the reference the (double tilt) branch that matched worst to the initial reference in the first round. The resulting averages were virtually identical to those using the initial references. For the data from frozen-hydrated samples, the alignment and averaging were repeated with the handedness of the reference inverted (by mirroring along the short axis). The resulting average was also virtually identical to the one with the initial reference.
We calculated resolution by dividing the final aligned volume sets into two random halves, averaging these halves separately, and calculating the 0.5 cut-off of the Fourier shell correlation between the two (Figs. S1 A and S2 C). To improve resolution, we repeated the processing of the negatively stained samples using spherical Butterworth filters with different cut-off radii. If the radius is chosen too large, the resolution deteriorates as a result of the flexibility of the actin filaments distal from the branch point. The largest cut-off radius that still gave the highest achievable resolution according to the Fourier shell correlation was used for the final reconstruction. A hierarchical descendant cluster analysis indicated that the dataset from negatively stained A. castellanii samples is homogeneous. However, we detected two subpopulations for negatively stained samples made with bovine and yeast Arp2/3 complex by image processing. These subpopulations correlate well with the two possible orientations (up or down) a branch can adopt on the sample support. The existence of these subpopulations could suggest that these structures were distorted by interaction with the substrate. As a consequence, we restricted the docking analysis to the samples made with A. castellanii Arp2/3 complex.
Docking and modeling
We docked atomic models of structural domains into the reconstructions using the CoAn software (Volkmann and Hanein, 1999
). All analyses were run on the original reconstructions without sharpening. The algorithm used in CoAn is based on a global evaluation of the 6D fitting space followed by statistical analysis. The target function used in the analysis is the real-space density correlation coefficient between the experimental density and the density calculated from the search model. In contrast to other approaches that return only a single best fit based on a single scoring function value, the CoAn algorithm relies on the identification of solution sets, which are small regions in fitting space (confidence intervals) that are compatible with the data and its associated error level at a predefined confidence level (we used 99.5%; P < 0.005). This concept efficiently captures variations caused by random errors, low reconstruction quality, and insufficient local resolution. Differences in fit using alternative models can then be tested for significance using standard statistical tests with these solution sets. All proposed conformational changes were tested using this concept and were found to be statistically significant at the 99.5% confidence level.
To allow for cross-validation and estimation of precision of the fit and confidence intervals, we divided the data in two random halves and generated two averages that were treated as two independent experimental realizations of the underlying structure. We repeated all docking and modeling experiments described in the following paragraphs for both averages as well as for the apo-Arp2/3, ADP, and ATP crystal structures (Robinson et al., 2001
; Nolen et al., 2004
) to compile the statistics and precision estimates as described previously (Volkmann and Hanein, 2003
). Thus, each modular fit is cross-validated by six different docking experiments. We used the 3D watershed transform (Volkmann, 2002
) to generate three segments for each reconstruction: one corresponding to the mother filament, one corresponding to the daughter filament, and one corresponding to the Arp2/3 complex. First, we performed docking of the Arp2/3 complex portion. We docked the unmodified, inactive crystal structures into the corresponding segments. The orientation of the complex was very similar to that obtained independently using labeling information (Egile et al., 2005
). Several areas showed severe mismatches between density and model ( and Video 2), clearly implicating the need for conformational changes even in the cryoreconstruction with its limited resolution (Fig. S2).
The resulting docking precision, which was compiled from the six independent docking experiments, was ~0.4 nm. The docking precision is related to the reproducibility of the fit within the error margin given by the data. This value is only a good estimate for the actual coordinate error (accuracy) if the model is near correct and no conformational changes occur that were not included in the modeling. Because the accuracy of the actin model used is limited (it was built using low resolution fiber diffraction and EM data), the docking precision value should be used as a lower limit of accuracy. An accuracy of 0.4–0.6 nm at an interface is high enough to make reliable predictions about residue patches being likely involved in interactions. Individual residues in the center regions of these patches are almost guaranteed to be close enough to interact; residues at the periphery are less certain. Because we have no knowledge about the side chain conformations at the protein surfaces, especially at the new interfaces, the data do not allow reliable predictions of specific residue–residue interactions.
Next, we modeled a filamentlike conformation of the Arps by superimposing individual subdomains of each Arp on an atomic model (Volkmann et al., 2005
) of the corresponding filamentous actin subunit subdomain. Disordered portions of Arp2 were replaced by the corresponding portions of filamentous actin. We then arranged the modified Arps using the actin filament symmetry to provide a short-pitch template for filament growth. We inserted this short-pitch Arp2–Arp3 heterodimer into the complex by superimposing subdomains 1 of the modified and unmodified Arp3. We performed an energy minimization with REFMAC (Murshudov et al., 1997
) to relieve minor clashes and strains. We then redocked this modified complex into the reconstructions, giving a significantly higher (P < 0.005) correlation with the density than the unmodified one and an improved visual fit ( and S2).
Then, we added four filamentous actin subunits to the modified Arp2/3 complex model using the actin filament symmetry. We docked this whole entity (corresponding to the modified Arp2/3 complex and the four subunits of the daughter filament) into the merged density segments corresponding to the Arp2/3 complex and the part of the daughter filament visible in the reconstruction. The root mean square deviation between the position of the Arp2/3 complex subunits except for Arp2 and Arp3 (which were modified) deduced from the unmodified crystal structure fitting and that for the branch model fitting are 0.45 nm, which is close to the cross-validated uncertainty of ~0.4 nm for the individual model fits. This close correspondence indicates that the absence or presence of the daughter filament and the remodeling of Arp2 and Arp3 does not introduce significant bias into the positioning of the rest of the Arp2/3 complex. In the configuration with the four daughter filament subunits added to the remodeled Arp2 and Arp3, all members of the Arp2/3 complex and the daughter filament are within the high density region of the reconstructions, indicating that there is no need for further large-scale conformational changes. The precision estimate of the fit in this region was 0.35 nm.
A crystal structure for filamentous actin is not yet available. Current models were obtained by refining monomeric actin against x-ray fiber diffraction data (0.8-nm resolution; Holmes et al., 1990
; Lorenz et al., 1993
; Tirion et al., 1995
) and by docking monomeric actin into electron microscopic reconstructions of actin filaments (~2-nm resolution; Volkmann et al., 2005
). We fitted an actin-filament model (Volkmann et al., 2005
) with 12 subunits into the density segment corresponding to the mother filament. A symmetry mismatch was visually apparent, but global changes of the filament symmetry parameters did not lead to improvements, indicating the existence of a local effect. A division of the filament model in two halves (breaking it between M3 and M4) and redocking of these individual subfilaments leads to a statistically significant increase in density correlation (P < 0.005). The accompanying change in the mother filament suggests an untwisting close to the central density bridge with the Arp2/3 complex. Insertion of a monomeric actin in the presence of ATP (PDB accession no. 1ATN
; Kabsch et al., 1990
) instead of a filamentous actin at this location accommodates the change. Replacement of this subunit with other monomeric actin structures also accommodates the change. Compared with the filamentous conformation, monomeric actin is bent ~15° between subdomains 1 and 2 and between subdomains 3 and 4 (Fig. S2 A). If the interactions of subdomains 3 and 4 of subunit M4 with its long-pitch neighbors M6 and M2 are maintained, this monomeric conformation of subunit M4 reproduces the observed symmetry mismatch within the estimated precision of the docking (~0.3 nm). We further modified the monomeric actin subunit (M4) and its barbed end neighbor (M2) to move the protruding subdomains 2 () into the density. We performed a global search subject to simultaneously maximize the correlation coefficient with the density, maintaining main-chain connectivity in the actin subunits and minimizing steric clashes. The resulting models were subjected to energy minimization to relieve minor strains and clashes. The precision estimate for the fit in the mother filament region is ~0.4 nm.
The Fourier shell correlation used to estimate the resolution of the reconstructions technically only measures the correlation between two noisy volumes in Fourier space. The resulting value may not correspond to the actual resolution, as it would occur in x-ray crystallography, where the Fourier space signal caused by the order of the underlying structure can easily be identified (diffraction spots). As a consequence, structural interpretations relying on resolution estimates may lead to overinterpretation, especially if Fourier space sharpening is used to boost high resolution features. In this study, no sharpening was used except in for visualization purposes. Furthermore, in the CoAn algorithm, Fourier amplitude effects such as those caused by the contrast transfer function or the resolution are explicitly accounted for by decorrecting the Fourier amplitudes of the search model density to (on average) match those of the experimental density. To further guard against resolution-related overinterpretation, we repeated the complete analysis after the reconstructions were truncated to a resolution of 4 nm. The need for changes in the conformation of Arp2 and Arp3 as well as in subdomain 2 of M2 and M4 are still apparent by visual inspection and are significant at a 99.5% confidence level. The untwisting of the mother filament is still highly significant, with a maximum confidence level of 99.2%. The cross-validated precision estimate drops to 0.59 nm, indicating that the higher resolution terms in the nontruncated reconstruction provide significant signal.
Online supplemental material
Fig. S1 shows quality criteria for the negative stain and ice averaged reconstructions. Fig. S2 contains information about remodeling of the mother filament and ARPC1. Fig. S3 contains information about remodeling of the Arp2/3 complex. Video1 shows a180° rotation of the atomic model of the branch junction inside the reconstruction density, and Video 2 shows an illustration of the conformational changes. Online supplemental material is available at http://www.jcb.org/cgi/content/full/jcb.200709092/DC1