Structural mechanism for bidirectional actin cross-linking by T-plastin

Significance To fulfill the cytoskeleton’s diverse functions in cell mechanics and motility, actin networks with specialized architectures are built by cross-linking proteins. How these cross-linkers specify cytoskeletal network geometry is poorly understood at the level of protein structure. Here, we introduce a machine-learning–enabled pipeline for visualizing cross-linkers bridging cytoskeletal filaments with cryogenic electron microscopy (cryo-EM). We apply our method to T-plastin, a member of the evolutionarily conserved plastin/fimbrin family, revealing a sequence of conformational changes that enables T-plastin to bridge pairs of actin filaments in both parallel and antiparallel orientations. This provides a structural framework for understanding how plastins can generate actin networks featuring mixed filament polarity.

Immediately prior to sample preparation, CF-1.2/1.3-3Au 300-mesh gold C-flat holey carbon cryo-TEM grids (Protochips) were plasma cleaned with a Hydrogen / Oxygen mixture for 5 seconds in a Gatan Solarus. Actin solution (3 μl) was first applied to the grid in the humidified chamber of a Leica EM GP plunge freezer and incubated for 60 s at 25°C. T-plastin solution (3 μl) was then applied and incubated for 30 s. Solution (3 μl) was then removed and an additional 3 μl of T-plastin solution was applied. After an additional 30 s, the grid was back-blotted for 5 s, plunge-frozen in ethane slush, and stored in liquid Nitrogen until imaging.
Cryo-EM data for the T-plastin-actin (-Ca 2+ ) complex were recorded on a Titan Krios (ThermoFisher / FEI) at the Rockefeller University operated at 300 kV equipped with a Gatan K2 Summit camera. SerialEM (2) was used for automated data collection. Movies were collected at a nominal magnification of 29,000X in super-resolution mode resulting with a calibrated pixel size of 1.03 Å / pixel (super-resolution pixel size of 0.515 Å / pixel), over a defocus range of -1.5 to -3.

High-resolution cryo-EM image processing
Unless otherwise noted, all image processing was performed within the RELION-3.0 package (3), following a recently described procedure (4). Movie frames were aligned and summed with 2 x 2 binning for the -Ca 2+ dataset, and no binning for the +Ca 2+ dataset, using the MotionCor2 algorithm (5) implemented in RELION (6) with 5 × 5 patches. The contrast transfer function (CTF) was estimated from non-doseweighted summed images with CTFFIND4 (7).
Bimodal angular searches with psi angle priors were used in all subsequent 2D and 3D alignment / classification procedures. Approximately 2,000 segments were initially manually picked, extracted, and subjected to 2D classification to generate templates for auto-picking. Helical autopicking was then performed with a step-size of 3 asymmetric units corresponding to a 27 Å helical rise. Segments were extracted from dose-weighted (8) sum images in 512 x 512 pixel boxes without down-sampling, followed by a second round of 2D classification and auto-picking with featureful class averages. Auto-picked segments were then extracted and subjected to 2D classification using a 200 Å tube diameter and 300 Å mask diameter. Segments that contributed to featureful class averages were selected for 3D analysis.
All subsequent 3D analysis steps were primed with estimates of helical rise and twist of 27.0 Å and -167.0°, respectively, utilizing an initial reference low-pass filtered to 35 Å resolution, with the outer tube diameter set to 200 Å, inner tube diameter set to -1, and the mask diameter set to 300 Å. The first round of 3D classification into 3 classes was performed with bare actin filament reconstruction (EMBD-7115) as the initial reference. A second iteration was then performed with a class featuring clear ABP density from the first round as the initial reference. For the T-plastin-actin (-Ca 2+ ) dataset, this second round of 3D classification yielded two classes with helical parameters similar to the initial estimates and well-resolved 3D features, and one junk class with aberrant helical parameters and distorted features. For T-plastin-actin (+Ca 2+ ) dataset, all three classes were good, featureful 3D classes, indicating the high quality of this dataset.
Segments contributing to the selected classes were then pooled for 3D auto-refinement.
The first round of auto-refinement was then performed using one good 3D class as the initial reference. All masks for subsequent post-processing steps were calculated with 0 pixel extension and a 6 pixel soft edge from the corresponding converged reconstruction, low-pass filtered to 15 Å and thresholded to fully encompass the density. First-round post-processing was performed with a 50 % z length mask, followed by CTF refinement without beam-tilt estimation and Bayesian polishing (6). A second round of auto-refinement was then performed, followed by post-processing with a 30 % z length mask, then a second round of CTF refinement with beam-tilt estimation and Bayesian polishing. Final auto-refinement was then performed, once again employing a 30 % z length mask for post processing.  Table S1.

T-plastin pre-bundling state cryo-EM image processing
The weak signal in Class I of the -Ca 2+ condition (SI Appendix, Fig. S2) suggested that in a subset of the particle images, signal from an additional CH domain was consistently present in a defined region. To isolate and refine those particles, symmetry expansion and subsequent rounds of focused 3D classification were used to generate a set of 322,743 segment images (SI Appendix, Fig. S6A). Using the asymmetric unit and helical parameters from 3D auto-refinement, particles were symmetry expanded. These particles were then re-extracted with recentering of 2D shifts with a box size of 128 pixels at a down-sampled pixel size of 4.12 Å and subjected to a consensus 3D auto-refinement. After this consensus refinement, a mask containing only density for the bound ABD and the additional density was used for an initial round of 3D classification with 14 classes and no image alignment. Classes with consistent density were selected and 6 subjected to 3D auto-refinement. The resultant map was used as an updated reference for another round of 3D classification of the original 322,743 segment images with 6 classes, using an updated mask. After selecting 147,851 segment images from two classes with similarly placed density, a local 3D auto-refinement was performed, which showed substantially increased density in the masked region beyond the bound ABD. Using an updated mask, another round of 3D classification was run with 3 classes and no image alignment. One class, composed of 50,077 particles, with good density was selected. These particles were re-extracted at a box size of 384 pixels without down-sampling and locally refined using a 70 % z mask. These particles were then re-extracted with the plastin molecule at the center of the box and re-refined with a 50 % z mask.
A subsequent round of 3D classification with 3 classes and no image alignment and was performed, and the best two classes were selected with 31,413 particles. After a local 3D autorefinement, the particles were re-centered on the actin filament to optimize CTF refinement and Bayesian Polishing performance for two rounds. During the second round of particle polishing, the particles were once again re-centered on the bound T-plastin protein. Polished particles were subjected to a final 3D auto-refinement, which yielded a map assessed at 4.4 Å resolution. A final, local 3D auto-refinement was performed by masking only the plastin and two actin protomers to give the final map at 6.9 Å resolution (SI Appendix, Fig. S6) within this masked region.

Measurement of inter-filament distances
To provide reasonable inter-filament distances for the synthetic particles used to train the neural network, the distances between obvious filament bundles lying perpendicular to the electron beam were measured. Specifically, 98 micrographs were selected, down-sampled by 4, and low pass-filtered to 25 Å. One to four filament bundles were analyzed per micrograph (total of 127 filament pairs), by drawing a line through the central axis of one filament and drawing a parallel line through the axis of the other filament in the bundle using FIJI (9). The distance was measured based on the system of equations: y1 = m*x + b1 ; y2 = m*x + b2 ; d = |b2-b1|/ √(m 2 + 1), where m is the shared slope between the two parallel lines, b1 and b2 are the intercepts for the lines, x and y are the coordinates along the lines, and d is the distance between the two lines.
These measurements revealed an average inter-filament distance of 159 Å with a standard deviation of 24 Å (SI Appendix, Fig S7B).

Synthetic dataset generation
We developed a convolutional neural network-based approach to specifically identify bundled filaments, while excluding individual filaments. To achieve full-micrograph segmentation, a neural network was first trained to recognize potential bundle configurations in synthetic data, and then used to predict on real data.
We first trained a denoising autoencoder on projections of plausible in silico bundle models and used the learned weights from this network to make a semantic segmentation network. The precise workflow for synthetic projection generation is outlined in SI Appendix, Fig.   S8. Briefly, projection images were generated of zero to three filament units, with each filament unit consisting of either an individual filament or a two-filament bundle. To approximately reflect the frequency of bundles in the actual dataset, each filament unit had a 65% chance of being a bundle and 35% chance of being an individual filament. If the filament unit was a bundle, the filament would be copied, and the copy would be rotated about its helical axis by a random, uniformly sampled integer between 0° and 359°, randomly tilted by a bimodal Gaussian centered at 0° and 90°, with standard deviations of 1.5°, and then randomly translated in the y-direction by 159 Å ± 24 Å (empirically measured from 127 bundles in real micrographs), and uniformly translated in the z-direction (along the helical axis) by ±181 Å. Subsequently, for both individual and bundled filament units, the filament unit would be rotated about the phi and rot angles by a random, uniformly sampled value between 0° and 359°, and the tilt by 0° with a standard deviation of 2.5°. The filament unit was then randomly translated around the box by ±250 Å and projected along the z-axis to generate a noiseless projection. The same map was used to generate a noisy image paired with this noiseless projection, by adding pink noise in Fourier space, as implemented in EMAN2's python package to generate realistic-looking synthetic data (10). Three-channel stacks of semantic maps associated with the noisy/noiseless projection pairs were generated by binarizing the filament unit and assigning it as a bundle or individual filament before projection.

Network architecture and training
A denoising autoencoder (DAE) was trained using the architecture outlined in SI Appendix, Fig Fig. S9C). As a separate estimate of the 2D denoising reconstruction's resolution, 10,000 noiseless synthetic particle images not used during network training or validation were compared to noisy particle images denoised using the trained DAE, and the Fourier Ring Correlation (FRC) was computed (SI Appendix, Fig. S9D); the average FRC curve fell below 0.5 at 12.2 Å (~1.5 times Nyquist resolution), and 100 example FRC curves are also shown.

Particle picking using neural network
To pick particles, motion-corrected micrographs down-sampled by 4 were converted into semantic maps by extracting 192-pixel boxes across the micrograph in a raster pattern with 48 pixels of overlap and stitching back the output into a semantic map by computing a maximum intensity projection of the overlapping regions. Only the bundle channel results were used, and they were binarized using a fixed threshold of 0.85. After binarization, central axes of the bundles were selected for by excluding pixels near object borders using empirically derived parameters, and small objects were removed. The remaining binarized image was skeletonized, and nonmaximum suppression was used to ensure all particle picks were spaced at least 148 Å away from each other. These particle coordinates were used for extraction in RELION.

Comparison to other particle pickers
To assess the capabilities of commonly used particle-picking software for the purpose of specifically picking bundles, we employed Topaz (11), crYOLO (12), and RELION's templatebased auto-picker (3) on a subset of micrographs from the -Ca 2+ dataset. Specifically, 50 micrographs with combinations of single-and multi-filament F-actin instances were selected; 25 micrographs were used for manual picking to train the networks, and 25 micrographs were used for assessment of picking quality and agreement with the neural network-based segmentation outlined in our work. A total of 3,377 particle coordinates of two-filament bundles, precisely centered, were hand-picked and used to train a crYOLO network using default parameters (PhosaurusNet architecture, batch size of 4, learning rate of 0.0001, trained for 200 epochs).
After a network was trained, picking was performed on the test set of micrographs, and a threshold of 0.25 was found to be optimal for picking performance. For Topaz training, the Topaz auto-picking feature as integrated into RELION4.0 (13) was used with the default parameters, except training was done for 15 epochs, instead of the default of 10, to improve results. After training, picking was performed on the test set of micrographs, and a -0.6 figure-of-merit threshold was found to be optimal for picking performance. Notably, in our experience, the crYOLO results had a large range of acceptable thresholds that produced similar results.
Similarly, for our picker, a wide range of binarization thresholds applied to the semantically segmented map yielded very similar results because of the high network confidence. Conversely, for both Topaz and RELION's auto-picker, changing the threshold slightly drastically changed the number of picks, and we found the defaults had extremely high false-positive picks.
Particle picks from the various software were compared by visual inspection and quantitative comparison to the semantic segmentation from our neural network (SI Appendix, Fig.   S11). Picks were visually assessed for their accuracy in centering the particle picks, and ability to discriminate single-from multi-filament bundles (SI Appendix, Fig. S11A). To assess agreement with our network's semantic segmentation of micrographs in the test set, picks were categorized based on which semantic bin the coordinate was positioned. For example, a picked coordinate that was in the bundle semantic channel was counted as a "true positive" in this case. Whereas a coordinate positioned on a pixel identified by the neural network as a single-filament was counted as a single filament pick, and the remaining picks were scored as background picks. Notably, the semantic maps are intentionally eroded to allow for good centering in our software, so many close, but un-centered picks were marked as background in this analysis (intentionally to account for pick centering).

Bundle processing
Visual inspection revealed that nearly all extracted particles had multiple filaments, but reference-free 2D classification with standard parameters produced classes with one filament or one well-resolved filament and one poorly resolved filament (SI Appendix, Fig. S7C). To prevent alignment from refining one filament at the expense of the other, particles images were extracted in a large, 256-pixel box downsampled to 4.12 Å pixel size, and multiple rounds of 2D classification were performed in cryoSPARC (14), limiting the reconstruction resolution to 45 Å and the alignment resolution to 50 Å (SI Appendix, Fig. S12A). The "Align filament classes vertically" option was used to determine in-plane rotation, and at each iteration, 2D class averages were re-centered using a binary mask with a low threshold to maintain the bundle in the class average center. Particles were re-extracted with shifted, re-centered coordinates and psi angles were rotated by 90° for RELION helical conventions. With these re-centered particles, 2D classification with small translations yielded high-quality, reference-free 2D class averages in RELION, where obvious parallel and antiparallel 2D classes were present (SI Appendix, Fig.   S12B). Despite exhaustive attempts at generating reasonable initial models using ab initio model generation, models would frequently be produced with one filament centered in the box or two poorly defined filaments.
Therefore, using the reference-free 2D class averages, initial models were generated using a custom projection-matching scheme. The map derived from single-filament helical analysis was rescaled and roughly positioned in the box to project onto one of the two filaments, then EMAN2's e2classvsproj script (10) was used on each filament to globally search Euler angles (with 10 degree sampling) and shifts (maximum of 84 Å) for an initial projection-match to the 2D class average. Finally, a custom projection-matching script using the EMAN2 python package was used to perform a finer, gradient-descent-based projection matching of each filament. While these projections had excellent correspondence to the 2D class averages, they did not have full 3D information to properly position filaments for an initial model. To properly zposition the filaments, a parallax-based approach using 2D classes that had clear side-views was used, and adjustments to the relative z-positions of the oriented filament maps was performed to maximize the cross-correlation between the side-view 2D class averages and the 3D models (SI Appendix, Fig. S12C). These maps were then lowpass filtered to 20 Å and used as initial references for 3D classification (SI Appendix, Fig. S12D).
3D classification required careful angular and translational searches to prevent head-on two-filament images from shifting to side-on views of one filament. An approach similar to previous 3D classification schemes of large, mixed-population filamentous structures was used (15). Briefly, the tilt prior was kept fixed at 90° and global searches of rot, local, bimodal searches of psi, and local searches of tilt with 3.7° sampling and a 20° search range were performed, with a translation search range of 41.2 Å. After one iteration with these parameters, the 3D classification was interrupted; the tilt prior was updated, and translations were deleted. The 3D classification then resumed for three iterations, with global searches of rot, local, bimodal searches of psi, and local searches of tilt with a fine 1.8° sampling and 5° search range, and a large 123.6 Å translational search range was used.
During this supervised 3D classification, there was an apparent preferred orientation for particular rot angles (SI Appendix, Fig. S13A). 2D classification of particles with rot angles of 0±15° or 180±15° revealed that some side-on view 2D class averages look similar to individual filaments, and their constituent particle images were either bundled particles shifted to have one filament in the box center or poorly picked individual filaments. 2D classes with one centered filament that was much better resolved than other 2D classes were excluded and the angular distribution improved (SI Appendix, Fig. S13A,B). After this winnowing, five more iterations of 3D classification were performed with a smaller translational search of 20.6 Å, leading to 87,980 particles sorted into the parallel class and 69,408 particles sorted into the antiparallel class (SI Appendix, Fig. S13C). A 45% z-mask was generated for both classes, and subsequent signal subtraction and focused 3D classification without alignment removed particles without clear bridging plastin density. Un-subtracted particles were re-substituted, and following a consensus 3D auto-refinement, particles overlapping within the masked region were removed. A final, asymmetric, consensus 3D auto-refinement was performed on 41,701 particles for the parallel class, which reached 9.0 Å, and 28,759 particles for the antiparallel class, which was assessed at 10.0 Å (SI Appendix, Fig. S13D).

Variability analysis
During 3D auto-refinement, it became apparent that filament density quickly smeared outside the masked region, presumably due to the relative flexibility of the complex. This observation, coupled with the composition of the system being two rigid bodies connected by a flexible cross-linking protein, led us to employ multi-body refinement, as implemented in RELION, to handle flexural heterogeneity (16). Using default 3D multi-body refinement parameters, with masks shown in SI Appendix, Fig. S13C, the resolution of the constituent filaments improved.
Furthermore, we reasoned we could utilize the multi-body refinement parameters to measure the relative motions of the two filaments to each other. Specifically, volumes were generated for each particle image using RELION's relion_flex_analyze script (41,701 volumes for parallel, and 28,759 volumes for antiparallel), and atomic models of two separate plastin-decorated actin filaments were procedurally rigid-body docked into each of the volumes using scripting functions in UCSF Chimera (17). Distances and measurements between these docked models were measured using custom scripts employing functions from the ProDy package (18). Plots were generated with GraphPad Prism (Fig. 3E-H).

Model building and refinement
To generate homology models of the actin-binding 'core' (Fig. 1B) of plastin, the Robetta server was used (19). The selected homology model for 'prebound' T-plastin was the model that had the highest score. Sharpened, local-resolution-filtered maps as described above were used for model building. The high-resolution density maps were of sufficient quality for de novo atomic model building. As structures of components were available, initial models of actin (PDB 3j8a) and the 'prebound' T-plastin homology model were fit into the density map using Rosetta (20).
Models were subsequently inspected and adjusted with Coot (21,22), and regions that underwent significant conformational rearrangements were manually rebuilt. The models were then subjected to several rounds of simulated annealing followed by real-space refinement in Phenix (23,24) alternating with manual adjustment in Coot. A final round of real-space refinement was performed without simulated annealing.
The pseudo-atomic models for the T-plastin prebundling state (Fig. 2) and both T-plastin bundle configurations (Fig. 3) were generated by rigid-body docking the high-resolution postbound T-plastin model and ABD1 fragments from the 'prebound' homology model, followed by flexible fitting with ISOLDE (25), then real-space refinement in Phenix with harmonic restraints on the starting model enabled. Key statistics summarizing model building, refinement, and validation are reported in SI Appendix, Tables S1 and S2.
Structural figures and movies were prepared with ChimeraX (26). Structural superpositions were performed with the MatchMaker function in ChimeraX. Pruned RMSD, which excludes poorly aligning regions such as flexible loops, was calculated with the default parameters. Per-residue RMSD analysis was performed with UCSF Chimera (17) as previously described (27). The surface area of actin-binding interfaces was calculated with PDBePISA (28) (EMBL-EBI). Model quality was assessed with EMRinger (29) and MolProbity (30) as implemented in Phenix.

Actin co-sedimentation assays
Mixtures of F-actin (5 µM) and the indicated T-plastin constructs (2 µM) were incubated at room temperature for 30 min in co-sedimentation buffer (10 mM Tris pH 7.5, 100 mM KCl, 2.5 mM MgCl2, and 2 mM DTT. The samples were then spun for 30 min. at 16,000 rpm (low-speed) in a TLA-100 rotor and polycarbonate centrifugation tubes (Beckman Coulter No. 343775). This pellet was the 'low-speed' pellet. The supernatant was taken out from the centrifugation tube and then spun in a fresh centrifugation tube for another 30 min, at 100,000 rpm (high-speed). This pellet was the 'high-speed' pellet, and the supernatant was also collected. All three fractions were subject to SDS-PAGE and Coomassie staining. The gels were scanned using LI-COR imaging system, and subsequently quantified with ImageJ. Plots were generated with GraphPad Prism, and statistical analysis was performed with Microsoft Excel.

Cell culture and transfections
All experiments using HUVEC were performed with an hTERT-immortalized HUVEC line previously described (32)    Analysis regions were subsequently defined in Slidebook to exclude the ends of actin stress fibers and focal adhesions, and the region intensities were measured to quantify their recovery after photobleaching. Intensities were background subtracted and normalized to their pre-bleach intensity, then analyzed in GraphPad Prism by fitting to an exponential one-phase association model, Y = Y0 + (Plateau -Y0) * (1 -e (-K * x) ). Wild-type and mutant data were compared using a two-way ANOVA with a mixed-effects model and Geisser-Greenhouse correction (SI Appendix, Table S3).          Network performance on top view particle from A. Top row, left: extracted particle image; middle: denoised particle image; right: binarized bundle channel used for particle picking. Bottom row, semantic segmentation channels denoting background, single filaments, and bundle. Scale bar, 30 nm. (C) Network performance on side view from A, which is not readily discriminated from single filaments by eye. Subpanels as in B, scale bar 30 nm. (D) Picked particle featuring a three-filament bundle, upon which the network was not explicitly trained. Scale bar, 30 nm.        Table S1. High resolution cryo-EM data collection, refinement, and validation statistics.

Data collection and processing
Movie S1. Interpolation along the first principal component from the multi-body refinement of the parallel bundle. The measured splay and skew angles are displayed at each frame.
Movie S2. Interpolation along the first principal component from the multi-body refinement of the antiparallel bundle. The measured splay and skew angles are displayed at each frame.

928
Colored as in Fig. 3A