Branched actin networks push against each other at adherens junctions to maintain cell–cell adhesion

Compartmentalization of tissues in an organism is mediated by cohesive monolayers of epithelial and endothelial cells. Both cell types create a barrier at the tissue boundary, but an endothelial barrier is more dynamic and permits solute exchange and leukocyte transmigration (Giannotta et al., 2013; Schnittler et al., 2014). Inadequate control of endothelial permeability leads to edema that accompanies inflammation, allergy, ischemia, and other disorders (Dejana and Giampietro, 2012).

Cells control monolayer permeability by forming various cell–cell adhesions. Among them, adherens junctions (AJs) carry most of the mechanical load (Huveneers and de Rooij, 2013; Twiss and de Rooij, 2013; Ladoux et al., 2015). AJs are formed by adhesion receptors, mainly of the cadherin family, and strengthened by the actin cytoskeleton, which interacts with cadherins through α- and β-catenins and other components of the cadherin adhesion complex (Twiss and de Rooij, 2013; Padmanabhan et al., 2015; Mège and Ishiyama, 2017). To build AJs, epithelial and vascular-endothelial cells use epithelial cadherin (E-cadherin) and vascular endothelial (VE)-cadherin, respectively. AJs are often classified into punctate (discontinuous) and linear (continuous) AJs that are typical for remodeling and cohesive cell sheets, respectively (Twiss and de Rooij, 2013). Because of more active junction remodeling, endothelial AJs exhibit greater polymorphism than epithelial cells, with a greater fraction of punctate AJs.

Actin cytoskeleton is important for stabilization, remodeling, and mechanosensitive properties of AJs (Michael and Yap, 2013; Schnittler et al., 2014; Hoffman and Yap, 2015; Ladoux et al., 2015; Röper, 2015). As the major force-generating machinery in the cell, the actin cytoskeleton can produce both pulling and pushing forces (Svitkina, 2018). Pulling (contractile) forces in nonmuscle cells are generated by sliding of bipolar filaments of nonmuscle myosin II (NMII) along actin filaments. Generation of pushing (protrusive) forces most commonly involves polymerization of actin filaments organized into branched networks by the Arp2/3 complex (Pollard, 2007; Svitkina, 2013).

The NMII-generated contractile force applied to AJ is required for their stabilization (Twiss and de Rooij, 2013; Ladoux et al., 2015). The best understood mechanosensor at AJs is α-catenin, which can unfold under force (Yonemura et al., 2010; Barry et al., 2014; Buckley et al., 2014; Yao et al., 2014), allowing for recruitment of additional adhesion complex components (Yonemura et al., 2010; Twiss et al., 2012; Thomas et al., 2013; Yao et al., 2014; Oldenburg et al., 2015), long-range clustering of cadherin (Chen et al., 2015), and stabilization and elaboration of AJs (Liu et al., 2010). The presence of contractile actin bundles at AJs is well established. In epithelial cells, mature linear AJs are typically flanked by tangential (parallel to the AJ) actin–NMII bundles located immediately next to the junctional plasma membranes (Hull and Staehelin, 1979; Hirokawa and Tilney, 1982; Yonemura, 2011). In assembling AJs, tangential bundles can be located at a distance from the cadherin-rich zone and connected to the AJs by small oblique bundles approaching the AJ at various angles (Yonemura et al., 1995). The latter organization is often exhibited by linear AJs of endothelial cells, although more mature configurations resembling linear AJs of epithelial cells are also present (Huveneers et al., 2012). Punctate AJs in both cell types are associated with oblique actin–NMII bundles (often referred to as “radial”) that form end-on attachments with the cadherin-rich foci (Millán et al., 2010; Huveneers et al., 2012).

The roles of Arp2/3 complex–dependent actin polymerization in AJ morphogenesis are mostly discussed in relation to AJ remodeling. Lamellipodia driven by polymerizing branched actin networks allow cells to come into physical contact (Krendel and Bonder, 1999; Ehrlich et al., 2002; Ayollo et al., 2009; Hoelzle and Svitkina, 2012), expand the area of cell–cell interaction (Ehrlich et al., 2002; Yamada and Nelson, 2007), and heal the intercellular gaps (Martinelli et al., 2013; Abu Taha et al., 2014). In assembling AJs, a faint population of dynamic F-actin is located at the cadherin-rich zone between tangential actin bundles (Yonemura et al., 1995; Adams et al., 1998; Krendel and Bonder, 1999; Zhang et al., 2005) and contains proteins typical for Arp2/3 complex–dependent protrusions (Kovacs et al., 2002; Helwani et al., 2004; Verma et al., 2004; Yamazaki et al., 2007). In mature AJs, however, actin polymerization is thought to become obsolete (Mège et al., 2006; Ryu et al., 2009) because both protrusive activity of cells and the signaling pathways that activate actin polymerization are down-regulated upon AJ maturation (Ehrlich et al., 2002; Yamada and Nelson, 2007; Ayollo et al., 2009). Nonetheless, functional experiments indicate that Arp2/3 complex–dependent polymerization machinery is required for maintenance of mature AJs (Sahai and Marshall, 2002; Ivanov et al., 2005; Furman et al., 2007; Gates et al., 2007; Yamazaki et al., 2007; Kovacs et al., 2011; Tang and Brieher, 2012; Lee et al., 2016).

Specific roles of Arp2/3 complex–dependent machinery at AJs are not clear. Conceptually, if only contractile forces acted at mature AJs, the AJs would be unstable and prone to separation. Indeed, both excessive activation and inhibition of cell contractility impairs mature AJs, although punctate AJs can be strengthened as a result of mechanostimulation (Krendel et al., 1999; Sahai and Marshall, 2002; Shewan et al., 2005; Ayollo et al., 2009; Huveneers et al., 2012; Naydenov et al., 2016). It is typically assumed that contractile forces at AJs can be counterbalanced just by cadherin-mediated adhesion (Twiss and de Rooij, 2013; Schnittler et al., 2014), but protrusive forces were also proposed as a possibility (Yap et al., 2015). However, direct evidence for the existence of such protrusive forces at mature AJs is lacking as is a conceptual understanding of the interplay between pushing and pulling at AJs.

One way to understand the relative roles of pulling and pushing forces at AJs is to determine spatial distribution of actin cytoskeleton components responsible for protrusion and contraction. EM is a powerful approach to attain this goal. Besides being able to precisely localize molecular markers by immunolabeling, EM has a potential to distinguish structural features of protrusive and contractile actin arrays. Protrusive actin networks formed by Arp2/3 complex consist of short branched filaments with a ∼70° angle between their barbed ends oriented in the direction of protrusion (Svitkina et al., 1997; Svitkina and Borisy, 1999). Conversely, contractile actin structures consist of long unbranched actin filaments that often form tight bundles (Svitkina et al., 1995; Verkhovsky et al., 1995). Traditional thin-section EM can visualize actin bundles reasonably well, especially large and stable ones (Sanger and Sanger, 1980; Tilney et al., 1980). However, visualization of branched actin networks by conventional EM techniques is challenging because of the fragility of these networks and their complex 3D architecture. Among various EM approaches, platinum replica EM (PREM), which combines the high resolution of EM with an ability to preserve three-dimensionality (Svitkina, 2017), has been most productive in revealing branched actin filament networks at various subcellular locations and in different cells (Svitkina and Borisy, 1999; Cameron et al., 2001; Korobova and Svitkina, 2010; Collins et al., 2011; Svitkina, 2017).

In this study, we used PREM in combination with immunogold labeling to determine actin cytoskeleton organization at cell–cell junctions of human umbilical vein endothelial cells (HUVECs). We show that Arp2/3 complex–positive branched actin networks represent the actin cytoskeleton component that is commonly associated with VE-cadherin at both linear and punctate AJs. In contrast, NMII-containing actin bundles are typically linked to the VE-cadherin–rich zone through branched networks. In mature linear AJs, a thin junctional strip of branched networks consists of two oppositely oriented network populations originating from two adjacent cells, whereas VE-cadherin is enriched at the interface between these colliding networks. The structure and dynamics of cell–cell junctions after inhibition of Arp2/3 complex activity or NMII-dependent contractility suggest that branched networks at AJs push against each other to promote VE-cadherin transinteraction, whereas contractile activity of NMII stabilizes cadherin complexes at the AJ cytoplasmic face.

HUVECs form both linear and punctate AJs (Huveneers et al., 2012). We stained HUVECs with phalloidin and antibodies to VE-cadherin and vasodilator-stimulated phosphoprotein (VASP) to characterize their AJs by fluorescence microscopy. VASP is an actin elongation factor that in cells localizes to sites of actin polymerization (Dominguez, 2009). In migrating cells, VASP is highly enriched in focal adhesions at the ends of contractile stress fibers (Reinhard et al., 1992). A lower level of VASP is present at the leading edge of lamellipodia (Rottner et al., 1999). This difference in VASP enrichment in the bundled and branched actin arrays can help to roughly discriminate these actin subsets based on the level of VASP immunofluorescence.

VE-cadherin staining in HUVECs revealed both linear and punctate AJs (Fig. 1). As expected, punctate AJs were associated with oblique actin bundles (Fig. 1, C and F), whereas linear AJs were often flanked by tangential actin bundles (Fig. 1, B and E), usually located at some distance from the VE-cadherin–positive line. Additionally, a faint F-actin signal colocalized with the linear VE-cadherin–rich zone (Fig. 1, B and E). VASP colocalized with VE-cadherin at both punctate and linear AJs, but exhibited bright staining at punctate AJs (Fig. 1 C), suggesting their association with bundled actin filaments, and faint labeling at linear AJs (Fig. 1 B), suggesting the presence of branched actin networks there.

To examine the contractile actin cytoskeleton at AJs, we costained HUVECs with phalloidin and an antibody to NMIIA, the most abundant isoform of NMII (Fig. 1 D). NMIIA staining was hardly detectable in linear AJs marked by faint continuous staining of VASP and F-actin (Fig. 1 E) but was prominently present in both tangential and oblique actin bundles (Fig. 1, E and F).

To test whether linear VE-cadherin distributions in HUVECs correspond with mature or assembling AJs, we performed live-cell imaging of GFP–VE-cadherin and mCherry-LifeAct (an F-actin marker) in HUVECs. We found that linear GFP–VE-cadherin patterns located between adjacent or more distally positioned tangential bundles overlap with a faint dynamic strip of mCherry-LifeAct fluorescence and exhibit subtle dynamics. However, they do not undergo gross changes over time periods up to 60 min, suggesting that such AJs are fully formed and stable (Videos 1, 2, and 6).

We next analyzed cytoskeleton structure at cell–cell junctions by PREM. Because contractile actin bundles are thought to be most intimately associated with AJs, we combined PREM with immunogold staining of NMIIA. We first focused on linear AJs, which were identified morphologically at low EM magnification (Fig. 2 A). A conventional signature of mature AJs in epithelial cells is the presence of tangential actin–NMII bundles closely flanking a continuous linear distribution of cadherin. Such conventional linear AJs are less common, but still present, in endothelial cells.

High-magnification views of linear AJs exhibiting a classic appearance showed three sets of actin structures: two bundles of long actin filaments labeled intermittently with NMIIA immunogold and a narrow strip of unlabeled actin network between these bundles (Fig. 2, B–D). The presence of Y-shaped actin filament branches, short filaments, and abundant filament ends in this network (Fig. 2, C and D) suggested an Arp2/3 complex–dependent actin array. Based on analysis of five linear AJs with a total length of 55 µm, the spacing between two flanking bundles could be as narrow as 100–300 nm. Sometimes, the bundles even appeared to touch each other, but even in such cases, a strip of branched actin network was present between the bundles (Fig. 2 C).

Not all linear junctions in HUVECs had this classic appearance, in which a branched actin strip was closely flanked by two parallel tangential bundles. In many cases, the tangential bundles were several micrometers apart and less perfectly aligned. Nonetheless, a narrow strip of branched actin network was consistently present between peripheral tangential bundles (Fig. S1). The connection between the tangential bundles and the strip of the branched network was mediated by oblique actin bundles or individual actin filaments, which merged with the branched strip at a variety of angles (Fig. S1, B and C).

These data show that a narrow strip of branched network is always present at continuous cell–cell contacts, which, based on morphological criteria, correspond with linear AJs.

The PREM technique involves detergent extraction of cells, which exposes the cytoskeleton but dissolves the plasma membrane, thus making imprecise the morphological identification of cell–cell boundaries. To determine the exact location of the AJ, we performed double-immunogold staining of HUVECs with antibodies to VE-cadherin (18 nm gold; actual diameter is 15.8 ± 1.0 nm; mean ± SD; n = 100) and NMIIA (12 nm gold; actual diameter is 10.1 ± 0.6 nm; n = 100). These gold sizes can be unambiguously distinguished from each other and from upright actin filament ends that also appear as bright dots in inverse contrast, but unlike gold particles, they have a dark core (Fig. S2 A). Strikingly, VE-cadherin immunogold at linear AJs colocalized predominantly with the branched actin network (80.4 ± 6.4%; n = 2,393 gold particles in 15 segments of linear AJs with a total length of 152.4 µm from two experiments). In contrast, NMIIA immunogold labeled tangential and oblique actin bundles and barely overlapped with the VE-cadherin–positive regions (Fig. 3). Even when a strip of the branched network shifted away from one tangential actin bundle to another, VE-cadherin immunogold was still associated with the branched network, whereas tangential actin bundles exhibited negligible VE-cadherin labeling (Fig. 3 B). We did not encounter examples in which a linear stretch of VE-cadherin immunogold would reside on tangential bundles without branched actin networks present in the vicinity. The remaining (19.6 ± 6.4%) VE-cadherin gold particles, which were not associated with branched networks, were typically found along the distal parts of oblique actin bundles near their junction with the AJs (Fig. 3, A and B). These additional VE-cadherin gold particles produced local enrichments of VE-cadherin labeling along the length of the linear AJ, which likely reflected force-dependent local augmentation of the VE-cadherin adhesion complex (Liu et al., 2010). Interestingly, the strips of branched networks that were densely labeled with VE-cadherin immunogold were usually quite narrow (230 ± 108 nm wide; 14 linear AJs from three independent experiments with total length of 163 µm measured every ∼1 µm). In contrast, when cell–cell contacts were made by broad overlapping lamellipodia, VE-cadherin labeling was sparse (Fig. S2, B–D), suggesting that AJs in such cases were not fully formed. Despite its narrowness, the branched actin strip at AJs was as dense as branched actin network in lamellipodia at the free edges of HUVECs, although lamellipodia were much wider (0.5–2 µm wide). Quantification of the actin density (see Materials and methods) showed indistinguishable (P = 0.08) occupancy of the network area by actin filaments in both cases (lamellipodia: 67.5 ± 2.3%; six regions with total area of 13.1 µm²; junctional strips: 64.5 ± 3.3%; seven regions with total area of 25.2 µm²). These similar densities suggest that branched networks at AJs can be as efficient in exerting pushing force as those assembled in lamellipodia (Cojoc et al., 2007) and in reconstituted systems (Bieling et al., 2016).

Because VE-cadherin is linked to actin filaments indirectly, we examined localization of α-catenin, an actin-binding component of the VE-cadherin adhesion complex. By both fluorescence microscopy (Fig. S3 A) and PREM (Figs. 3 C and S3 B), VE-cadherin and α-catenin at linear AJs colocalized with each other predominantly on branched networks.

For PREM experiments, we mostly analyzed AJs formed by cohesive cells in the middle of cell colonies in subconfluent HUVEC cultures. Such cells are better spread and exhibit fewer cell overlaps than cells in dense confluent cultures, which makes PREM of subconfluent cultures more interpretable. Nonetheless, in linear AJs formed in dense confluent HUVEC monolayers, VE-cadherin immunogold also localized to branched actin networks, and not to the adjacent actin bundles (Fig. S4). At other cell–cell boundaries in confluent cultures, VE-cadherin immunogold had broad distributions caused by significant overlapping of adjacent cells and assembly of adhesion complexes in the x–y plane between cells.

Although the presence of Y-shaped branches with an ∼70° angle between actin filaments (Fig. 2) is a reliable reporter of Arp2/3 complex–dependent actin polymerization in the junctional actin strip, we further validated this interpretation by double-immunogold staining of the Arpc2 subunit of the Arp2/3 complex (12-nm gold) and VE-cadherin (18-nm gold; Fig. 4). Both markers localized predominantly to the morphologically recognizable branched actin network and were scarce in adjacent cell regions (Fig. 4, B–D). A small fraction of VE-cadherin immunogold found outside the junctional strip (Fig. 4 D) was usually associated with ends of oblique actin bundles (Fig. 4 C). A small fraction of nonjunctional Arpc2 immunogold (Fig. 4 D) corresponded with nonjunctional branched network.

The orientation of actin branches shows direction of pushing force generated by branched networks. Close examination of the actin strip at linear AJs revealed two sets of branches oriented roughly against each other (Figs. 5 and S5). VE-cadherin immunogold tended to localize to the region of overlap between these two subsets (Fig. 5, C and D; and Fig. S5 C), indicating that they belonged to different contacting cells. Quantification of branch orientation showed that 55.8 ± 13.0% of branches were oriented toward the VE-cadherin–rich zone, 19.6 ± 10.1% were oriented away and 18.2 ± 11.0% were oriented parallel to the AJ (n = 1,222 branches in 15 segments of linear AJs with a total length of 35.7 µm from two independent experiments). Very rarely, we could observe a segment of the linear AJ, where a branched network from one cell collided with a tangential bundle of the neighboring cell and VE-cadherin localized approximately at the interface between the branched network and the bundle (Fig. S5 D). Typically, even when low-magnification images gave an impression that a lamellipodium of one cell is pushing against a bundle of the other cell (Fig. 5 B), high-magnification images showed that the “attacked” cell “defended” itself by sending a smaller lamellipodium with oppositely oriented actin filament branches to meet the attacker and that VE-cadherin was localized approximately to the area of collision (Fig. 5 D).

Punctate AJs are a diverse group. Their common feature is discontinuous VE-cadherin (Figs. 1 and 6) and α-catenin (Fig. S3, A and C–E) distribution and association with oblique actin-NMII bundles. Using the existing nomenclature with slightly modified definitions, we categorized punctate AJs into three groups: focal AJs, intercellular bridges, and engulfed fingers. Focal AJs have been previously defined as VE-cadherin clusters attached to radial (“oblique” in our terminology) actin–NMII bundles (Huveneers et al., 2012). We limit this category to cases in which cells still maintain an extended cell–cell contact without obvious gaps, even if VE-cadherin is discontinuous. If VE-cadherin clusters are flanked by intercellular gaps, we define them as intercellular bridges (Hoelzle and Svitkina, 2012). Engulfed fingers were described previously as VE-cadherin–rich linear extensions of one cell that appeared to be partially internalized by the neighboring cell (Hayer et al., 2016).

Focal AJs were associated with actin bundles terminating in the VE-cadherin–rich regions (Fig. 6 A). However, branched actin filaments of opposite orientation were also typically detected at the junction. These branched networks were often contiguous with the lamellipodia formed along the edge of each cell between focal AJs, but VE-cadherin labeling was low in these interstitial regions.

In intercellular bridges, VE-cadherin immunogold (Figs. 6 B and S6) could form a compact cluster (Fig. S6), but more commonly it was distributed along the associated actin bundles (Fig. 6 B). This distribution probably corresponded with a junction along the lateral surfaces of filopodium-like protrusions from neighboring cells (Hoelzle and Svitkina, 2012). Again, branched actin filaments often could be seen in the area labeled by VE-cadherin (Figs. 6 B and S6) as also confirmed by Arpc2 immunogold staining (Fig. 6 B).

Engulfed fingers (Fig. 6 C) appeared as a thin bundle of actin filaments extending from one cell underneath the neighboring cell. These extensions were labeled by VE-cadherin immunogold (Fig. 6 C). The edge of the engulfing cell formed a funnel-shaped lamellipodium that extended over the surface of the adjacent cell. In addition to labeling the finger-like extension, a significant amount of VE-cadherin immunogold was also present along the sides of the “funnel,” apparently, where the engulfing cell encountered the base of the engulfed finger. Branched actin filaments were not readily apparent along the length of the engulfed finger. The funnel-shaped lamellipodium, however, contained abundant branched actin filaments oriented toward the neighboring cell. Branched actin filaments oriented in the opposite direction could be detected in the engulfed cell (Fig. 6 C).

Endothelial cells often exhibit so-called reticular AJs (Cain et al., 2010; Seebach et al., 2015) with a polygonal distribution of VE-cadherin, which likely correspond with regions where neighboring cells overlap and present the lateral view of an AJ. Such AJs are particularly abundant in dense confluent HUVEC monolayers. By fluorescence microscopy, VE-cadherin at reticular AJs formed a network-like pattern, whereas phalloidin staining did not show any conspicuous distribution (Fig. S7 A). By PREM (Fig. S7, B–D), VE-cadherin distributions that appeared to correspond with reticular AJs colocalized with a loose network of actin filaments, which consisted of a complex arrangement of thin bundles, individual long actin filaments and Arpc2 immunogold labeled branched filaments (Fig. S7, C and D).

Localization of VE-cadherin between colliding branched actin networks at linear AJs suggested that two cells might use Arp2/3 complex–dependent actin polymerization to push against each other in order to maintain contact. To test this idea, we treated HUVECs with the Arp2/3 complex inhibitor CK666. Fluorescence staining with phalloidin and cortactin antibody, a marker of branched actin networks, showed much weaker, but still detectable, staining of cortactin after treatment with 100 µM CK666 compared with control cells (Fig. S8 A). The presence of remaining lamellipodia in CK666-treated cells also indicated a partial inhibition of the Arp2/3 complex in these conditions. Immunofluorescence staining with VE-cadherin antibody (Fig. 7, A and B) revealed formation of multiple gaps between CK666-treated cells, with bright VE-cadherin puncta marking remaining cell–cell contacts (Fig. 7 B). Faint VE-cadherin signals apparently localizing to gap edges could sometimes be seen between these puncta (Fig. 7 B).

Live-cell imaging of HUVECs expressing GFP–VE-cadherin and mCherry-LifeAct showed that addition of CK666 caused cell–cell separation, which occurred either abruptly (44 of 65 cell–cell boundaries) or gradually (21 of 65 cell–cell boundaries), probably depending on intrinsic differences in cell contractility and/or adhesion in individual cells. Addition of DMSO as control did not affect cell–cell junctions (Video 2). Abrupt cell–cell separation was accompanied by overall cell retraction, with the formation and subsequent breakage of long intercellular bridges (Video 3). Gradual separation often began with thinning of the GFP–VE-cadherin–rich linear zone and was followed by slow formation of narrow intercellular gaps crossed by intermittent short VE-cadherin–rich bridges; a fainter VE-cadherin signal also transiently remained at the gap edges (Fig. 7, C and D; and Videos 4 and 5).

PREM also showed incomplete inhibition of branched actin nucleation by CK666. Prominent lamellipodia were still present at free cell edges, but they contained a sparser actin network. We focused on cell–cell interfaces that still preserved some interaction and appeared to correspond with the slow cell–cell separation mode induced by CK666 treatment (Fig. 8). In these cases, boundaries of adjacent cells ran almost parallel to each other at 0.3–2.5-µm distances (1.1 ± 0.6 µm; n = 10 intercellular gaps with a total length of 111 µm measured every 1–2 µm; Fig. 8, B and C), which could make a subset of these gaps undetectable by light microscopy. The gaps were sometimes crossed by VE-cadherin–positive bridges (Fig. 8, F and G). Consistent with a significant decrease of branched actin networks in CK666-treated cells, VE-cadherin immunogold was more prominently associated with actin bundles or loose unbranched actin filaments (75.7 ± 7.8%; n = 2,282 gold particles in 11 segments of adjacent cell–cell boundaries with total length of 118.7 µm) compared with control cells (19.6 ± 6.4%; see above). The rest of the gold particles (24.3 ± 7.8%) were associated with remaining branched networks. Besides labeling intercellular bridges, VE-cadherin could be often seen at the gap edges (Fig. 8, B–E). The unusual preservation of VE-cadherin at free cell edges after detergent extraction suggests that VE-cadherin–catenin complexes are still connected to the cytoskeleton despite the loss of adhesion. Indeed, in some cases, it was possible to observe a nonfilamentous material associated with VE-cadherin immunogold at gap edges (Fig. 8, D and E), which could correspond with VE-cadherin–catenin complexes.

Because of incomplete inhibition of the Arp2/3 complex in our conditions, we labeled CK666-treated cells with cortactin immunogold to evaluate the degree of local inhibition of branched nucleation. In control HUVECs, cortactin immunogold, as expected, labeled the junctional strip of the branched actin network (Fig. S8, B–D). In CK666-treated cells, cortactin localized to the remaining cell–cell junctions, whereas free gap edges were less efficiently labeled (Fig. S8, E–G). These data suggest a correlation between down-regulation of branched networks and loss of cell–cell adhesion. Overall, the effects of CK666 suggest that after Arp2/3 complex inhibition, linear AJs split along the cell–cell interface as if resulting from impaired VE-cadherin transinteraction in the intercellular space, whereas cytoplasmic VE-cadherin–containing adhesion complexes transiently remain associated with the cytoskeleton despite cell–cell separation.

To determine how cell contractility affects cytoskeletal architecture of AJs, we treated HUVECs with the NMII inhibitor blebbistatin (Fig. 9, A and B) or the Rho-associated protein kinase (ROCK) inhibitor Y-27632 (Fig. 9, C–E; and Fig. S9), which prevents ROCK-mediated activation of NMII. Both inhibitors produced roughly similar phenotypes. Their most obvious effect was a significant reduction of actin bundles and formation of prominent overlapping lamellipodia at cell–cell junctions. Some lamellipodia formed a ruffle above the neighboring cell. More commonly, a lamellipodium of one cell extended underneath a nondynamic (lacking lamellipodia) edge of the neighboring cell (Figs. 9 A and S9 A). Both types of cell–cell overlaps exhibited sparse VE-cadherin labeling. Local accumulations of VE-cadherin were seen only in cases when both cells formed small colliding lamellipodia (Fig. S9 C) or when one cell formed a small lamellipodium that collided with an actin bundle in the other cell (Fig. S9 B).

Some linear AJs in blebbistatin- (Fig. 9, A and B) or Y27632-treated (Fig. 9, C–E) cells were partially preserved, but often exhibited round or oval-shaped gaps. The preserved segments of linear AJs (Fig. 9 D) were flanked with thin tangential actin bundles and contained branched actin networks decorated by VE-cadherin immunogold (71.2 ± 14.5% of gold particles on branched networks and 28.8 ± 14.5% on loose unbranched filaments). Oblique actin bundles as well as local VE-cadherin enrichments were mostly lost as expected after inhibition of NMII activity. In contrast with gaps generated by CK666 treatment, VE-cadherin immunogold did not label the edges of oval gaps in cells treated with blebbistatin (Fig. 9 B) or Y-27632 (Fig. 9 E) but was concentrated at the sites where cell edges preserved interaction (Fig. 9, B, D, and E).

Live-cell imaging of GFP–VE-cadherin and mCherry-LifeAct showed that shortly after Y-27632 addition, punctate and reticular AJs acquired linear morphology, with GFP–VE-cadherin forming a thin sharp line. Over time, gaps began to interrupt this linear VE-cadherin distribution (Fig. 9 F, Video 7, and control Video 6), thus recapitulating the phenotype observed by PREM.

In this study, we answer two important outstanding questions regarding organization and functions of the actin cytoskeleton at AJs in endothelial cells. First, we show that VE-cadherin–containing adhesion complexes predominantly interact with Arp2/3-containing branched actin networks, whereas NMII-containing contractile actin bundles are located more distally and often interact with AJs via branched networks. These findings significantly change the common thinking that cadherin complexes directly interact with contractile actin arrays. Second, we demonstrate that branched actin networks function at mature AJs and are necessary for their maintenance, whereas branched actin was previously thought to function only during establishment and expansion of AJs (see Introduction). We propose that branched actin networks of contacting cells push against each other to maintain transinteraction of VE-cadherins. Our data also provide additional insights into interplay between pushing and pulling forces at AJs.

The major challenge for understanding relative roles of protrusion and contraction at AJs was spatial limitations imposed by AJ geometry, which confines both actin cytoskeleton activities to a very tight space. To solve this technical problem, we took advantage of unique features of PREM (Svitkina, 2017). First, because of single-filament resolution and preservation of the sample three-dimensionality, PREM can discriminate structural features of protrusive and contractile actin filament arrays. Second, PREM is easily compatible with immunochemistry and thus can precisely localize specific proteins. In this study, we used VE-cadherin as a reporter of AJs, NMIIA as a marker of contractile actin structures and Arpc2 or cortactin as markers of Arp2/3 complex–dependent protrusions.

In all AJ types examined in this study, VE-cadherin was mostly associated with the branched actin network, whereas more distal actin-NMII bundles were linked to VE-cadherin–rich regions through these branched networks. Although VE-cadherin binds actin filaments indirectly, the length of the bridge formed by adapter proteins (30 nm maximum; Bertocchi et al., 2017) is insufficient to cross the observed distances between VE-cadherin and actin-NMII bundles. However, it does not mean that VE-cadherin is unable to bind actin bundles because a fraction of VE-cadherin immunogold is found on tips of oblique bundles in linear AJs, but rather suggests that branched actin networks preferentially interact with VE-cadherin–catenin complexes.

Although the mutual arrangement of pushing and pulling actin arrays varies among AJ types (Fig. 10), it can be viewed as a continuum. In the most classic-looking linear AJs (Fig. 10 A), two tangential actin-NMII bundles closely flank a narrow strip of the branched network formed by two colliding subsets with VE-cadherin at their interface. In a modified version of linear AJs (Fig. 10 B), tangential bundles are located at a distance from the VE-cadherin–positive branched strip, but connected to the strip by oblique actin-NMII bundles or individual filaments. VE-cadherin is continuously distributed in these linear AJs, but it is also enriched at the sites of oblique bundle attachment, probably because of mechanical stimulation transmitted by these bundles. Similar cadherin clusters were detected in linear AJs by subdiffraction fluorescence microscopy (Indra et al., 2013; Truong Quang et al., 2013; Strale et al., 2015; Wu et al., 2015).

Among punctate AJ types, focal AJs (Fig. 10 D) most closely resemble linear AJs. At focal AJs, cell edges contact each other over significant distance despite discontinuous VE-cadherin distribution. Focal AJs could be formed from linear AJs through greater clustering and subsequent individualization of local VE-cadherin accumulations at the tips of oblique bundles. In turn, intercellular bridges (Fig. 10, E and F) closely resemble focal AJs except that the cell edges between bridges are retracted and form gaps. Branched actin networks in intercellular bridges can be present at the tips (Fig. 10 E) or along the length (Fig. 10 F) of filopodium-like protrusions. The engulfed fingers (Fig. 10 C) are likely formed from intercellular bridges, when one of the contacting cells resumes lamellipodial protrusion and runs over the bridge.

The apparent structural continuity among different AJs is consistent with the known ability of cells to interconvert AJs during their remodeling. We suggest that the whole range of observed AJ morphologies (Fig. 10) depends on a balance between the pushing force, which is generated by junctional branched networks along the length of the cell–cell contact, and the pulling force, which is produced by actin-NMII bundles. If cell contractility is reduced, the surviving AJs have morphology of linear AJs with sparse homogeneous VE-cadherin distribution as observed in this study and previously (Krendel et al., 1999; Sahai and Marshall, 2002; Shewan et al., 2005; Ayollo et al., 2009; Huveneers et al., 2012). Conversely, excessive contractility disrupts linear AJs and promotes focal AJs and intercellular bridges (Essler et al., 1998; Krendel et al., 1999; Gavard and Gutkind, 2008; Millán et al., 2010; Huveneers et al., 2012).

The consistent presence of branched actin networks at the VE-cadherin–positive AJs suggests important functional implications. We propose that branched actin networks of two cells constantly push against each other at the AJ in order to maintain transinteraction between VE-cadherins of contacting cells. Consistent with this idea, inhibition of Arp2/3 complex causes cell–cell separation, indicating that cells are unable to maintain AJs in these conditions probably because of the absence of adequate pushing forces from both cells. Arp2/3 inhibition may also act indirectly, for example, by up-regulating RhoA signaling and cell contractility (Korobova and Svitkina, 2008), which might explain the fast mode of cell–cell separation under CK666 treatment. The transient presence of VE-cadherin at the gap edges in CK666-treated cells is unusual because unengaged cadherin is normally endocytosed (Le et al., 1999) and becomes detergent soluble. Thus, inhibition of the Arp2/3 complex does not immediately disrupt the interaction of VE-cadherin with the cytoskeleton but leads to cell–cell separation through a different mechanism, most likely by compromising cadherin transinteraction in the intercellular space.

The branched junctional strip is distinct from the junctional lamellipodia, which repair intercellular gaps (Martinelli et al., 2013; Abu Taha et al., 2014). In our samples, the overlapping junctional lamellipodia show low VE-cadherin labeling and most likely function in establishing rather than maintaining AJs, in contrast with the branched junctional strip. The narrowness of the junctional strip may result from an interplay between two opposite processes induced by cadherin engagement: Arp2/3 complex–dependent protrusion (Kovacs et al., 2002; Helwani et al., 2004; Verma et al., 2004) and contact inhibition of protrusion (Abercrombie and Heaysman, 1954; Gloushankova et al., 1998; Kitt and Nelson, 2011).

NMII-mediated contractility is also important for AJ maintenance. Similar to previous studies (Shewan et al., 2005; Zhang et al., 2005; Smutny et al., 2010; Breyer et al., 2012), we found that inhibition of contractility leads to some loss of cadherins at AJs and formation of intercellular gaps. In contrast with Arp2/3 complex inhibition, we could not detect VE-cadherin at the gap edges after NMII inhibition, suggesting disassembly of cadherin complexes at the cytoplasmic face of the AJ, which makes cadherins susceptible to endocytosis and detergent extraction. A likely mechanism for cadherin complex disassembly is refolding of α-catenin into an autoinhibitory conformation in absence of force and subsequent uncoupling of VE-cadherin from the cytoskeleton.

In the absence of adequate structural data, it was previously assumed that contractile bundles are directly linked to the cadherin complexes (see Introduction). In this context, the linkage of oblique actin-NMII bundles to punctate AJs seems optimal for force transductions because it resembles stress fibers linked to focal adhesions. In contrast, the mechanism by which tangential bundles are linked to AJs and transmit force appeared much less obvious. Our findings that actin-NMII bundles are predominantly connected to VE-cadherin complexes at AJs by branched networks suggest that NMII-mediated contraction is transmitted to cadherin-based adhesion complexes through branched networks. With this geometry, orientation of actin filaments is optimal for efficient application of load to the adhesion complexes. Indeed, it has been shown recently that vinculin forms a stronger catch bond with F-actin when force is directed toward the pointed end of the actin filament (Huang et al., 2017). Because most actin filaments in branched networks are oriented with their barbed ends to the membrane, they are optimally positioned to promote stronger attachment of adhesion complexes to the cytoskeleton.

If branched actin networks act as a force transmitter, they should be physically linked to contractile bundles. We indeed observed that branched networks frequently originate from actin filaments in tangential bundles, which thus serve as “mother” filaments for Arp2/3 complex–dependent nucleation. The “daughter” filaments that originate at an ∼70° angle from these mother filaments continue to branch further in the direction of the AJ, where they apparently make contacts with adhesion complexes. However, if contractile forces are transmitted to AJs by protrusive actin networks, the pushing and pulling forces are expected to cancel each other or at least diminish each other’s contribution to AJ maintenance. A similar problem, however, also exists in lamellipodia. Lamellipodia use a branched actin network to generate protrusion and NMII-dependent contractility to stimulate retrograde flow and form cell–matrix adhesions. This analogy suggests existence of intricate regulation of contractile and protrusive forces at AJs, possibly through their alternation, similar to the oscillatory behavior of protrusion and contraction in lamellipodia (Giannone et al., 2004).

We also propose that besides generating contraction, tangential actin-NMII bundles can improve functionality of branched actin networks at AJs. First, these bundles help to initiate the junctional branched networks by supplying mother filaments for Arp2/3 complex–dependent nucleation. Second, they can serve as a stiff base that limits the retrograde flow of the branched network and thereby enhances the pushing force in the forward direction. Notably, retrograde flow still should exist because branched actin networks polymerize continuously but remain largely stationary in the context of AJs. The retrograde flow of the branched network in this narrow space can be an additional source of pulling force to stimulate adhesion at linear AJs. Finally, the tangential bundle can form a stiff barrier against which the branched network of the opposite cell exerts effective pressure in order to bring VE-cadherins of two cells within an interaction distance. This idea can potentially explain the unrestrained protrusion of overlapping lamellipodia and inefficient AJ formation after inhibition of cell contractility.

Although we conducted this study in endothelial cells, our conclusions can also apply to epithelial cells because molecular mechanisms of AJ formation are significantly conserved between two cell types and largely rely on different paralogs of the same protein families. However, there are also important differences between epithelial and endothelial cells. AJs in epithelial cells are more stable and uniform, whereas endothelial AJs are more dynamic. Epithelial AJs are associated with thicker tangential actin bundles that are thought to be positioned too close to each other to accommodate branched networks in between. Direct investigation of epithelial AJs is challenging because of extremely high density of their cytoskeleton. Future studies may advance our understanding of similarities and differences of the AJ cytoskeleton between endothelial and epithelial cells.

In conclusion, we show that Arp2/3 complex–dependent branched actin filament networks are an essential component of all types of AJs including mature linear AJs. Branched actin networks represent the actual actin cytoskeleton component that is linked to cadherin-based adhesion complexes. We propose that the key role of junctional branched networks is to push plasma membranes of contacting cells against each other and thus maintain transinteraction between cadherins. The contractile actin–NMII bundles that are associated with AJs mechanically stimulate assembly of adhesion complexes by generating pulling force that is transmitted to cadherins via branched actin networks. Additionally, tangential actin bundles serve as a launch pad for assembly of the junctional branched networks and as a structural base that restrains pushing force generated by the branched network and its retrograde flow within a narrow space. Our data provide an important conceptual framework to explain the results of functional experiments that target various components and regulators of different actin cytoskeleton subsets at AJs.

HUVECs (CC-2519; Lonza) were cultured in endothelial cell basal medium (CC-3121; Lonza) with supplements (CC-4133; Lonza) and maintained for no longer than six passages. For experiments, HUVECs were plated on coverslips or glass-bottom dishes (MatTek Corporation) coated with ∼50 µg/ml (5 mg/cm²) collagen from rat tail (354236; BD). Primary mouse monoclonal antibodies were used for cadherin-5 (VE-cadherin; 610251; BD), cortactin p80/85 (clone 4F11; 05-180; EMD Millipore), and NMIIA (ab55456; Abcam). Primary rabbit monoclonal antibody was used for VASP (clone 9A2; 3132; Cell Signaling Technology). Primary rabbit polyclonal antibodies were used for ArpC2 (p34-Arc; 07-227; EMD Millipore), α–E-catenin (3236T; Cell Signaling Technology), and NMIIA (BT-567; Biomedical Technologies). Secondary antibodies and phalloidin fluorescently labeled with Alexa Fluor 488, 594, and 680 were from Invitrogen; secondary anti–rabbit and anti–mouse IgG antibodies conjugated to 12- or 18-nm colloidal gold were from Jackson ImmunoResearch Laboratories, Inc. Blebbistatin (B592500; Toronto Research Chemicals) was prepared from 10-mM stock in DMSO, and Y-27632 (Y100500; Toronto Research Chemicals) was prepared from 10-mM stock in water. For fluorescence microscopy experiments, cells were treated with 50 µM blebbistatin or 50 µM Y-27632 in culture medium for 1 h and then processed for immunofluorescence or PREM. CK666 (SML0006; Sigma-Aldrich) was prepared from 10-mM stock in DMSO and used at a concentration of 100 µM for 40–60 min.

The following plasmids were used in live-cell imaging experiments: GFP–VE-cadherin in adenoviral vector (gift from Y. Komarova, University of Illinois, Chicago, IL) and mCherry-LifeAct in pLL-5.0 lentiviral vector (the vector is a gift from A. Efimov, Fox Chase Cancer Center, Philadelphia, PA; Chia et al., 2016). For lentivirus production, HEK293T cells were cultured in DMEM supplemented with 10% FBS for 24 h before transfection. Cultures at 50–70% confluence were transfected with a mixture of packaging (Pax2), envelope (MD2G), and transfer (pLL–mCherry-LifeAct) plasmids using Fugene6 reagent (Promega). After incubation for 5–6 h with the transfection mixture, the medium was replaced with DMEM with 10% FBS. The medium containing viral particles was collected 48 h later. For infection, HUVECs were incubated for 5–6 h with the virus-containing medium supplemented with 8 µg/ml protamine sulfate (Sigma-Aldrich).

Immunofluorescence staining of HUVECs was performed after simultaneous fixation and extraction with the mixture of 4% formaldehyde (15710; Electron Microscopy Sciences) and 0.3% Triton X-100 in PBS for 15 min. For cortactin staining, cells were extracted with 1% Triton X-100 in PEM buffer (100 mM piperazine-N,N′-bis(2-ethanesulfonic acid)-KOH, pH 6.9, 1 mM MgCl₂, and 1 mM EGTA) containing 2% polyethylene glycol (M_r 35,000), 5 µM phalloidin, and 2 µM taxol for 3 min. After washing with PEM buffer, extracted unfixed cells were incubated with the primary antibody, washed with the PEM buffer, and fixed with 0.2% glutaraldehyde. After quenching with 2 mg/ml NaBH₄ in PBS for 10 min, cells were stained with secondary Alexa Fluor antibody. Coverslips were mounted to slides with ProLong Gold antifade mountant (P36941; Molecular Probes). Light microscopy was performed using an Eclipse TE2000-U inverted microscope (Nikon) equipped with Plan Apochromat 20× 0.75-NA and Apochromat 100× 1.3-NA objectives and Cascade 512B CCD camera (Photometrics) driven by MetaMorph imaging software (Molecular Devices).

For live-cell imaging, cells were maintained at 37°C in humidified atmosphere with 5% CO₂ using an UNO stage-top incubator (Okolab) with 20–30 min allowed for cell accommodation before imaging. Time-lapse sequences were acquired every 10 s for 15 min (Video 1) or every 1 min for 60–70 min (Videos 2, 3, 4, 5, 6, and 7). Vehicle (DMSO for CK666 and culture medium for Y-27632) or drugs were added during a pause between the ninth and 10th frames at the same concentration as for experiments with fixed cells. Imaging was performed using an Eclipse Ti inverted microscope (Nikon) equipped with a CSUX1 spinning disk (Yokogawa Electric Corporation), CFI60 Plan Apochromat Lambda 20× 0.75-NA and CFI60 Apochromat total internal reflection fluorescence 100× 1.49-NA oil immersion objectives, and a QuantEM 512SC digital camera (Photometrics) driven by NIS Elements software (Nikon). Multiple positioning in the x–y plane and single-slice spinning-disk confocal mode with 488- and 561-nm laser wavelengths and FITC and TRITC filters were used for imaging. FIJI software (ImageJ; National Institutes of Health) was used to adjust image contrast and create montages. Gaussian blur filter with 1-pixel radius was applied to Video 1 to reduce noise.

Sample preparation for regular and immunogold PREM was performed as described previously (Svitkina, 2007, 2016). In brief, cells were extracted as described for cortactin staining, washed with PEM, and fixed with 2% glutaraldehyde in 0.1 M Na-cacodylate, pH 7.3. Fixed cells were sequentially treated with 0.1% tannic acid and 0.2% uranyl acetate in water, critical-point dried, coated with platinum and carbon, and transferred onto EM grids for observation.

For immunogold staining, extracted unfixed cells were incubated for 30 min with a primary antibody in PEM buffer containing 5 µM unlabeled phalloidin and 2 µM taxol, washed with PEM buffer, and fixed with 0.2% glutaraldehyde. After quenching with 2 mg/ml NaBH₄ in PBS for 10 min, cells were blocked with 1% BSA in buffer A (20 mM Tris-HCl, pH 8, 0.5 M NaCl, and 0.05% Tween-20), stained with secondary colloidal gold–conjugated antibody, and postfixed with 2% glutaraldehyde.

PREM samples were analyzed using a JEM 1011 transmission electron microscope (JEOL) operated at 100 kV. Images were captured by an ORIUS 832.10W charge-coupled device camera (Gatan) and presented in inverted contrast. Stitching of PREM images was done using the Photomerge tool in Photoshop (Adobe) with the Reposition option. Automatic alignment of layers was manually corrected before applying the Auto-Blend Layers tool. Identification of gold particles in PREM samples was performed at high magnification after contrast enhancement to distinguish them from other bright objects in the samples such as actin filament tips. Color labeling was performed using Photoshop as described previously (Shutova et al., 2012). In brief, gold particles were highlighted using the Brush tool in Photoshop with 50% opacity. Pseudocolors were applied using the Hue/Saturation tool in Photoshop to avoid obscuring the structural details.

For quantification of branch orientation relative to VE-cadherin distribution in linear AJs, actin filament branches were identified and colorized at high magnification to distinguish branches from filament intersections or separating filament pairs. Branch orientation was scored according to three categories—toward, away, or parallel to a line drawn through middle of the VE-cadherin–labeled zone—and expressed as a percentage of all detected branches. For quantification of gold labeling of different actin structures, gold particles were categorized as bound to branched actin networks or unbranched actin filaments, either bundled or unbundled. For quantification of VE-cadherin and Arpc2 immunogold colocalization, 12- and 18-nm gold particles were scored within the visually recognizable junctional branched networks at linear AJs and separately in the rest of the image. Densities of each label were determined by per-unit area occupied by the cytoskeleton in respective cell regions. The area covered by the cytoskeleton was determined using FIJI software by thresholding after increasing the image contrast between the cytoskeleton and the background and applying a mean filter of 20 pixels (∼10 nm) to reduce granularity of the background. Statistical significance was determined by Tukey–Kramer multiple comparisons test after evaluating data distribution normality by Kolmogorov–Smirnov normality test using Instat software (GraphPad Software).

Fig. S1 shows the cytoskeleton of linear AJs associated with tangential and/or oblique actin-NMII bundles. Fig. S2 shows high magnification of gold particles and VE-cadherin immunogold labeling of cell–cell contacts made by overlapping lamellipodia. Fig. S3 shows immunofluorescence and immunogold labeling of AJs with α-catenin antibody. Fig. S4 shows immunofluorescence and immunogold staining of VE-cadherin in confluent HUVEC monolayers. Fig. S5 shows orientation of branched actin filaments at linear AJs. Fig. S6 shows branched actin networks in intercellular bridges. Fig. S7 shows reticular AJs. Fig. S8 shows cortactin immunostaining of control and CK666-treated HUVECs. Fig. S9 shows AJs after treatment with Y27632. Video 1 shows dynamics of GFP–VE-cadherin and mCherry-LifeAct at linear AJs. Videos 2, 3, 4, 5, 6, and 7 show dynamics of AJs after addition of DMSO (Video 2), CK666 (Videos 3, 4, and 5), culture medium (Video 6), or Y-27632 (Video 7).

We thank Dr. Yulia Komarova for the gift of GFP–VE-cadherin construct and Dr. Andrey Efimov for pLL-mCherry vector.

This work was supported by National Institutes of Health grant R01 GM 095977 to T.M. Svitkina.

The authors declare no competing financial interests.

Author contributions: N. Efimova and T.M. Svitkina conceived the project, designed research, performed experiments, analyzed the data, and wrote the paper.