New Insights into the Conformational Activation of Full-Length Integrin

Integrin binding to extracellular matrix proteins is regulated by conformational transitions from closed, low affinity states to open, high affinity states. However, the pathways of integrin conformational activation remain incompletely understood. Here, by combining all-atom molecular dynamics simulation, coarse-graining, heterogeneous elastic network modeling, and experimental ligand binding measurements, we test the effect of integrin β mutations that destabilize the closed conformation. Our results support a “deadbolt” model of integrin activation, where extension of the headpiece is not coupled to leg separation, consistent with recent cryo-EM reconstructions of integrin intermediates. Moreover, our results are inconsistent with a “switchblade-like” mechanism. The data show that locally correlated atomistic motions are likely responsible for extension of integrin headpiece before separation of transmembrane legs, without persistence of these correlations across the entire protein. By combining modeling and simulation with experiment, this study provides new insight into the structural basis of full-length integrin activation.


INTRODUCTION
In order to explain the intrinsically multiscale mechanism of integrin opening, two conceptual 80 models based on experimental observations have been proposed. In the "switchblade" model (8), 81 opening of the βA/hybrid hinge and separation of transmembrane legs occur in a coordinated 82 fashion. In contrast, the "deadbolt" model proposes more conservative changes around the bent 83 structure with progressive loss of constraining contacts between βA domain and β tails that occur 84 before leg separation (23). Thus, without dynamic, nonequilibrium information about how 85 structural changes in the headpiece are coupled to the separation of the legs, major questions 86 remain concerning the pathway of integrin opening.

88
Recently, structures captured in various degrees of opening of full-length α IIb β 3 integrin showed 89 headpiece extension without leg separation (24). However, even in this scenario, it is still 90 possible that electrostatic interactions in the β 3 helix decrease upon headpiece extension and that 91 a coordinated structural change occurs at the interface between the two legs without detectable 92 separation, consistent with coordinated structural reconfiguration between headpiece and legs at 93 short length and time scales. 94 95 In order to understand the relationship between headpiece extension and legs separation, we have 96 combined here a multiscale simulation approach with experimental ligand binding 97 measurements. We tested the effect of activating mutants on the molecular structure and their 98 impact on the long-range structural rearrangements of integrin. Our results support the notion 99 that headpiece extension occurs before legs separation, consistent with a deadbolt model of 100 integrin activation and inconsistent with a switchblade model. This mechanism is mediated by 101 local, correlated atomistic motions within and between neighboring subdomains of the receptor, 102 and independent from long-range interactions. 103 104

RESULTS 105
All atom molecular dynamics simulations of integrin mutants 106 Inspections of root mean square deviation (RMSD) plots of WT, single, and double integrin 107 mutants showed that all structures reached local equilibrium states within 1 µs of MD 108 simulations (Figure 2a,d). Analysis of root mean square (RMS) fluctuations also showed that 109 some of the most flexible regions of integrin are at the interface between the b-propeller and bA 110 domains, together with the Linker 1 and EGF motifs, which form the a and b genu, respectively 111 (Figure 2b-c, e-f). There was no significant difference in RMS fluctuations between WT and the 112 mutant integrins considered here. These data collectively support that the EGF domain region is 113 relatively plastic, especially between EGF1 and EGF2, at the β knee, and at the PSI/hybrid and 114 hybrid/I-EGF1 junctions. It was previously reported that the flexibility of the bA domain would 115 also facilitate such interdomain interactions (15). 116 117 Over the last 100 ns of the AA MD simulations, the average angle < ϑ > of the mutants was 118 about 5-7% different from WT ( Figure 3b); D 12 was enhanced in the S243E mutant relative to 119 both WT and other single/double mutants (figure 3c); D EM was about 15% different for the single 120 mutants and about 8% different in the double mutants relative to WT ( Figure 3d). Accordingly, 121 the single mutants showed enhanced persistency of high values of D EM (Figure 3e). Time 122 evolution of D 12 , D EM , and corresponding probability distributions, are reported in Figure S1 and 123 FigureS2 of the Supplemental Information. Time evolutions of ϑ 1 and ϑ 2 are reported in Figure  124 S3. Both maximum values and standard deviations of ϑ 1 and ϑ 2 over the 1µs-long MD 125 simulations were higher in S243E mutant than the other integrin single/double mutants ( Figure  126 S4a-b). Maximum value and standard deviation of ϑ 1 , over 1µs-long MD, were among the 127 highest for D723R with respect to the other single mutants ( Figure S4a-b). High standard 128 deviations in the kink angles of S243E and D723R with respect to the other single mutants 129 indicate that the configurations of the two headpiece hinges were far from their mean. Thus, the 130 angles were more flexible in these mutant relative to the other single mutants examined. Also, 131 maximum value and standard deviation for D 12 were higher in the S243E and D723R (Figures 132 S4c) relative to the other single mutants. Among the double mutants, D723R-E206T also showed 133 high values of peak and standard deviation for kink angle, leg separation, and headpiece 134 extension ( Figure S4a-d). Taken together, these results support that the mutants here tested 135 destabilize the integrin closed conformation by enhancing variability of kink angles in both 136 subunits, leg separation, and/or overall distance of the headpiece from the transmembrane legs.

138
We next addressed the molecular mechanism by which S243E, one of the most activating 139 mutant, affects integrin conformation. In WT integrin, neutral histidine 244, negatively charged 140 aspartic acid 113, and positively charged arginine 352 surround S243, with Arg352 at a distance 141 (Figure 4a). In MD simulations of systems with protonated glutamic acid in S243E, a salt-bridge 142 formed between the oxygen atom of arginine and the hydrogen atom of glutamic acid ( Figure  143 4b). Salt bridge formation was accompanied by reorientation of the positively charged arginine, 144 which moved closer to the negative charged glutamic acid and re-oriented towards the negatively 145 charged aspartic acid (figure 4b). These molecular rearrangements resulted in greater headpiece 146 extension, flattening of kink angles and separation of the transmembrane legs in the MD 147 simulations. Time evolution and probability density functions of distances of ARG352 from the 148 center of mass of residues 244, 243 and 113 in WT and mutant conditions are in Figure S5. 149

150
Taken together, the analysis of MD simulations showed local destabilization of the closed 151 configuration in integrin mutants. However, structural quantities which are representative of 152 integrin conformational activation at the level of the whole receptor, such as D EM , ϑ 1 and ϑ 2 , did 153 not convergence within 1 µs. In order to allow for enhanced sampling of integrin long-range 154 interactions, we instead used the MD trajectories to build CG models.

156
Coarse-grained (CG) simulations 157 Atomistic integrin structures were converted into CG systems, without the lipid bilayer included 158 explicitly in the CG model, to detect large-scale motions of the receptor. As described earlier, we 159 used a combination of ED-CG and heteroENM to create the CG model and systematically 160 removed harmonic bonds or converted them into Morse potentials, motivated by the assumption 161 that weak CG effective harmonic correlations between domains are likely to dissociate upon 162 integrin activation. We analyzed our results to identify structural differences between WT and 163 mutant integrins. Our goal was to sample multiple integrin states underlying the equilibrium 164 conformers and to provide insight concerning which mutants most potently destabilize integrin 165 closed states. We mapped the atomistic WT system to a CG ED-CG model of 200 CG sites or 166 "beads", with average resolution 8 ± 3 residues per CG site, which is of the same order used 167 previously (32). We compared the CG root mean square fluctuations from ED-CG-heteroENM at 168 different initial cutoffs, spanning 3-5 nm, with those from atomistic simulations converted into 169 CG fluctuations (Figure 5a). Using cutoffs of 3, 4, or 5 nm, the average differences in RMS 170 fluctuations from the all-atom fluctuations were 0.11, 0.12 and 0.16 nm, respectively. We 171 therefore chose a 3 nm cutoff for our CG systems that best reproduced atomistic fluctuations, and 172 built ED-CG-heteroENM models for each mutant integrin ( Figure 5b). The fraction of intra-173 domain springs was about 0.6 in all systems, with inter-domain springs that connected non-174 consecutive subdomains along the primary sequence below 0.2. Intra-domain connections had 175 the highest spring constants, up to 25 kcalmol -1 A 2 (Figure 5d), while inter-domain springs had 176 about 3-fold lower characteristic spring constants, below 8 kcalmol -1 A 2 (Figure 5e-f). Snapshots 177 from a representative CG simulation showed average kink angles between 120-140 deg ( Figure  178 6a), depending on the particular mutant. Theses angles are about twice those from MD, showing 179 that conformations far from equilibrium were sampled with this CG method and that the 180 structures are extended. All mutants showed higher kink angles than WT αvβ3 ( Figure 6a). Also, 181 comparison of the fraction of time that D EM is above 95% of its maximum showed that all the 182 mutants were above 15%, whereas WT integrin was below 10% (Figure 6b). Thus, in the CG 183 simulations, the mutants were in extended conformations more often than WT. By capturing the 184 effect of point mutations, our CG models of integrin were able to enhance sampling 185 conformations from MD trajectories.

187
Integrin affinity measurements 188 Next, we directly measured integrin affinity for the well-characterized Fab fragment, Wow-1, 189 whose RGD-dependent binding to αvβ3 increases dramatically after activation (25). Wow1 is 190 monomeric, minimizing possible effects of integrin clustering on its binding. Cells expressing 191 WT or mutant αvβ3 were incubated with Wow-1 in standard, Mg 2+ -and Ca 2+ -containing buffer 192 or in the presence of Mn 2+ , which is commonly used as a positive control for maximal activation 193 (34). αvβ3-null cells were used as a negative control. We found that in Ca 2+ /Mg 2+ , all of the 194 point mutants showed significantly increased binding compared to WT, with no significant 195 differences detected among the activated mutants ( Figure 7). Further, the activating mutants 196 showed no further increase in the presence of Mn 2+ . This last result implies that all of the 197 mutants are maximally activated in these assays.

199 CG simulations reveal intermediate states 200
In all of our CG simulations, the integrin headpiece extended away from the legs, with the degree 201 of extension depending on the mutant examined and the threshold of removed or converted 202 harmonic interactions. Snapshots from a representative simulation of S243E, with springs 203 preserved for k > 0.1 kcalmol -1 A 2 , are shown in Figure 8. We next computed the RMSD of the 204 simulated systems for all of the single point mutants relative to the four cryo-EM structures (24). 205 In particular, we looked at the configurations of wild type and mutant integrins that were closest 206 to each cryo-EM structure, using the minimum RMSD from CG trajectories, RMSD MIN , and 207 compared how much they differed. Our results showed that for k > 0.001 kcalmol -1 A 2 , the single 208 mutants deviate from the cryo-EM conformer more than the WT in the following cases: Furthermore, the simulations show that this opening mechanism results from weakening low 216 stiffness long-range correlations while preserving local dynamics correlations within each 217 integrin subdomain and between pairs of neighboring subdomains along the primary protein 218 sequence. 219 220 221

223
In this study, we have addressed how integrin headpiece extension coordinates with leg 224 separation during integrin opening. We utilized a novel combination of AA MD simulation, ED-225 CG modeling, and a modified heteroENM approach that allows for large conformational change 226 on single and double point mutants of full-length α V β 3 integrin. We investigated whether 227 destabilization of the closed conformation occurs as localized structural rearrangements or as 228 more cooperative changes in the receptor headpiece and legs. Also, we used ligand binding 229 experiments and comparison of our CG models with single molecule cryoEM integrin 230 reconstructions to validate our modeling configurations. 231 232 The AA MD simulations, albeit of limited large-scale sampling, showed that β3 mutations 233 induce molecular structural rearrangements in both headpiece and legs of α and β subunits. In 234 mutants, these rearrangements are generally enhanced relative to wild type integrins, by initiation 235 of headpiece opening, flattening of the kink angle, and separation of the transmembrane helices 236 ( Figure 3). All mutants here analyzed destabilize the linkers within the α and β subunits, 237 including Linker 1 and the EGF motifs ( within 1 µs ( Figure S1 and Figure S3), the mutants that maximally destabilize the closed state are 244 S243E and D723R ( Figure S4). The hydrophilic surface area is known to be significantly larger 245 for S243E compared to wild type αvβ3 and other mutants. This reflects greater extension of the 246 headpiece away from the lipid bilayer, flattening of both kink angles and initiation of separation 247 of transmembrane helices (Figure 3c and Figure S4). The effect of S243E on the global structural 248 reorganization of integrin is triggered by the formation of a salt bridge at the point of the 249 mutation, which induced local reorientation of a histidine, aspartic acid and arginine ( Figure 4). 250 In the case of the double mutant D723R_S243E, a salt bridge still forms in the βA domain 251 ( Figure S6), but the altered electrostatic interactions induced by mutation of aspartic acid to 252 arginine in the β transmembrane helix and its higher pKa generate more stable interactions 253 within the two helices ( Figure S7).

255
The result that a single mutation in the βA domain of the β subunit can initiate structural opening 256 is consistent with a number of previous MD studies of the integrin headpiece. For example, 257 simulations of fibronectin-bound integrin headpiece showed that the ligand binding pocket at the 258 interface between α and β subunits together with the hinge between the βA and hybrid domain of 259 the β subunit are allosterically linked to initiate opening (18). In MD simulations of both α IIb β 3 260 and α V β 3 integrin headpieces, a common transition pathway for propagation of conformational 261 changes within the βA domain was identified as the precursor of structural opening (36), 262 consistent with our results from modified heteroENM that opening of the βA/hybrid junction can 263 act as a hinge. Molecular simulations of the integrin headpiece were also previously performed 264 in combination with experimental headpiece mutation to show that reorientation of the hybrid 265 domain in the β subunit is required for structural activation (37). Forced unbending of the 266 integrin αvβ3 headpiece was simulated using SMD, which showed that pulling the head readily 267 induced changes starting from the headpiece (38). This supports that headpiece extension is very 268 critical in integrin opening. In the current study, we used the full-length atomistic receptor 269 which, unlike previous efforts, allowed us to characterize motional correlations between 270 headpiece and legs and to further detect the impact of short versus long range correlations on 271 integrin extension. With this study, we were also able to analyze the effect of the D723R 272 mutation in the cytoplasmic tail of the β subunit, showing that propagation of structural 273 activation can also be reproduced in this mutant.

275
In order to set our MD results into a broader context and test whether headpiece extension and 276 leg separation are correlated at longer time scales, we used multiscale CG methods on individual 277 integrins (with effect of lipid bilayer implicitly included in the AA MD input), based on ED-CG 278 and HeteroENM methods, which were developed in (32, 33). This approach reduced the number 279 of integrin sites from 27215 to 200, and reduced the computational cost by >100-fold. We 280 modified our standard heteroENM model so that it can undergo large scale conformational 281 changes by systematically removing low stiffness springs or converting them into softer, 282 dissociable Morse interactions to facilitate realistic conformational flexibility. The predictive 283 validity of standard ED-CG-hENM approaches was therefore significantly extended to enable 284 sampling of multiple integrin conformations outside of the AA MD used to parameterize aspects 285 of the model.

287
Results from the CG simulations confirmed that all mutants destabilized the integrin closed 288 conformation via enhancement of kink angles and persistence of open configurations ( Figure 6). 289 This result was consistent with the experimental finding that WOW-1 binding was maximal for 290 all of the mutants examined ( Figure 7). However, this maximal activation for all mutants 291 obscures possible differences. Dynamic measurements that are sensitive to the kinetics of 292 activation will be an interesting target for future studies.

294
We also compared our αvβ3 CG structures with cryo-EM reconstructions of αIIbβ3 integrins at 295 different states of activation, which had observed headpiece extension before leg separation (24). 296 Integrin extension in the CG models resulted from preserving local atomistic dynamics 297 correlations within each integrin subdomain and between pairs of neighboring subdomains along 298 the primary protein sequence, while removing weak molecular long-range correlations. This 299 suggests that, in the case of limited sampling, certain correlations present in the AA MD 300 simulations were not representative of correlations in the global conformational landscape. 301 302 303

CONCLUDING REMARKS 304 305
Our results from AA MD simulations show that coordinated atomistic motions within and 306 between headpiece and legs destabilize integrin closed conformation on time scales on the order 307 of µs (8). However, at longer time scales, accessed via CG simulations, headpiece extension 308 precedes leg separation. Also, this activation mechanism for αvβ3 integrin is consistent with 309 recent cryo-EM reconstructions of α IIb β 3 integrin (20). In our simulations, extension of the 310 headpiece occurs upon reduction of heteroENM effective harmonic interaction connectivity by 311 maintaining connections within and between consecutive subdomains and modifying the low-312 frequency connections between distant subdomains. This implies that local contacts can persist 313 during integrin opening and that long-range, low-frequency motional correlations are not 314 consistent between closed and extended integrin states. Our model therefore supports the notion 315 integrin extension results from disruption of weak, long-range interactions. In order for the legs 316 to move apart in the model, stronger correlations between non-consecutive subdomains should 317 also be reduced in the CG model, but this would lead to an overall loss of structural integrity and 318 not only legs separation. Stated differently, our model is inconsistent with a switchblade 319 mechanism. 320 321 To conclude, the main findings from this study are: (1)

345
In order to characterize integrin conformational activation, we computationally reconstructed 346 full-length α V β 3 integrins embedded in a lipid bilayer. We mutated the WT conformer using 347 single or double point activating mutations in the βA domain of the β subunit, in the 348 transmembrane β helix or in both. We tested whether initiation of integrin opening is, on the 349 timescale of 1µs, a local effect, involving only the headpiece or the legs, or a global 350 phenomenon, with simultaneous structural rearrangements in both the headpiece and the legs. 351 Then, we built coarse-grained models of each integrin system to sample a wider range of 352 conformations and test correlations between headpiece extension and leg separation over longer 353 times. With both MD and CG simulations, we tested which activating mutations favor structural 354 opening. This result was verified with experiments using the monovalent ligand WOW-1Fab 355 (25), which direct assesses α V ß 3 affinity state. Last, we compared our CG conformations with 356 available cryo-EM reconstructions of α IIb β 3 integrin, which is closely related to αvβ3 (24). 357 358 All-atom simulations 359 To examine the effect of mutation on the conformation of full length integrin a V b 3 , µs length all-360 atom molecular dynamics (AA MD) simulations were performed on the integrins embedded in a 361 lipid square patch (see Figure 1b). We assembled the bent headpiece of a V b 3 integrin [from 362 3IJE.pdb (26)] with its transmembrane helical and cytoplasmic parts [taken from 2KNC.pdb 363 (27)], using homology modeling to reconstruct missing residues (28). Point mutations were 364 selected based on studies that identified mutants that increased affinity for RGD ligands (12). 365 Starting from this initial configuration, we used the VMD software (29) to build five single 366 mutants and four double mutants: D723R, L138I, E206T, S243E, K417E, D723R-L138I, 367 D723R-E206T, D723R-S243E and D723R-K417E. For each integrin, we generated a 368 multicomponent model lipid bilayer with 80% DOPC and 20% DOPS lipids, using CHARMM-369 GUI membrane builder (30). We then removed lipids in the center of the lipid patch to make a 370 hole and inserted the integrin (Figure 1b). Last, we placed the membrane/integrin system within 371 a rectangular box and filled its space with TIP3P water molecules and 150 mM NaCl, for a total 372 of about 1.2 million atoms. In order to reorganize the lipids around the integrin, energy 373 minimization was run for 5000 steps of steepest decent algorithm, followed by 50 ns of position 374 restraint in a constant NPT ensemble, using the Berendsen thermostat. Production AA MD 375 simulation were carried out, using Gromacs 5.0.4 (31), for 1µs in the NPT ensemble using Nose-376 Hoover thermostat and Parrinello-Rahman barostat to keep the temperature at 310 K and 377 pressure at 1 atm. Long-range electrostatic interactions were incorporated through the PME 378 method with a cut-off of 1 nm. The same cut-off value was used for Lennard-Jones interactions.

380
In order to identify differences in headpiece versus leg arrangements between WT integrin and 381 the mutants, we defined metrics that report both headpiece extension and legs separation ( Figure  382 3a). We quantified the following: kink angles, ϑ 1 and ϑ 2 , for a and b subunits, respectively; their 383 average, ϑ; transmembrane legs separation; D 12 , and headpiece extension, D EM (see schematics in 384 Figure 3A). The angle ϑ 1 was defined in the a subunit as the angle between points A1 (center of 385 mass of residues 82-85 in the b-propeller domain), A2 (center of mass of residues 599-602, 386 between tight domain and Linker 1) and A3 (center of mass of restudies 963-966 in the a 387 transmembrane helix); for the b subunit, ϑ 2 was defined as the angle between points B1 (center of 388 mass of residues 236-239 in the bA domain), B2 (center of mass of residues 480-484, between 389 the hybrid domain and the EGF-1/EGF-2 motifs) and B3 (center of mass of restudies 696-699 in 390 the b transmembrane helix); E is a point at the interface between the two headpiece subunits and 391 was defined as the center of mass between points A1 and B1; M was defined as the center of 392 mass between points A3 and B3 (corresponding residues are 963-966 and 696-699 in the two 393 helices); points A and B, whose distance was indicated with D 12 , were given by the midpoints of 394 each transmembrane helix.

396
Coarse-grained model 397 In order to fully become active, integrins must sample multiple intermediate conformational 398 states (24). Many of these conformational states were not accessed by AA MD simulations 399 owing to the relatively short timescales that can be sampled at that level. We therefore built CG 400 models based on the observed motional correlations of atoms in MD simulations and used them 401 to identify structural differences between the WT and mutants on effectively longer timescales. 402 We first developed Essential Dynamics Coarse-graining (ED-CG) (32) and heterogeneous elastic 403 network (HeteroENM) models (33) of each integrin starting from the AA MD trajectories 404 (without explicit inclusion of the lipid membrane in these model-it is there in the AA MD data, 405 however). The ED-CG approach was chosen in order to select CG sites which preserve 406 independent motion in the CG protein, and because it is constructed from the primary protein 407 sequence without distorting it when exploring a wide conformational space. The heteroENM 408 approach was used to create effective harmonic interactions between the CG sites which directly 409 capture nanoscale correlations from the AA MD simulations. This approach was critical for 410 creating CG models that maintain molecular differences between the integrin mutants studied 411 here. changes. We either systematically removed a varying fraction of the effective harmonic 422 potentials or converted some of the inter-domain harmonic interactions into "softer" Morse 423 potentials. In particular, we modified only those inter-domain interactions between non-424 consecutive subdomains along the subunits in order to maintain connections along the primary 425 sequence of the proteins. We tested conditions where harmonic interaction potentials with 426 equilibrium stiffness k < 0.0005-0.1 kcalmol -1 A 2 were modified, using 200 CG sites and an 427 enforced cutoff 3 nm for each integrin. By modifying a fraction of harmonic potentials with k 428 below 0.0005 kcalmol -1 A 2 , no significant structural reconfiguration of the receptor was observed. 429 By increasing the upper limit above 0.