Membrane-bound model of the ternary complex between factor VIIa/tissue factor and factor X

Blood clotting is triggered when the integral membrane protein, tissue factor (TF), is exposed to blood after vascular damage or pathologic TF expression.¹ TF binds factor VII (FVII) and promotes its conversion to the active serine protease, FVIIa, which in turn recognizes membrane-bound FX as its favored substrate, activating it to FXa (Figure 1A-B). FXa propagates the clotting cascade, leading to fibrin deposition and platelet activation. Thus, formation of the cell-surface TF:FVIIa:FX Michaelis complex, termed here the extrinsic complex (EC), is the event that triggers clotting in normal hemostasis and many thrombotic diseases.² In this study, used molecular dynamics (MD) simulations to develop, to our knowledge, the first atomic structure of the EC on the surface of a phospholipid bilayer.

X-ray crystal structures are available for the isolated TF ectodomain, termed soluble TF (sTF), and also for the sTF:FVIIa complex.^7,8 However, none include the phospholipid membrane, which recruits clotting proteins to the surface of activated or damaged cells and greatly enhances their enzymatic activities.⁹ Furthermore, no structures of the EC have yet been solved using X-ray crystallography or cryogenic electron microscopy, with or without the membrane. Understanding the ternary structure of the EC, and in particular how membrane bilayers containing anionic phospholipids modulate the formation and activity of the EC, would provide critical insights into the triggering of clotting in normal hemostasis and in thrombosis.

Mutagenesis has identified FVIIa and TF residues that contribute to recognizing FX as a substrate, providing clues as to how FX interacts with TF and FVIIa within the EC. Thus, mutating certain residues in the serine protease domain (SPD) of FVIIa (FVIIa-SPD) diminishes FX activation without affecting assembly of the TF:FVIIa complex.^10,11 Strikingly, mutagenesis experiments have also defined a putative substrate-binding “exosite” region on TF (comprising K165, K166, and nearby residues; highlighted in Figure 1C), which is located far from the FVIIa catalytic site but near the membrane interface. It contributes strongly to FX activation but plays no role in FVIIa binding to TF.^4,12-14 Furthermore, anti-TF monoclonal antibodies whose epitopes overlap the TF exosite strongly block FX activation while not interfering with FVIIa binding to TF.^15-18 A number of approaches have implicated FX’s γ-carboxyglutamate-rich (GLA) domain as the part of FX that interacts with the TF exosite.^5,14,19,20 Additional evidence also implicates the FVIIa-GLA domain as contributing to FX recognition within the EC.^5,14,21 The role of the first epidermal growth factor (EGF)–like domain of FX (FX-EGF1) in binding to TF:FVIIa is somewhat less clear.^19,22,23 

In 2003, 2 models of the EC were developed in protein-protein docking studies in the absence of membranes.^24,25 Multiple studies have shown, however, that protein-membrane interactions involving the GLA domains of FVIIa and FX determine the orientation and positioning of these proteins on the membrane surface,^4,26,27 and are therefore highly likely to make major contributions to the interactions between FVIIa, FX, and TF within the EC complex.

MD simulations can provide concurrent spatial and temporal resolutions that cannot be achieved experimentally or through rigid-body docking. MD previously has enabled the identification of protein–phospholipid interactions for several coagulation proteins,^26-29 including FVIIa-GLA²⁶ and FX-GLA,²⁷ and has also enabled the characterization of membrane interactions and dynamic behavior of the TF:FVIIa complex.⁶ Combined with nonequilibrium techniques, the method has successfully captured conformational changes in complex membrane proteins such as channels and transporters.^30-34 Custom membrane models have been developed to overcome traditional difficulties in membrane simulations, including the use of the highly mobile membrane-mimetic (HMMM) membrane modeling approach that our laboratory has pioneered.^27,28,35-55 

We now report the atomic structure of the EC assembled on the surface of a phosphatidylcholine (PC)/phosphatidylserine (PS) membrane lipid bilayer, using specialized MD simulations. This is, to our knowledge, the most detailed and complete model of the EC to date, including details of TF exosite interactions with residues in the FX- and FVIIa-GLA domains.

Our strategy for constructing the atomic model of EC was first to generate information on the interactions between the individual domains of the protein components, either from docking of individual domains from available experimental data or from mechanistic implications, and then to use these interdomain interactions as targets to guide a simulation in which membrane-bound models of FX and of sTF:FVIIa are brought together while they remain bound to the membrane. More details on construction and equilibration of the models and their analysis are provided in supplemental Data. Here, we briefly mention the main steps in our approach.

The full-length sTF:FVIIa complex was previously modeled.⁶ In this study, full models of FX and sTF:FVIIa were constructed in their bound form to a 50:50 PC:PS HMMM³⁵ membrane, starting from membrane-bound FVIIa-GLA^26,35 and FX-GLA,²⁷ and aligning the sTF:FVIIa and FX X-ray structures. Residue numbering here is based on the polypeptide sequences of the mature proteins (Figure 1A). Each full model was simulated for 100 nanoseconds.

Two sets of rigid-body protein-protein docking using DOT2⁵⁶ were performed to evaluate binding of various components of the EC: (1) FX-SPD/EGF2 docking to FVIIa-SPD/EGF2, and (2) FX-GLA docking to sTF:FVIIa (supplemental Figure 1). FX and sTF:FVIIa structures used for docking were derived from the 100-nanosecond simulations. Each protocol was repeated 5 times, with 1 000 000 docked structures each. Docked structures were filtered, and only those structures satisfying experimental constraints were retained.

Simulations for this part were guided (biased) by specific coordNum collective variables (CVs; see supplemental Data) to induce binding of FX and sTF:FVIIa on the surface of a membrane in steps (Figure 2). The system was guided using moving harmonic potentials.

After the EC formed on the membrane, all biases were phased out over a 50-nanosecond span, after which the EC was simulated freely for an additional 200 nanoseconds. To ensure that the final sTF:FVIIa:FX ternary complex was not unduly affected by the use of HMMM, all systems were converted to conventional (full-tailed) membranes, using the last frame of each 200-nanosecond HMMM simulation in which the EC was formed. Each EC complex was further simulated for another 150 nanosecond in the full membrane.

All MD simulations were performed with NAMD^57-59 and the CHARMM36 force-field parameters for lipids⁶⁰ and proteins.⁶¹ Parameters for GLA residues were those previously developed in our laboratory after atom types were renamed from CHARMM27 to CHARMM36.²⁶ Parameters for the special residue BHD (β-hydroxyaspartate) were developed by analogy based on aspartate and serine parameters in CHARMM36. System setup, visualization, and analysis were carried out in VMD (Visual Molecular Dynamics).⁶² 

Our model of the EC was developed using specialized MD simulation driven to form the EC on the surface of a PC/PS bilayer. We used a novel protocol guided by protein-protein docking results as an approximate target for nonequilibrium MD, performed on the surface of a specialized HMMM membrane model with enhanced lipid dynamics.^35,63 Formation of the EC was slowly induced on the bilayer surface using target-docked poses filtered according to the best experimental information available. These were guided particularly by interactions between the TF exosite and the FX-GLA domain, which, as discussed in “Introduction,” are strongly supported by experimental data.

We initially constructed multiple full-length models of separate sTF:FVIIa and FX on a 50:50 PC:PS bilayer. Our previous work using nanodiscs to quantify FX binding to membranes showed that each GLA domain binds to a cluster of 6 to 8 PS molecules,^64,65 and also that Ca²⁺ induces PS clustering.⁶⁶ For this reason, we used 50:50 PC:PS bilayers to mimic PS-rich membrane nanodomains that favor binding of GLA domains. Each model was simulated for 100 nanoseconds, and 5 snapshots from time (t) = 60 to 100 nanoseconds were selected as starting conformations for the 5 EC-formation simulations and for rigid-body docking calculations. Construction of the individual FX and sTF:FVIIa models is described in detail in the supplemental Information. Our sTF:FVIIa complex model is constructed similarly to our previous membrane-bound sTF:FVIIa model,²⁶ except that a 50:50 PC:PS HMMM was used. Construction of the FX model required particular care, because no complete FX or FXa structure is currently available. Our FX model is membrane-bound and near complete, lacking only the activation peptide (S143-R194, corresponding to residues 183-234 of the zymogen sequence; Figure 1A). A partial FXa structure (PDB: 1XKA⁶⁷) was used to help construct our model, which has an activated SPD with the activation peptide cleaved off. We determined that the available loop-modeling methods would not be sufficient to accurately represent the activation peptide’s conformation, especially considering its extensive N-/O-linked glycosylation. We therefore included a truncated activation peptide, adding only the P1 residue, R194, then docked that into the S1 pocket of FVIIa.⁶⁸ 

Both FX and sTF:FVIIa models were relaxed for 100 nanoseconds, with characteristics assessed as described in the supplemental Information. We confirmed that the binding poses of both FVIIa-GLA and FX-GLA were stable, that the orientations of FX-GLA and sTF:FVIIa were consistent with our previous studies, and that a sufficient variety of conformations of FX would be represented (see supplemental Information for extended discussion).

Five independent biased simulations were performed using different starting configurations obtained from snapshots of FX and sTF:FVIIa taken at 10-nanosecond intervals from t = 60 to 100 nanoseconds of membrane-bound simulations (see supplemental Information for a detailed description). First, the simulations were used to bring FX-GLA near the TF exosite using both distance and contact CVs. Nonequilibrium work and CV values were assessed during these steps. Both distance and contact CVs progressed linearly toward their targets. Once the distance CV brought FX-GLA to a distance sufficiently close for the contact CV to take effect, the distance CV was gradually eliminated between t = 50 and 60 nanoseconds (Figure 2).

During t = 75 to 100 nanoseconds, FX-SPD is rotated so that the groups on the biased domains have a favorable alignment and an unobstructed line of sight for binding. During t = 100 to 125 nanoseconds, interactions are established between FX-SPD and FVIIa-SPD using both contact and distance CVs. Contacts between FX-SPD and FVIIa-SPD were gradually increased over t = 100 to 150 nanoseconds, toward a target value of 5. This CV reached a value of 4 to 4.5 in 3 of the 5 simulations, and ≤3 in the other 2. In t = 125 to 150 nanoseconds, distance and contacts are maintained whereas 3 other CVs are used to dock FX-SPD residue R194 to the FVIIa-SPD S1 pocket.

During this time, the number of contacts is restrained for the GLA-exosite interaction, fluctuating within 0.5 of its target value. Distance is not restrained for the FX-GLA during this period, and gradually converges to a similar value for all simulation copies except the 1 in which FX-GLA gradually disassociated from TF. No further simulations were performed with this copy, whose analysis showed that FX-GLA was bound to a totally different location from the stably bound complex copies. The difference in binding mode between stable and unstable complexes may be because of a difference in the initial orientation of TF with respect to the membrane surface. The initial value of θ_CM for TF in the unstable simulation copy is higher by ∼10° than any other TF orientation used. During unbinding, θ_CM decreases from 110° to 90°. It may be that the increased tilt allowed an FX-GLA interaction that was not consistent with a stable EC.

The orientations of each GLA domain, as well as the number of lipids within 5 Å of FX-GLA were also tracked during the biasing simulations. By the end, all tilt angles are within the 90° to 120° range as expected based on our previous simulations.²⁷ 

Fluctuations are seen in the orientations of individual domains but by the end all FX-GLA domains have orientation values within 10° relative to their respective initial orientations. The number of lipids in the vicinity of FX-GLA is relatively constant throughout the simulations, with an average of 6 lipids within 5 Å of FX-GLA C_α atoms and Ca²⁺ ions. This is consistent with variation in the last 10 nanoseconds of our equilibrium simulation of isolated FX-GLA.²⁷ Thus, orientation and binding mode of FX-GLA after association with sTF:FVIIa are consistent with previously characterized FX binding orientations.

After gradually removing all biases from the simulations over the course of 50 nanoseconds, each complex was equilibrated for an additional 200 nanoseconds in HMMM membranes (short-tailed lipids), which were then converted to full membranes (conventional, full-tailed lipids) and simulated for another 150 nanoseconds. Values of contact and distance CVs for FX-GLA and the exosite remained stable and close to their nonequilibrium simulation target values: the distance between SPD domains remained within 5 Å of its target value, whereas its contact number showed only a small drop (2-3) in each simulation. Some residues even increased their contact number after the equilibrium simulations.

Throughout the simulation trajectory, FVIIa-SPD residues in contact with FX-SPD included V154, R290-G291, S320, D338-K341, and W364-V371. Of these, only R290, K341, and Q366 were included in the pulling simulation. FX-SPD residues in contact with FVIIa-SPD included R194-G197, Q199, K202, T327, H328, R332, R336, E341, Y367, K370-E372, and R386. Of these, only I195 and R332 were part of the pulling protocol.

Orientations of FX-GLA, FVIIa-GLA, and TF were assessed during the last steps of the protocol and were consistent with orientations from the end of their equilibrium simulations (within 20°). Figures tracking these orientations, as well as number of lipids in the vicinity of FX-GLA and FVIIa-GLA domains, are provided in supplemental Information.

The final model of the EC assembled on a 50:50 PC:PS bilayer with full-length lipids is shown in Figure 3. One of the most striking features is that the sTF:FVIIa complex interacts with FX primarily via the latter’s protease and GLA domains; essentially all of the rest of the FX light chain (ie, both EGF-like domains) are in the solvent, located away from sTF:FVIIa. Thus, the FX protease domain interacts exclusively with the FVIIa protease domain (∼825 Å² interfacial surface area or ∼600 Å² discounting R194; plus ∼200 Å² interfacial area contributed from the accompanying peptide C-terminal to EGF2), whereas the FX-GLA domain interacts with both the TF exosite region (∼400 Å² interfacial area) and with the FVIIa-GLA domain (∼20 Å²). As discussed below, this orientation of FX within the EC is a departure from previously published models of the EC, which were assembled in the absence of the membrane surface.

At the end of the simulation (totaling 550 nanoseconds), the P1 residue of FX (R194) was positioned within the S1 pocket of FVIIa (Figure 4A), near the catalytic triad residues H193, D242, and S344 (equivalent to H57, D102, and S195 in chymotrypsinogen numbering). Additionally, residue D339 (D189 in chymotrypsinogen numbering) at the base of the primary specificity pocket was ∼2 Å from R194. Figure 4B shows a 90° rotated view of the interaction between the 2 protease domains, showing that FVIIa residue R290 (in loop 140s) may interact with FX residue E341, emphasizing the critical role of this loop in FVIIa-mediated FX activation.^69,70 We note, however, the following limitations of this model concerning the interactions between FVIIa and FX SPDs. The model is based on the membrane equilibrated crystal structures of sTF:FVIIa⁷ and FXa,⁶⁷ which implies the assumption of insignificant conformational changes in the FX-SPD after the activation cleavage. However, the FX activation peptide linking FX-EGF2 to FX-SPD is not modeled because of the absence of a reliable template. Yet, the precleaved activation peptide should provide additional interactions to FVIIa-SPD, for example, occupying the S2, S3 pockets, etc. Moreover, the N-terminus of the mature FXa (residue I195) is folded inside a hydrophobic pocket of the SPD, suggesting this pocket would be an exposed cleft in the zymogen.

Moving to the other main area of contact, the FX-GLA domain makes stable contacts with the TF exosite region and, to a lesser extent, with the FVIIa-GLA domain. Specifically, FX-GLA residues make significant contacts with TF residues S162, S163, G164, and K166 (Figure 5). Of these, TF residues S163, G164, and K166 are experimentally identified to impair FX activation when mutated.^4,74 All these residues show greater number of contacts at the end of the simulation (Figure 6), supporting a proper initial positioning through the biased simulations so that FX-GLA spontaneously develops more contacts with TF later on. The increased TF-FX contacts during the unrestrained simulations agree with a previous model and mutagenesis data.²⁸ K79 of the FX-EGF1 and A51 in FVIIa show no interactions in the final complex, although the position of E51 in FX is not in line with the previous model.²⁵ 

TF residues K165 and K166 are the 2 canonical residues whose mutagenesis originally defined the TF exosite.¹⁵ Further mutagenesis studies⁷ identified additional critical TF residues (Y157, K159, S163, G164, and Y185) as important in supporting FX activation, thereby extending the definition of the TF exosite region. We previously showed that the length of the nearby 4 × Ser loop in TF (S160-S163) is crucial for FX activation,⁷⁴ implicating its participation in the TF exosite as well.

The TF exosite region makes stable contacts with several residues in the FX-GLA domain and, to a lesser extent, with residues in the FVIIa-GLA domain (summarized in supplemental Table 3). FX-GLA residues F31 and γ32 make significant contacts with TF residues S163, G164 (backbone), and K166 (Figure 5). In particular, FX-GLA residue γ32 interacts with TF residues K159, K166, S163, and Y185, with K159 and K166 being the closest (Figure 5; supplemental Table 3), suggesting a pivotal role for this FX-GLA residue in TF:FX interaction. Additionally, FX residue D33 contacts TF K166, whereas FX residue F31 interacts with G164 backbone in TF. FVIIa residue R36 makes a salt bridge to FVIIa residue D33 and a hydrogen bond to the backbone carbonyl of TF residue S160 (part of the 4 × Ser loop). Interestingly, although mutation Y157A of TF significantly decreased FX activation,⁴ Y157 makes no contact with FX or FVIIa and is positioned behind K166 (4.2 Å away; Figure 5B), indicating an indirect, more structural role for this residue. Thus, mutating Y157 likely affects the conformation of nearby exosite residues (previously mentioned) that directly interact with FX.

Although most of the known TF exosite residues interact with the FX-GLA domain, the side chain of K165 projects nearly 180° away from that of K166 and interacts with the FVIIa-GLA domain instead. Specifically, TF residue K165 forms electrostatic interactions with the negatively charged residue γ35 in FVIIa (Figure 5A). Additionally, FVIIa-GLA residue D33 interacts with TF 4 × Ser loop residue S162 (Figure 5). Overall, the FX-GLA domain exhibits more extensive contact with the TF exosite whereas the FVIIa-GLA domain has only minor interactions, and there appears to be minimal direct contact between the GLA domains of FVIIa and FX. Instead, TF exosite residues (Y157, K159, K165, K166, Y185, and S163), along with S162 from the TF serine loop, seem to form a bridge that brings both GLA domains closer. K166 participates in at least 2 interactions (with FX), whereas the other exosite residues are involved in at least 1 interaction each. This may account for the particularly significant role of K166. Interestingly, we observed that the TF 4 × Ser loop defines an interface, with TF K165 engaging FVIIa-GLA on 1 face and FX-GLA on the other (Figure 5).

In summary, we now identify a binding surface for FX within the EC, consisting of residues in the C-terminal portion of the FX-GLA domain. We propose that these residues bind to the TF exosite in a “double-zipper” formation of alternating electrostatic and hydrophobic interactions close to the surface of the lipid bilayer. This interaction mode is consistent with previous experimental studies.^14,75 

No experimental structure for the EC complex has been reported to compare our model with. There are however 2 previous computational models of the EC using protein-protein docking, both in the absence of membrane: 1 with minimal MD refinement (3 nanoseconds), and the other accompanied with mutagenesis data.^24,25 In a later computational study,⁷⁶ the model by Norlege et al²⁵ was used as a template to build an EC model and study the effect of membrane cholesterol and posttranslational palmitoylation on its dynamics with MD simulations, concluding that the presence of cholesterol and palmitoylation may contribute to structural rigidity, stability, and compactness of key domains of the EC by augmenting protein-protein and protein-lipid interactions. In addition, a cryogenic electron microscopy structure for another coagulation protein complex, the prothrombin-prothrombinase complex,⁷⁷ provides relevant insights when assessing our model.

In the Norledge et al²⁵ model, FX forms an extended conformation, with all 4 of its domains making extensive interactions with TF and FVIIa, whereas the GLA and EGF-like domains of FVIIa are not significantly involved in the interaction. Our model shows a very different pattern of interactions in that most of the FX light chain isn’t docked onto the sTF:FVIIa complex. Instead of an extensive contact surface between FX and TF/FVII, we observe contacts limited primarily to the membrane-proximal domains and the catalytic domains (Figure 3). This “loose” mode of interaction is reminiscent of the structure reported for the prothrombin-prothrombinase complex.⁷⁷ As opposed to an extensive contact interface, a loose structure can be more readily formed because of thermal motions of the individual elements on the surface of the membrane. This binding mode can be rationalized in terms of enzyme/substrate interactions. FX binds to TF:FVIIa in order to be recognized as its substrate, but if the binding interface were very extensive, the tightly bound product may have an unacceptably slow off rate. Therefore, this loose mode of interaction may help allow the product, FXa, to egress from the EC more readily.

The other model by Venkateswarlu et al²⁴ was briefly (3 nanoseconds) relaxed in solution after protein-protein docking, which also showed an extensive contact interface severalfold larger than in our model. Although a structure is not publicly available, it is clear from their figures that there is a significant misalignment between the lines of Ca²⁺ ions in FX-GLA and FVIIa-GLA. Ca²⁺ ions are shown to “glue” the GLA domains of coagulation proteins to the membrane surface.^26,27,78 These ions in FX and FVIIa must both align with the membrane surface to accomplish their tasks, an aspect that could not be captured in the reported model because of missing the explicit membrane.

X-ray crystal structures have been published for sTF bound to Fab fragments of 2 monoclonal anti-TF antibodies while retaining their FVIIa binding.^17,79,80 One of the antibodies (10H10) is completely noninhibitory, with the rate of FX activation unaltered by antibody binding to TF. Thus, any plausible EC model must accommodate 10H10 bound to TF without significant clashes to FX. The other antibody (5G9) allows FVIIa to bind but completely blocks FX activation. In the crystal structure of the complex of 5G9-Fab with sTF,¹⁷ the 5G9 epitope overlaps with the TF exosite. These sTF-5G9¹⁷ and sTF-10H10^79,80 complex structures are instructive for validating our model against the previous models. Figure 7A-B shows our sTF:FVIIa:FX model and that of Norledge et al (1NL8),²⁵ respectively. Superimposing our model (Figure 7C) or the Norledge model (Figure 7D) with the sTF-5G9 structure reveals, as predicted, strong steric clashes between FX and 5G9-Fab. In our model, the clash primarily involves the EGF-like domains of FX, whereas in the Norledge model, both the GLA and EGF1 domains of FX clash with 5G9-Fab. In contrast, the 2 models differ greatly with regard to steric clash with 10H10-Fab. Our model readily accommodates the simultaneous binding of both FX and 10H10-Fab to sTF:FVIIa without any steric clashes (Figure 7E). Although no structure is available for the model developed by Venkateswarlu et al,²⁴ the superposition of 10H10-bound sTF with the model of Norledge et al (1NL8)²⁵ results in substantial steric clashes between FX and the 10H10-Fab (Figure 7F). This includes a large overlap between FX-SPD and 10H10, and a lesser but still substantial overlap between the EGF-like domains of FX and 10H10. Thus, the location of FX in the Norledge model conflicts with the published sTF-10H10 structure, whereas the location of FX in the present model does not.

The authors acknowledge support from National Institutes of Health, National Institute of General Medical Sciences awards P41 GM104601, R24 GM145965, and R01 GM123455, and National Institutes of Health, National Institute of General Medical Sciences awards R35 HL135823, and R35 HL171334. M.P.M. acknowledges support of the National Heart, Lung, and Blood Institute (award F31 HL136155).

The content is solely the responsibility of the authors and does not necessarily represent the official views of the National Institutes of Health.

Contributions: M.P.M., J.H.M., and E.T. conceived the project; M.P.M. performed the modeling, docking, and simulations; P.-C.W., A.M., and J.C.S. assisted with preparation of figures and contributed to the results and discussion; and all authors contributed to the writing and revision of the manuscript.

Conflict-of-interest disclosure: The authors declare no competing financial interests.

Correspondence: Emad Tajkhorshid, Department of Biochemistry, Center for Biophysics and Quantitative Biology, University of Illinois Urbana-Champaign, 318 S Mathews Ave, Urbana, IL 61801; email: emad@illinois.edu; and James H. Morrissey, Department of Biological Chemistry, University of Michigan Medical School, 4301B MSRB III, 1150 West Medical Center Dr, Ann Arbor, MI 48109-5606; email: jhmorris@umich.edu.