Decrypting Integrins by Mixed-Solvent Molecular Dynamics Simulations

Integrins are a family of α/β heterodimeric cell surface adhesion receptors which are capable of transmitting signals bidirectionally across membranes. They are known for their therapeutic potential in a wide range of diseases. However, the development of integrin-targeting medications has been impacted by unexpected downstream effects including unwanted agonist-like effects. Allosteric modulation of integrins is a promising approach to potentially overcome these limitations. Applying mixed-solvent molecular dynamics (MD) simulations to integrins, the current study uncovers hitherto unknown allosteric sites within the integrin α I domains of LFA-1 (αLβ2; CD11a/CD18), VLA-1 (α1β1; CD49a/CD29), and Mac-1 (αMβ2, CD11b/CD18). We show that these pockets are putatively accessible to small-molecule modulators. The findings reported here may provide opportunities for the design of novel allosteric integrin inhibitors lacking the unwanted agonism observed with earlier as well as current integrin-targeting drugs.


■ INTRODUCTION
Integrins are a 24-membered family of α/β heterodimeric cell surface receptors which are expressed in cell lineage defined arrays throughout the organism. They mediate cell adhesion, migration, differentiation, and proliferation. The activity of integrins is regulated by coordinated global conformational rearrangements. Signals from inside the cells convert integrins from their inactive bent conformation to their active, extended conformation, which allows ligands to bind (inside-out signaling). Conversely, ligand binding stabilizes the active integrin state and conveys signals back into the cell (outside-in signaling). 1,2 This bidirectional signaling capability is unique to integrins and enables cells to dynamically respond to microenvironmental changes. 3 Integrins can be categorized into two major subfamilies based on the presence or absence of a globular domain inserted in the top part of the integrin α subunit, termed α I domain. Nine out of the 18 known integrin α chains contain an α I domain ( Figure 1). If present, this domain serves as the primary binding domain for integrin ligands. Located at the upper surface of the α I domain a single divalent cationbinding site, the so-called MIDAS (metal ion-dependent adhesion site) has been demonstrated to directly interact with integrin ligands. At the distal bottom face, the C-terminal α 7 helix of the α I domain is central to integrin affinity regulation. A downward axial displacement of the α 7 helix determines the transition of the α I domain from a low-affinity state to a high-affinity state, thereby enhancing its ligand binding affinity by up to 10,000-fold. 4 Integrins play crucial roles in diseases of high unmet medical need including cardiovascular disease, thrombosis, inflammation, and cancer, rendering them attractive therapeutic targets. 6 6 β-strand (orange), helices α 1 (pink) and α 7 (red). The classical α I allosteric site (herein also referred to as β 6 −α 7 pocket) and the newly identified β 6 −α 7 −α 1 pocket are shown in gray. substantial challenges reiteratively defeating therapeutic expectations. 8−10 A particularly profound challenge with integrin inhibition is the potential of unwanted agonist-like effects (i.e., effects opposite to the intended effects). Such unwanted partial agonism has been mostly observed with ligand mimetics interacting with the native ligand binding site of the integrin. The unintended triggering of these agonist-like effects is considered a main reason for reiterative late clinicalstage development failures of integrin-targeting interventions. 8,9,11 Allosteric integrin modulation stabilizing integrins in the desired affinity state holds the promise of resolving this pharmacological challenge. The feasibility of allosteric integrin modulation by small molecules was demonstrated early for the α I domain-containing integrin lymphocyte function-associated antigen-1 (LFA-1, α L β 2 , CD11a/CD18). 12,13 The respective allosteric site (termed α I allosteric site or lovastatin (L)-site) has been discovered by serendipity. It resides within the α I domain of LFA-1 and is located underneath the C-terminal α 7 helix of the α I domain (Figure 1). Small molecules binding to this site inhibit LFA-1 function by locking LFA-1 in the inactive bent conformation. 4 To date, chemically diverse smallmolecule LFA-1 inhibitors have been described, which are active in the low nanomolar range in vitro assay systems and efficacious in vivo in experimental disease models. 14−17 Moreover, as predicted from their mechanism of action, these allosteric inhibitors do not induce unwanted agonist-like effects observed with previous integrin-targeting modalities. Further, they display high selectivity for LFA-1 versus other integrins, differentiating them from most orthosterically acting integrin inhibitors. 17−19 Setting out for the current study, we hypothesized that α I allosteric sites similar to the LFA-1 α I allosteric site may be accessible in other α I domain-containing integrins. For the identification of such therapeutically employable allosteric sites, several computational and experimental methods are available to date. 20 For our study, we employ mixed-solvent molecular dynamics (MD) simulations, which use small probes (mostly hydrophobic organic molecules) mixed with water to detect cryptic sites in proteins. The application of this method has several advantages: (1) it allows for protein flexibility, (2) it takes competition between the probe and water into account, and (3) it is validated. 21−25 In particular, a novel cryptic pocket with druggable properties was identified in the SARS-CoV-2 spike glycoprotein via mixed-solvent MD simulations and validated via hydrogen−deuterium exchange mass spectrometry experiments. 23 When combined with virtual screening, better performance in docking to ensembles extracted from mixed-solvent simulations was seen. 21 The druggability of the major pocket of the nonstructural protein 1 of SARS-CoV-2 (cryptic) was recently demonstrated by mixed-solvent MD simulations and validated by fragment-based screening via Xray crystallography, 24 further strengthening the potential of the approach.
As integrin model systems for the mixed-solvent simulations, we chose the α I domains of the integrins LFA-1, macrophage-1 antigen (Mac-1, α M β 2 , CD11b/CD18), and very late antigen-1 (VLA-1, α 1 β 1 , CD49a/CD29). All three integrins have been associated with autoimmune and malignant diseases. 26,27 Thus, the search for novel, potentially targetable allosteric sites within their α I domains holds substantive therapeutic promise. Their α I domains have been extensively characterized at molecular levels. 2,28,29 Structurally, the three α I domains consist of six βstrands and seven α-helices which form a compact conformation (LFA-1: PDB ID: 2ICA; 5 Mac-1: PDB ID:1NA5 30 and VLA-1: PDB ID 1PT6 31 ). The overall structure is conserved with some changes in the loops, particularly in the G283 to T292 loop of VLA-1, which is longer than the The pink, blue, and white backgrounds indicate residue similarity, conservation more than 50%, and no conservation, respectively, across the three systems. (b) Average root-mean-square deviations (RMSD) of α 7 with respect to the rest of the I domain. The reference structure is the closed conformation. For LFA-1, the reference structure is a representative closed conformation post-equilibration. First structural alignment of the individual snapshots saved along the MD simulations is carried out on the Cα-atoms of all residues except those comprised in α 7 , i.e., F292-I306, F302-I316, and E317-I331 for LFA-1, Mac-1, and VLA-1, respectively. Then for each MD snapshot, the α 7 Cα-RMSD is calculated as = r r ( ) corresponding ones of LFA-1 and Mac-1. This loop is not in contact with β 6 −α 7 segment, where inhibitors were previously identified for LFA-1. 5 The sequence alignment focusing on the β 6 −α 7 segment reveals partial conservation of the residues in the individual sequences (blue background in Figure 2a). Essentially, the previously identified allosteric pocket is stabilized by the E284-K305 salt bridge in LFA-1 and D294-K315 in Mac-1. The analogue contact in VLA-1 is K309-R330, which can form hydrogen bonds between the K309 backbone oxygen atom and the R330 side chain. By applying mixed-solvent MD simulations to integrins, we report here the opening of α I allosteric sites analogous to the classical α I allosteric site of LFA-1 in the integrins Mac-1 and VLA-1. Moreover, the MD simulations identify a novel pocket in all three integrin α I domains studied (termed α 6 −β 7 −α 1 pocket) shown to be accessible to small molecules by molecular docking studies. Potential implications of these findings for the design of next-generation integrin-targeting drugs are discussed.

α I Allosteric Site Closes in the Absence of Inhibitors and Becomes Flexible in the Presence of Fragment
Probes. To identify new α I allosteric pockets and to implicitly probe the effects of hydrophobic compounds on their structure and dynamics, we used the α I domains of the integrins LFA-1, Mac-1, and VLA-1 as model systems. We started from existing crystal structures of the selected α I domains (LFA-1: PDB ID: 2ICA; 5 Mac-1: PDB ID:1NA5 30 and VLA-1: PDB ID 1PT6 31 ). We first removed the allosteric ligand co-crystallized with the LFA-1 α I domain. The crystal structures of the Mac-1 and VLA-1 α I domains did not contain any ligands (referred to as closed structures). Using these α I domains, we performed long molecular dynamics simulations in the presence and absence of fragments (benzene). In brief, six different sets of simulations were carried out for each system, using three different concentrations of benzene (0, 30, and 65 mM, respectively) and temperature values of 300 and 350 K. Five independent 1-μs simulations were carried out for each combination of benzene concentration and temperature for a cumulative sampling of 30 μs as described in the Methods section. In this report, we focus the analysis on the production simulations at 300 K. Similar results were obtained at 350 K with the main difference shown in the Supporting Information.
Upon removal of ligand and in the absence of fragments, the LFA-1 α I allosteric site between β 6 and the C-terminal α 7 helix closed within the first 5 ns of equilibration (not shown). This rapid closing is consistent with previously reported NMR solution phase structures of the LFA-1 α I domain. These structures reveal that the β 6 −α 7 region where α I allosteric ligands bind to is occluded. In consequence, a segmental (i.e., rigid body) movement of the C-terminal α 7 helix is mandatory to allow ligands to interact in solution. 32,33 Additionally, we show here that in the absence of fragments, the average rootmean-square deviations (RMSDs) of α 7 from the closed conformation of the LFA-1 α I allosteric site and the respective putative sites in Mac-1 and VLA-1 (β 6 −α 7 pockets) ( Figure  2b, empty circles) are small (below 2.7 ± 0.1 Å), indicating a preserved closed conformation (Movies S1−S3).
These simulation results motivated the use of mixed-solvent MD simulations to uncover putative α I allosteric pockets in Mac-1 and VLA-1. To prompt the opening of these pockets, the proteins were solvated in an aqueous mixture with benzene molecules at different concentrations. The choice of molecules was prompted by studies supporting the hypothesis of fragment-probe-induced pocket formation. 23 The RMSD analysis revealed that increased benzene concentration leads to higher deviations from the closed conformation of the β 6 −α 7 pocket, with the most pronounced effect observed for LFA-1 (Figure 2b). The effect was marginal for Mac-1 and VLA-1, which can be ascribed to the polarity of the residues at the β 6 −α 7 interface. For LFA-1, charged residues interact with bulk water and contribute to a large extent to the flexibility of the pocket (Figure 2c, left). Hydrophobic residues are also present, but their side chains in α 7 are far from apolar residues in β 6 ( Figure 2c, right). In contrast, the Mac-1 and VLA-1 interfaces are rich in hydrophobic residues pointing toward each other, contributing to their closed conformation ( Figure  2d,e, right snapshots). In addition, the closed conformation is also stabilized by the D294-K315 salt bridge or the K309-R330 contact in Mac-1 and VLA-1, respectively.
Higher temperatures lead to increased deviations for Mac-1 and VLA-1 and comparable deviations for LFA-1 ( Figure S1). For the three integrin α I domains, the high error bars indicate A salt bridge is considered to form when the distance between the E284 C δ and K305 N ζ atoms and the distance between the D294 C γ and the K315 N ζ atoms is below 0.5 nm in LFA-1 and Mac-1, respectively. Analogously, a contact is preserved in VLA-1 when the distance between the backbone oxygen atom of K309 and the C ζ of R330 is below 0.7 nm. (b) Solvent-accessible surface area of α 7 , i.e., the F292-I306, F302-I316, and E317-I331 segments for LFA-1, Mac-1, and VLA-1, respectively. (c) Interaction frequencies between the α I domains and the benzene molecules mapped on the sequence of the proteins and normalized by the number of benzenes in the system. A contact is formed between the benzene molecules and a residue if the distance between any of their atoms is below 0.5 nm. The error bars represent the standard error calculated as the standard deviation of the five independent values. more frequent transitions from open to closed conformations of the pocket.
The analysis, focusing on the flexibility of the α 7 helix of the α I domains on two different timescales (10 and 100 ns, Figure  3, top and bottom, respectively) revealed that the C-terminus of α 7 is more susceptible to perturbations induced by the probe molecules on the short timescale than on the long timescale, while the plasticity of other segments is similar at both timescales. Independently of the analyzed system, the profiles show the least fluctuations in the last eight residues of α 7 in the absence of benzene molecules. Increasing the benzene concentration leads to higher fluctuations of the C-terminus of α 7 in LFA-1. This effect was more pronounced on the shorter timescale. Interestingly, intermediate benzene concentrations (30 mM) increase the fluctuations of the N-and Ctermini of α 7 in Mac-1, and high concentrations have only a minor impact on the fluctuations compared to the system devoid of benzene. The C-terminus of VLA-1 shows higher plasticity only at high benzene concentration (65 mM) on the short timescale, while negligible differences in flexibility are observed as a function of benzene concentration on the long timescale. The more pronounced effects on the short timescale show the dynamic response of the α I domain to benzene, which is an indication of the opening and closing of the β 6 −α 7 pocket.
High Benzene Concentration Reduces the Population of Stabilizing Salt Bridges. The E284-K305 salt bridge has been previously identified to play an important role in preserving the closed conformation of the β 6 −α 7 pocket in LFA-1. 34 The corresponding salt bridge in Mac-1 is D294-K315 and the analogue contact in VLA-1 is K309-R330, which can form hydrogen bonds (H-bonds) between the K309 backbone oxygen atom and the R330 side chain. Results show that these polar contacts are mostly populated in the absence of benzene molecules and are differently affected with increasing benzene concentration across the distinct α I domains ( Figure 4a, Movies S1−S9). Visual inspection reveals a clear trend toward more pronounced pocket opening effects at 65 mM of benzene (Movies S7−S9) compared to a benzene concentration of 30 mM (Movies S4−S6). The LFA-1 salt bridge is the most prone to break, showing a decrease of occupancy of more than 60% at 65 mM compared to physiological conditions. Note that compared to the analogous salt bridge in Mac-1 and the H-bond in VLA-1, the LFA-1 salt bridge is less populated in the absence of benzene molecules. Adding benzene to the solvent has a comparable effect on the stabilizing contact in Mac-1 and VLA-1, with a decrease of about 30% in occupancy at 65 mM benzene. Interestingly, at intermediate benzene concentrations, VLA-1 is predominantly more stable than at high concentrations (yet remains within the error bar at 65 mM). The susceptibility of the LFA-1 α I allosteric site to open more easily is also reflected by the solvent-accessible surface area of its α 7 , which is the largest across the three systems independently of the benzene concentration ( Figure 4b). The α 7 of VLA-1 is the least solvent exposed and is only marginally influenced by the addition of benzene molecules, while for Mac-1 the α 7 solvent exposure is reduced. The large error bars for Mac-1 are an indication of the variability of the system and correlate well with the flexibility of α 7 .
The contact map analysis ( Figure 4c) shows that the benzene molecules rarely interact with the residues involved in the salt bridge in LFA-1 and Mac-1 (black stars in Figure 4c) but rather engage with preponderance in nonpolar interactions with leucine and phenylalanine side chains. In contrast, in VLA-1 the benzene molecules interact with the alkyl (i.e., hydrophobic) moiety of the K309 and R330 side chains, though they also contact residues L327-G328. The disruptive effect of benzene on the stabilizing contact arises from the more frequent interactions of the molecules with the residues in the core of α 7 and its surrounding. In practice, this translates in LFA-1 into the insertion of benzene molecules between α 7 and β 6 , and α 7 and α 1 , thereby pushing α 7 away from β 6 and opening the allosteric pockets. Secondary effects include the displacing of α 7 away from α 1 (Figure 5a, Movies S4 and S7), which could give rise to a new allosteric pocket. Similarly, in Mac-1 and VLA-1, the interactions of the benzene rings with L312 and F302, and L327 and G328, respectively, loosen the contacts between α 7 and α 1 (Figure 5b,c, Movies S8 and S9).

SAPPHIRE Analysis Reveals a Novel Pocket.
To gain structural insight into the conformational modifications induced by the organic solvent, we employ SAPPHIRE (States and Pathways Projected with High Resolution) analysis. 35,36 The SAPPHIRE analysis (Figures 6−11) gives an overview of the cumulative sampling for LFA-1, Mac-1, and VLA-1, respectively, at 300 K. In brief, the idea consists of reordering all trajectory snapshots based on geometric similarity as defined by a metric given as input (see Tables S1−S3 for the full definition of the metric). The data is partitioned into basins consisting of similar snapshots without any a priori clustering or overlap between the distinct states. The resulting sequence of snapshots is referred to as the progress index. The cut functions (black lines in the lower part of the plots) represent pseudo-free energy profiles that separate the individual states and help identify the barriers between different basins. The progress index is annotated with suitable variables, which highlight the conformationally and/or kinetically homogeneous states and the dynamics between them. The dot patterns Highlighted are the residues (sticks) in α 7 (dark purple helix) that interact with the benzene rings (gray van der Waals spheres).  35,36 are annotated with a set of interatomic distances, which contribute to the integrity of the allosteric pocket, i.e., donor−acceptor α 7 hydrogen-bond distances α 7 H-bonds, α 7 −β 6 distances (α 7 −β 6 dist), and α 7 −α 1 distances (α 7 −α 1 dist), and the secondary structure assignment of α 7 (top annotation). The barriers in the cut function (i.e., local maxima on the profile in black at the bottom of Figures 6, 8, and 10) separate individual metastable states, whose weights are quantifiable by the progress-index segment between two consecutive barriers. Importantly, recurrence across the individual simulations and systems shows that most basins are sampled several times and in different simulation systems (dark blue, light blue, and green dots above the cut function).

Journal of Chemical Information and Modeling
The SAPPHIRE analysis finds multiple metastable states of the LFA-1 α I allosteric pocket. As the interest of this study is to identify and characterize the open conformation of the pocket, the following analysis will focus mainly on the states that describe this arrangement and not address in detail every basin. The most recurrent state across all simulations, but (Middle) Intramolecular distances that contribute to the stability of the allosteric pocket, i.e., donor−acceptor α 7 -intramolecular distances (α 7 Hbonds), α 7 −β 6 distances (α 7 −β 6 dist), and α 7 −α 1 distances (α 7 −α 1 dist). The complete list of distances used for the progress index metric is in Table S1. Each set of distances has values in nm grouped into seven bins (color legend on the top of each set of distances). (Top) Secondary structure assignment of α 7 with its corresponding legend on top.
predominantly in the systems devoid of benzene molecules, is the closed state, in which no potential ligand can access the α I allosteric pocket. For instance, the basin between 1.8 × 10 5 and 2.3 × 10 5 highlights states characterized by the presence of the E284-K305 and K287-E301 salt bridges both of which stabilize the closed conformation of the LFA-1 allosteric pocket ( Figure   6, top center). In contrast, the following basin (between 2.3 × 10 5 and 2.7 × 10 5 ) highlights states that are primarily identified in the systems with benzene molecules and only minimally present in the simulations devoid of fragment probes. Furthermore, none of the above-mentioned salt bridges are formed and the structural overlap to the crystal structure (PDB    In contrast to LFA-1, the identification of open allosteric pockets in Mac-1 and VLA-1 is a more challenging task. The SAPPHIRE analysis of Mac-1 reveals that the allosteric β 6 −α 7 pocket is predominantly closed, independently of the benzene concentration ( Figure 8). This is consistent with the distance analysis in Figure 4a, which indicates the long lifetime of the D294-K315 salt bridge. We identify a limited number of basins, in which open conformations of the β 6 −α 7 pocket are sampled (Figure 9 and Movie S8). These are present in the systems with benzene molecules (e.g., between 5.2 × 10 5 and 5.8 × 10 5 and marked by stars in Figure 8) and are characterized by the breakage of the D294-K315 salt bridge, which is induced by the insertion of benzene molecules between α 7 and β 6 (Movie S8). Rarely, the quasi-opening of the pocket is owed to the unwinding of the C-terminus of α 7 , which leads to a loss in the secondary structure. Like LFA-1, the opening of the new pocket coincides with the rearrangement of the loops at the β 6 −α 7 −α 1 interface, which arises due to the frequent interactions of the benzene molecules with the α 7 −β 6 loop (specifically residues N300 and F302, see Figure 4c, Figure  9e), the N-terminus of α 1 (residues D149 and F150, see Figure  9e) and residues D140 and D242 in the surrounding loops ( Figure 9e). Consequently, the E303-R152 salt bridge linking α 7 to α 1 is broken, the helices are pushed away from each other, and the β 6 −α 7 −α 1 pocket forms (Figure 9, e.g., the basin between 1.5 × 10 5 and 2 × 10 5 and Movie S8).
In VLA-1, the SAPPHIRE analysis discriminates less between β 6 −α 7 open and closed conformations as most states are populated across most basins ( Figure 10). In fact, the K309-R330 contact remains populated over 70% across the simulations independently of the benzene concentration ( Figure 4a) and dissociates in one of the simulations at high benzene concentrations (last basin in Figure 10 and Movie S9).
Aside from this, the open conformation is short-lived and sampled only in small basins (marked along the progress index by stars). On the other hand, like in Mac-1, the β 6 −α 7 −α 1 pocket is more prone to open through loop and concurrent helix rearrangements at the interface (Figure 11d,e and Movie S9). The displacement of the α 7 helix is achieved through the intercalation of benzene molecules between α 1 and α 7 , facilitated by frequent interactions of benzene molecules with residues Y156, P157, and E317 (Figure 11e). The benzene molecules trigger a rigid body motion of the N-terminus of α 7 (captured also in Figure 3 by the increased flexibility of the Nterminus on the short timescale), which leads to a larger separation between residues E317 and Y156 and implicitly a helix rearrangement. Additionally, benzene intercalation between the α 1 and α 7 helices stimulates at times the parallel downward sliding of α 7 . Interestingly, this conformational change has been previously induced in LFA-1 by introducing artificial disulfide-bond ratchets. Moreover, this artificial pulling down of the α 7 helix led to a progressive increase in α I domain binding affinity. 37 Potential Druggability of the β 6 −α 7 and β 6 −α 7 −α 1 Pockets. To assess the druggability of the newly discovered pockets, i.e., the β 6 −α 7 pockets in Mac-1 and VLA-1, and the β 6 −α 7 −α 1 pockets in all three α I domains, small rigid fragments were docked into the sites using the open source program SEED. 38 The fragments are mainly hydrophobic yet contain an amide extension as a possible linker for the addition of hydrophilic groups, with the exception of cyclohexane (Figure 12a). We provide a first overview of the binding sites, as the de novo small-molecule design is beyond the scope of the current study.
Five open conformations are selected from the SAPPHIRE analysis based on the size of the pocket (Figure 12). For comparison, SEED docking was also applied to the LFA-1 crystal structure, used as a starting conformation for the MD simulations. In the preparation of the docking process, the structures are energy-minimized with restrains applied on the C α -atoms. The number of fragment poses with a SEED energy of less than −12 kcal/mol is summarized in Figure 12b. Further, to visualize the size of each pocket available for drug design, the combined surfaces of fragment poses at the energy cutoff are shown together with the protein surfaces ( Figure  12c). Only for the LFA-1 β 6 −α 7 −α 1 site a cutoff of −13 kcal/ mol was used for the combined fragment surface (Figure 12c). As expected, the fragments interact with the pockets through their hydrophobic part. In most cases, the amide linker moieties of the fragments point toward the outside of the pocket. These positions of the linker moieties may allow the attachment of solubilizing groups or other groups which could improve the druggability of potential ligands.
Moreover, the results indicate that the β 6 −α 7 pocket of VLA-1 is larger than the analogous pocket in the LFA-1 crystal structure. However, it needs to be taken into account that the LFA-1 pocket has adapted its size to the ligand present in the crystal structure. Thus, it could possibly provide space for larger ligands. The β 6 −α 7 pocket in Mac-1 is the smallest, but could still provide enough space to bind an organic molecule. Further, in the case of the β 6 −α 7 −α 1 pocket, the one in LFA-1 is clearly the largest relative to those detected in Mac-1 and VLA-1. The LFA-1 β 6 −α 7 −α 1 pocket has the shape of an open tunnel that traverses the whole protein. In comparison, the β 6 −α 7 −α 1 pocket of Mac-1 is rather broad but not very deep such that bigger fragments do not achieve the fit as observed for fragments binding to the LFA-1 pocket. The VLA-1 β 6 −α 7 −α 1 pocket is the smallest but is nevertheless suited to provide enough space for a ligand (Figure 12c). Whether ligands binding to the newly identified sites in the α I domains will have similar effects on the whole integrin and indeed modulate its function remains to be tested.

■ DISCUSSION
Allosteric modulation of receptors creates opportunities for drug discovery and development which are potentially superior to classic orthosteric modulation. Specifically, allosteric modulation avoids the risk of inadvertently eliciting effects naturally triggered by ligand binding to the orthosteric site. 39,40 Integrins represent a family of therapeutically relevant receptors which are of particular interest for allosteric rather than orthosteric modulation. This is because pharmacologic interference with the orthosteric ligand binding site is known already to have the potential of inducing major unwanted effects, such as unintended agonist-like effects of integrin inhibitors. A first prototypic example establishing that allosteric modulation can be applied to members of the integrin family exists with the integrin LFA-1. Small molecules targeting the α I allosteric pocket of LFA-1 have been shown to stabilize LFA-1 in a distinct conformational state, translating into inhibition 14 or, observed less frequently, into activation of LFA-1 function. 41 Despite this promise, no rigorous attempts have been made to date towards identifying molecules targeting putative α I allosteric sites similar to the LFA-1 α I allosteric site in other integrins. A few in silico studies provide evidence that such sites may exist in integrins closely related to LFA-1, e.g., αxβ 2 (CD11c/CD18) 42 and Mac-1. 43−45 In addition, there is evidence suggesting that the putative α I allosteric site of Mac-1 can accommodate inhibitors 44 as well as activators 43,45 of Mac-1 function. However, these investigations did not prompt further research into the accessibility and druggability of putative α I allosteric pockets in other integrins. One important reason for this lack of progress is the difficulty in identifying therapeutically tractable allosteric pockets. 24,25 It is worth noting that the LFA-1 α I allosteric pocket itself was discovered by serendipity rather than by prospectively defined research. Moreover, our results show that the LFA-1 α I allosteric pocket closes promptly in the absence of ligands, i.e., can be missed easily unless searched for specifically, underlining the difficulty of uncovering hidden allosteric sites. In fact, all LFA-1 α I domain crystal structures published to date are ligand-bound and capture the open conformation of the α I allosteric site, further suggesting that the open configuration may be unstable in the absence of ligands and hence difficult to access. 32,33 By applying mixed-solvent MD simulations to integrins, the current study demonstrates and characterizes the potential opening of cryptic α I allosteric sites of different integrins. We investigated the concentration-dependent effects of benzene molecules on the α I domain conformational state of three therapeutically relevant integrins, i.e., LFA-1 (as a control), Mac-1, and VLA-1. Our results show that the putative α I allosteric sites of Mac-1 and VLA-1 are predominantly closed under physiological conditions yet are susceptible to open at the β 6 −α 7 interface (region analogous to the α I allosteric site of LFA-1) in benzene-rich solvent. Unlike LFA-1, the closed conformations in Mac-1 and VLA-1 are stabilized mainly by hydrophobic contacts between β 6 and α 7 . By enriching the solvent with benzene molecules, the populations of stabilizing contacts decrease, which in turn leads to the opening or quasiopening of the α I domain pockets of Mac-1 and VLA-1, respectively. Our results are in line with previous findings showing that the salt bridge interaction of α 7 with other parts of the α I domain differs for distinct members of the β 2 integrin subfamily [i.e., LFA-1, Mac-1, and αxβ 2 ] and is associated with diverse α 7 flexibility. 34 A limitation intrinsic to the cryptic site opening approach described here is that the degree of site opening is dependent on the chemical nature of the fragment probe selected. 25 In the current study we use benzene (rather than alternative fragments such as pyridine, imidazole, indole, or pyrimidine) for several reasons. First, the molecular structure of benzene mimics lovastatin, the first inhibitor described to bind to the α I allosteric pocket of LFA-1 12,13 Second, benzene is the most common hydrophobic ring fragment present in approved drugs. 46 Thirdly, aromatic rings are constitutive to broad spectra of natural biomolecules, including molecules of potent immunomodulatory capacity.
Mechanistically three different modes of action of the α I allosteric pocket opening are to be considered, i.e., conformational selection, induced-fit, or a mixed mechanism. 25 In the first case, a cryptic pocket opens transiently, yet spontaneously, and is then occupied by the ligand. In the second case, the pocket does not open spontaneously and needs the help of ligand for opening in an induced-fit style. The mixed mechanism is characterized by a slight spontaneous opening, which requires ligand to reach the fully opened state. 25 In our mixed-solvent MD simulations we observe the occasional openings of the α I allosteric pocket of LFA-1 in water. In the presence of solvent/benzene mixtures, the intercalation of benzene molecules between the β 6 and α 7 interface is observed, enhancing pocket opening. This result indicates that the α I allosteric site of LFA-1 opens according to the mixed mechanism model. In contrast, Mac-1 and VLA-1 appear to require benzene for the opening of their putative α I allosteric pockets (β 6 −α 7 pocket), thus following the induced-fit model.
We project that the Mac-1 and VLA-1 β 6 −α 7 pockets can be utilized to design small-molecule modulators of these two integrins, as has been accomplished successfully in the past with the analogous site within LFA-1. 14 Indeed, SEED analysis of the respective integrin α I domains revealed that allosteric modulation by small molecules may be possible through these pockets. For Mac-1, this finding is in agreement with earlier molecular docking studies referred to above. 43−45 For VLA-1 the current study is the first study to suggest that this integrin may be targetable via the β 6 −α 7 pocket. Although VLA-1 is a promising therapeutic target, 27 pharmacological progress in designing VLA-1 inhibitors has been slow, to date. To the best of our knowledge SAN-300, a humanized monoclonal antibody apparently no longer in development, was the sole modality explored clinically for the treatment of rheumatoid arthritis. 27 As anti-integrin antibodies are known to be associated with substantive limitations, 18 allosteric targeting of VLA-1 by small molecules may emerge as an important translational advance. Target indications of high medical need under consideration for future VLA-1 inhibitors include inflammatory, fibrotic, and malignant diseases. 47−49 Unexpectedly, our results show the formation of a second potentially druggable pocket (termed here β 6 −α 7 −α 1 pocket), located at the interface between the β 6 −α 7 loop and the Nterminus of α 1 in all three α I domains. This novel pocket opens independently of the β 6 −α 7 conformation. It forms through the concomitant motion of the helices, driven by the intercalation of benzene molecules between α 1 and α 7 and stimulated by frequent interactions with aromatic residues. Earlier simulations of the α I domains of several integrins found that the β 6 −α 7 loop of the β 6 −α 7 −α 1 pocket can adopt closed, open or intermediate conformations, with the first being observed more frequently in VLA-1 than in LFA-1 and Mac-1. 50 The authors attribute the stability of the closed conformation to the R152-E303 salt bridge connecting α 1 to α 7 in Mac-1. Our results are in line with this hypothesis. We additionally show that the breakage of the salt bridge can be induced and occurs as a consequence of the intercalation of benzene molecules between α 1 and α 7 , thereby opening the β 6 −α 7 −α 1 allosteric pocket. In VLA-1, the interaction between α 1 and α 7 is mainly hydrophobic with no salt bridge stabilizing the interfacial contact between the β 6 −α 7 loop and α 1 . Intriguingly, our SEED analysis provides the first evidence that molecules interacting with the novel β 6 −α 7 −α 1 pocket of the integrin α I domains can be designed, as also shown for the newly identified β 6 −α 7 pockets of Mac-1 and VLA-1. However, it remains to be explored whether molecules targeting these newly identified allosteric pockets will exhibit effects similar to the small-molecule modulators which bind to the classical α I allosteric site of LFA-1. Moreover, further structural characterization of the newly described β 6 −α 7 and β 6 −α 7 −α 1 pockets in integrin α I domains, e.g., crystallization of the respective α I domains with benzene and/or other ligands identified by molecular docking, are required for verification.
In conclusion, by applying mixed-solvent MD simulations to integrins, we demonstrate that pockets similar to the α I allosteric pocket of the integrin LFA-1 can be rendered accessible by adjusting solvent conditions. For the integrins Mac-1 and VLA-1, specifically, we show that compounds interacting with their putative α I allosteric sites can be identified readily using molecular docking. In addition, we identify a novel potentially druggable cryptic pocket at the β 6 −α 7 −α 1 interface. Taken together, our findings indicate that integrin allosteric site hopping might be possible and might become a strategy for extending the discovery of nextgeneration α I allosteric integrin modulators. Specifically, it appears conceivable that the mixed-solvent MD simulation approach pioneered for three integrins in the current study may become extended to systematically assess the allosteric targetability of 9 of 24 integrins carrying α I domains. As several of these integrins, including the three integrins analyzed here, play decisive roles in diseases for which there is no cure, to date, substantive therapeutic benefit may derive from these future investigations.
Simulation Protocol. The simulations were carried out using the GROMACS 2020.5 simulation package. 51 All simulations were performed using the all-atom CHARMM36m force field 52 and the TIP3P water model. 53 The N-and Ctermini of the proteins were uncapped. Each complex was first solvated in a cubic box (edge length of 8 nm) with TIP3P water molecules to which 150 mM NaCl were added, including neutralizing counterions. Additionally, in the simulations in an organic solvent, benzene molecules were randomly distributed in the box at noninteracting distances from the protein.
Following the steepest descent minimization, the systems were first equilibrated under constant pressure for 5 ns. The temperature and pressure were maintained constant at 300 K and 1 atm, respectively, by using the modified Berendsen thermostat (0.1 ps coupling 54 ) and barostat (2 ps coupling 55 ). The production simulations were performed in the NVT ensemble in the absence of restraints. The short-range interactions were cut off beyond a distance of 1.2 nm, and the potential smoothly decays to zero using the Verlet cutoff scheme. Periodic boundary conditions were used and the Particle Mesh Ewald (PME) technique 56 was employed (cubic interpolation order, real space cutoff of 1.2 nm, and grid spacing of 0.16 nm) to compute the long-range electrostatic interactions. Distances of covalent bonds involving H atoms were constrained by means of a fourth-order LINCS algorithm with 2 iterations. 57 Six different sets of simulations were carried out for each system, at three concentrations of benzene (0, 30, and 65 mM) and temperature values of 300 and 350 K. Five independent 1-μs simulations were carried out for each set for a total cumulative sampling of 30 μs. In all simulations the time-step was fixed to 2 fs and the snapshots were saved every 25 ps. The higher temperature is employed to enhance the sampling; the density of water is kept at the value of the 300 K simulations (i.e., same volume of the box and same number of water molecules) to perturb the free energy surface as little as possible. 58−60 The benzene concentrations were chosen such that the proteins are not denatured and that no aggregation of benzene molecules is observed in bulk. 61 Nine movies resulting from this work are shown in the Supporting Information. SEED Analysis. SEED calculations were carried out with version 3.3.6 of the program 38,62 (https://gitlab.com/ CaflischLab/SEED). The force field-based energy function used in SEED consists of four terms, namely, electrostatic and van der Waals interactions between the protein and the fragment, and the electrostatic desolvation penalty of the protein and the fragment. CHARMM36m partial charges and van der Waals parameters were assigned to proteins and fragments using the molecular graphics program Wit!P (http://www.biochem-caflisch.uzh.ch/download). For all calculations, dielectric constants of 2.0 and 78.5 were used for solute and solvent, respectively. All polar fragments were placed according to the option "both", i.e., in optimal polar and apolar regions. Pose clustering was carried out individually for each fragment, and for each cluster, only the member with the most favorable SEED total energy was saved. ■ ASSOCIATED CONTENT Data Availability Statement PDB files were downloaded from the RCSB Protein Data Bank (https://www.rcsb.org). GROMACS (https://www.gromacs. org/) was used to perform the molecular dynamics simulations and SEED for the docking step (https://gitlab.com/ CaflischLab/SEED). Campari (https://campari.sourceforge. net) was used for analyses. Input files and analysis scripts are available at https://github.com/ilieim/DecryptingIntegrins.