Analysis of aligned PD-1 structures (Figure 2) led us to classify the bound-like and non-bound-like conformational states using two binary order parameters defined by the ‘open’ or ‘closed’ states of Asn66 and Ile126. Namely, for a non-bound-like interface Asn66 is closed and Ile126 is open; for the PD-L1-specific bound-like state Asn66 is open and Ile126 is closed; and for the PD-L2-specific bound-like state both Asn66 and Ile126 are open (Figure 2e). In the PD-L1–bound state, the interface exhibits a large hydrophobic patch that interacts with the side chain of ligand interface residue Ala121 (Figure 2b). In the PD-L2–bound state, the interface exhibits a deep hydrophobic cavity that buries ligand residue Trp110 (Figure 2c). Neither this hydrophobic patch nor deep cavity is sampled in the apo PD-1 NMR ensemble, where, instead, the closed state of Asn66 blocks the Trp110-binding pocket by exposing its NH2 group (Figure 2a,e, Figure 2—figure supplement 2), making a hydrophilic site. MDs of apo PD-1 confirm that Asn66 remains closed (Figure 4a), stabilized by a hydrogen bond with Lys78 that is also present in NMR structures (Figure 5a). These findings suggest that specific interactions between apo PD-1 and a nearby ligand might be required to open Asn66 and reshape the hydrophilic interface into its hydrophobic bound-like states.

Figure 4 with 3 supplements see all

Download asset Open asset

Figure 4—source data 1

Excel workbook containing all the simulation trajectory data plotted in Figure 4, Figure 4—figure supplement 2, and Figure 4—figure supplement 3.

The first three sheets of the workbook, titled ‘Asn66 position,’ ‘cavity volume,’ and ‘Ile126 dihedral’ contain the time-series data plotted in Figure 4a,b and c, respectively. They also contain the data for the replicate simulations shown in Figure 4—figure supplement 2. The columns labeled with a ‘1’ (apo1, GDY1, etc…) correspond to the data from the first replicate simulation, i.e. the data shown directly in Figure 4. Columns labeled with ‘2’ and ‘3’ correspond to data plotted in Figure 4—figure supplement 2. The final sheet of the workbook is titled ‘alternate anchors,’ and contains the time series data plotted in Figure 4—figure supplement 3.

https://doi.org/10.7554/eLife.22889.009

Download elife-22889-fig4-data1-v2.xlsx

Figure 5

Download asset Open asset

For both induced fit and conformational selection, the association of the apo receptor and ligand is driven mainly by diffusion (DeLisi, 1980; Northrup and Erickson, 1992). It has been shown that often protein-protein interactions stabilize the initial encounter complex through the burial of a bound-like anchor motif on the ligand (Rajamani et al., 2004), which allows subsequent, longer timescale intermolecular interactions to take shape. Co-crystal structures, MDs and docking studies of PD-L1/2 suggest that the homologous interface residues Tyr123/112 (see Figure 2b,c) may serve as anchors. Specifically, MDs of apo PD-L1/2 show that Tyr123/112 remain within 0.5 Å heavy atom RMSD of their bound-like conformations 88 ± 16% of the time. Furthermore, the Tyr123/112-binding pocket is unobstructed in the apo PD-1 NMR ensemble (Figure 2d), facilitating immediate burial of the side chain upon association. Docking exercises also point to the stabilizing role of the Tyr anchor. Namely, ClusPro (Comeau et al., 2004) successfully re-docked the wild-type human PD-1 – PD-L1 co-crystal (Zak et al., 2015), but it failed for single-residue PD-L1 mutants Y123G and Y123A (Table 1). Collectively, these results suggest an anchor role for Tyr123/112 that facilitates molecular recognition between non-bound-like apo PD-1 and its ligands (as sketched in Figure 1).

Co-crystal structures of bound PD-1 exhibit an open Asn66 that forms two hydrogen bonds: the first with the backbone oxygen of homologous PD-L1/2 Ala121/Trp110 and the second with either PD-1 Tyr68 (human PD-1 - PD-L1 complex) or a crystal water (murine PD-1 – PD-L2 complex) (Figure 5b,c). MDs of PD-1 in complex with a GGG peptide positioned to mimic the backbone of PD-L1/2 residues ADY123 and WDY112, respectively, show that Asn66 fluctuates back and forth between a bound-like open state, where it makes the aforementioned backbone hydrogen bond to the GGG peptide, and the non-bound-like closed state, where it is bonded to PD-1 Lys78 (Figure 4a). On the other hand, simulations with a GDG peptide show that the Asp122/111 mimic forms a hydrogen bond to the Tyr68 OH group, stabilizing a Tyr68 rotamer that can simultaneously hydrogen bond to the NH2 of Asn66 (Figure 5d). Together, this Asn66 – Tyr68 hydrogen bond and the aforementioned Asn66 – backbone hydrogen bond stabilize the bound-like open state of Asn66 (Figure 4a).

The robust, four-membered hydrogen bond network between the Ala121/Trp110 backbone mimic, Asn66, Tyr68, and the Asp122/111 mimic that we observe in GDG MDs is fully consistent with all available structures and mutagenesis experiments. Namely, the hydrogen bonds rationalize the conservation of Asp122/111 in all known PD-L1/2 sequences and explain PD-L2 mutagenesis studies showing that the D111A mutation abolishes binding to PD-1 (Lázár-Molnár et al., 2008). MDs of apo PD-L1/2 further support the importance of Asp122/111 interactions in the encounter complex by showing that this side chain remains within 0.4 Å RMSD of its bound-like conformation 82 ± 25% of the time. The stabilization of the bound-like Asp122/111 side chain in simulation is achieved via hydrogen bonds with the neighboring Lys124/113, bonds which are also observed in bound cocrystal structures of PD-1 (Figure 2b,c). The importance of this stabilizing interaction is underscored by the fact that the K124S and K113A point mutations in PD-L1 and PD-L2, respectively, both abolish binding to PD-1 (Lázár-Molnár et al., 2008; Lin et al., 2008).

PD-1 ligands open Asn66 by offering two novel hydrogen bonds (with the Ala121/Trp110 backbone and Tyr68) that out-compete the single Lys78 hydrogen bond that stabilizes the closed state. Interestingly, the one known PD-1 sequence that diverges at the Tyr68 position is murine PD-1, which has a Y68N mutation. The murine PD-1–PD-L2 co-crystal shows that although the shorter Asn68 side chain cannot hydrogen bond directly to Asn66 or Asp111, it hydrogen bonds to a crystal water molecule that forms the same hydrogen bond network as Tyr68 (Figure 5c). MDs of a human Y68N PD-1 mutant and the GDG peptide suggest a functional equivalence of Asn68 to Tyr68: the Asn68 side chain spontaneously recruits a stable water to the co-crystal position that then opens Asn66 via a specific hydrogen bond network analogous to that formed by Tyr68 (Figure 5e).

While GDG MDs show an open Asn66 (Figure 4a) that exposes a hydrophobic surface, this surface remains flexible and fluctuates between a deep open cavity and closed shallow patch (Figure 4b). Contrary to the GGG MDs that exhibited open-closed fluctuations of Asn66 (Figure 4a), the pocket instability observed in GDG MDs is caused by open-closed fluctuations of PD-1 residue Ile126 (Figure 4c). In contrast, MDs show that the GDY peptide stabilizes the open states of both Asn66 and Ile126 and maintains the open hydrophobic interface cavity seen in the PD-L2 bound-like state of PD-1 (Figure 4b,c,d). Comparison of the GDG and GDY MDs reveal that the Tyr side chain serves as a ‘wedge’ that stabilizes the flexible loop surrounding Ile134 into a bound-like configuration that is observed in both the PD-L1 and PD-L2 co-crystal structures (Figure 6). In the presence of the GDY peptide, the bound-like Ile134 makes a hydrophobic contact with the long arm of Ile126, which pulls the latter residue out of the pocket and stabilizes its open state (Figure 4—figure supplement 1).

Figure 6

Download asset Open asset

Figure 6—source data 1

Excel workbook with a single sheet containing the numerical RMSD data shown in the Figure 6 bar chart.

https://doi.org/10.7554/eLife.22889.016

Download elife-22889-fig6-data1-v2.xlsx

Although the PD-L1 interface exhibits the GDY scaffold, Ile126 is closed in the PD-L1-specific EC_BL state, suggesting that an additional ligand motif not contained in the GDY scaffold is responsible for closing the pocket. MDs with an ADY peptide that mimics Ala121 show that the extra C_β carbon of the Ala side chain out-competes Ile134 for the long arm of Ile126, stabilizing its closed state (Figure 4c,e). Interestingly, MDs with GDG and ADG peptides both show similarly unstable open-closed fluctuation of Ile126 (see Figure 7 below), which suggests that the effect of the Ala121 C_β carbon on Ile126 dynamics only emerges in the presence of the anchor Tyr123/112. Thus, in addition to facilitating molecular recognition, stabilization of the Ile134 loop by the burial of Tyr123/112 is shown to enable ligand-specific induced fit responses by the PD-1 interface.

Figure 7

Download asset Open asset

Figure 7—source data 1

Excel workbook with a single sheet containing the numerical ${\displaystyle \mathrm{\Delta }{G}_{open}}$ data plotted in Figure 7.

https://doi.org/10.7554/eLife.22889.018

Download elife-22889-fig7-data1-v2.xlsx

We applied Maxwell-Boltzmann statistics to our peptide simulations (see Materials and methods) to quantify the role played by each trigger in the structural transitions at the PD-1 interface. We evaluate ${\displaystyle \mathrm{\Delta }{G}_{open}}$, that is, the free energy differences between the open and closed states of Asn66/Ile126 for PD-1 in isolation and PD-1 interacting with nine different peptides representing distinct PD-L1/2 interface motifs (Figure 3, Table 2; note that ${\displaystyle \mathrm{\Delta }{G}_{open}}$ and ${\displaystyle \mathrm{\Delta }{G}_{BL}}$ are trivially related). These ${\displaystyle \mathrm{\Delta }{G}_{open}}$ values are plotted in Figure 7. Remarkably, the ADY and GDY motifs, respectively, shift the ratio of our predefined bound-like to non-bound-like states from 1: 44 ± 24 (based on ${\displaystyle \mathrm{\Delta }{G}_{open}^{apo}\left(Asn66\right)}$) to 7.4 ± 2.8: one for the PD-L1 bound-like state (based on ${\displaystyle \mathrm{\Delta }{G}_{open}^{ADY}\left(Ile126\right)}$) and 12 ± 9.6: one for the PD-L2 bound-like state (based on ${\displaystyle \mathrm{\Delta }{G}_{open}^{GDY}\left(Ile126\right)}$). More importantly, we show that each triggering contact monotonically lowers the relative free energy of ligand-specific bound-like states starting from no contacts (apo), to the first, conserved contact with the anchor (Y), to the second, conserved contact with Asp122/111 (DY), to the final, unconserved contact with the backbone O of A/G in the complete triggering motifs (ADY/GDY) (Figure 7). The fact that these downhill binding pathways do not encounter energy barriers strongly suggests that the PD-1 binding mechanism is primarily one of induced fit (see Figure 1).

In the apo simulation Asn66 is closed (${\displaystyle \mathrm{\Delta }{G}_{open}^{apo}\left(Asn66\right)\approx 3.6k_{B}T}$), repelling Ile126 into an open conformation (${\displaystyle \mathrm{\Delta }{G}_{open}^{apo}\left(Ile126\right)\approx -4.4k_{B}T}$). Docking of the Tyr anchor (Y) and formation of the encounter complex destabilizes the non-bound-like apo PD-1 interface, causing increased open-closed fluctuations in both Asn66 and Ile126. The subsequent docking of Asp122/111 (DY) allows Tyr68 to compete with Lys78 to form one hydrogen bond with Asn66, causing it to swap back and forth between open and closed (${\displaystyle \mathrm{\Delta }{G}_{open}^{DY}\left(Asn66\right)\approx 0}$). Fluctuations of Asn66 correlate with simultaneous fluctuations of Ile126 (${\displaystyle \mathrm{\Delta }{G}_{open}^{DY}\left(Ile126\right)\approx 0}$). Adding the adjacent Ala121/Trp110 backbone from PD-L1/2 (ADY/GDY) provides the second hydrogen bond for the NH2 of Asn66 that fully stabilizes its open state (${\displaystyle \mathrm{\Delta }{G}_{open}^{GDY/ADY}\left(Asn66\right)\approx -3.0k_{B}T}$). With Asn66 open, the Cβ atom of Ala121 modulates Ile126 dynamics. When present (ADY), the C_β hydrophobically recruits Ile126 into the closed pocket state (${\displaystyle \mathrm{\Delta }{G}_{open}^{ADY}\left(Ile126\right)\approx 2.0k_{B}T}$). Without C_β (GDY), Ile126 remains open (${\displaystyle \mathrm{\Delta }{G}_{open}^{GDY}\left(Ile126\right)\approx -2.3k_{B}T}$).

Our ${\displaystyle \mathrm{\Delta }{G}_{open}}$ calculations also quantify the critical role of the anchor residue Tyr123/112 in ensuring the ligand specificity of PD-1 interface deformations. This is demonstrated by the fact that GDY and ADY peptides impose clear differential influence on the dominant rotamer state of Ile126, while for both GDG and ADG, Ile126 fluctuates about evenly between the open and closed state (${\displaystyle \mathrm{\Delta }{G}_{open}^{GDG/ADG}\left(Ile126\right)\approx 0}$) (Figure 7).

We ran MDs of the PD-L1/2 encounter complexes starting from docked poses of apo PD-1 and the interacting domains of PD-L1/2 that anchored Tyr123/Y112 (see Supporting Materials and methods). Encounter complex MDs recapitulated the triggering mechanisms we identified in our peptide simulations and their resulting PD-1 interface transitions from the EC_NBL to the ligand-specific EC_BL states. The chronology for these interactions (Table 3) is the same for both ligands. Consistently, the first interaction to take place after docking the conserved anchor is the formation of the hydrogen bond between receptor residue Tyr68 and ligand residue Asp122/111. This is followed by stabilization of Asn66 in the open pocket state via hydrogen bonds with neighboring Tyr68 and the ligand Ala121/Trp110 backbone. The Ala121/Trp110 side chains then proceed to stabilize a closed/open hydrophobic pocket. Note that the Trp in the WDY motif of PD-L2 readily fills the hydrophobic pocket as the XDY motif latches and opens Asn66 (Figure 8). Consistent with a downhill free energy induced fit mechanism, the realization of these four contacts takes less than 10 ns total. On a longer timescale, encounter complex simulations demonstrate the formation of secondary hydrogen bonds at the interface periphery that are also observed in co-crystal structures of human and murine PD-1. These secondary hydrogen bonds, including the bond from PD-1 Lys78 to PD-L1/2 Phe19/21 and from Gln75 to Arg125/Tyr114 (Figure 9), were consistently observed to form approximately 10 nanoseconds after the aforementioned Asn66 and Tyr68 hydrogen bonds (Table 3), suggesting that EC_BL contacts shaped by the triggers of induced fit are enough to drive the subsequent transition to the HAC.

Figure 8

Download asset Open asset

Figure 8—source data 1

Excel workbook with a single sheet containing the time-series Trp110 atom overlap data from the encounter complex simulations plotted in Figure 8.

https://doi.org/10.7554/eLife.22889.021

Download elife-22889-fig8-data1-v2.xlsx

Figure 9

Download asset Open asset

Recently, two FDA-approved PD-1–targeting antibodies have emerged as part of a new generation of anticancer immune checkpoint inhibitors. Published crystal structures of one of these antibodies, pembrolizumab, bound to extracellular PD-1 show a hydrophobic receptor binding surface that overlaps that which binds PD-L1/2 (Figure 10b) (Horita et al., 2016; Lee et al., 2016; Na et al., 2017). Comparison of the pembrolizumab – PD-1 interface to the PD-L1 – PD-1 interface using the FastContact server (Champ and Camacho, 2007) highlights several differences in the main contacts that characterize the two binding modes (Figure 10a, Tables 4 and 5). Remarkably, the pembrolizumab-bound crystal structures reveal that the antibody stabilizes the same open state of Asn66 as PD-L1/2 using an analogous hydrogen bond network (Figure 10c). The fact that this antibody, designed via a distinct evolutionary pathway, shares PD-L1/2’s mechanism for opening Asn66 and revealing a hydrophobic binding surface (Figure 2a,b,c, Figure 10b) underscores the role of this specific interaction in PD-1 interface remodeling.

Figure 10 with 2 supplements see all

Download asset Open asset

Although pembrolizumab’s interaction with Asn66 mimics the native-like contacts of PD-L1/2, the antibody-bound receptor exhibits a novel configuration of Ile134, with both Ile126 and Ile134 in inward-flipped, ‘closed’ states (Figure 10b). The result is a large hydrophobic surface where, like in the PD-L1–bound state, the closed Ile126 occludes the Trp110-binding pocket, but where, unlike the PD-L1/2–bound states, a closed Ile134 partially fills the Tyr/123/112 anchor cavity. In fact, pembrolizumab has no anchor analog. Instead, the Arg102 side chain extends along the PD-1 interface such that the C_Z carbon overlaps the C_γ position of Tyr123/112 (Figure 10—figure supplement 1), and the NH1/2 groups hydrogen bond to a crystal water above the receptor interface (Figure 10b). In this configuration, the hydrophobic carbon chain of Arg102 forms a ‘cap’ above the closed Ile126 and Ile134, desolvating their hydrophobic interactions with each other and the neighboring Gly124 and stabilizing a flat hydrophobic surface (Figure 10b).

A similar closed conformation of Ile134 is observed in our MDs of PD-1 interacting with the GDG peptide (Figure 4—figure supplement 1). This is unsurprising: like pembrolizumab, the GDG peptide has the necessary machinery to trigger the opening of Asn66, but lacks an anchor ‘wedge’ that prevents the resulting inward collapse of Ile134. Results of the GDG MDs thus rationalize the pembrolizumab binding mode and suggest an anchor-independent induced fit PD-1 binding pathway: one in which the antibody opens Asn66 using the canonical hydrogen bond network and stabilizes the resulting flat hydrophobic interface by ‘capping’ the closed states of Ile126/134 with the carbon chain of Arg102.

Although two PD-1-targeting antibodies already exist on the market, there are no small-molecule PD-1 inhibitors in clinical trial, despite the enormous interest in this blockbuster immunotherapy target (Dömling and Holak, 2014; Couzin-Frankel, 2013; Zarganes-Tzitzikas et al., 2016). Given that ligand-binding sites tend to be concave (Laskowski et al., 1996; Liang et al., 1998) and largely hydrophobic (Cheng et al., 2007), the undruggability of PD-1 might be due to the closed Asn66 and the resulting flat polar interface in the apo form (Figure 2a). However, the highly specific hydrogen bond network presented by PD-L1/2 and pembrolizumab strongly suggests a path to open Asn66 and transform the hard to drug hydrophilic patch into a hydrophobic one. Interestingly, Brystol-Myers-Squibb recently patented a 1.03 nM macrocyclic inhibitor of the PD-1–PD-L1 interaction (Miller et al., 2014). Although no mechanism of action has been described, the macrocycle includes a peptidic mGDV motif that is structurally similar to the aforementioned ADY induced fit trigger, with an N-methylated Gly and an Asp side chain that resemble PD-L1’s Ala121 and Asp122, respectively (Figure 10—figure supplements 1 and 2). This alignment puts the mGDV motif’s short Val side chain at the position of the much longer Tyr123 anchor, where it aligns with the C_∆ side chain carbon of pembrolizumab residue Arg112 (Figure 10—figure supplement 1).

Given the resemblance of the mGDV motif to the interface residues of both PD-L1 and pembrolizumab, we used our MDs method to evaluate whether this motif was capable of remodeling the apo, non-bound-like PD-1 interface into a bound-like state. We observed that mGDV opened Asn66 using a native-like hydrogen bond network analogous to those seen in previous simulations (Figures 5, 7 and 10d). However, Ile126 and Ile134 dynamics mirrored those seen in the pembrolizumab cocrystal, with both sidechains favoring inward-flipped ‘closed’ configurations (Figure 11). Simulation trajectories showed that the short Val side chain of the mGDV motif, unlike the cognate Tyr123/112 anchors, did not penetrate deep enough into the PD-1 interface to be a ‘wedge’ stabilizing an open Ile134. Instead, like the carbon chain of pembrolizumab residue Arg102, the Val ‘capped’ stable hydrophobic interactions between a closed Ile134, a closed Ile126, and the neighboring Gly124.

Figure 11

Download asset Open asset

Figure 11—source data 1

Excel workbook with a single sheet containing the 2D histogram data for the heatmaps shown in Figure 11.

https://doi.org/10.7554/eLife.22889.030

Download elife-22889-fig11-data1-v2.xlsx

Our GDG, ADG, GDY and ADY simulations demonstrated that precise regulation of the closed/open states Ile126 via the Ala121 C_ß is realized only when the Tyr123/112 anchor is buried (Figure 7). Thus, given that mGDV lacks an anchor, a natural question to ask is whether a Ile126 would be opened by a GDV peptide without the N-methyl group. Interestingly, MDs of PD-1 interacting with a GDV peptide revealed identical Ile126 and Ile134 dynamics to mGDV simulations (Figure 11), indicating that the N-methyl group was not recruiting Ile126 into the closed state in the style of Ala121 C_ß. These results help to further illuminate the role of the conserved anchor Tyr123/112, which in its absence does not wedge Ile134 into the open state, disabling the capability of PD-1 to stabilize an open Ile126 and form a hydrophobic cavity at that site.

Compared to GDG simulations in which Ile126 fluctuated between open and closed (Figure 4b,c), in GDV simulations it remained closed, suggesting a stabilizing role for the Val side chain. The overlap of (m)GDV’s Val with the carbon chain of pembrolizumab’s Arg102 (Figure 10—figure supplement 1) and the similarity between the (m)GDV-induced PD-1 interface and the pembrolizumab–bound interface supports the ‘capping’ role of Arg102 in stabilizing the flat hydrophobic surface of PD-1. This mechanism is also consistent with models of macrocycle conformations generated by Balloon (Vainio and Johnson, 2007) docked to PD-1, which readily identify poses that align the mGDV motif to corresponding PD-L1 and pembrolizumab interface residues (Figure 10—figure supplements 1 and 2), rationalizing the potency and specificity of the compound.

Our studies show that apo PD-1 does not sample bound-like hydrophobic interface conformations, but instead presents a non-bound-like hydrophilic patch around Asn66 at the core of its binding interface (Figure 2). By mapping the effect of specific ligand contacts on the apo PD-1 interface, we identify a highly conserved subset of PD-L1/2 motifs responsible for coordinating Asn66 and triggering the transition from the hydrophilic to hydrophobic interface. Namely, Asp122/111 and the backbone O of PD-L1/2 Ala121/Trp110 form a robust, four-membered hydrogen bond network with Tyr68 and Asn66 that neutralizes the latter residue into a bound-like open state. Simultaneously, the conserved anchor Tyr123/112 stabilizes Ile134 into a bound-like state that, with Asn66 open, creates a hydrophobic surface that fluctuates between a patch and a cavity modulated by Ile126. These three linear ligand motifs (XDY), shared by both PD-L1/2, comprise the molecular key that unlocks the promiscuity of PD-1 by revealing a flexible hydrophobic binding surface (Figure 4).

With XDY triggering the transition to the flexible hydrophobic surface, specificity toward the two PD-1 ligands is actualized by the formation of the hydrophobic patch when binding PD-L1 vs. the formation of hydrophobic cavity when binding PD-L2. These two states can be distinguished by the conformation of Ile126 (Figure 2e). For PD-L1, we show that the ADY motif is sufficient to stabilize the hydrophobic patch (Figure 4c). Specifically, the Ala121 Cβ atom, which does not overlap with PD-1 apo NMR structures (Figure 2d), recruits Ile126 into the closed (patch) state. On the other hand, in the absence of C_β, the GDY trigger stabilizes the open state of Ile126, producing a large hydrophobic interface cavity consistent with the pocket that buries PD-L2 Trp110. Note that the Trp in the WDY motif of PD-L2 readily fills the hydrophobic pocket as the XDY motif latches and opens Asn66 (Figure 8).

The pre-arrangement of PD-L1/2 motifs XDY in bound-like conformations in the absence of the receptor is important for efficient ligand recognition and binding. Docking studies and peptide MDs highlight a critical role for the conserved Tyr123/112 anchor both in both molecular recognition and in modulating Ile134 during induced fit, both of which require the Tyr side chain to maintain a stable bound-like rotamer. Furthermore, simulations demonstrate that peptides such as GDG, mGDV, and GDV, which either lack or have a modified anchor analogue, cannot stabilize an open state of Ile126, highlighting an allosteric role for Tyr123/112 in splitting the PD-1 induced fit binding pathway.

Several anchors substitutes were tested in simulation starting in bound-like configurations similar to the cognate Tyr112/123. These MDs produced three broad types of PD-1 interface dynamics (Figure 4—figure supplement 3): (1) aromatic substitutions XDF and XDW stabilized either an open (X=G) or closed (X=A) pocket like the cognate XDY motif. (2) Polar substitutions XDH, XDR, and XDK were not accommodated in the hydrophobic anchor pocket and their side chains laid along the receptor surface, consistent with pembrolizumab’s bound Arg102 (Figure 10b), producing a closed pocket like that of (m)GDV. (3) XDG or XDA resulted in open-closed fluctuations of both Ile134 and Ile126 (Figure 4b,c). These observations suggest that certain anchor mutations are tolerated by PD-1 and are consistent with mutagenesis studies showing that the Y112A PD-L2 point mutation slightly reduces, but does not abolish, binding to PD-1 (Lázár-Molnár et al., 2008). However, the observed conservation of Tyr123/112 in mammalian species (Lázár-Molnár et al., 2008) might suggest specific kinetic constraints on ligand recognition arising from hydrophobic contacts with Ile134 and the hydrogen bond with Glu136 (Figure 9), which are not shared by other sidechains.

In addition to the anchor residue, our peptide MDs also suggest an essential role for the conserved Asp122/111 in erecting a stable hydrogen bond network that opens PD-1 Asn66, which can only be achieved by a bound-like Asp side chain. The primacy of these intermolecular interactions to PD-1 binding is reinforced by our MDs of apo PD-L1/2, which reveal that Tyr123/112 and Asp122/111 all remain in bound-like conformations in the absence of the receptor, primed to interact immediately upon interface association. Equally important is the fact that apo PD-1 structures all accommodate (i.e. do not block) any of contacts of the XDY scaffold, ensuring a rapid recognition process that facilitates subsequent induced fit transitions.

Our MDs demonstrate that the set of consecutive intermolecular interactions triggered by ADY and GDY peptides lead to energetically downhill binding pathways with no opposing energy barriers. These pathways strongly suggest that PD-1 occurs mostly by induced fit (Figure 1). Specifically, simulations and estimated ${\displaystyle \mathrm{\Delta }{G}_{open}}$ values show that apo_BL states of PD-1 are rare, which undermines a conformational selection mechanism. On the other hand, ligand-specific triggers are shown to efficiently shift the PD-1 interface conformational ensemble from a non-bound-like: bound-like ratio of roughly 44: 1 (in the apo ensemble) to roughly 1: 7 (in the encounter complex ensemble) (Figure 7). Unconstrained MDs of PD-L1/2 encounter complexes show that the geometry and chronology of triggering contacts is highly optimized, driving the transition from the non-bound-like to the bound-like states in less than 10 ns. This time scale promotes rapid recognition and ensures fast activation of this important T-cell checkpoint.

Although regulatory proteins are promiscuous in that they bind multiple targets, they must also be specific so as to limit binding to just those targets. Our analysis of the binding mechanism to PD-1 reveals how these two seemingly contradictory requirements can be simultaneously achieved. Here, we show that apo PD-1 samples an ensemble of non-bound-like conformations that present an obstructive Asn66 on its interface, which likely prevents non-specific binding. The apo PDL1/2 interfaces feature a conserved, bound-like, XDY binding motif that holds the key to opening Asn66 and forming a flexible hydrophobic surface, which completes the first binding step. In the second step, the ligands then attune the flexible interface via specificity-determinant contacts (X=A for PD-L1, X=W for PD-L2) that modulate Ile126, splitting the binding pathway and stabilizing either a hydrophobic patch or a binding pocket (Figures 2, 4 and 7). The key structural properties in this pathway are: (a) a flexible, non-bound-like apo receptor interface ensemble that presents an unfavorable binding surface, (b) a core subset of shared ligand binding motifs clustered about an anchor residue that latch the receptor interface but allow it to remain flexible, and (e) ligand-specific motifs that split the binding pathway by stabilizing different conformations of the flexible interface.

Bound cocrystal structures of the PD-1–targeting antibody pembrolizumab reveal that it exploits an evolutionarily designed induced fit trigger: the four-membered hydrogen bond network that opens Asn66 and makes the receptor interface hydrophobic. This same principle can be applied to design smaller molecular weight PD-1 inhibitors. We have shown that the mGDV motif of the Brystol-Myers-Squibb PD-1 inhibitor combines key pharmacophore features of both PD-L1 and pembrolizumab interfaces: the backbone O of the Gly resembles that of PD-L1’s Ala121, the Asp side chain resembles PD-L1’s Asp122, and the Val side chain resembles pembrolizumab’s Arg102. Simulations suggest that this structural resemblance produces functionally similar dynamics by displacing receptor residue Asn66 (Figure 10d) and stabilizing a bound-like, flat hydrophobic surface formed by closed Ile126 and Ile134 (Figures 7 and 11). Docked conformations of the full inhibitor recapitulate most secondary native-like contacts in addition to the core triggering interactions (Figure 10—figure supplements 1 and 2). Taken together, these results support the idea that nature’s mechanisms for modulating receptor surfaces might be exploited to design novel chemistries capable of transforming hard to drug targets into more druggable candidates.

Promiscuous regulatory proteins must optimize binding kinetics for multiple ligands by exploiting structural flexibility. Given nature’s general mechanisms for flexibility-mediated binding (Figure 1), specificity toward multiple ligands could, in principle, be conferred either through conformational selection, by evolving the receptor to intrinsically sample different ligand-specific apo_BL states, or by induced fit, by evolving interface interactions that efficiently drive transitions to the ligand-specific EC_BL states. If conformational selection is used to achieve multi-ligand specificity, the binding pathway flux will de facto be limited by ${\displaystyle \mathrm{\Delta }{G}_{BL}^{apo}}$, the free energy difference between each ligand-specific bound-like states and other states in the apo ensemble. In this scenario, a natural bottleneck would emerge as an increasing number of ligands would lead to lower association rates.

On the other hand, if selective promiscuity is conferred through induced fit, binding pathway flux will not depend on the fractional populations of apo ensemble microstates, but instead will be determined by the ligand-specific triggering mechanisms. We show here that induced fit can efficiently reshape the shallow polar interface of a flexible receptor into a hydrophobic interface amenable to binding multiple ligands by co-evolving a common set of intermolecular contacts. From an evolutionary perspective, this is an efficient approach to spawning novel protein interactions, since these core contacts can be designed just once. Selectivity to novel ligands can then be achieved by evolving relatively small sequence modifications around these core contacts. Perhaps more importantly, we note that contrary to conformational selection, the induced fit approach to selective promiscuity is in principle not limited by the total number of ligands.

It is interesting to note that many well-characterized eukaryotic regulatory domains (Pawson and Scott, 1997) bind to several linear binding sequences that share common motifs around an anchor residue and differ in other nearby regions. This trend suggests that the selective promiscuity via induced fit mechanism proposed here for PD-1 might apply elsewhere in nature. This possibility is currently being studied by analyzing the triggers of induced fit in other systems.

Molecular dynamics simulations (MDs) of the extracellular domain of PD-1 were run in triplicate using the first three solution NMR structures of apo human PD-1 (PDB ID: 2M2D [Cheng et al., 2013]). Before simulating specific receptor-ligand interactions, MDs of apo PD-1 were evaluated to ensure that the resultant dynamics are consistent with the experimentally derived apo NMR ensemble. As shown in Figure 12, apo MDs stabilize within about 2.0 Å backbone RMSD of their respective NMR starting points, suggesting that we can successfully sample native-like unbound states.

Figure 12

Download asset Open asset

Figure 12—source data 1

Excel workbook with a single sheet containing the time-series RMSD-to-unbound data from the apo PD-1 simulations plotted in Figure 12.

https://doi.org/10.7554/eLife.22889.032

Download elife-22889-fig12-data1-v2.xlsx

Available co-crystal structures of human PD-1/human PD-L1 (PDB ID: 4ZQK [Zak et al., 2015]), murine PD-1/human PD-L1 (PDB ID: 3BIK [Lin et al., 2008]) and murine PD-1/murine PD-L2 (PDB ID: 3BP5 [Lázár-Molnár et al., 2008]) complexes were used as templates for placement of peptides in bound-like loci at the receptor interface, and the dynamics of the PD-1 binding interface in response to interactions with different structural motifs on the ligands were analyzed. We focus on interactions relevant for the opening and closing of the pocket around Asn66. Based on co-crystals, we noticed that the core interacting residues of PD-L1 (Ala121, Asp122, Tyr123) and the homologous residues on PD-L2 (Trp110, Asp111, Tyr112) form critical hydrogen bonds (hydrogen bonds) shaping this pocket. Thus, to dissect the contribution of each contact, we simulate the effects of the receptor interacting with a diverse set of peptide derivatives of these specific ligand residues.

Ten distinct PD-1 systems were simulated in order to dissect the ligand groups that trigger induced fit interface deformations on the receptor. These systems included the apo receptor in isolation and in complex with nine different peptides that mimic cognate ligand backbone and side chain interactions with the receptor (Figure 3): the anchor residue Tyr, the backbone peptide GGG, five peptides to probe role of ligand side chain contacts DY, GGY, GDG, ADG, GDY, the PD-L1 peptide ADY, and the mGDV peptide, which mimics a patented PD-1 inhibitor.

To generate initial structures for our receptor-peptide MDs, NMR models 1–3 of the human PD-1 were backbone aligned to the murine receptor co-crystal (Lin et al., 2008) and peptides were modeled after the corresponding human PD-L1 interface residues Ala121–Tyr123, homologous to PD-L2 interface residues W110-Tyr112. Systems are simulated for 200 ns, resulting in three replicate MDs per system (including the apo PD-1 system, which does not include any peptide), and receptor interface dynamics are compared across systems to identify the ligand motifs and interactions responsible for structural transitions toward the bound-like receptor state. Harmonic restraints (100.0 kcal/mol) on all heavy atoms of ligand-mimicking peptides were used in simulation to prevent dissociation of the peptide from the receptor interface.

In peptide MDs, harmonic restraints (100.0 kcal/mol) were also placed on backbone atoms of non-interface PD-1 beta sheets residues 50–55, 80–81, 96–98, 106–109 and 120–122. These residues exhibit <0.35 Å backbone RMSD in the apo NMR ensemble, and previous studies have also shown that the conformational changes induced by ligand binding do not propagate through the major fold of PD-1 (Cheng et al., 2013). Hence, these restraints should not prevent our ability to sample the native-like binding dynamics of the receptor interface in biological conditions. The resolved portion of the N-terminal tail of PD-1 (residues 33–36), which in NMR models has <0.65 Å backbone RMSD, was also restrained so as to limit artificial mobility that might result from the fact that residues 1–32 were missing from simulation.

Human PD-1–PD-L1 and PD-1–PD-L2 encounter complexes were modeled and then simulated in triplicate to probe induced fit trajectories and determine the chronology of inter-molecular interactions and specific interface deformations. We modeled encounter complexes by rigid body docking the extracellular domain of the apo receptor and the Ig-like V-type domains of the apo ligands, allowing no structural overlaps. Docked models of PD-L1 had an average backbone RMSD of 5.7 ± 1.2 to the human PD-1–PD-L1 cocrystal. Docked models of PD-L2 had an average backbone RMSD of 4.8 ± 1.8 to the murine PD-1–PD-L2 cocrystal (no human cocrystal is currently available for the PD-1–PD-L2 complex).

Structural models of apo human PD-L1 and PD-L2 that we used when building encounter complexes were generated by simulating the ligands in solution for 400 ns, using a VMD (Humphrey et al., 1996) clustering plugin (https://github.com/luisico/clustering) to cluster frames by backbone RMSD using a 3 Å cutoff, and taking the centroid frame of the largest cluster for each ligand. The initial structure for the PD-L1 clustering MDs was taken as the structure of the bound human ligand from the co-crystal complex with murine PD-1 (PDB ID: 3BIK). As there are currently no available crystal structures of human PD-L2, a homology model was built as a starting point for the clustering simulation by manually mutating the bound structure of murine PD-L2 (PDB ID: 3BP5) and minimizing the resulting structure. We used the ClusPro protein-protein docking server (Comeau et al., 2004) to dock the top apo PD-L1 and PD-L2 centroid structures from their respective MDs to the first three NMR structures of apo human PD-1 (all three receptor structures are non-bound-like). Three bound-like candidate models for the PD-1–PD-L1/2 encounter complexes that correctly anchored Tyr123/112 were chosen from the ClusPro output. We then simulated these encounter complexes for 400 ns to probe the dynamics of the induced fit binding pathway.

We ran MDs using AMBER14’s (Case et al., 2014) pmemd.cuda module (Götz et al., 2012) and the AMBER ff12SB force field. The cutoff for non-bonded interactions was set at 10 Å. Systems were simulated in an octahedral TIP3P water box with periodic boundary conditions and a 12 Å buffer around the solute. Cl ions were added to the solvent to neutralize the charge of the systems. We minimized each system twice and then equilibrated them before beginning production runs. In the first minimization, solute atoms were held fixed through 500 steps of steepest descent and 500 steps of conjugate gradient minimization. In the second minimization, only the solute backbone atoms were restrained through 2000 steps of steepest descent and 3000 steps of conjugate gradient. After minimization, system temperatures were raised to 300 K over the course of a 200 ps constant volume simulation (with an integration step of 2 fs) during which the solute was fixed with weak (10.0 kcal/mol) restraints. Bonds involving hydrogens were held at constant length. For the production MDs, the 200–400 ns simulations were held at 300 K under constant pressure with the constraints as listed above for each system and an integration step size of 2 fs.

The PyMOL Molecular Graphics System v1.7.4.0 was used for structure preparation and analysis (Schrödinger, 2010). Trajectories were analyzed using VMDv1.9.2 (Humphrey et al., 1996) and the MDpocket software package v2.0 (Le Guilloux et al., 2009; Schmidtke et al., 2010) for cavity detection and volume/surface area measurement. Measurements of PD-1 binding pocket occlusion, shown in Figures 2d and 4d,e, were calculated from molecular dynamics simulations of PD-1 using a Python script (available at https://github.com/npabon/md_pocket_occlusion; a copy is archived at https://github.com/elifesciences-publications/md_pocket_occlusion [Pabon, 2016]). Briefly, the script takes a molecular dynamics trajectory and a set of static reference atoms and identifies which reference atoms are overlapped in each frame of the simulation. Overlap occurs when any simulated atom crosses the ‘clash radius’ of a reference atom, the clash radius being equal to the sum of the van der Waals radii of the two atoms. The output of the script is the fractional occlusion of each reference atom position, equal to the percentage of simulation frames in which that reference atom is overlapped by simulated atoms. This script was used to evaluate the extent to which the Trp110 and Tyr112/123 binding cavities are open in simulations of PD-1 interaction with various peptides, simulations of apo PD-1, and the apo NMR ensemble of PD-1.

We classified PD-1 interface conformations using two binary order parameters that define whether interface residues Asn66 and Ile126 are in their ‘open’ or ‘closed’ rotamer states. These parameters are used to distinguish the non-bound-like interface, where Asn66 is closed and Ile126 is open, from the PD-L1-specific bound-like state, where Asn66 is open and Ile126 is closed, and the PD-L2-specific bound-like state, where both Asn66 and Ile126 are open (Figure 2e). We estimated the energy differences ${\displaystyle \mathrm{\Delta }{G}_{BL}^{apo}}$ and ${\displaystyle \mathrm{\Delta }{G}_{BL}^{EC}}$ (Figure 1) using Maxewell-Boltzmann statistics by assessing the bound-like (BL) and non-bound-like (NBL) state population distributions in the apo and encounter complex (EC) receptor ensembles:

In the above equations, ${\displaystyle ⟨n_{BL}^{apo/EC}⟩}$ and ${\displaystyle ⟨n_{NBL}^{apo/EC}⟩}$ denote fractional equilibrium populations of the apo / encounter complex receptor ensembles in the bound-like and non-bound-like macrostates, and k_BT is the product of the Boltzmann constant and temperature. We used MDs to generate the equilibrium ensembles of receptor conformations and analyzed the trajectories to calculate ${\displaystyle ⟨n_{BL/NBL}^{apo/EC}⟩}$ values.

MDs trajectories were analyzed as follows. Reference structures for the open and closed states of Asn66 were defined using its side chain configuration in the first apo NMR model and PD-L1-bound human cocrystal, respectively (Asn66 has <0.2 Å heavy atom RMSD between PD-L1 and PD-L2 cocrystals 4ZQK and 3BP5). Each frame of the MDs trajectory is labeled with the state to which the simulated Asn66 had the smaller side chain RMSD to the reference structure. Reference structures for the open and closed states of Ile126 were defined using its X₁ rotamer angle in the murine PD-L2 and human PD-L1 cocrystals, respectively, this angle being the main distinguishing feature between the two different ligand-bound interfaces (Figure 2e). Each frame of the MDs was labeled with the state to which the simulated Ile126 had the closest rotamer angle. The free energy changes of opening Asn66 and Ile126 are calculated using Equation (1) and then compared across different simulations in order to identify triggers of interface deformations.

The funders had no role in study design, data collection and interpretation, or the decision to submit the work for publication.

This work was funded by the National Institute of Health (grant GM097082 to CJC) and the National Science Foundation (grant 1247842 to NAP).

© 2017, Pabon et al.

This article is distributed under the terms of the Creative Commons Attribution License, which permits unrestricted use and redistribution provided that the original author and source are credited.