Structural characterization of AlphaFold prediction
AlphaFold2-multimer predicted with high confidence the full-length dimeric structure of E. coli PhoQ, achieving an average predicted local distance difference test (pLDDT) score of 0.8. [47] (Supplementary Figure S1 A). The generated model formed a symmetrical dimer, with a root mean square deviation (RMSD) of only 0.75 Å between the two monomers. The dimeric interface was characterized by a continuous helix (Figure 1B, C). Notably, there were marked drops in the pLDDT score in regions aligned well with the expected boundaries between the different domains of PhoQ (Figure S1 B). These discernible drops were leveraged to demarcate the start and end points of each domain for subsequent analyses.
The predicted PAS domain structure closely resembled the known crystallographic structure of E. coli (PDB ID: 3bq8) at the monomeric level, with a low RMSD of 0.6 Å [31]. However, significant deviations emerged in the predicted dimer (RMSD: 5.25 Å, Supplementary Figure S2). The interfacial helices in the model formed a shallower angle (10°) compared to the experimental structure (41°), impacting the orientation of the critical acidic patch (residues D136-D150) [48]. To quantify this difference, we embedded the Alphafold model into a lipid bilayer, using the orientations of proteins in membranes (OPM) server [49], and measured angles of 50.21° and 45.25° between the membrane surface and the acidic patch helices for chains A and B, respectively. To compare the orientations of the acidic patches in the experimental structure, we aligned the sensor interfaces (residues T48-K64) with the oriented Alphafold model. The angle for chain A remained comparable (47.02°), but was more parallel for chain B (25.85°), due to the inherent asymmetry observed in the experimental structure. Additionally, the crucial salt bridge between Asp179 and Arg50' [31] present in the experimental structure was absent in the prediction, despite a comparable Cα-Cα distance (approximately 11.4 Å for both).
The transmembrane domain of PhoQ formed a four-helical bundle with antiparallel helices per monomer, defined as TM1 (S12 - V42) and TM2 (Y189 - A214). This structural arrangement is strongly corroborated by prior research [31], [32], [50]. Both TM1–TM1’ and TM2–TM2’ assembled in coiled coils with single-residue insertions disrupting the typical helical packing at positions V25 and S200, respectively (accommodation index of 1.0 [51]). To assess the novelty of this fold, we employed the FoldSeek search tool to query the Protein Data Bank (PDB) (Table S1) [52]. The four-helix bundle aligned with various structures; however few were transmembrane proteins (best PDB ID: 6sss_C, TM-score: 0.46, RMSD: 3.90 Å). This probably reflects the scarcity of such structures in the database. In contrast, comparison with the commonly used HTrII transmembrane bundle [36] using the TM-align mode [53] (RCSB.org [54]), yielded a significantly higher TM-score of 0.55 and a reduced RMSD of 2.56 Å. Additionally, we compared the Alphafold model to a previously published atomic model of the transmembrane domain [39] (Supplementary Figure S3A, B). Despite some variations (increased TM1 separation and closer TM2 helices in the published model, RMSD: 3.8 Å, TM-score: 0.56), the overall folds aligned well, enabling a seamless connection between the sensor interface helices and the HAMP domain (Supplementary Figure S3C).
The predicted HAMP domain adopted the expected two-antiparallel helix configuration, connected by a 10-residue loop. The FoldSeek search against the PDB database identified three HAMP domains from sensor proteins among the top hits but with low structural similarity (TM-Score: 0.6). The key difference resided in the interhelical angle. This variability was further confirmed by a statistical analysis (principal components analysis) of the identified HAMP domains, which revealed a wide distribution of angles, suggesting a lack of a single preferred conformation for the HAMP domain in these related sensor proteins.
The predicted DHp domain consisted of two alpha helices (a1 and a2) connected by a hairpin loop, forming a left-handed four-helix bundle in the dimer. This structure closely resembles other experimentally determined DHp domains (best TM-score: 0.84). Notably, the conserved histidine residue (H277), critical for the phosphorylation cascade, resides within the a1 helix with its side chain exposed to the cytoplasm. Moreover, the handedness of the bundle places the intra-protomeric catalytic domain in close proximity to this conserved histidine, strongly suggesting a cis-autophosphorylation mechanism [55].
The catalytic domain adopted the expected Bergerat fold and aligns, closely matching the experimentally solved structure of E. coli (PDB id: 1id0, RMSD=1.7Å) [56]. However, a key structural difference resided in the loop forming the ATP-lid, which connects helix α3 and beta-strand β3. In the 1id0 structure, the ATP-lid extended further, exposing the ligand to the solvent (Supplementary Figure S4). Although Alphafold was not trained with ligands, it has been shown that the configuration of binding pockets can often accommodate them [57]. We successfully positioned the ATP molecule and Mg2+ cation from the experimental structure (1id0) into the predicted binding pocket by aligning the main helices of the catalytic domains. In both structures, the Mg2+ cation was chelated by three oxygens of the triphosphate head and an additional interaction with N388 or N389, N385, and Q442 in the predicted and experimental structures, respectively. Conserved interactions were observed between K392 and N389 with the triphosphate head in both structures. However, the predicted structure lacks interactions with the ATP-lid residues R434 and R439. Furthermore, the adenosine moiety of the ATP engaged in pi-pi interactions with Y393 and formed a hydrogen bond with D415 in both structures
Structural refinement with all-atom MD simulations
In order to further refine the predicted structure of the full-length PhoQ dimer, we carried out a set of all-atom molecular dynamics simulations. The AlphaFold model () was inserted into a bacterial mimic membrane bilayer and three replicas of all-atom MD simulations were conducted. These simulations were performed under near-physiological conditions (310K and 1 atm pressure), with each replica extending up to approximately 1.4 microseconds (Supplementary Table S2). The models remained stable throughout the simulations (average RMSD: 8.76 2.57 Å, Supplementary Figure S5). The major difference was associated with the rearrangement of the catalytic domains, moving slightly away from the DHp. The protein's core exhibited only marginal deviation from the predicted model, maintaining an average RMSD of 4.23 ± 1.28 Å (Figure 2A and Supplementary Figure S6A). The cross angle between the helices of the sensor interface remained consistently stable, averaging 7.38° ± 3.76° (Supplementary Figure S7). During the simulations, we also observed the dynamic formation of a salt bridge between D179 and R50’ (42% of simulation time with a cut-off distance of 3.6 Å) and R50 and D179’ (22% of simulation time).
We attempted to refine the AlphaFold model's sensor interface using targeted molecular dynamics (TMD) simulations () to match the conformation observed in the experimental dimer structure (PDB ID: 3bq8). However, this introduced significant structural distortions (RMSD of full protein: 10.32 3.76Å and core: 5.19 2.03Å). In particular, the helix connecting the sensor to TM1 partially unfolded (residues D45 to T47), potentially leading to the decoupling of the sensor domain from the transmembrane region (Supplementary Figure S6B). We concluded that the TMD simulation employed was insufficient to induce the desired conformational alteration in the AlphaFold model, suggesting a more intricate structural coupling between the sensor and transmembrane domains that may hinder the desired conformational change in the model.
Given the good structural alignment between the previously modeled transmembrane [39] and , we constructed a hybrid model () by combining the modeled transmembrane domain with the AlphaFold-predicted sensor and cytoplasmic domains (Supplementary Figure S6C). We then performed all-atom MD simulations in triplicate to refine the structure and assess its stability. The core of PhoQ remained stable throughout the simulations (4.16 0.47Å, Supplementary Figure S5) and the sensor interfacial helices maintained a parallel alignment (cross angle: 13.10 ± 4.00°, Supplementary Figure S7). The simulations also captured the release of one of the catalytic domains.
Encouraged by the stability of the , we conducted a TMD simulation () to induce an interface angle resembling the experimental dimer (PDB ID: 3bq8). Even though the structure remained stable during the subsequent 1.5 µs unrestrained MD simulations, conducted in three replicas, (RMSD of 7.46 3.24 Å, Supplementary Figure S5), the resulting conformation deviated significantly from (Figure 2B and Supplementary Figure S6D). The equilibrated conformation of was characterized by a highly asymmetric and curved structure, with several kinks along the core. Notably, the sensor domain tilted significantly (49.12° ± 5.11°) relative to the membrane compared to its perpendicular orientation in . Furthermore, a kink at position F44 in chain B allowed the sensor interface to retain a cross-angle more closely aligned with the experimental structure, approximately 27.22° ± 4.63°. This conformation was similar to the experimental structure (PDB ID: 3bq8, RMSD: 4.07 0.44 Å). The interface helices pivoted around E55, while the E55 (chain A) and R53 (chain B) formed a distance of 3.7 Å for 33% of the simulation time, agreeing with the experimental structure. The increased cross-angle led to a near vertical alignment of the acidic patch helix on chain B (~ 71.2°), bringing the acidic cluster close to the membrane surface. In contrast, the acidic patch helix of the other protomer aligned almost parallel to the membrane surface, forming an angle of approximately 7°. Although the conformations of the HAMP and DHP domains remained very similar to (1.82 ± 0.30 Å RMSD), the catalytic domain of chain A was released and pivoted towards the DHp region.
Conformational change of the transmembrane bundle
To understand how the altered transmembrane domain between and affected the protein structure, we investigated structural deviations propagated along this domain. We divided the transmembrane domain into three slices, each defined by two alpha carbons per protomer and oriented parallel to the membrane (Figure 3 and Supplementary Figure S8). The top slice (closest to the sensor domain) comprised residues A36 and W194, the middle slice (center of the membrane) residues L28 and N202, and the bottom slice (closest to the HAMP domain) residues T21 and L210. Compared to , exhibited consistently larger TM1-TM1' distances and smaller TM2-TM2' distances. This resulted in TM1-TM1' helices forming a cone shape, narrowing towards the sensor domain, whereas TM2-TM2' helices adopted an inverted cone shape. Interestingly, despite the overall asymmetry of , the arrangement of residues within each slice displayed a more symmetrical, diamond-like shape compared to the distorted appearance in (Figures 3 and Supplementary Figure S8). The configuration of was accompanied by a bent at I207 of approximately ~3.5 degrees proximal to P208 in chain A (compared to ~6.8 degrees in ), and a pistonning movement between transmembrane helices (Supplementary Figure S9). A partial unfolding on the connecting region between TM2 and the sensor domain (I181 – Y189) was also observed in . In both models, the amphipathic N-terminal helices of TM1 (M1 to P8) were folded back and bound to the membrane surface.
Comparison to experimental cross-linking data
When compared to available experimental cross-linking data for the sensor interface, transmembrane (TM), and HAMP domains [38], all structural models exhibited generally good agreement between the cross-linking fractions and average inter-protomeric Cα-Cα distances (Figure 4). However, some key discrepancies emerged. Notably, the absence of cross-links for residues A20-L30 and the decreased periodicity for residues S43-D45 align with the larger interatomic distances and kink in the TM1 helix as it exits the membrane observed within the model.
Solvation of the transmembrane bundle
Previous experiments have shown that the polarity of N202 is essential for the function of PhoQ [37], [39]. This residue, located at the center of the transmembrane bundle, is thought to be crucial for solvating the bundle and enabling transitions between PhoQ signaling states. Consistent with this hypothesis, the and models exhibited extensive transmembrane solvation throughout the simulations, contrasting with the complete absence of water molecules in the model (Figure 5). Detailed analysis revealed a dynamic pattern of water occupancy within the transmembrane bundle. On average, 1.7 water molecules were present within the TM bundle of . The most frequent configuration involved two water molecules occupying the bundle (45 11 %). Occasionally, only one water molecule remained (23 26 %), and rarely, the bundle became dehydrated (14 12 %). The average residence time for an individual water molecule was around 287 156 ns, with frequent exchanges occurring. Interestingly, some water molecules resided within the bundle for significantly longer durations, up to 1 microsecond (μs).
The transmembrane domain of displayed a greater degree of water occupancy, with an average of 3.5 water molecules per frame. The primary configurations involved either one (20 30%) or two water molecules (19 8 %) or a more dynamic distribution of three to seven water molecules (27 24 %). The average water residence time was about 70 50 ns, with one instance exhibiting a residence time of 1.4 μs.
Interactions with Mg2+ cations
Considering the crucial role of PhoQ in Mg²⁺ sensing, we investigated the presence of Mg2+ ions in the vicinity of the acidic patch (residues: L131 - D150), known to be important for sensing cations. Minimal interactions with Mg2+ were observed for the and models (less than 1% of the total simulation time for both chains). In striking contrast, the acidic patch of the model exhibited extensive Mg²⁺ interactions with the acidic patch, particularly residues D149 and D152 of chain A (100% simulation time). Interestingly, each bound Mg²⁺ cation also interacted with a lipid phosphate group of the membrane, either directly or indirectly through water molecules (Figure 6). Although positioned further from the membrane surface, residue D125 in the acidic patch of chain B for still bound to a Mg²⁺ ion for 95% of the simulation time.
Dynamics of the cytoplasmic region
Despite overall similar appearances across simulations, principal component analysis (PCA) identified subtle but distinct conformational changes within the cytoplasmic domain of PhoQ (Supplementary Figure S10). The first principal component (PC1) captured a dominant bending motion in the B-chain of the S249-E296 helix (Supplementary Figure S10B, D), whereas the second principal component (PC2) described a scissoring motion of the N-terminal helices of the HAMP domain (R219-H234) (Supplementary Figure S10C, E). These small conformational shifts within the HAMP and DHp domains resulted in the dissociation of the catalytic domain from the core region in several simulations. To characterize these diverse conformations, we employed two key metrics: (i) the distance between a catalytic residue (H277) and the center of mass of the catalytic domain (excluding the lid region), and (ii) the angle formed between the Dhp domain core and the catalytic domain's gripper helix (G445-E457). This analysis identified seven distinct poses adopted by the catalytic domain (Figure 7). The most prevalent conformation (red in Figure 7) closely resembled the Alphafold prediction, with the gripper helix bound to the DHp domain. Notably, at least one cytoplasmic domain adopted this conformation in almost all simulations (Supplementary Figure S11). Two additional conformations (light orange and green in Figure 7) exhibited an orientation that resembled a trans-phosphorylation state. These conformations positioned the catalytic domain near the core, with the gripper helix perpendicular to the Dhp domain. However, they crucially lacked interaction with the key residue H277. The light orange conformation, observed primarily in chain A of the model, positions the domain slightly closer to the core compared to the green conformation. The remaining clusters (yellow, purple, cyan, and orange) all represented detached conformations of the catalytic domain lacking interaction with the Dhp core. Notably, the cyan cluster corresponded to a conformation where the domain moved towards and directly interacted with the membrane surface.