Impact of the Cellular Environment on Adenosine Triphosphate Conformations

Abstract

The cytoplasm is an environment crowded by macromolecules and filled with metabolites and ions. Recent experimental and computational studies have addressed how this environment affects protein stability, folding kinetics, protein-protein and protein-nucleic acid interactions, though its impact on metabolites remains largely unknown. Here we show how a simulated cytoplasm affects the conformation of adenosine triphosphate (ATP), a key energy source and regulatory metabolite present in high concentration in cells. Analysis of our all-atom model of a small volume of the E. coli cytoplasm when contrasted with ATP modelled in vitro or resolved with protein structures deposited to the Protein Data Bank, reveals that ATP molecules bound to proteins in-cell form specific pitched conformations that are not observed at significant concentration in the other environments. We hypothesize that these interactions evolved to fulfil functional roles when ATP interacts with protein surfaces.

Graphical Abstract

Proteins in vivo perform their biological functions in a complex environment of water, inorganic ions, metabolites, and other macromolecules. The cell’s interior contains ~10-40% macromolecules by volume,¹ creating a crowded environment where nonspecific, nonfunctional interactions between various components are inevitable. Experimental and computational studies have demonstrated that the cellular environment impacts proteins’ stability, folding kinetics, structural ensembles, binding, and quinary structure.^2–11 However, the impact of the cellular environment on small molecules is largely unknown.

Adenosine triphosphate, or ATP, is one of the most abundant metabolites in the cell, with a concentration in the ~1-10 mM range across a variety of cell types and organisms.^12–14 ATP is canonically known as the ‘energy source of the cell’.¹⁵ However, recent studies have found that the presence of ATP has a variety of effects on nearby macromolecules, including increasing solubility,^16,17 stabilizing nearby proteins,¹⁸ and modulating the liquid-liquid phase separation of intrinsically disordered proteins.¹⁹ These findings point to the evolution of additional functions enabled by the close proximity of ATP molecules and macromolecules in the dense environment of the cell.

Previous studies have also shown that ATP’s geometry in water-box MD simulations differs from the structures found in the PDB.²⁰ However, it is unknown whether these differences are due to crystal packing, or whether the unique steric and electrostatic environment of the cytoplasm also affects ATP’s conformational distribution in the proximity of proteins or other macromolecules.

Here, we study the following question: are all the structures observed for ATP bound to proteins in the PDB (mostly crystal structures) representative of ATP bound to proteins in-cell, or does the even stronger crowding in crystals compared to cells produce structures that, while corresponding to local minima in the ATP energy landscape, are not populated biologically? By comparing over 3,500 ATP structures in the Protein Data Bank²¹ (PDB) with atomistic simulations of ATP in a model E. coli cytoplasm, and in a water box, we find that ATP’s conformational space is less impacted by the cellular environment than by crystal packing: the PDB contains many ATP structures at high resolution that do correspond to actual local minima in the energy landscape of ATP, but are not significantly represented in-cell in our simulations. This is important when selecting allosteric or binding sites based on PDB structures: it allows crystal-packing induced structures to be differentiated from ones that are likely to occur in the cell.

To determine the role of the local environment on ATP dynamics, we opted to compare our in-cell simulations to two datasets—a water-box simulation of ATP to compare ATP’s dynamics with a test-tube like in vitro environment, and high-resolution structures in the PDB, which represent the structures adopted by ATP when bound to a protein in a crystal for the most part (Figure 1).

Figure 1:

Overview of the ATP structural datasets used in this study: (a) Representative snapshot of ATP (color) bound to a protein (gray) in the PDB dataset, PDB ID: 1ZYD. (b) Snapshot of the ATP in the cellular simulation. (c) Snapshot of water-box in vitro simulation. (d) Drawing of the structure of ATP, with the geometric coordinates of interest highlighted: dihedral angle Φ₁, Φ₂, Φ₃, Φ₄, and Φ₅; and the PCN angle: the atoms are highlighted in orange.

To study the effects of the cellular environment on ATP conformational distribution, we used data from the CHARMM 36m with CUFIX corrections model of the E. coli cytoplasm (EC36mCU) that we previously developed to study protein-protein sticking and folding in the cytoplasm^7,22,23. This force field was selected due to observations that E. coli cytoplasmic models simulated with both CHARMM 36m (EC36m) and C36mCU had less macromolecular sticking than a similar model simulated with CHARMM 22* (EC22*).⁷ Therefore, we hypothesized that the C36m and C36mCU models more accurately model the in-cell environment.

With 11 ATP copies included in the EC36mCU model, we analyzed the ATP conformations sampled in 906 frames (snapshots every 24 ns for ~21.7 μs), resulting in 9,966 unique ATP structures from the cellular simulations. To further separate the microenvironments within the cellular model, we applied filters based on the identity of ATP’s neighbors. These filters sort ATP structures by whether or not they are within a 5 Å cutoff of macromolecules using VMD’s pbwithin function.²⁴

Water-box simulations of ATP were performed using parameters identical to the cellular simulations.²⁵ These simulations consisted of one ATP molecule, four sodium ions to balance charge, and TIP3P water molecules.²⁶ This system was simulated for 100 ns in the NPT ensemble (T = 303 K, P = 1 atm), using NAMD 3.0 Alpha software.²⁷ Because experimental studies have shown that similarly-sized free amino acids and short peptides form intermolecular contacts on a similar timescale, we hypothesized that 100 ns would be sufficient to sample ATP’s conformational space.²⁸ Snapshots from the trajectory were taken every 100 ps, resulting in 1,000 unique ATP structures.

Structures in the PDB dataset were obtained by searching “Chemical ID in ATP” in the PDB’s Advanced Search Query Builder.²¹ This resulted in 1,522 unique PDB structures, of which 1,499 were from crystal structures (98.5%). Because multiple copies of ATP were often present in a single unit cell, this analysis contained 3,686 unique ATP structures. Altogether, between the cellular simulations, water-box simulations, and the PDB dataset, 14,652 unique ATP structures were analyzed (Table S1).

Given recent results that both the nucleoside base and the phosphate tail are necessary for ATP’s function as a protein hydrotrope,¹⁶ we hypothesize that the interactions between and geometries of the base and tail play an important role in ATP dynamics and function. To better understand the relative positioning of the nucleoside base and phosphate tail, we calculated five geometric coordinates of interest: the Pγ-C5’-N1 angle (PCN angle); the C3’-C4’-C5’-O5’ dihedral (Φ₁), the C5’-O5’-Pα-O3α dihedral (Φ₃), the O3α-Pβ-O3β-Pγ dihedral (Φ₄), and the Pβ-O3β-Pγ-O3γ dihedral (Φ₅) (Figure 1). The C4’-C5’-O5’-Pα dihedral (Φ₂) also was calculated (Figure S2). The results of these calculations are shown in Figure 2.

Figure 2:

Comparison of the angular probability distributions of Φ₁, Φ₃, Φ₄, Φ₅, and PCN angles arranged from the least crowded to most crowded conditions: the water-box simulation (top row, blue); two different conditions in the cellular model, not within 5 Å of a macromolecule (second row, light green) and within 5 Å of macromolecules (third row, dark green); and in the PDB dataset (bottom row, gray). Long tick mark on the y-axis is 0.1. Compare these to MD data in the presence of Mg²⁺ (Figure S4) and the PDB dataset (Figure S6).

Some of the angular distributions are quite similar, whether ATP is simulated in water, in-cell far from macromolecules, in-cell near a macromolecule, or found in a PDB structure. The Φ₂ distribution (Figure S2) looks nearly identical between all cases. The distributions of Φ₃ and Φ₅, the dihedral angles between the first and between the third phosphorus atom to its connecting oxygen atom, are relatively similar, although the distribution of Φ₅ in the PDB is broader. Binding to the protein surface accesses Φ₅ angles higher in free energy than are populated in unbound ATP.

In contrast, the PDB dataset (Figure 2) populates the Φ₁ angle, between the furanose ring and methylene group, very differently above 120° from any of the MD simulations (Figure 2, green and blue). Although in-cell ATP near macromolecules (Figure 2) has a small population of structures with Φ₁ peaked at 300° as well (290/7776 structures, 3.7%, between 250° and 360°), the PDB data set has over four times more structures in the 250° to 360° range (567/3686 structures, 15.4%) than the macromolecular in-cell data, and an additional peak at 180° that is only minimally observed in the water box simulations. Indeed, we observed that the PDB data set was significantly different (p < 0.01) from all other conditions across all of the geometric coordinates of interest (Tables S2–S6).

The additional peaks of Φ₁ values in the PDB dataset were not observed in earlier studies, which sampled far fewer PDB structures.²⁰ To ensure that the Φ₁ > 120° conformations were not an artifact of poorly resolved PDB structures, we compared the resolution of the PDB structures with Φ₁ > 120° to the entire PDB dataset (Figure S1). Although there was a slightly higher proportion of resolution > 5 Å structures that had high Φ₁ values, overall, the two distributions were very similar. Additionally, we observed no correlation between structure resolution and Φ₁ value (Figure S1). Therefore, we conclude that the Φ₁ distribution in the PDB dataset accurately reflected ATP poses bound to macromolecules. Typical conformations at Φ₁ = 180° and 300° are shown (Figure 3e, f) for PDB and in-cell near protein data. Similar additional angular peaks were also observed for Φ₄, the angle between the second phosphorus atom and its outward neighboring oxygen.

Figure 3:

2-D distributions of dihedral angles: (a-d) Φ₃ vs. Φ₁ in the PDB dataset (a), in-cell simulation near proteins (b), in-cell near not near proteins (c), and in water (d). Color indicates the density of structures at each Φ₁, Φ₃ value (10° bins). Representative snapshot of the 180°, 300° (Φ₁, Φ₃) conformation from the PDB dataset (PDB ID: 4AB2) (e). Representative snapshot of the 300°, 300° (Φ₁, Φ₃) conformation near protein in the in-cell simulations (f).

The PCN angle, which measures the overall bending of the ATP from relatively elongated (180°) to strongly bent (60°), is best correlated with increased crowding as we move from water to in-cell far from proteins, to near proteins, and finally to the highly crowded environment of a crystal (1499/1522 PDB structures were crystal structures). The ATP bend is reduced in the presence of macromolecules, with the bimodal population in water peaked at 60° and 140° moving over to a single peak at 140° in the PDB structures. Our examination of the water box and far-from-protein ATP data shows that these low-PCN angle structures are the result of ATP molecules coordinating neighboring inorganic ions. In contrast, ATP tends to lie flatter on macromolecular surfaces, particularly when wedged between two proteins in a crystal. When we repeated our in-water simulations with Mg²⁺ ions instead of Na⁺, we observed fewer low-PCN angles than we did in the in-water Na⁺ simulation. This indicates that the presence of both divalent ions and macromolecules may induce a ‘flatter’ ATP conformation.

When Φ₁ and Φ₃ 2-dimensional distributions are plotted, one can see the correlation between different angles in the data sets. As noted above, we observe structures that occur only in the PDB and in the in-cell macromolecular datasets, two examples of which are shown (Figure 3e, f). To compare the ATP energy landscape with the PDB populations in Figure 2, we mapped out the potential energy of ATP as a function of the dihedral angles Φ₁, Φ₃, Φ₄ and Φ₅, while keeping the rest of the molecule rigid. (Initial ATP structures for the four scans were randomly selected from the PDB dataset as PDB IDs 1A0I, 2AQX, 3CJB, and 4H0T, respectively), and the ATP’s charge was set to −4. The presence of solvent was estimated by using the polarizable continuum model.²⁹ Calculations were performed with Gaussian 16 software³⁰, using the B3LYP method³¹ with the 6-31G(d) basis set.³² In each calculation, the dihedral of interest was rotated fully over 36 steps (10 degree increments) and the potential energy at each dihedral value calculated (Figure 4).

Figure 4:

Potential Energy Surfaces: (a-d) The dihedral angle scans of the potential energy surface for Φ₁, Φ₃, Φ₄ and Φ₅ (right axis, black line) overlaid on the distribution of the respective dihedral angles in the PDB dataset (gray). PDB ID’s 1A0I, 2AQX, 3CJB, and 4H0T, chosen randomly, were used to hold the remaining bond distances and angles constrained in a-f respectively). Relative energy was calculated by adding 2542 to absolute values output by Gaussian in Hartrees, and then converting to kJ/mole.

The minima of the dihedral scan calculations agree well with the peaks in the PDB angular distributions. However, the potential energy differences between these minima (up to 20 kJ/mol) do not match the PDB distributions, whose peak areas suggest a potential energy difference on the order of 4 kJ/mol and much smaller barriers, if one converts populations ρ into potentials of mean force using PMF(Φ) = k_BTln(ρ). This observation reinforces the idea that the PDB geometries not observed, or minimally observed, in water and in-cell correspond to local potential energy minima, but these minima are high in energy in the absence of the strong crowding present in a crystal. Other interactions, such as those between ATP and neighboring amino acids, may also stabilize geometries that are energetically inaccessible in other environments.

To determine whether ATP had any characteristic structures that belonged to a specific environment, we performed k-means clustering on the ATP structures using the Python package pyclustering.³³ Because the dihedral angle values were represented between 0 and 360, we modified the Euclidean distance function in the package to reflect that the angles 0 and 360 are equivalent to avoid artifacts in the clustering analysis. Due to its narrow and consistent distribution throughout all four local environments, the Φ₂ data was not used in the clustering analysis.

We found that five clusters, 0 to 4 (Figure 5), best described the data. With one exception, the clustering did not isolate structures by different environments, but mainly by correlated groups of angles (Figure S3 and Figure 3). For example, clusters 0, 1, 2, and 4 are high-high, low-high, low-low and high-low in the Φ₃ and Φ₅ angles. There is one major exception cluster 3, corresponding to the angular region Φ₁∈(120°,300°), Φ₃∈(120°,300°), Φ₄∈(120°,360°), PCN∈(90°,360°), was almost exclusively populated by PDB structures (Figure 5), and is thus a sufficient but not necessary identifier of ATP in a crystal environment.

Figure 5:

Clustering data (a) Results of k-means clustering analysis by data type (blue = water box, light green = in-cell simulation not near protein, dark green = near protein, gray = PDB). The distributions in terms of dihedral angles are shown in Figures S3 and S5. Cluster 3 represents a dihedral combination unique to the PDB structures. (b) Two representative structures from cluster 3, taken from PDB ID’s 4NDO (top, Φ₁ = 203°, Φ₃ = 271°, structure resolution = 1.35 Å) and 2OGX (bottom, Φ₁ = 204°, Φ₃ = 274°, structure resolution = 1.5 Å)

We again checked that the resolution of the PDB structures in cluster 3 is not different from the others and found that to be the case (3.35±1.89 Å vs. 3.07±2.03 Å). Thus, the representative structures shown in Figure 5b likely represent real geometries, but ones that can only be observed in the highly constrained environment of a crystal. In Figure 4, this cluster does correspond to a combination of population maxima near (Φ₁ = 197°, Φ₃ = 265°, Φ₄ = 225°, and Φ₅ = 246°), again supporting that it is a local minimum in the ATP structural landscape.

The in-cell environment away from macromolecules does not modulate ATP structure far away from the structural ensemble observed in water with dissolved ions. This observation is correlated with these two environments having a higher coordination with small inorganic ions than the environment near proteins, and such coordination promotes small PCN (bending) angles of ATP. In contrast, the distribution shifts towards a flatter ATP geometry near protein surfaces in our all-atom simulation, albeit not as extreme as in the PDB: crowding and reduction of ion coordination are most extreme in crystal structures.

The structures seen in the PDB but not elsewhere do correspond to actual minima in the solvated ATP energy landscape. However, crystal packing lowers the free energy of ATP conformations not populated significantly in aqueous solution or in-cell simulations. Thus, the ATP conformational distribution, given our extensive multi-microsecond sampling, is likely much narrower even in the crowded cytoplasm than in crystals. In particular, clustering identified a portion of the angular distribution that is unique to the PDB. When we repeated the clustering analysis with the PDB dataset only, we were unable to completely isolate these structures from the larger PDB dataset as a whole.

Such information is important when selecting sites for drug targeting or studying allosteric binding: to select targets for small molecule binding that can occur in cells, clusters such as 3 (Figure 5) due to the crystal environment, should not be chosen, but rather ones that are likely to be populated in vivo. Studies have shown that ATP binds to proteins in a wider range of conformations when it is acting as an allosteric modulator than it does when it binds as an enzyme substrate.³⁴ Additionally, it is well-established that enzymes place strain on their substrates in order to catalyze reactions.³⁵ Nonetheless, some allosterically bound conformations from crystal structures are not representative of the in-cell environment and are likely to be induced by extreme crowding in the crystal, even if they correspond to local minima in the ATP energy landscape.

Our simulations here are in an E. coli cytoplasm, and it will be interesting to see if conformational bias is different in a model of the human cytoplasm: it is known that the two present different electrostatic environments that strongly affect protein mobility.^8,36,37