Protonation and magnesium ions shape the transition state diversity of phosphoanhydride hydrolysis in water Authors: Adrián García-Martínez, Miguel de la Puente, Marion Cunin, Kirill Zinovjev, J Javier Ruiz-Pernía, Damien Laage, Iñaki Tuñón Published: 2026 ABSTRACT AbstractPhosphoanhydride hydrolysis is a central reaction in biochemistry, powering processes from biosynthesis to molecular motors. Yet, its solution mechanism and the molecular origins of the catalytic effects of protonation and magnesium ions remain elusive both to experiments and simulations. Here we use machine learning potentials trained at density functional theory accuracy combined with extensive reaction path sampling and free-energy methods to dissect pyrophosphate hydrolysis in solution, a model for ATP and GTP reactivity. Our simulations reveal multiple mechanistic pathways and identify a dominant mechanism involving synchronous P-O bond formation and cleavage, followed by solvent-assisted proton transfer. Protonation and Mg²⁺ coordination both stabilize tighter transition-state ensembles and lower activation barriers, although our simulations show that the intrinsic catalytic effect of the metal ion has been overestimated. These results establish the key drivers of phosphate hydrolysis.Subject terms: Homogeneous catalysis, Computational chemistry, Biocatalysis Introduction The hydrolysis of phosphoanhydrides such as adenosine and guanosine triphosphate (ATP and GTP) powers nearly every process in biology, from biosynthesis to molecular motion1–3. Yet, despite decades of experimental and theoretical efforts the molecular mechanism of phosphate hydrolysis remains debated1,2,4–6. Key unresolved questions include whether bond cleavage and formation proceed through an associative, dissociative, or concerted pathway; whether the proton is transferred from nucleophile to nucleofuge through solute or solvent assistance7,8; and how these pathways are modulated by protonation and by the ubiquitous Mg2+ cofactor that apparently accelerates hydrolysis in both enzymes and solution. Classical experimental mechanistic analyses, based on linear free-energy relationships (LFERs) and kinetic isotope effects (KIEs), suggest that phosphate hydrolysis proceeds through a concerted but dissociative transition state reminiscent of a metaphosphate species2,9. However, these experimental observables provide averaged information over multiple molecular routes, masking potential mechanistic diversity. Computational studies have offered valuable insights but have reached divergent conclusions5,10–12, and remain hindered by the non-reactive treatment of solvent in continuum or hybrid QM/MM models, the limited sampling in computationally costly ab initio molecular dynamics13,14, and the arbitrary choice of reaction coordinates for these highly multidimensional reactions. Another unresolved aspect is the origin of catalysis itself. While the intrinsic stability of P-O bonds necessitates the assistance of catalysts to facilitate their hydrolysis15, the molecular basis of this acceleration remains debated. Protonation is known to enhance reaction rates, but whether it merely reduces electrostatic repulsion or fundamentally changes the mechanism is unclear16. The role of Mg²⁺ ions, present in nearly all enzymatic phosphoryl transfers, has been studied for decades9,17–20, but their catalytic function is still uncertain: do they stabilize the nascent charge on the leaving group, polarize the electrophilic phosphorus, or facilitate nucleophile deprotonation? Furthermore, although a continuum of transition states between tight (five-coordinate phosphorane-like) and loose (three-coordinate metaphosphate-like) structures has long been proposed2,9,21, its quantitative characterization, and how it shifts with pH or with metal ion binding, remains unexplored. Here, we address these open questions by combining recent advances in machine learning potentials (MLPs)22–25 trained on high-level density-functional theory with extensive, unbiased transition-path sampling (TPS)26,27 to map the molecular mechanisms of pyrophosphate hydrolysis in water and biased enhanced sampling methods to obtain converged free energy landscapes. This approach addresses key limitations of previous simulations and combines quantum-level accuracy with microsecond-scale sampling28,29, while employing a potential that allows every solvent molecule to participate in proton exchange and nucleophilic attack. This contrasts with QM/MM or cluster-based simulations where only a limited number of solvent molecules are included in the reactive region. Recent MLP studies of phosphoester bond formation demonstrated that this approach can capture both associative and dissociative pathways in water28. Here, we extend it to pyrophosphate hydrolysis, a model for ATP/GTP hydrolysis, where multiple competing reaction channels emerge. Our simulations reveal five distinct mechanisms and quantify their relative free-energy barriers. The computed activation free energies are in close agreement with experiment, supporting the reliability of our approach. Our simulations reveal that pyrophosphate hydrolysis proceeds through a solvent-assisted mechanism, in which bond formation and cleavage occur synchronously across a broad range of transition-state tightness, quantifying the previously proposed continuum of transition states9,21. Protonation and Mg²⁺ coordination tighten this transition-state ensemble, lowering activation barriers without altering the concerted character of the reaction. Moreover, our results show that the experimentally observed rate acceleration by Mg²⁺—over 2 × 10³ fold19— is largely attributable to accompanying changes in protonation, rather than to intrinsic catalytic stabilization of the transition state. By providing a quantitative and molecularly detailed description of phosphate hydrolysis in solution, our work resolves long-standing mechanistic ambiguities and clarifies the role of metal ions and protonation in phosphoanhydride chemistry. The resulting picture establishes the molecular baseline for understanding enzymatic ATP and GTP hydrolysis. Results Baseline mechanisms of pyrophosphate hydrolysis We first examine the hydrolysis of fully deprotonated pyrophosphate (P2O74-), which serves as an excellent model for the ATP4- and GTP4- forms found in enzymatic active sites, and also plays a key role in metabolism30. Kinetic measurements have determined that the hydrolysis reaction rate increases upon protonation and in the presence of Mg²⁺ ions19, providing an experimental benchmark for our simulations. The hydrolysis formally involves the attack of a water molecule on a phosphorous atom, departure of a phosphate anion as the leaving group, and transfer of a proton from the nucleophilic water to the leaving group (Fig. 1a). To characterize this process, we mapped the free energy surface (FES, see Fig. 1b) using on-the-fly probability enhanced sampling (OPES)31 simulations enabled by the efficiency of our MLP calculations (see “Methods”). The FES was projected along two key collective variables designed to capture the progress of bond formation and cleavage, and the accompanying proton rearrangement. The first, Δd = dpℓ – dpw, is the antisymmetric combination of the distances from the electrophilic phosphorous to the leaving group and nucleophilic water oxygens. The second, ΔC = Cℓ – Cw, is the difference between the hydrogen coordination numbers of the leaving phosphate and the nucleophilic water, reporting on their respective protonation states (see Fig. 1a, “Methods” and SI). Pyrophosphate (P₂O₇⁴⁻) hydrolysis reaction. a Schematic representation of the reaction, highlighting the four collective variables employed in the enhanced sampling simulations, similar to those employed for phosphoester bond formation28. For clarity, the proton transfer is illustrated with a single water molecule, although the simulations reveal that the proton is typically transferred through a solvent-assisted relay involving multiple water molecules. b Simulated free energy surface obtained from On-the-fly Probability Enhanced Sampling31 simulations, projected onto the antisymmetric combinations of distances (Δd = dpℓ – dpw) and coordination numbers (ΔC = Cℓ – Cw). Different reaction pathways identified by Transition Path Sampling27 may overlap when projected onto this reduced two-dimensional collective-variable space. c Schematic overview of the five distinct mechanistic pathways identified through Transition Path Sampling simulations and their respective transition state (TS). Source data are provided as a Source Data file. The FES displays two well-defined basins corresponding to the reactants (Δd ≈ −2.0 Å, ΔC ≈ −2), and the products (Δd ≈ 2.0 Å, ΔC ≈ 0). A striking feature of the FES is the presence of two saddle points along the dividing region separating these basins near Δd ≈ 0 Å, at distinct ΔC values. This reveals that reaction pathways can proceed through transition states (TS) characterized by different protonation states of the nucleophile and/or the leaving group: one (ΔC ≈ −2) in which no proton rearrangement has yet occurred, and another (ΔC ≈ −1) where deprotonation of the nucleophile, or protonation of the leaving group, has already taken place. OPES provides an efficient initial exploration of the free energy landscape. However, the projected FES cannot resolve the full diversity of TS structures and unambiguously determine the exact protonation states of the nucleophile, leaving group, and central (α) phosphorous atom. The restricted set of collective variables employed to project the free energy surface may fail to capture all relevant protonation patterns, while their arbitrary selection can omit critical reactive coordinates, hinder convergence, and potentially lead to hysteresis effects. When projected onto a limited set of collective variables, distinct reaction pathways may partially overlap in the free energy surface, potentially masking their individual contributions. To overcome these limitations, we performed an agnostic full-dimensional dynamical exploration of reactive trajectories using TPS. Starting from reactant and product state definitions and from two initial pathways identified on the FES during the OPES exploration (one with nucleophilic water still protonated and the other with hydroxide already formed at the TS), we generated an extensive ensemble of more than 50,000 reactive trajectories. To characterize these transition-path ensembles, we employed a histogram-based approach, projecting the path densities onto key pairs of distance and proton coordination variables to track the timing of chemical events across the trajectories. This analysis unveiled the full mechanistic diversity of pyrophosphate hydrolysis and revealed five distinct mechanisms (see Fig. 2 and S1), differing by the protonation state of the nucleophile, leaving group and central α-phosphate (see TS1–TS5 structures in Fig. 1c). Analysis of transition path reactive trajectories for P₂O₇⁴⁻ pyrophosphate hydrolysis. a Committor probability distributions from the transition path ensemble of paths involving a water nucleophile at the transition state (TS) and b paths involving a hydroxide ion at the TS. c Path probability densities for the coordination number of the nucleophilic water (Cw) along the reaction progress Δd = dpℓ – dpw for paths involving a water molecule at the TS and d paths involving a hydroxide ion at the TS; e Path probability densities for the coordination number of the leaving group (Cℓ) along the reaction progress Δd for paths involving a water molecule at the TS and f paths involving a hydroxide ion at the TS; g Path probability densities for the coordination number of the α-phosphate group (Cpα) along the reaction progress Δd for paths involving a water molecule at the TS and h paths involving a hydroxide ion at the TS. i Average value of the coordination of hydrogen atoms initially bonded to nucleophilic water to its oxygen atom (CH,w), to surrounding solvent molecules (CH,sol) and to α-phosphate oxygen atoms (CH,Op) along Δd for transition path sampling27 trajectories involving a water molecule at the TS and j trajectories involving a hydroxide ion at the TS. The shaded area represents the standard deviation. Source data are provided as a Source Data file. TS structures were identified from TPS trajectories through committor analysis (i.e., the probability that a trajectory initiated from a given configuration reaches products before reactants). Configurations with a committor value of 0.5 were taken as true dynamical TSs. The average committors projected along the Δd coordinate (Fig. 2a, b and Fig. S2) show that TSs are located around Δd ≈ 0 Å, with the equicommitor surface lying within the range −0.2 to 0.2 Å. While simple geometric coordinates suffice to describe the committor here, more complex systems may require machine learning approaches that build committor-based collective variables for enhanced transition-state sampling32,33. To dissect the underlying mechanisms, we projected the reactive trajectories onto key protonation coordinates that report on the protonation states of the nucleophile, leaving group, and α-phosphate (Fig. 2c–h). The timing of nucleophile deprotonation is tracked by the probability density along the Δd reaction progress of the hydrogen coordination number of the nucleophilic water oxygen (Cw), which measures the number of hydrogens bound to the nucleophile (Fig. 2c, d). In the first ensemble (water deprotonated after TS), Cw ≈ 2 at the TS and decreases only after Δd ≈ 0 Å, indicating that nucleophile deprotonation follows bond formation. In contrast, in the second ensemble (water deprotonated before TS), Cw ≈ 1 already at the TS, confirming that deprotonation precedes the attack. The coordination number of the leaving group oxygen atom (Cℓ) provides the complementary information for the leaving phosphate. These distributions show that protonation of the leaving group can occur either before or after the TS for both nucleophiles (Fig. 2e, f). Finally, the α-phosphate coordination number (Cpα) which reports on its protonation state, shows that the α-phosphate is always unprotonated at the TS when the nucleophile is water but can be either unprotonated or protonated when the nucleophile is hydroxide (Fig. 2g, h). This systematic analysis (see also section S1) identifies five mechanistic pathways summarized in Fig. 1c. For reactions initiated with water at the TS, TS1 and TS2 correspond respectively to protonation of the leaving group after and before the TS. For reactions initiated with a hydroxide ion already formed at the TS, three additional mechanisms arise: TS3 and TS4, in which the leaving group is protonated after the TS and the α-phosphate is respectively unprotonated and protonated at the TS; and TS5, in which the leaving group is protonated before the TS while the α-phosphate remains unprotonated. As our focus is on the TS structure, we use short TPS trajectories (the average time length is 0.64 ps, see Fig. S14) which do not allow for full protonation equilibrium in the product state, and our simulations lead to different intermediate products, P1 (HPO42- and PO43-), P2 (H2PO4- and PO43-) and P3 (two HPO42- ions), depending on the proton transfer events occurring around the TS. We next exploit this extensive transition path ensemble, generated with a potential in which all solvent molecules are treated as reactive, to resolve the long-standing question of whether proton transfer in pyrophosphate hydrolysis is solvent or substrate assisted. Figure 2i, j report the evolution of the coordination numbers of the two protons initially bonded to the nucleophilic water with the nucleophilic oxygen (CH,w), the α-phosphate oxygens (CH,Op), and the surrounding water solvent molecules (CH,sol). A clear concerted exchange is observed between the nucleophile and the solvent (as CH,w decreases, CH,sol increases correspondingly), while CH,Op remains close to zero in the reactive trajectories. This demonstrates that leaving group protonation does not occur mediated by substrate atoms but is exclusively carried out by solvent molecules. When a hydronium ion is formed following deprotonation of the nucleophilic water, it remains near the pyrophosphate until a proton is transferred to the substrate—a process that involves, on average, two solvent molecules. Thus, proton rearrangement during pyrophosphate hydrolysis proceeds through a solvent-assisted relay rather than a substrate-assisted pathway (the schematic representation in Fig. 1a illustrates the overall reaction but does not explicitly depict the multiple solvent molecules involved in the proton relay). To quantify the relative probabilities of the five different mechanisms revealed by TPS, we calculated the relative free energies of the TS ensembles (TS1-TS5). Once the critical coordinates have been identified from the unbiased TPS simulations, we return to enhanced sampling methods biased along these coordinates to compute converged free energies. In particular, free energy differences between pairs of TSs were obtained from Umbrella Sampling (US) simulations biased along the relevant coordination number (Cw, Cℓ or Cpα), with Δd restrained near 0 Å, and averaged over three independent replicas to ensure convergence (section S2). When the transformation is governed by a single dominant coordinate, US allows efficient sampling and robust convergence of the corresponding free energy profile. The resulting free energy differences (Figs. S5–S7 and Table S1) identify TS1 as the lowest free energy TS. The absolute activation free energy for this mechanism was then determined by US simulations along the antisymmetric coordinate Δd, keeping Cw ≈ 2 and Cℓ ≈ 0, because TPS trajectories indicated that only the distances to the nucleophile and leaving group change significantly between reactants and TS1. This result was independently confirmed by the Minimum Free Energy Path (MFEP) obtained with the Adaptive String Method (ASM) (see Fig. S8), which provides a multidimensional description of the reaction pathway, enabling us to confirm the relative timing of different coordinates along the process. The resulting free energy profile (Fig. 3b) shows that the computed TS1 activation barrier of 35.2 ± 0.1 kcal·mol−1 is in excellent agreement with the experimental value of 35.0 ± 0.2 kcal·mol−1 at 25 °C19. Combining this absolute barrier for TS1 with the relative free-energy differences among TSs provides activation free energies for all mechanisms (Fig. 3b). The free energy differences obtained from US calculations show that the next most favorable mechanism, via TS4, lies 6.1 ± 0.4 kcal·mol−1 higher in free energy, implying a contribution to the overall reaction rate constant at least four orders of magnitude smaller. Therefore, under the present conditions, pyrophosphate hydrolysis is effectively dominated by the TS1 mechanism. Activation free energies for pyrophosphate hydrolysis. a Free energy profiles along the (Δd = dpℓ – dpw) coordinate for P2O74-, HP2O73-, H2P2O72- and MgP2O72- hydrolysis through TS1. Note that, due to the restraints imposed on Cw and Cℓ, the products state cannot be reached in the Potential of Mean Force.; b Activation free energy barriers for the five hydrolysis mechanisms presented in Fig. 1c for P2O74-, H2P2O72- and MgP2O72- (note that the full mechanistic breakdown was not performed for HP2O73- after confirming that TS1 is favored in all the other cases); c Schematic representation of the most favored transition state for P2O74-(TS1), HP2O73- (TS1’α,β), H2P2O72- (TS1”) and MgP2O72- (TS1Mg). Source data are provided as a Source Data file. These results provide a quantitative characterization of the TS continuum in P2O74- hydrolysis. However, the relative free energies of the five identified pathways are not fixed and depend on reaction conditions such as pH, temperature, and the nature of the nucleophile and leaving groups. We next examine how protonation and coordination by Mg2+ ions modulate the mechanism and TS ensemble of pyrophosphate hydrolysis. Effect of substrate protonation We next examine how substrate protonation impacts the mechanism and energetics of pyrophosphate hydrolysis. Pyrophosphoric acid can undergo four sequential deprotonations, with pKa values of 0.85, 1.96, 6.60, and 9.4134. Experimental measurements show a marked rate enhancement at lower pH, interpreted as due to the decreasing hydrolysis activation free energies from 35.0 ± 0.2 kcal·mol−1 for P2O74- to 31.2 ± 0.4 kcal·mol−1 for HP2O73- and 28.5 ± 0.1 kcal·mol−1 for H2P2O72-19. We first investigated the hydrolysis mechanisms of H2P2O72- (Fig. 4a). The overall free-energy landscape (Fig. 4b) and mechanistic features remain qualitatively similar to those of the unprotonated substrate (P₂O₇⁴⁻), indicating that protonation does not alter the fundamental concerted nature of the reaction but modulates the relative stability of the transition states. In this case, the △C coordinate ranges from ~−1 in the reactant state to ~1 in the product state because the leaving group is initially protonated. Analysis of the TPS trajectories (section S1) also reveals five distinct mechanistic pathways (see Figs. S4 and S7), corresponding to the same families of TSs identified for the unprotonated system. As before, US simulations show that the doubly protonated analog of TS1 (denoted TS1”; Fig. 3c) has the lowest free energy among all pathways. The activation free energy for this mechanism, obtained from US simulations along Δd (Fig. 3a), is 28.3 ± 0.2 kcal·mol−1, again in excellent agreement with the experimental value of 28.5 ± 0.1 kcal·mol−119. The activation free energies for the remaining mechanisms (TS2” to TS5”) are reported in Fig. 3b. Impact of protonation and Mg2+ ion on pyrophosphate hydrolysis. a Schematic representation of the H2P2O72- hydrolysis reaction; b Free energy surface for H2P2O72- hydrolysis obtained from On-the-fly Probability Enhanced Sampling31 simulations, projected onto the antisymmetric combinations of distances and coordination numbers; c Schematic representation of the MgP2O72- hydrolysis reaction; d Free energy surface for MgP2O72- hydrolysis, projected onto the antisymmetric combinations of distances and coordination numbers. Different reaction pathways identified by Transition Path Sampling27 may overlap when projected onto this two-dimensional collective-variable space. Source data are provided as a Source Data file. We next simulated the hydrolysis reaction for HP2O73-. In this case we considered two possibilities, differing by the nature of the protonated phosphate group (electrophilic α-phosphate group or β-phosphate leaving group). Based on our previous results for H2P2O72- and P2O74-, we considered that the most favourable mechanisms take place through TS1α’ and TS1β’ (see Fig. 3c). Figure 3a shows the free energy profiles corresponding to these two mechanisms. The hydrolysis mechanism involving the nucleophilic attack on the unprotonated phosphate group (TS1β’) presents an activation free energy of 29.0 ± 0.1 kcal·mol−1, while the attack on the protonated phosphate group (TS1α’) results in a slightly larger activation free energy barrier of 31.7 ± 0.8 kcal·mol−1. This difference reflects the greater reduction of the negative charge on the leaving group in the former case, which correlates with a lower activation barrier (see section S5). Note that the experimental value reported by Wolfenden et al. (31.2 ± 0.4 kcal·mol−1)19 lies in between these two values and probably correspond to a combination of both reaction channels. These calculated barriers show that protonation on the leaving group is a better catalytic strategy than protonation on the attacked phosphate, because of the stabilization of the leaving group growing negative charge9,35. Effect of magnesium ion coordination Magnesium ions are ubiquitous cofactors in enzymes that catalyze phosphoanhydride hydrolysis, yet their precise catalytic role has remained controversial for decades. They have been proposed to accelerate the reaction by stabilizing negative charge at the transition state, by polarizing the electrophilic phosphorus center, or by promoting nucleophile deprotonation9,17–19. Like ATP and GTP, pyrophosphate is known to form a strong complex with a Mg2+ ion, which was suggested to be responsible for the measured hydrolysis reaction rate enhancement, corresponding to a reduction of the activation free energy by 4.6 kcal·mol−119. However, the molecular mechanism underlying this acceleration has remained unresolved. The nature of Mg2+ coordination to phosphate species, whether mono-, bi- or tri-dentate, has been a topic of long-standing debate36,37. Our simulations, which combine high-level quantum accuracy and extensive sampling, show that under aqueous conditions pyrophosphate forms a stable bidentate complex with Mg2+ (see Fig. 4c), consistent with structures observed in ATP/GTP-hydrolyzing enzymes37. We then investigated the hydrolysis of MgP2O72- using the same combination of MLP-MD-driven enhanced sampling and TPS approaches applied above. The resulting FES obtained from OPES simulations (Fig. 4d) also reveals possible competing pathways partially masked by the projection on two coordinates. Analysis of the reactive trajectories and US simulations (see Figs. S3 and S6) confirms that the five mechanisms identified previously (Fig. 1c) remain accessible, with TS1Mg (in which the protonation states of the nucleophile and leaving group remain unaltered, see Fig. 3c) being the lowest in free energy (Fig. 3b). Thus, an important result is that Mg2+ coordination does not change the preferred reaction mechanism, nor does it significantly promote nucleophile deprotonation. The computed activation free energy for TS1Mg is 33.0 ± 0.1 kcal·mol−1 (Fig. 3a), slightly higher than the experimental estimate of 30.4 ± 0.4 kcal·mol−119. While our simulations show that the hydrolysis barrier is lowered by Mg2+ coordination, they predict a barrier reduction of 2.3 kcal·mol−1 relative to the uncomplexed substrate, smaller than the 4.6 kcal·mol−1 decrease proposed from kinetic data19. However, re-examining the experimental conditions reveals that this apparent discrepancy likely arises from pH effects. While kinetics of uncomplexed pyrophosphate are measured at pH 14, kinetics with Mg2+ were determined near pH 8, under the assumption that MgP2O72- was the sole reactive species and neglecting the contribution of the acidic complex MgHP2O7-, which is expected to be populated at this pH38. Our simulations show that this protonated complex has a much lower barrier of 28.4 ± 0.1 kcal·mol−1 (see Fig. S9). Consequently, protonation of the complex contributes significantly to the rate enhancement even at low concentrations. The experimental barrier (30.4 kcal·mol−1) lies between our predicted values for MgP2O72- (33.0 kcal·mol−1) and MgHP2O7- (28.4 kcal·mol−1). Our results thus demonstrate that, contrary to the traditional view, the observed rate enhancement upon Mg²⁺ addition arises primarily from protonation rather than from any intrinsic catalytic effect of the Mg²⁺ ion. The larger energetic effect of protonation on the activation free energy of pyrophosphate hydrolysis, compared with Mg²⁺ complexation, is further supported by an analysis of the electronic structure changes induced by these two factors. Protonation produces a significantly stronger perturbation of the electronic structure than Mg²⁺ binding, leading to a larger reduction of the scissile P-Ol bond order and to a marked stabilization of the leaving phosphate group charge (see section S5). As discussed there, these quantities, in particular the leaving group charge, correlate very well with the computed activation free energies, indicating that stabilization of the leaving group is a key factor controlling the reaction barrier and explaining the stronger catalytic effect of protonation relative to Mg²⁺ coordination. Note also that recent experiments show that Mg²⁺ ions can even slow ATP hydrolysis at low pH, suggesting that protonation may be more effective than metal complexation in promoting the reaction20. Impact of protonation and Mg2+ on the TS continuum We have shown that our simulations quantitatively reproduce the experimental rate enhancement observed upon increased protonation (H2P2O72- > HP2O73- > P2O74-) and upon Mg2+ coordination (MgP2O72- > P2O74-). Although many pathways are accessible, our results establish that the preferred mechanism remains the same in all cases: water acts as the nucleophile, with both nucleophile deprotonation and leaving group protonation occurring after crossing the TS and proceeding through solvent assistance. However, within this single mechanism, a broad range of TS geometries, from loose to tight, are possible, leading to the TS continuum concept, which reflects how the position of the TS ensemble shifts along this loose-tight spectrum depending on chemical conditions, such as pH or meta coordination9,21. We now quantify how the rate acceleration induced by protonation and Mg2+ coordination is connected to their effects on the tightness of this continuum. Figure 5 compares the reaction pathways and TS geometries for P2O74-, MgP2O72- and H2P2O72-, as obtained from TPS trajectories. We first consider the unprotonated P2O74- situation as a reference. Figure 5a shows the probability density distribution of the dpw and dpl distances along the trajectories passing through TS1, similar to a More O’Ferrall-Jencks correlation diagram2. The reaction proceeds first with the approach of the nucleophilic water to the α-phosphate group (initial decrease of dpw), followed by an intermediate stage where the high-probability path moves diagonally across distance space; the distance to the leaving group (dpℓ) increases simultaneously as the nucleophile approaches, and finally the leaving group departs. This concurrent evolution along both distances shows that the mechanism is neither purely stepwise associative nor stepwise dissociative, but rather concerted and synchronous, consistent with recent findings for phosphoester bond formation28. This is further supported by the plot of the probability-weighted average distances along the reaction progress (Δd) (see Fig. S15) that confirms a concerted evolution of dpw and dpℓ for the unprotonated, Mg2+-coordinated and deprotonated species. Path probability densities and transition state distance distributions for P2O74-, MgP2O72- and H2P2O72-. a Path probability density obtained from Transition Path Sampling27, projected onto the dpw and dpℓ distances for the hydrolysis of P2O74- through TS1.; b Probability density distribution of these distances at TS1 ensemble. Dashed vertical lines represent the average distances in the reactant state; c Path probability density for the hydrolysis of MgP2O72- through TS1Mg; d Probability density distribution of dpw and dpℓ distances at TS1Mg. Dashed vertical lines represent the average distances in the reactant state; e Path probability density for the hydrolysis of H2P2O72- through TS1”; f Probability density distribution of dpw and dpℓ distances at TS1”. Dashed vertical lines represent the average distances in the reactant state; g Probability density distribution of Mg-O distances at the reactant state and at TS1 of MgP2O72-. Source data are provided as a Source Data file. The TS1 structures span a broad range of dpw and dpℓ distances (Fig. 5b), with average values of 2.30 ± 0.17 and 2.55 ± 0.17 Å, respectively, and ranges of 1.9–2.8 and 2.2–3.1 Å, respectively. This broad distribution demonstrates that TS1 is formed by an ensemble of structures including both loose and tight structures, and the reaction mechanism is better defined by the protonation states of the involved species than by the loose or tight character of the TS ensemble, which is composed of a variety of structures. For the Mg2+-coordinated and protonated pyrophosphates (MgP2O72- and of H2P2O72-), the reactive trajectories passing through TS1Mg (Fig. 5c) and TS1” (Fig. 5e) likewise proceed through concerted and synchronous bond cleavage and formation. However, both systems display tighter TS ensembles than the P2O74- reference (Fig. 5d, f). The average dpw and dpℓ P-O distances at the TS decrease from 2.30 ± 0.17 and 2.55 ± 0.17 Å for P2O74- to 2.15 ± 0.10 and 2.37 ± 0.10 Å upon Mg2+ coordination, and 1.91 ± 0.07 and 2.16 ± 0.08 Å for the protonated H2P2O72-. This progressive tightening of TS geometries correlates directly with the reactivity trend (H2P2O72- > MgP2O72- > P2O74-) and is consistent with experimental observations indicating that Mg²⁺ perturbs the TS ensemble less than protonation does9,35. Partial compensation of the substrate’s negative charge, through either protonation or Mg2+ complexation, reduces electrostatic repulsion, stabilizes the increasing charge on the leaving group, and increases the α-phosphate electrophilicity, enabling a closer approach at the TS. The effect, more noticeable with protonation, is consistent with experimental KIEs from phosphate monoesters to triesters, which point to progressively tighter TS structures as the net phosphate charge decreases39. In contrast, the interaction between Mg2+ and pyrophosphate changes little from reactants to TS1Mg, with average Mg-O distances of 2.06 ± 0.08 Å and 2.04 ± 0.08 Å, respectively (see Fig. 5g), indicating that Mg2+ does not preferentially stabilize the TS with respect to the reactants but acts primarily through overall charge compensation. Discussion By combining MLPs and Transition Path Sampling (TPS), we have achieved a quantitative description of pyrophosphate hydrolysis mechanisms in explicit, reactive water. This approach associates the quantum accuracy required to describe P-O bond making/breaking and proton transfer with the extensive sampling necessary to capture the mechanistic diversity, showing a promising direction toward addressing elusive mechanisms using ML-driven simulations. Our simulations reveal the presence of five distinct reaction mechanisms, differing in the protonation state of the nucleophile, leaving group and transferred phosphate group. Although several pathways are accessible, the reaction is effectively dominated by a single mechanism. In this mechanism, the reaction proceeds through a concerted, synchronous P-O bond cleavage and formation, while proton transfer occurs after the transition state through solvent assistance. The activation free energies computed for these mechanisms quantitatively reproduce experimentally derived values across protonation states, supporting the accuracy of our machine learning—TPS framework. We show that protonation and Mg²⁺ coordination both tighten the transition state ensemble and correlate with rate enhancement, the former having the stronger effect. A key result is that Mg²⁺ coordination alone has only a modest catalytic effect on pyrophosphate hydrolysis: the experimentally observed acceleration largely arises from formation of the more reactive acidic complex MgHP2O7- rather than direct transition-state stabilization. These results challenge the traditional view of Mg²⁺ as a strong catalytic ion favoring nucleophile deprotonation and instead suggest a modest electrostatic contribution without affecting the reaction mechanism. This conclusion is confirmed by the analysis of the effects of protonation and Mg2+ on the electronic structure of pyrophosphate. The mechanistic conclusions reported here are supported not only by quantitative agreement with experimental activation free energies, but also by the consistent reproduction of systematic trends across protonation states and Mg²⁺ coordination, rendering them robust with respect to the uncertainty of the reference functional. The diversity of protonation-dependent pathways uncovered in our simulations provides a molecular interpretation for the complex behavior of LFERs used to infer the mechanism of phosphate transfer2. In these studies the pKa of the nucleophile and/or the leaving group can be varied by up to 6 pKa units2, which translates into a free energy change for the deprotonation/protonation of the nucleophile/leaving groups of more than 8 kcal·mol−1. Our simulations reveal that multiple mechanistic channels could be operative within this free energy window, depending on the protonation states of these groups, and contribute to the experimentally observed reactivity trends, making the interpretation of linear free energy relationships challenging. The solution mechanism we establish for pyrophosphate hydrolysis reproduces many features of enzymatic ATP and GTP hydrolysis40,41: a reaction involving water as the nucleophile and a metaphosphate-like structure whose charge stabilization is key for lowering the reaction barrier. In enzymes, positively charged residues (Arg and Lys) forming strong hydrogen bonds with the phosphate leaving group mirror the catalytic effect of substrate protonation in solution. In contrast, coordination by the Mg²⁺ ion, a common cofactor in enzymatic hydrolysis, likely has a smaller catalytic effect and appears to act primarily as a structural organizer that promotes binding and stabilizes tighter, less fluctuating transition state structures for more effective recognition and stabilization by active site residues. Our results therefore establish the roles of protonation and metal ions in phosphate chemistry and provide a quantitative foundation for understanding and engineering phosphoryl-transfer catalysis in enzymes and synthetic systems alike. Beyond the specific case of pyrophosphate hydrolysis, the present study illustrates how the combination of machine-learning potentials with complementary enhanced sampling techniques can provide a general framework for investigating complex reaction mechanisms in condensed phase systems with biochemical relevance. Methods Simulated systems The simulated systems comprise four different forms of solvated pyrophosphate: the fully deprotonated substrate P2O74-, the magnesium-bicoordinated substrate (MgP2O72-), and the partially protonated forms HP2O73- and H2P2O72-. In all cases, the substrate is solvated by 169 water molecules in a cubic box of 17.20 Å per side, with periodic boundary conditions applied in all three dimensions. Two Mg2+ ions are included in the simulation box to ensure overall charge neutrality. When a magnesium ion is not coordinated to the substrate, it is kept at a distance of at least 6.5 Å from the substrate using a repulsive wall. In the monoprotonated pyrophosphate system, a hydroxide ion is generated in the solvent and constrained within the coordination sphere of the neutralizing Mg2+ ion. For the diprotonated pyrophosphate, a second hydroxide ion is generated and similarly constrained around the second Mg2+ ion. This setup ensures that all systems have the same number of atoms and maintain overall charge neutrality, facilitating the generation of the data set. A visual representation of the simulation box is shown in Fig. S12 of SI. Neural network potential We trained a neural network potential (NNP) to describe the full-system potential energy surface at the hybrid revPBE0-D342 density functional level of theory using a triple-ζ (TZV2P) basis set, thereby treating all solvent molecules as potentially reactive. This functional provides accurate predictions of pyrophosphate reaction energies when benchmarked against coupled-cluster CCSD(T) calculations (see Table S4) and has been shown to offer a reliable description of liquid water structure and proton transfers in solution43–45. Reference calculations were performed with CP2K (v2023.2)46 while NNP training employed DeePMD-kit v247. Exploration of the potential energy surface was carried out via molecular dynamics simulations in the canonical ensemble at 298.15 K using LAMMPS (release – update 3)48, driven by the NNP forces, in combination with biased-sampling techniques as implemented in PLUMED (v2.8.2)49. These explorations were performed across all chemical systems (P2O74-, MgP2O72-, HP2O73- and H2P2O72-), with the goal of producing a single NNP capable of describing all pyrophosphate chemistry. The training dataset was generated iteratively through an active learning scheme, under the ArcaNN framework50, in which new configurations were added until the NNP reproduced the regions of the potential energy surface relevant for the reaction. Specifically, the final training set includes configurations from unconstrained trajectories initiated in the reactant and product basins, guessed transition state structures and biased trajectories propagated using enhanced-sampling techniques such as OPES31. The full training set contains over 51,000 configurations. The converged NNP is able to produce stable MD, reproducing the revPBE0 energies and forces, with a speed-up of approximately 6 orders of magnitude relative to ab initio MD. The NNP was validated on an independent set of 4000 configurations collected along the main reaction path of the different systems. The root-mean-square error on the atomic forces was below 0.05 eV·Å−1, supporting the accuracy of the NNP description. Further details on the NNP architecture, iterative training procedure and validation are provided in section S7 of the SI. Enhanced sampling methods and free energy calculations In this work, a variety of enhanced sampling methods were employed, specifically OPES31, TPS27, US51 and the ASM52. All procedures described below were carried out for the systems used to investigate the hydrolysis mechanism of pyrophosphate in different chemical conditions: P2O74-, MgP2O72-, HP2O73- and H2P2O72-. As a first step, multiple replicas of 10 ns OPES simulations were performed, biasing four collective variables (CVs): the distance between the nucleophilic water oxygen and the electrophilic phosphorus (dpw), the distance between the electrophilic phosphorus and the bridging oxygen of pyrophosphate (dpℓ), and the coordination numbers of the nucleophilic oxygen (Cw) and of the leaving group oxygens (Cℓ) with solvent hydrogens. The reaction free energy surface was reweighted in two dimensions, using the antisymmetric combinations of distances (Δd) and coordination numbers (ΔC) as projection coordinates. These simulations were performed with LAMMPS using the OPES implementation in PLUMED. Further details on the CV definitions and simulation parameters are provided in section S8 of the SI. Starting from TS geometries extracted from the OPES free-energy surfaces, we performed TPS simulations. TPS generates an ensemble of unbiased dynamical trajectories connecting the reactant and product basins without the need to define collective variables. We used the TPS implementation in OpenPathSampling (v1.5.2)53,54 combined with the OpenMM (v7.6.0)55 molecular dynamics engine, interfaced with DeePMD-kit to compute NNP forces during the MD propagation. For each system and reaction mechanism (water-mediated or hydroxide-mediated), initial reactive trajectories were generated from 20–30 independent guesses of transition states obtained from the OPES simulations. These guesses were followed by unbiased dynamics from the TS to identify successful committors. The resulting reactive trajectories served as starting points for 60–90 independent TPS simulations per case. In total, more than 104,000 uncorrelated transition paths were generated, corresponding to over 0.55 μs of aggregated simulation time. The trajectories were analyzed to determine the sequence of molecular rearrangements along the reaction coordinate and to locate the transition state through committor analysis. The resulting transition path ensemble was analyzed using a histogram-based approach, from which five distinct transition states were identified. Free energy differences for the five TSs identified from the TPS ensemble were computed via US with WHAM reweighting. US targeted the relevant C coordinate with Δd restrained near zero. All results were averaged over three independent replicas. For TS1, US was performed along the Δd coordinate while restraining coordination CVs to isolate the mechanism. Finally, the ASM was employed to obtain the full reaction profile for the TS1 mechanism, which exhibited the lowest free energy barrier. Calculations were performed using an in-house interface between emle-engine56 and the sander module of AMBER2457 to be able to use the NNP generated by DeePMD. The main distance and coordination numbers were selected as CVs, and the MFEP was converged using 46 string nodes. Further details of the enhanced sampling and free energy calculations are provided in section S8 of the SI. Supplementary information Supplementary Information Transparent Peer Review file Source data Source Data Acknowledgements A.G.-M. acknowledges the warm hospitality of the Département de Chimie (École Normale Supérieure, Paris). The authors acknowledge Guillaume Stirnemann and Elise Duboué-Dijon for helpful discussions about magnesium ion catalysis. Authors also acknowledge Rolf David for his assistance with ArcaNN and TPS simulations. The authors thank financial support from grant PID2024-157213NB-I00 funded by MCIN/AEI/10.13039/501100011033/ and by “ERDF A way of making Europe” and from grant PROMETEO CIPROM/2021/079 of Generalitat Valenciana. This work was performed using HPC resources from GENCI–IDRIS and GENCI-TGCC (Grant A0170707156) and from the Barcelona Supercomputing Center (project QH-2025-2-0001). K.Z. acknowledges EUR2025-165005 funded by MCIN/AEI/10.13039/501100011033/. Author contributions A.G.-M., I.T. and D.L. were responsible of the design of the work. A.G.-M. and M.C. performed the calculations, K.Z. and M.P. assisted in the preparation of simulations and all the authors participated in the discussion and analysis. A.G.-M., I.T., J.J.R.-P. and D.L. wrote the draft, while all the authors participated in the revision and editing of the manuscript. Peer review Peer review information Nature Communications thanks the anonymous reviewer(s) for their contribution to the peer review of this work. A peer review file is available. Data availability Inputs and outputs from the simulations and NNP have been deposited in the Zenodo database58, with 10.5281/zenodo.1804429 at https://zenodo.org/records/18044294. Source data are provided with this paper. Competing interests The authors declare no competing interests. Footnotes Publisher’s note Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations. Contributor Information Damien Laage, Email: damien.laage@ens.psl.eu. Iñaki Tuñón, Email: ignacio.tunon@uv.es. Supplementary information The online version contains supplementary material available at 10.1038/s41467-026-73697-7. References Associated Data Supplementary Materials Supplementary Information Transparent Peer Review file Source Data Data Availability Statement Inputs and outputs from the simulations and NNP have been deposited in the Zenodo database58, with 10.5281/zenodo.1804429 at https://zenodo.org/records/18044294. Source data are provided with this paper.