Abstract
<jats:p>Metal ion–ligand interactions govern reactivity, bioavailability, and catalytic function across chemistry and biology, yet the many-body effects underlying water exchange around multivalent metal ions remain difficult to capture with classical force fields. Ab initio molecular dynamics (AIMD), by contrast, provides accurate results but is limited to picosecond timescale sampling by its high computational cost. Machine learning force fields (MLFFs) offer a promising route to quantum-mechanics-level accuracy at force-field-level cost, but we show that existing MACE foundation models which were trained on materials predict an incorrect coordination number for Zn2+ in aqueous phase. To address this gap, we developed a CREST-based enhanced-sampling workflow to generate diverse quantum mechanical reference data and trained specialized E(3)-equivariant MACE models for Zn2+, Ca2+, and La3+ in aqueous solution. Systematic benchmarking of DFT functionals, model hyperparameters, data precision, and water geometry shows that model performance is sensitive to several of these choices, underscoring the need for careful validation. Encouragingly, the resulting specialized MACE models accurately reproduce hydration structures, coordination number distributions, and water-exchange kinetics in excellent agreement with experiment and AIMD, outperforming classical force fields (12-6, 12-6-4, and AMOEBA) in simultaneously capturing structure, mechanism, and energetics. We further extend this approach to ML/MM simulations of the metalloenzyme thermolysin, reproducing the monodentate/bidentate switching of Glu166 and capturing Zn2+ coordination number flexibility over a 2 ns trajectory—a timescale inaccessible to conventional QM/MM. These results demonstrate that specialized MLFFs provide a practical and transferable strategy for bridging quantum mechanical accuracy with molecular mechanics speed for metal-ligand systems, enabling long-timescale conformational sampling, mutant screening, and mechanistic studies of metalloenzyme catalysis that were previously out of reach.</jats:p>