Chemical Accuracy for Carbapenem Breakdown by Class A beta-Lactamases Using a Transferable Embedded Machine-Learned Potential

Bacterial resistance to carbapenems (potent β-lactam antibiotics) is a critical global health threat. The primary cause is β-lactamase enzymes that hydrolyze antibiotics. Previously, carbapenems resisted breakdown, making them ‘last resort’ antibiotics, but enzymes with carbapenemase activity have become widespread. Understanding and predicting the activity of β-lactamases will help in developing antibiotics and inhibitors, and combating antibiotic resistance. Effective prediction requires methods that are both accurate and computationally efficient. Multiscale combined quantum mechanics/molecular mechanics (QM/MM) molecular dynamics simulations of reactions can discriminate between class A β-lactamases (the most widely distributed group) with or without carbapenemase activity, and identify determinants of activity. but face an accuracy/efficiency tradeoff: approximate semiempirical QM methods do not predict reaction barriers accurately, but higher level QM calculations require extensive computer time and resources, making them impractical for rapid predictions. Here, we develop and test a transferable ML/MM framework for the deacylation reaction of the carbapenem meropenem, using the electrostatic machine learning embedding (EMLE) scheme. This model gives free energy barriers to within 1 kcal/mol of values derived from experiment, for four different class A β-lactamases. Crucially, the model was trained only on active site structures from QM/MM simulations of a single enzyme. It did not require retraining to generalize across Class A β-lactamases. It is both faster and more accurate than semi-empirical QM/MM simulations. The ML/MM simulations capture the structural and electrostatic determinants of carbapenemase activity, including oxyanion stabilization and active site electric fields, that differ between enzymes. This work establishes transferable EMLE-based ML/MM simulation as a practical route for accurate, efficient modelling of enzyme-catalyzed reactions.