In this post, we highlight recent research in electrostatic embedding published in the Journal of Chemical Theory and Computation by researchers at the Universities of Edinburgh, Newcastle and Valencia, along with OpenBioSim, that advances the accuracy of hybrid machine learning and molecular mechanics simulations for drug discovery applications.
The Challenge of Polarization in Molecular Simulations
Machine learning potentials (MLPs) have emerged as a promising approach to molecular simulation, offering quantum mechanical accuracy at a fraction of the computational cost. When combined with traditional molecular mechanics (MM) force fields in hybrid ML/MM schemes, they enable efficient modelling of complex biomolecular systems. However, most ML/MM simulations to date have relied on mechanical embedding approaches that fail to capture a critical physical phenomenon: polarization—the way molecules respond to the electric fields created by their environment.
This limitation is significant. Recent benchmarking studies have shown that mechanical embedding ML/MM methods often perform no better than well-parametrized classical force fields for predicting binding affinities and hydration free energies. A missing ingredient to improve accuracy is electrostatic embedding, where polarization effects on the ML region by the surrounding MM environment are explicitly incorporated.
Achieving Transferable Chemical Accuracy with EMLE
The research team, including OpenBioSim’s Lester Hedges, developed robust training methodologies for the electrostatic machine learning embedding (EMLE) model to address this challenge. By computing absolute hydration free energies for a set of 20 small organic molecules, they demonstrated that properly trained EMLE models can achieve chemical accuracy—predictions within 1 kcal/mol of experimental values.
The study introduces two key innovations. First, a “bespoke” training protocol that fits EMLE parameters to quantum mechanical data, significantly improving agreement with reference QM/MM calculations. Second, a “patching” procedure that further refines the model by directly fitting to the components of QM/MM electrostatic interactions. When combined with an empirical correction in the form of an effective dielectric constant, these EMLE models match or exceed the performance of state-of-the-art classical force fields.
Importantly, the team demonstrated that well-calibrated EMLE models are transferable—they work across different machine learning potentials (including ANI-2x and MACE-OFF) and can be applied to chemically related molecules outside the training set, and deliver accurate hydration free energies, even for molecules where accuracy degrades significantly with traditional forcefields such as OpenFF. This transferability is crucial for practical drug discovery applications.
OpenBioSim’s Contributions
The team has released two new open-source software packages as part of the scientific study: fes-ml for running hybrid ML/MM free energy calculations, and emle-bespoke for training and refining EMLE models. fes-ml builds directly on OpenBioSim’s Sire framework for molecular modelling, and OpenBioSim co-author Lester Hedges assisted the team to develop these capabilities.
Both packages also integrate seamlessly with OpenMM for molecular dynamics, OpenFF for force field parameters, and the emle-engine package for electrostatic embedding, demonstrating the benefits of interoperable open-source tools in computational chemistry.
Looking Ahead
This work represents an important step toward robust electrostatic embedding in ML/MM simulations. By providing well-validated training protocols and accessible software tools, the research enables the broader community to apply these advanced methods to challenging drug discovery problems where traditional force fields fall short. Future developments will focus on extending these approaches to diverse solvent and protein environments, opening new possibilities for accurately modelling protein-ligand interactions in drug design.
Drafted by Claude and edited by Julien Michel
