Improved electrostatic machine learning embedding ML/MM scheme for a phosphoryl transfer reaction
- 1Université Paris Cité, CNRS, Laboratoire de Biochimie Théorique, 13 rue Pierre et Marie Curie, Paris, France, 75005
- 2Departament de Química Física
- 3Universitat de València
Abstract
The promise of hybrid machine learning/molecular mechanics (ML/MM) simulations is to provide QM/MM-quality free energy profiles at a fraction of the cost. One of the challenges is to accurately treat the interaction between the ML subsystem and the MM environment. Here we test the Electrostatic ML Embedding (EMLE) scheme on a minimal model of RNA cleavage in water, which is already challenging due to a doubly charged reactant dissociating into two anions. The original scheme, a MACE potential embedded by an independent EMLE model, is compared with a new EMLE–MACE architecture, in which a single equivariant network predicts the gas-phase energy together with the atomic charges, valence widths, polarizabilities and static dipoles required by the embedding. The models were trained by active learning against M06-2X/def2- TZVP data, with structures selected by query-by-committee from the ML/MM dynamics itself. About 1500 single point calculations were sufficient for stable 100 ps umbrella sampling simulations in 73 windows along the reaction pathway. The EMLE–MACE profile reproduces the reference DFT/MM profile within its sampling error, including a metastable pentacoordinate intermediate. Energy decomposition shows that the static dipoles are essential: without them the static error stays above chemical accuracy even for an ideally trained model, due to the irreducible error of approximating electron density with spherical atoms.
Related
- A Hybrid Message-Passing and Electric-Field Representation for Electrostatic QM/MM Embedding with Machine Learning Interatomic Potentials
- ¹⁸ O-Informed Metabolomics Reveals Microbial Phosphoryl-Transfer Dynamics and Stress-Induced Changes
- GPU infrastructure for ML/MM with electrostatic embedding for enzyme chemistry
- Machine learned potentials with electrostatic embedding accurately capture Kemp eliminase reactivity
- Ensemble Molecular Surface Electrostatic Potential Descriptors as Informative Features for Active Learning
BibTeX
@misc{kantin2026improved,
title = {{Improved electrostatic machine learning embedding ML/MM scheme for a phosphoryl transfer reaction}},
author = {Laetitia Kantin and Iñaki Tuñón and Kirill Zinovjev and Elise Duboué Dijon},
howpublished = {ChemRxiv},
year = {2026},
month = oct,
doi = {10.26434/chemrxiv.15010047/v1},
url = {https://chemrxiv.org/doi/full/10.26434/chemrxiv.15010047/v1},
note = {CC BY},
}
RIS
TY - UNPB
TI - Improved electrostatic machine learning embedding ML/MM scheme for a phosphoryl transfer reaction
AU - Laetitia Kantin
AU - Iñaki Tuñón
AU - Kirill Zinovjev
AU - Elise Duboué Dijon
PY - 2026
DA - 2026/10/07
PB - ChemRxiv
DO - 10.26434/chemrxiv.15010047/v1
UR - https://chemrxiv.org/doi/full/10.26434/chemrxiv.15010047/v1
AB - The promise of hybrid machine learning/molecular mechanics (ML/MM) simulations is to provide QM/MM-quality free energy profiles at a fraction of the cost. One of the challenges is to accurately treat the interaction between the ML subsystem and the MM environment. Here we test the Electrostatic ML Embedding (EMLE) scheme on a minimal model of RNA cleavage in water, which is already challenging due to a doubly charged reactant dissociating into two anions. The original scheme, a MACE potential embedded by an independent EMLE model, is compared with a new EMLE–MACE architecture, in which a single equivariant network predicts the gas-phase energy together with the atomic charges, valence widths, polarizabilities and static dipoles required by the embedding. The models were trained by active learning against M06-2X/def2- TZVP data, with structures selected by query-by-committee from the ML/MM dynamics itself. About 1500 single point calculations were sufficient for stable 100 ps umbrella sampling simulations in 73 windows along the reaction pathway. The EMLE–MACE profile reproduces the reference DFT/MM profile within its sampling error, including a metastable pentacoordinate intermediate. Energy decomposition shows that the static dipoles are essential: without them the static error stays above chemical accuracy even for an ideally trained model, due to the irreducible error of approximating electron density with spherical atoms.
ER -