Hierarchical Quantum Embedding by Machine Learning for Large Molecular Assemblies

Moritz Bensberg, Marco Eckhoff, Raphael T. Husistein, Matthew S. Teynor, Valentina Sora, William Bro‐Jørgensen, F. Emil Thomasen, Anders Krogh, Kresten Lindorff‐Larsen, Gemma C. Solomon, Thomas Weymuth, Markus ReiherJournal of Chemical Theory and Computation 2025 · DOI

TL;DR

This paper introduces a novel "quantum-in-quantum" embedding strategy combined with machine learning to achieve highly accurate descriptions of large molecular systems. By focusing high-accuracy quantum calculations on specific "quantum cores" within a larger quantum region, and then using transfer learning to integrate this into a machine learning potential, they significantly improve binding free energy predictions for protein-ligand complexes.

What Problem Does This Paper Address?

Describing large molecular assemblies, like proteins interacting with drug molecules, with high accuracy is computationally prohibitive. Traditional quantum-classical hybrid (QM/MM) methods embed a quantum mechanical (QM) region within a classical molecular mechanics (MM) environment. However, as the QM region grows to capture important interactions, only approximate QM methods (like DFT) can be used, limiting accuracy. This paper tackles the challenge of improving the accuracy of the QM description within these large QM regions, especially for complex problems like calculating binding free energies, where even small errors can be critical.

Methods

The core innovation is a hierarchical approach:

  • Quantum-In-Quantum (QM/QM) Embedding: Instead of treating the entire QM region with a single QM method, they define smaller "quantum cores" within the QM region. These cores are treated with highly accurate, correlated quantum chemical methods (like DLPNO-CCSD(T)), while the rest of the QM region is described by a less computationally expensive method (like DFT). This is achieved using Huzinaga-type projection-based embedding.
  • Machine Learning Potential (MLP): The energies from these QM/QM/MM calculations are then used to train a machine learning potential.
  • Transfer Learning: To overcome the challenge of limited high-accuracy QM/QM/MM data, they employ transfer learning. An initial MLP is trained on more abundant QM/MM data, and then its parameters are fine-tuned using the more accurate QM/QM/MM energies. This allows the MLP to learn the corrections introduced by the higher-accuracy QM/QM embedding.
  • Alchemical Free Energy Calculations: The final binding free energies are calculated using alchemical free energy methods, specifically non-equilibrium switching simulations, which are made more efficient and accurate by the improved MLPs.

Datasets

The study uses data from a well-studied protein-ligand complex: the inhibitor 19G binding to the myeloid cell leukemia 1 (MCL1) protein.

  • Training Data: Structures were sampled from MM trajectories and active learning. QM/QM/MM energies were calculated for these structures.
  • Reference Data: The paper refers to previous work (Ref. 15) for the QM/MM potential energy surface and the initial DFT calculations.
  • No new benchmarks were introduced. The focus was on improving the accuracy of existing QM/MM approaches.

Key Insights and Ideas

  • Targeted High Accuracy: The "quantum core" concept is crucial. By focusing the most expensive, accurate quantum calculations on the most critical parts of the system (e.g., the ligand's reactive centers), significant accuracy gains can be achieved without prohibitively increasing computational cost.
  • Transfer Learning for Efficiency: The success of this method hinges on transfer learning. It allows leveraging existing, less accurate QM/MM data to build a robust ML potential, which is then refined with a smaller amount of highly accurate QM/QM/MM data. This is essential because generating high-accuracy QM/QM/MM data is very demanding.
  • QM/QM Embedding Improves MLPs: The QM/QM embedding step, by providing more accurate energies, leads to MLPs with significantly lower RMSEs compared to MLPs trained solely on QM/MM data. This directly translates to more reliable predictions.
  • Small QM/QM Corrections Matter: Even though the QM/QM embedding introduces a relatively small energy shift compared to the overall QM/MM energies, this shift is critical for achieving close agreement with experimental binding free energies.

Key Results

  • The QM/QM/MM approach, combined with transfer learning, provided a binding free energy for the 19G-MCL1 complex that closely matched the experimental estimate (-37.2 kJ/mol ± 1.0 kJ/mol vs. -37.3 kJ/mol ± 0.1 kJ/mol). This is an improvement over their previous QM/MM approach, which underestimated the binding.
  • The RMSEs of the QM/QM/MM MLPs were 5% and 14% lower than those of the base QM/MM MLPs for the protein-ligand complex and the solvated ligand, respectively.
  • The energy differences introduced by the QM/QM embedding (∆EEmb) accounted for a significant portion of the total energy difference (∆Etot) between the QM/QM/MM and QM/MM approaches.

Limitations and Future Work

  • Quantum Core Selection: The paper notes that the specific selection of quantum cores can influence the results. While they manually selected cores, they suggest that automated selection procedures could be beneficial.
  • Energy Derivatives: The current implementation does not utilize energy gradients (forces) for the QM/QM/MM calculations, which are known to improve MLP training. The authors suggest that developing analytical expressions for energy derivatives from highly accurate electronic structure methods would be a valuable future direction.
  • Active Learning: They propose investigating active learning strategies to further reduce computational costs by intelligently selecting structures for QM/QM/MM calculations where the QM/QM/MM and QM/MM potential energy surfaces differ significantly.
  • Broader Applicability: While demonstrated for binding free energies, the approach is presented as general and could be applied to other physico-chemical properties.
  • Quantum Computing: The paper hints at future work where quantum computation could be used to generate even higher-accuracy quantum energies.

Summary updated

Abstract

We present a quantum-in-quantum embedding strategy coupled to machine learning potentials to improve on the accuracy of quantum-classical hybrid models for the description of large molecules. In such hybrid models, relevant structural regions (such as those around reaction centers or pockets for binding of host molecules) can be described by a quantum model that is then embedded into a classical molecular-mechanics environment. However, this quantum region may become so large that only approximate electronic structure models are applicable. To then restore accuracy in the quantum description, we here introduce the concept of quantum cores within the quantum region that are amenable to accurate electronic structure models due to their limited size. Huzinaga-type projection-based embedding, for example, can deliver accurate electronic energies obtained with advanced electronic structure methods. The resulting total electronic energies are then fed into a transfer learning approach that efficiently exploits the higher-accuracy data to improve on a machine learning potential obtained for the original quantum-classical hybrid approach. We explore the potential of this approach in the context of a well-studied protein-ligand complex for which we calculate the free energy of binding using alchemical free energy and nonequilibrium switching simulations.
BibTeX
@article{bensberg2025hierarchical,
  author = {Moritz Bensberg and Marco Eckhoff and Raphael T. Husistein and Matthew S. Teynor and Valentina Sora and William Bro‐Jørgensen and F. Emil Thomasen and Anders Krogh and Kresten Lindorff‐Larsen and Gemma C. Solomon and Thomas Weymuth and Markus Reiher},
  title = {{Hierarchical Quantum Embedding by Machine Learning for Large Molecular Assemblies}},
  journal = {Journal of Chemical Theory and Computation},
  year = {2025},
  doi = {10.1021/acs.jctc.5c00389},
}
Figure