Hybrid ML/MM Molecular Dynamics Engine in AMBER

This repository contains the implementation of a hybrid Machine Learning / Molecular Mechanics (ML/MM) molecular dynamics simulation engine, based on the AMBER software suite. This tool enables multiscale simulations by integrating machine-learned interatomic potentials (MLIPs) into the AMBER MD engine, specifically SANDER, to perform accurate and efficient free energy calculations.

Features

• Built into AMBER's SANDER engine with Fortran/C++ interfacing
• Integrated support for 9 popular MLIPs:
  • ANI series (ANI-2x, ANI-1xnr)
  • MACE series (MACE-OFF23 Small/Medium/Large)
  • AIMNet2-NSE (with charge and spin support)
  • SpookyNet
  • EANN/FIRANN
  • Egret-1S
• Link Atom Scheme: Complete QM/MM boundary handling mechanism for covalent bonds crossing the ML/MM boundary
• PME Electrostatics Integration: Accurate long-range electrostatics treatment for ML/MM interactions
• Supports inference + auto-gradient in C++ for force calculation via LibTorch
• ML computation on GPU, MD on CPU for parallel asynchronous computing
• Validated with energy & momentum conservation tests
• Supports thermodynamic integration (TI) for hydration free energy calculations
• Compatible with MM-PBSA for protein-ligand binding free energy estimations

New Features (v2.0)

• Multi-model support: Runtime model selection via mlp_model parameter
• AIMNet2-NSE charge/spin: Full support for mlp_charge and mlp_multi parameters for spin-polarized calculations
• Link Atom Scheme: Position link atoms, distribute forces, and manage charges at QM/MM boundaries
• Flexible configuration: GPU device selection, QM region constraints, SHAKE options
• Improved stability: JIT optimization guard for in-place operation compatibility

References

This work is based on the methods described in the following papers:

Article

"Accurate Free Energy Calculation via Multiscale Simulations Driven by Hybrid Machine Learning and Molecular Mechanics Potentials"

Xujian Wang, Xiongwu Wu, Bernard R. Brooks, and Junmei Wang JCTC, 2025 DOI: 10.1021/acs.jctc.5c00598

"Redefining Computational Enzymology with Multiscale Machine Learning/Molecular Mechanics Metadynamics: Deciphering Catalytic Mechanism and Stereoselectivity in Diels–Alderases"

Xujian Wang, Haocheng Tang, Xiongwu Wu, Bernard R. Brooks, Junmei Wang and Wan-Lu Li ChemRxiv, 2025 DOI: [10.1021/acs.jctc.5c00598](https://chemrxiv.org/doi/full/10.26434/chemrxiv-2025-x2k6j)

Documentation

• INSTALL.md - Installation guide
• TUTORIAL.md - Usage tutorial

Supported MLP Models

mlp_model	Model Name	Elements	Description
0	ANI-2x	H,C,N,O,S,F,Cl	General organic molecules
1	MACE-OFF23-S	H,C,N,O,F,...	Small MACE model
2	ANI-1xnr	H,C,N,O	Fast screening
3	MACE-OFF23-M	H,C,N,O,F,...	Medium MACE model
4	MACE-OFF23-L	H,C,N,O,F,...	Large MACE model
5	AIMNet2-NSE	H,C,N,O,S,F,Cl,Br	Spin-polarized support
8	SpookyNet	H,C,N,O,F,...	Long-range interactions
9	EANN	Configurable	Custom trained models
10	Egret-1S	H,C,N,O,...	Latest MACE variant

Quick Start

&cntrl
  ifmlp = 1,           ! Enable MLP
  ...
/
&mlp
  animask = '@1-50',   ! QM atom selection
  mlp_model = 0,       ! ANI-2x
  gpu_id = 0,          ! GPU device
  lnk_atomic_no = 1,   ! Link atom element (H)
  lnk_dis = 1.09,      ! Link atom distance (Angstrom)
  mlp_shake = 1,       ! Constrain QM-H bonds
/

Contact

For questions or collaboration, please contact:

• hsuchein0126@outlook.com
• juw79@pitt.edu

---

© 2025. Released under CC BY 4.0 License