Scalable Hybrid Deep Neural Networks/Polarizable Potentials Biomolecular Simulations including long-range effects

TL;DR

Deep-HP integrates PyTorch/TensorFlow neural networks into the Tinker-HP MD engine to couple ANI-2X neural network potentials with the AMOEBA polarizable force field via a hybrid energy V_HYB(P∪W)= V_AMOEBA(P∪W) + V_ML(P) − V_AMOEBA(P), incorporating long-range effects through PME. The platform enables ns-scale simulations of systems up to 10^5–10^7 atoms on multi-GPU infrastructures, demonstrated by solvation free energies across 4 solvents and absolute binding free energies for SAMPL host–guest challenges, achieving chemical accuracy (errors ≲ 1 kcal/mol). The work shows strong scalability to millions of atoms (e.g., 7.7M atoms on 68 GPUs) and significant acceleration via multi-timestep RESPA-like schemes, reducing computational cost while preserving accuracy. By unifying ML potentials with physics-based polarizable models, Deep-HP paves the way for large-scale, chemically accurate, hybrid MD simulations at force-field cost, with future directions including adaptive timestepping, reweighting strategies, and extension to broader materials and catalysis contexts.

Abstract

Deep-HP is a scalable extension of the \TinkerHP\ multi-GPUs molecular dynamics (MD) package enabling the use of Pytorch/TensorFlow Deep Neural Networks (DNNs) models. Deep-HP increases DNNs MD capabilities by orders of magnitude offering access to ns simulations for 100k-atom biosystems while offering the possibility of coupling DNNs to any classical (FFs) and many-body polarizable (PFFs) force fields. It allows therefore to introduce the ANI-2X/AMOEBA hybrid polarizable potential designed for ligand binding studies where solvent-solvent and solvent-solute interactions are computed with the AMOEBA PFF while solute-solute ones are computed by the ANI-2x DNN. ANI-2X/AMOEBA explicitly includes AMOEBA's physical long-range interactions via an efficient Particle Mesh Ewald implementation while preserving ANI-2X's solute short-range quantum mechanical accuracy. The DNNs/PFFs partition can be user-defined allowing for hybrid simulations to include biosimulation key ingredients such as polarizable solvents, polarizable counter ions, etc... ANI-2X/AMOEBA is accelerated using a multiple-timestep strategy focusing on the models contributions to low-frequency modes of nuclear forces. It primarily evaluates AMOEBA forces while including ANI-2x ones only via correction-steps resulting in an order of magnitude acceleration over standard Velocity Verlet integration. Simulating more than 10 $μ$, we compute charged/uncharged ligands solvation free energies in 4 solvents, and absolute binding free energies of host-guest complexes from SAMPL challenges. ANI-2X/AMOEBA average errors are within chemical accuracy opening the path towards large-scale hybrid DNNs simulations, at force-field cost, in biophysics and drug discovery.

Figures (3)

• Figure 1: a) Performance comparison between ANI-1ccx(1NN), ANI-2x(1NN) and AMOEBA models in ns per day, over increasing system size, on a single Nvidia Tesla A100. b) Strong scaling logarithmic scale plot of ANI-2x model on benchmark systems. Simulations are performed in the NVE ensemble using a Velocity-Verlet integrator 0.2fs time-step.
• Figure 2: Solvation free energies of molecules in different solvents computed with AMOEBA (orange) from refs essex_free_energiespoltype1 versus hybrid model ANI-2X/AMOEBA (blue) and experiment (red). The blue domain correspond to the so-called chemical accuracy: error of 1 kcal/mol w.r.t experiment.
• Figure 3: Binding free energies of host-guest systems of the SAMPL4 and SAMPL8 blind challenges with AMOEBA (orange) from refs sampl_hostbind versus hybrid model ANI-2X/AMOEBA (blue) and experimental (red). The blue domain correspond to the so-called chemical accuracy: error of 1kcal/mol w.r.t experiment.