[Paper Review] Scalable Hybrid Deep Neural Networks/Polarizable Potentials Biomolecular Simulations including long-range effects

Théo Jaffrelot Inizan, Thomas Plé+6 more|arXiv (Cornell University)|Jul 28, 2022

Protein Structure and Dynamics4 citations

TL;DR

This paper introduces Deep-HP, a scalable molecular dynamics framework that hybridizes deep neural network (DNN) potentials (ANI-2X) with polarizable force fields (AMOEBA), enabling accurate, long-timescale simulations of large biomolecular systems (up to 100,000 atoms). By coupling DNNs for solute-solute interactions with AMOEBA’s explicit long-range electrostatics and using a multiple-timestep strategy, the method achieves ~10× speedup and computes solvation and binding free energies within chemical accuracy.

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.

Motivation & Objective

• To develop a scalable molecular dynamics framework that integrates deep neural network potentials with polarizable force fields to model both short-range quantum mechanical effects and long-range electrostatics.
• To enable large-scale biomolecular simulations (100k atoms, microseconds) with high accuracy by combining the strengths of ANI-2X (DNN) and AMOEBA (PFF).
• To accelerate simulations using a multiple-timestep integration scheme that reduces computational cost by evaluating DNN forces only in correction steps.
• To compute solvation and binding free energies for ligands in diverse solvents and host–guest complexes with chemical accuracy.
• To provide a flexible, user-defined partitioning scheme for hybrid simulations including polarizable solvents, counterions, and solutes.

Proposed method

• Extends the Tinker-HP MD package to natively support PyTorch/TensorFlow DNN models via Deep-HP, enabling GPU-accelerated hybrid simulations.
• Implements a hybrid ANI-2X/AMOEBA potential where solute-solute interactions are modeled by the ANI-2X DNN and solvent-solvent/solute-solvent interactions by the AMOEBA polarizable force field.
• Uses an efficient Particle Mesh Ewald (PME) method to explicitly include long-range electrostatics from AMOEBA while preserving ANI-2X’s quantum mechanical accuracy for short-range interactions.
• Applies a multiple-timestep integration strategy: AMOEBA forces are computed at every step, while ANI-2X forces are applied only in correction steps to reduce computational cost.
• Employs a user-defined partitioning scheme to allow flexible hybrid simulations, such as including polarizable counterions or solvents.
• Uses Bennett's acceptance ratio (BAR) method to compute free energies from alchemical alchemical transformations in SAMPL challenges.

Experimental results

Research questions

• RQ1 Can a hybrid DNN/polarizable force field approach achieve chemical accuracy in solvation and binding free energy calculations for large biomolecular systems?
• RQ2 How can long-range electrostatic effects be accurately and efficiently included in DNN-based MD simulations without sacrificing computational scalability?
• RQ3 To what extent can a multiple-timestep integration scheme accelerate DNN-based MD simulations while maintaining accuracy?
• RQ4 Can the ANI-2X/AMOEBA hybrid potential model complex biomolecular environments, including solvation and host–guest binding, with high fidelity?
• RQ5 What is the performance and accuracy of the Deep-HP framework in simulating systems with 100,000 atoms over microsecond timescales?

Key findings

• The ANI-2X/AMOEBA hybrid potential achieves solvation free energy predictions within chemical accuracy (≤1 kcal/mol) for both charged and uncharged ligands across four solvents.
• Absolute binding free energies for host–guest complexes from SAMPL challenges were computed with average errors within chemical accuracy, validating the method’s predictive power.
• Simulations of up to 10 μs were successfully performed, demonstrating the framework’s capability for long-timescale biomolecular dynamics.
• The multiple-timestep strategy reduced computational cost by approximately an order of magnitude compared to standard Velocity Verlet integration, enabling ns-scale simulations for 100k-atom systems.
• Deep-HP enables scalable, multi-GPU simulations of large biomolecular systems (100k atoms) with hybrid DNN/PFF accuracy at the cost of classical force fields.
• The user-defined partitioning scheme allows flexible modeling of key biomolecular features such as polarizable solvents and counterions, enhancing simulation realism.