Graph-convolutional neural networks for (QM)ML/MM molecular dynamics simulations.

Area of Science:

• Computational Chemistry
• Molecular Dynamics
• Machine Learning

Background:

• Accurate condensed-phase simulations require quantum mechanics (QM) and extensive sampling, but QM/MM methods are computationally expensive for large systems and long timescales.
• Machine learning (ML) models can replace QM calculations, but struggle with long-range interactions crucial for condensed-phase systems.
• A hybrid approach using density functional tight binding (DFTB) and high-dimensional neural network potentials (HDNNP) in a Δ-learning scheme has shown promise in capturing these interactions.

Purpose of the Study:

• To investigate the efficacy of graph-convolutional neural networks (GCNNs) for QM/MM molecular dynamics (MD) simulations, both with and without a Δ-learning scheme.
• To evaluate the performance of GCNN-based ML models in accurately describing condensed-phase systems, particularly concerning long-range interactions.
• To validate the developed GCNN Δ-learning approach through prospective simulations of biologically relevant systems.

Main Methods:

• Development and application of GCNN models, with and without a Δ-learning scheme, for QM/MM MD simulations.
• Utilizing DFTB as a baseline method within the Δ-learning framework to incorporate long-range interactions.
• Benchmarking the GCNN models against established QM/MM methods using solutes and chemical reactions in water.
• Performing prospective QM/MM MD simulations of retinoic acid and S-adenosylmethionine-cytosine interactions in aqueous environments.

Main Results:

• The Δ-learning approach employing GCNNs and DFTB demonstrated competitive performance compared to traditional QM/MM methods.
• The GCNN-based Δ-learning model effectively captured long-range interactions, a key challenge in ML for condensed-phase systems.
• Prospective simulations of retinoic acid and S-adenosylmethionine-cytosine systems in water validated the accuracy and applicability of the developed model.

Conclusions:

• The Δ-learning GCNN model presents a valuable and computationally efficient alternative for QM/MM MD simulations of condensed-phase systems.
• This approach successfully addresses the limitations of standard ML models in describing long-range interactions.
• The validated GCNN Δ-learning method holds significant potential for studying complex chemical processes in solution and biological environments.