Transferable Molecular Charge Assignment Using Deep Neural Networks

Ben Nebgen, Nick Lubbers, Justin S. Smith, Andrew E. Sifain, Andrey Y. Lokhov, Olexandr Isayev, Adrián E. Roitberg, Kipton Barros, Sergei TretiakarXiv 2018 · DOI · arXiv

Abstract

We use HIP-NN, a neural network architecture that excels at predicting molecular energies, to predict atomic charges. The charge predictions are accurate over a wide range of molecules (both small and large) and for a diverse set of charge assignment schemes. To demonstrate the power of charge prediction on non-equilibrium geometries, we use HIP-NN to generate IR spectra from dynamical trajectories on a variety of molecules. The results are in good agreement with reference IR spectra produced by traditional theoretical methods. Critically, for this application, HIP-NN charge predictions are about $10^4$ times faster than direct DFT charge calculations. Thus, ML provides a pathway to greatly increase the range of feasible simulations while retaining quantum-level accuracy. In summary, our results provide further evidence that machine learning can replicate high-level quantum calculations at a tiny fraction of the computational cost.

Introduction

In quantum mechanics, "charge" is not a simple discrete value but a continuous density field. To make this data useful for chemical applications—such as drug-protein binding or infrared (IR) spectrum prediction—scientists use charge partitioning schemes to condense this density into discrete atomic charges. Traditionally, the "gold standard" for these calculations is Density Functional Theory (DFT). However, DFT's computational cost scales cubically with the number of atoms, making it impractical for large biological systems or long-duration dynamical simulations. This research addresses the need for a method that maintains quantum-level accuracy while operating at the speed of a simple mathematical model.

Methods: The HIP-NN Architecture

The researchers utilized the Hierarchically Interacting Particle Neural Network (HIP-NN). Unlike standard neural networks, HIP-NN treats molecules as collections of interacting points rather than static images.

HIP-NN architecture for atomic charge prediction

The model's architecture is built on several key principles:

  • Feature Representation: It requires only atomic numbers and pairwise distances, ensuring the model is invariant to rotation and translation.
  • Hierarchical Interaction: It uses interaction layers to share information between neighboring atoms and on-site layers for local processing.
  • Summation Strategy: A unique feature is that the final charge prediction for an atom is a sum of contributions from multiple layers, allowing it to capture both local and broader structural contexts.

Key Insights and Results

The study revealed that the underlying physics of the target property significantly impacts learnability. The model excelled at "local" charge schemes like Hirshfeld or CM5, where an atom's charge is primarily determined by its immediate neighbors.

Prediction error across different charge partitioning schemes

As shown in Figure 3, the model struggled more with MSK charges, which are constrained to match the overall molecular dipole, introducing non-local dependencies. This suggests that while deep learning is powerful, the underlying physics of the target property dictates how easily a model can learn it.

Performance Highlights:

  • Quantum Accuracy: The model achieved a Mean Absolute Error (MAE) of approximately 0.005 $e^-$, surpassing the typical 0.01 $e^-$ variation between high-level quantum methods.
  • Efficiency: HIP-NN is roughly 10,000 times faster than direct DFT. For example, an IR spectrum simulation that would take 3,000 CPU hours via DFT was completed in just 5 minutes on a single GPU.
  • Scalability: The model successfully predicted charges for a 50kDa protein (glucoamylase) in under two minutes, a feat virtually impossible with standard quantum methods.

Conclusion and Future Directions

While HIP-NN demonstrates remarkable speed and accuracy, its reliance on local interactions within a certain cutoff distance limits its performance on non-local charge assignments like MSK. Future research may benefit from bypassing artificial partitioning schemes entirely, focusing instead on predicting experimentally observable quantities like dipole moments or bond orders directly. This approach would move machine learning models closer to pure physical prediction while maintaining the massive computational advantages demonstrated here.

Summary updated

Abstract

We use HIP-NN, a neural network architecture that excels at predicting molecular energies, to predict atomic charges. The charge predictions are accurate over a wide range of molecules (both small and large) and for a diverse set of charge assignment schemes. To demonstrate the power of charge prediction on non-equilibrium geometries, we use HIP-NN to generate IR spectra from dynamical trajectories on a variety of molecules. The results are in good agreement with reference IR spectra produced by traditional theoretical methods. Critically, for this application, HIP-NN charge predictions are about 104 times faster than direct DFT charge calculations. Thus, ML provides a pathway to greatly increase the range of feasible simulations while retaining quantum-level accuracy. In summary, our results provide further evidence that machine learning can replicate high-level quantum calculations at a tiny fraction of the computational cost.
BibTeX
@misc{nebgen2018transferable,
  author = {Ben Nebgen and Nick Lubbers and Justin S. Smith and Andrew E. Sifain and Andrey Y. Lokhov and Olexandr Isayev and Adrián E. Roitberg and Kipton Barros and Sergei Tretiak},
  title = {{Transferable Molecular Charge Assignment Using Deep Neural Networks}},
  year = {2018},
  eprint = {1803.04395},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/1803.04395},
}
Figure