Machine learning implicit solvation for molecular dynamics

Yaoyi Chen, Andreas Krämer, Nicholas E. Charron, Brooke E. Husic, Cecilia Clementi, Frank NoéRefubium 2022 · DOI

TL;DR

This paper introduces ISSNet, a machine learning model that learns implicit solvent models for molecular dynamics simulations. By leveraging graph neural networks and coarse-graining principles, ISSNet can accurately reproduce the thermodynamic behavior of explicit solvent systems, outperforming traditional implicit solvent models like GBSA-OBC. This advancement promises more efficient and accurate simulations for biological and drug design applications.

What Problem Does This Paper Address?

Accurately simulating the behavior of biological molecules in their solvent environment is critical for fields like computational biology and drug design. Explicitly modeling every solvent molecule (like water) is computationally expensive, limiting the length and size of simulations. Implicit solvent models offer a computationally cheaper alternative by approximating solvent effects as a continuous field. However, existing implicit models often struggle to capture the complex, many-body interactions of solvents, leading to inaccuracies in predicting molecular behavior, such as protein folding or conformational changes. This paper aims to bridge this gap by developing a machine learning approach that can learn highly accurate implicit solvent models directly from explicit solvent simulation data.

Methods

The core of this work is the ISSNet (Implicit Solvent SchNet) architecture, a graph neural network designed to learn the implicit solvent potential of mean force. It builds upon previous machine learning coarse-graining (ML-CG) models like CGnet and CGSchnet.

Here's a breakdown of the key methodological aspects:

  • Multi-scale Coarse-Graining (MCG) Framework: The authors frame the problem of learning an implicit solvent model as a "bottom-up" coarse-graining problem. Instead of explicitly modeling solvent molecules, the solute molecule is treated as a coarse-grained system, and the solvent's influence is learned as a potential.
  • Variational Force Matching (FM) Loss Function: ISSNet is trained using a loss function derived from the MCG framework. This loss function minimizes the difference between the forces predicted by the implicit solvent model and the forces experienced by the solute in explicit solvent simulations. This allows for direct optimization of the implicit solvent model against explicit solvent data.
  • Graph Neural Network (SchNet): The core of ISSNet is the SchNet architecture, a graph neural network well-suited for learning molecular properties. SchNet represents atoms as nodes in a graph and uses continuous-filter convolutions to learn interactions based on pairwise distances. This architecture inherently respects translational and rotational symmetries.
  • Embedding Strategies: The paper explores three ways to represent input atoms for the SchNet:
  • Type-only (t-ISSNet): Uses only the atom's element type (nuclear charge).
  • Charge-only (q-ISSNet): Incorporates the atomic partial charge from the force field.
  • Type-and-charge (qt-ISSNet): Combines both element type and partial charge information.
  • Training and Validation: Models are trained on data from explicit solvent simulations. The training process involves shuffling solute coordinates and forces, batching the data, and updating model parameters using stochastic gradient descent. Cross-validation is used to select optimal hyperparameters.
  • Simulation: Trained ISSNet models are then used to perform molecular dynamics simulations, generating implicit solvent trajectories.

Datasets

The paper utilizes data from existing, well-characterized molecular dynamics simulations:

  • Capped Alanine: A small peptide system. The authors performed a 1-µs all-atom MD simulation with TIP3P explicit solvent to generate training and validation data (solute coordinates and forces).
  • Chignolin: A miniprotein. The authors used existing explicit solvent all-atom simulation trajectories and force data (provided by the authors of Ref. 46) for training and validation. This dataset contained 1.8 × 10⁶ coordinate-solvation force pairs, from which 2 × 10⁵ pairs were randomly selected for training.

The paper does not introduce new benchmark datasets but rather uses these existing datasets to evaluate their novel method against established implicit solvent models.

Key Insights and Ideas

  • Learning Electrostatics is Crucial: The study highlights that incorporating partial atomic charges (q- and qt-ISSNet) into the neural network's input features significantly improves accuracy compared to using only atom types (t-ISSNet). This suggests that explicitly learning electrostatic interactions is vital for accurate implicit solvent modeling, rather than relying solely on the network to infer them from element types.
  • Force Matching is a Powerful Loss Function: The variational force matching (FM) loss function, derived from the coarse-graining framework, proves effective for training implicit solvent models. It directly optimizes the model to reproduce the forces experienced by the solute in explicit solvent.
  • ISSNet Outperforms Traditional Models: The developed ISSNet models consistently show better performance than the widely used GBSA-OBC model in reproducing the thermodynamic properties of the studied systems. This is evidenced by lower divergence and error metrics.
  • Potential for Transferability: The underlying SchNet architecture has shown promise for transferability across different molecular systems, suggesting that ISSNet models could potentially be generalized to describe solvent effects for a wider range of molecules.
  • FM Error Limitations: The authors note that while FM error is useful for hyperparameter optimization, it may not always correlate perfectly with the accuracy of free energy landscapes, especially for complex systems with rare events. This is because high-energy regions, crucial for barrier heights, contribute less to the overall FM loss.

Key Results

  • Improved Thermodynamic Accuracy: For both capped alanine and chignolin, ISSNet models (particularly q-ISSNet and qt-ISSNet) significantly reduced the Kullback-Leibler (KL) divergence, Jensen-Shannon (JS) divergence, and Mean Squared Error (MSE) of free energy landscapes compared to the GBSA-OBC model and vacuum simulations. For chignolin, the qt-ISSNet model achieved the lowest divergence metrics.
  • Accurate Free Energy Landscapes: ISSNet models were able to qualitatively and quantitatively reproduce the free energy landscapes of explicit solvent simulations, capturing metastable states and their relative populations more accurately than GBSA-OBC. For chignolin, GBSA-OBC tended to over-stabilize the unfolded state.
  • Realistic Folding Temperatures: For chignolin, ISSNet models predicted melting temperatures (Tm) that were much closer to experimental values and explicit solvent simulations than the GBSA-OBC model, which significantly underestimated Tm.
  • Comparable Structures: Visual inspection of representative structures from explicit and implicit solvent simulations showed good agreement, with ISSNet models producing overlays that were less displaced than those from GBSA-OBC.

Limitations and Future Work

  • Computational Cost: While ISSNet models are more accurate, the neural network evaluations for forces can be computationally demanding, limiting their application to very long simulations or extremely large systems.
  • Parameterization Specificity: Current ISSNet models are parameterized for a specific molecular system at a fixed thermodynamic state. They may not generalize well to different temperatures, pressures, or solute molecules without retraining.
  • Transferability: While the SchNet architecture has potential for transferability, achieving accurate and generalizable implicit solvent models across diverse systems and thermodynamic states remains a challenge.
  • Future Directions:
    • Reducing Computational Cost: Investigating methods like multiple time-step MD, optimizing communication between MD and deep learning frameworks, and exploring computationally cheaper neural network architectures.
  • Improving Transferability: Training on larger, more diverse datasets of molecules and thermodynamic states to develop more generalizable models.
  • Alternative Loss Functions: Exploring alternative or complementary loss functions (e.g., relative entropy) alongside force matching to better capture free energy landscapes and thermodynamic properties.
  • Modeling Temperature/Pressure Dependence: Developing methods to accurately model the dependence of solvation free energy on temperature and pressure.

Summary updated

Abstract

Accurate modeling of the solvent environment for biological molecules is crucial for computational biology and drug design. A popular approach to achieve long simulation time scales for large system sizes is to incorporate the effect of the solvent in a mean-field fashion with implicit solvent models. However, a challenge with existing implicit solvent models is that they often lack accuracy or certain physical properties compared to explicit solvent models as the many-body effects of the neglected solvent molecules are difficult to model as a mean field. Here, we leverage machine learning (ML) and multi-scale coarse graining (CG) in order to learn implicit solvent models that can approximate the energetic and thermodynamic properties of a given explicit solvent model with arbitrary accuracy, given enough training data. Following the previous ML–CG models CGnet and CGSchnet, we introduce ISSNet, a graph neural network, to model the implicit solvent potential of mean force. ISSNet can learn from explicit solvent simulation data and be readily applied to molecular dynamics simulations. We compare the solute conformational distributions under different solvation treatments for two peptide systems. The results indicate that ISSNet models can outperform widely used generalized Born and surface area models in reproducing the thermodynamics of small protein systems with respect to explicit solvent. The success of this novel method demonstrates the potential benefit of applying machine learning methods in accurate modeling of solvent effects for in silico research and biomedical applications.
BibTeX
@inproceedings{chen2022machine,
  author = {Yaoyi Chen and Andreas Krämer and Nicholas E. Charron and Brooke E. Husic and Cecilia Clementi and Frank Noé},
  title = {{Machine learning implicit solvation for molecular dynamics}},
  booktitle = {Refubium},
  year = {2022},
  doi = {10.17169/refubium-33988},
}
Figure