Transferable Implicit Solvation via Contrastive Learning of Graph Neural Networks

Abstract

Implicit solvent models are essential for molecular dynamics simulations of biomolecules, striking a balance between computational efficiency and biological realism. Efforts are underway to develop accurate and transferable implicit solvent models and coarse-grained (CG) force fields in general, guided by a bottom-up approach that matches the CG energy function with the potential of mean force (PMF) defined by the finer system. However, practical challenges arise due to the lack of analytical expressions for the PMF and algorithmic limitations in parameterizing CG force fields. To address these challenges, a machine learning-based approach is proposed, utilizing graph neural networks (GNNs) to represent the solvation free energy and potential contrasting for parameter optimization. We demonstrate the effectiveness of the approach by deriving a transferable GNN implicit solvent model using 600,000 atomistic configurations of six proteins obtained from explicit solvent simulations. The GNN model provides solvation free energy estimations much more accurately than state-of-the-art implicit solvent models, reproducing configurational distributions of explicit solvent simulations. We also demonstrate the reasonable transferability of the GNN model outside of the training data. Our study offers valuable insights for deriving systematically improvable implicit solvent models and CG force fields from a bottom-up perspective.

General Graph Neural Network Based Implicit Solvation Model for Organic Molecules in Water

“…As computational efficiency is crucial for the practical usability (i.e., the implicit solvent model must be faster than explicit-solvent simulations, which is not achieved by many of the proposed ML models [11][12][13]), we first investigated the effect of the complexity of the GNN architecture on simulation speed. For this, we varied the number of parameters in the GNN (i.e., the size of the hidden layers in the multi-layer perceptrons (MLPs)) from 128 to 96, 64, and 48 when training on our set of 369'486 diverse small molecules with molecular weight < 500 Da.…”

“…While many of them are indeed much faster than their explicit counterpart, the major drawback of these methods is that they do not describe the local solvation effects correctly. Based on recent successes of applying machine learning (ML) in the field of chemistry [8], ML-based approaches have been developed to learn the effects of a given environment (solvent) on a solute [9][10][11][12][13]. These models are either too slow and/or not sufficiently transferable between different molecules to be practically usable in MD simulations, leaving explicit solvent simulations as the only reliable solution for generating accurate conformational ensembles in solution.…”

“…ML approaches have also been applied to model the effects of a given solvent environment on molecules [9][10][11][12][13]. While these models are indeed capable of reproducing desired effects (e.g., reproduction of conformational ensembles or free-energy profiles) and dramatically outperform the standard implicit solvent methods in reproducing explicit solvent reference calculations, they currently have two major limitations for practical usability.…”

Differentiable simulation to develop molecular dynamics force fields for disordered proteins

“…7 Despite the lack of interactions between the solute and individual solvent molecules being a fundamental limitation of implicit solvent models, they are regularly used across a range of biomolecular simulations 8 and have been the target of machine learning approaches. 9,10 Small proteins can be folded in GPU-days with implicit solvent models and enhanced sampling. 11…”

Differentiable simulation to develop molecular dynamics force fields for disordered proteins

“…This speeds up simulations by reducing the number of atoms and by giving faster exploration of conformational space due to the lack of friction with solvent, with speedups of up to 100x over explicit solvent [7]. Despite the lack of interactions between the solute and individual solvent molecules being a fundamental limitation of implicit solvent models, they are regularly used across a range of biomolecular simulations [8] and have been the target of machine learning approaches [9,10]. Small proteins can be folded in GPU-days with implicit solvent models and enhanced sampling [11].…”