Scaling Graph Neural Networks to Large Proteins

Journal of Chemical Theory and Computation, Journal Year: 2025, Volume and Issue: unknown

Published: Feb. 6, 2025

Graph neural network (GNN) architectures have emerged as promising force field models, exhibiting high accuracy in predicting complex energies and forces based on atomic identities Cartesian coordinates. To expand the applicability of GNNs, machine learning fields more broadly, optimizing their computational efficiency is critical, especially for large biomolecular systems classical molecular dynamics simulations. In this study, we address key challenges existing GNN benchmarks by introducing a dataset, DISPEF, which comprises large, biologically relevant proteins. DISPEF includes 207,454 proteins with sizes up to 12,499 atoms features diverse chemical environments, spanning folded disordered regions. The implicit solvation free energies, used training targets, represent particularly challenging case due many-body nature, providing stringent test evaluating expressiveness models. We benchmark performance seven GNNs emphasizing importance directly accounting long-range interactions enhance model transferability. Additionally, present novel multiscale architecture, termed Schake, delivers transferable computationally efficient energy predictions Our findings offer valuable insights tools advancing protein modeling applications.

Language: Английский

---

Neural Network Potential with Multiresolution Approach Enables Accurate Prediction of Reaction Free Energies in Solution

Journal of the American Chemical Society, Journal Year: 2025, Volume and Issue: unknown

Published: Feb. 17, 2025

We present the design and implementation of a novel neural network potential (NNP) its combination with an electrostatic embedding scheme, commonly used within context hybrid quantum-mechanical/molecular-mechanical (QM/MM) simulations. Substitution computationally expensive QM Hamiltonian by NNP same accuracy largely reduces computational cost enables efficient sampling in prospective MD simulations, main limitation faced traditional QM/MM setups. The model relies on recently introduced anisotropic message passing (AMP) formalism to compute atomic interactions encode symmetries found systems. AMP is shown be highly terms both data costs can readily scaled sample systems involving more than 350 solute 40,000 solvent atoms for hundreds nanoseconds using umbrella sampling. Most deviations predictions from underlying DFT ground truth lie chemical (4.184 kJ mol–1). performance broad applicability our approach are showcased calculating free-energy surface alanine dipeptide, preferred ligation states nickel phosphine complexes, dissociation free energies charged pyridine quinoline dimers. Results this ML/MM show excellent agreement experimental reach most cases. In contrast, calculated static calculations paired implicit models or simulations cheaper semiempirical methods up ten times higher deviation sometimes even fail reproduce qualitative trends.

Language: Английский

---

Rapid Access to Small Molecule Conformational Ensembles in Organic Solvents Enabled by Graph Neural Network-Based Implicit Solvent Model

Journal of the American Chemical Society, Journal Year: 2025, Volume and Issue: unknown

Published: April 10, 2025

Understanding and manipulating the conformational behavior of a molecule in different solvent environments is great interest fields drug discovery organic synthesis. Molecular dynamics (MD) simulations with molecules explicitly present are gold standard to compute such ensembles (within accuracy underlying force field), complementing experimental findings supporting their interpretation. However, conventional methods often face challenges related computational cost (explicit solvent) or (implicit solvent). Here, we showcase how our graph neural network (GNN)-based implicit (GNNIS) approach can be used rapidly small 39 common solvents reproducing explicit-solvent high accuracy. We validate this using nuclear magnetic resonance (NMR) measurements, thus identifying conformers contributing most observable. The method allows time required accurately predict reduced from days minutes while achieving results within one kBT values.

Language: Английский

---

Free energy profiles for chemical reactions in solution from high-dimensional neural network potentials: The case of the Strecker synthesis

The Journal of Chemical Physics, Journal Year: 2025, Volume and Issue: 162(17)

Published: May 6, 2025

Machine learning potentials (MLPs) have become a popular tool in chemistry and materials science as they combine the accuracy of electronic structure calculations with high computational efficiency analytic potentials. MLPs are particularly useful for computationally demanding simulations such determination free energy profiles governing chemical reactions solution, but to date, applications still rare. In this work, we show how umbrella sampling can be combined active high-dimensional neural network (HDNNPs) construct systematic way. For example first step Strecker synthesis glycine aqueous provide detailed analysis improving quality HDNNPs datasets increasing size. We find that, addition typical quantification force errors respect underlying density functional theory data, long-term stability convergence physical properties should rigorously monitored obtain reliable converged solution.

Language: Английский

---

Predicting solvation free energies with an implicit solvent machine learning potential

The Journal of Chemical Physics, Journal Year: 2024, Volume and Issue: 161(23)

Published: Dec. 16, 2024

Machine learning (ML) potentials are a powerful tool in molecular modeling, enabling ab initio accuracy for comparably small computational costs. Nevertheless, all-atom simulations employing best-performing graph neural network architectures still too expensive applications requiring extensive sampling, such as free energy computations. Implicit solvent models could provide the necessary speed-up due to reduced degrees of freedom and faster dynamics. Here, we introduce Solvation Free Energy Path Reweighting (ReSolv) framework parameterize an implicit ML potential organic molecules that accurately predicts hydration energy, essential parameter drug design pollutant modeling. Learning on combination experimental data vacuum, ReSolv bypasses need intractable explicit bulk does not have resort less accurate data-generating models. On FreeSolv dataset, achieves mean absolute error close average uncertainty, significantly outperforming standard force fields. Compared potential, offers speedup four orders magnitude attains closer agreement with experiments. The presented paves way deep more yet computationally cost-effective than classical atomistic

Language: Английский

---