Learning neural network potentials from experimental data via Differentiable Trajectory Reweighting

Nature Communications, Journal Year: 2021, Volume and Issue: 12(1)

Published: Nov. 25, 2021

In molecular dynamics (MD), neural network (NN) potentials trained bottom-up on quantum mechanical data have seen tremendous success recently. Top-down approaches that learn NN directly from experimental received less attention, typically facing numerical and computational challenges when backpropagating through MD simulations. We present the Differentiable Trajectory Reweighting (DiffTRe) method, which bypasses differentiation simulation for time-independent observables. Leveraging thermodynamic perturbation theory, we avoid exploding gradients achieve around 2 orders of magnitude speed-up in gradient computation top-down learning. show effectiveness DiffTRe learning an atomistic model diamond a coarse-grained water based diverse observables including thermodynamic, structural properties. Importantly, also generalizes coarse-graining methods such as iterative Boltzmann inversion to arbitrary potentials. The presented method constitutes important milestone towards enriching with data, particularly accurate is unavailable.

Language: Английский

---

A review of machine learning in geochemistry and cosmochemistry: Method improvements and applications

Applied Geochemistry, Journal Year: 2022, Volume and Issue: 140, P. 105273 - 105273

Published: March 23, 2022

Language: Английский

---

Graph neural networks accelerated molecular dynamics

The Journal of Chemical Physics, Journal Year: 2022, Volume and Issue: 156(14)

Published: April 13, 2022

Molecular Dynamics (MD) simulation is a powerful tool for understanding the dynamics and structure of matter. Since resolution MD atomic-scale, achieving long timescale simulations with femtosecond integration very expensive. In each step, numerous iterative computations are performed to calculate energy based on different types interaction their corresponding spatial gradients. These repetitive can be learned surrogated by deep learning model, such as Graph Neural Network (GNN). this work, we developed GNN Accelerated (GAMD) model that directly predicts forces, given state system (atom positions, atom types), bypassing evaluation potential energy. By training variety data sources (simulation derived from classical density functional theory), show GAMD predict two typical molecular systems, Lennard-Jones water system, in NVT ensemble velocities regulated thermostat. We further GAMD's inference agnostic scale, where it scale much larger systems at test time. also perform comprehensive benchmark comparing our implementation production-level software, showing competitive performance large-scale simulation.

Language: Английский

---

Accelerators for Classical Molecular Dynamics Simulations of Biomolecules

Journal of Chemical Theory and Computation, Journal Year: 2022, Volume and Issue: 18(7), P. 4047 - 4069

Published: June 16, 2022

Atomistic Molecular Dynamics (MD) simulations provide researchers the ability to model biomolecular structures such as proteins and their interactions with drug-like small molecules greater spatiotemporal resolution than is otherwise possible using experimental methods. MD are notoriously expensive computational endeavors that have traditionally required massive investment in specialized hardware access biologically relevant scales. Our goal summarize fundamental algorithms employed literature then highlight challenges affected accelerator implementations practice. We consider three broad categories of accelerators: Graphics Processing Units (GPUs), Field-Programmable Gate Arrays (FPGAs), Application Specific Integrated Circuits (ASICs). These comparatively studied facilitate discussion relative trade-offs gain context for current state art. conclude by providing insights into potential emerging platforms MD.

Language: Английский

---

Deep Learning Model for Efficient Protein–Ligand Docking with Implicit Side-Chain Flexibility

Journal of Chemical Information and Modeling, Journal Year: 2023, Volume and Issue: 63(6), P. 1695 - 1707

Published: March 14, 2023

Protein-ligand docking is an essential tool in structure-based drug design with applications ranging from virtual high-throughput screening to pose prediction for lead optimization. Most programs are optimized redocking existing cocrystallized protein structure, ignoring flexibility. In real-world applications, however, flexibility feature of the ligand-binding process. Flexible protein-ligand still remains a significant challenge computational design. To target this challenge, we present deep learning (DL) model flexible based on intermolecular Euclidean distance matrix (EDM), making typical use iterative search algorithms obsolete. The was trained large-scale data set complexes and evaluated independent test sets. Our generates high quality poses diverse ligand structures outperforms comparable methods.

Language: Английский

---