Machine learning heralding a new development phase in molecular dynamics simulations

Artificial Intelligence Review, Journal Year: 2024, Volume and Issue: 57(4)

Published: March 29, 2024

Abstract Molecular dynamics (MD) simulations are a key computational chemistry technique that provide dynamic insight into the underlying atomic-level processes in system under study. These insights not only improve our understanding of molecular world, but also aid design experiments and targeted interventions. Currently, MD is associated with several limitations, most important which are: insufficient sampling, inadequate accuracy atomistic models, challenges proper analysis interpretation obtained trajectories. Although numerous efforts have been made to address these more effective solutions still needed. The recent development artificial intelligence, particularly machine learning (ML), offers exciting opportunities MD. In this review we aim familiarize readers basics while highlighting its limitations. main focus on exploring integration deep simulations. advancements by ML systematically outlined, including ML-based force fields, techniques for improved conformational space innovative methods trajectory analysis. Additionally, implications intelligence discussed. While potential ML-MD fusion clearly established, further applications needed confirm superiority over traditional methods. This comprehensive overview new perspectives MD, has opened up, serves as gentle introduction phase development.

Language: Английский

---

AIMNet2: A Neural Network Potential to Meet your Neutral, Charged, Organic, and Elemental-Organic Needs

Published: Oct. 12, 2023

Machine learned interatomic potentials (MLIPs) are reshaping computational chemistry practices because of their ability to drastically exceed the accuracy-length/time scale tradeoff. Despite this attraction, benefits such efficiency only impactful when an MLIP uniquely enables insight into a target system or is broadly transferable outside training dataset, where models achieving latter seldom reported. In work, we present 2nd generation our atoms-in-molecules neural network potential (AIMNet2), which applicable species composed up 14 chemical elements in both neutral and charged states, making it valuable model for modeling majority non-metallic compounds. Using exhaustive dataset 20 million hybrid quantum calculations, AIMNet2 combines ML-parameterized short-range physics-based long-range terms attain generalizability that reaches from simple organics diverse molecules with “exotic” element-organic bonding. We show outperforms semi-empirical GFN-xTB on par reference density functional theory interaction energy contributions, conformer search tasks, torsion rotation profiles, molecular-to-macromolecular geometry optimization. Overall, demonstrated coverage significant step toward providing access MLIPs avoid crucial limitation curating additional data retraining each new application.

Language: Английский

---

Data Generation for Machine Learning Interatomic Potentials and Beyond

Chemical Reviews, Journal Year: 2024, Volume and Issue: 124(24), P. 13681 - 13714

Published: Nov. 21, 2024

The field of data-driven chemistry is undergoing an evolution, driven by innovations in machine learning models for predicting molecular properties and behavior. Recent strides ML-based interatomic potentials have paved the way accurate modeling diverse chemical structural at atomic level. key determinant defining MLIP reliability remains quality training data. A paramount challenge lies constructing sets that capture specific domains vast space. This Review navigates intricate landscape essential components integrity data ensure extensibility transferability resulting models. We delve into details active learning, discussing its various facets implementations. outline different types uncertainty quantification applied to atomistic acquisition correlations between estimated true error. role samplers generating informative structures highlighted. Furthermore, we discuss via modified surrogate potential energy surfaces as innovative approach diversify also provides a list publicly available cover

Language: Английский

---

The Potential of Neural Network Potentials

ACS Physical Chemistry Au, Journal Year: 2024, Volume and Issue: 4(3), P. 232 - 241

Published: March 21, 2024

In the next half-century, physical chemistry will likely undergo a profound transformation, driven predominantly by combination of recent advances in quantum and machine learning (ML). Specifically, equivariant neural network potentials (NNPs) are breakthrough new tool that already enabling us to simulate systems at molecular scale with unprecedented accuracy speed, relying on nothing but fundamental laws. The continued development this approach realize Paul Dirac's 80-year-old vision using mechanics unify physics providing invaluable tools for understanding materials science, biology, earth sciences, beyond. era highly accurate efficient first-principles simulations provide wealth training data can be used build automated computational methodologies, such as diffusion models, design optimization scale. Large language models (LLMs) also evolve into increasingly indispensable literature review, coding, idea generation, scientific writing.

Language: Английский

---

Analyzing Atomic Interactions in Molecules as Learned by Neural Networks

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

Published: Jan. 10, 2025

While machine learning (ML) models have been able to achieve unprecedented accuracies across various prediction tasks in quantum chemistry, it is now apparent that accuracy on a test set alone not guarantee for robust chemical modeling such as stable molecular dynamics (MD). To go beyond accuracy, we use explainable artificial intelligence (XAI) techniques develop general analysis framework atomic interactions and apply the SchNet PaiNN neural network models. We compare these with of fundamental principles understand how well learned underlying physicochemical concepts from data. focus strength different species, predictions intensive extensive properties are made, analyze decay many-body nature interatomic distance. Models deviate too far known physical produce unstable MD trajectories, even when they very high energy force accuracy. also suggest further improvements ML architectures better account polynomial interactions.

Language: Английский

---

Accelerating Fourth-Generation Machine Learning Potentials Using Quasi-Linear Scaling Particle Mesh Charge Equilibration

Journal of Chemical Theory and Computation, Journal Year: 2024, Volume and Issue: unknown

Published: Aug. 16, 2024

Machine learning potentials (MLPs) have revolutionized the field of atomistic simulations by describing atomic interactions with accuracy electronic structure methods at a small fraction cost. Most current MLPs construct energy system as sum energies, which depend on information about environments provided in form predefined or learnable feature vectors. If, addition, nonlocal phenomena like long-range charge transfer are important, fourth-generation need to be used, include equilibration (Qeq) step take global into account. This Qeq can significantly increase computational cost and thus become bottleneck for large systems. In this Article, we present highly efficient formulation that does not require explicit computation Coulomb matrix elements, resulting quasi-linear scaling method. Moreover, our approach also allows calculation derivatives, explicitly consider structure-dependence charges obtained from Qeq. Due its generality, method is restricted applied within variety other force fields.

Language: Английский

---

Molecular Gas-Phase Conformational Ensembles

Journal of Chemical Information and Modeling, Journal Year: 2024, Volume and Issue: 64(3), P. 749 - 760

Published: Jan. 11, 2024

Accurately determining the global minima of a molecular structure is important in diverse scientific fields, including drug design, materials science, and chemical synthesis. Conformational search engines serve as valuable tools for exploring extensive conformational space molecules identifying energetically favorable conformations. In this study, we present comparison Auto3D, CREST, Balloon, ETKDG (from RDKit), which are freely available engines, to evaluate their effectiveness locating minima. These employ distinct methodologies, machine learning (ML) potential-based, semiempirical, force field-based approaches. To validate these methods, propose use collisional cross-section (CCS) values obtained from ion mobility–mass spectrometry studies. We hypothesize that experimental gas-phase CCS can provide evidence likely have minimum given molecule. facilitate effort, used our conformation library (GPCL) currently consists full ensembles 20 small be by community any engine. Further members GPCL readily created molecule interest using standard workflow compute values, expanding ability validation exercises. innovative techniques enhance understanding landscape insights into performance generation engines. Our findings shed light on strengths limitations each engine, enabling informed decisions utilization various where accurate determination crucial biological activity designing targeted interventions. By facilitating identification reliable conformations, study significantly contributes enhancing efficiency accuracy determination, with particular focus metabolite elucidation. The research also developing effective workflows predicting structures unknown compounds high precision.

Language: Английский

---

Graph to Activation Energy Models Easily Reach Irreducible Errors but Show Limited Transferability

The Journal of Physical Chemistry A, Journal Year: 2024, Volume and Issue: 128(13), P. 2543 - 2555

Published: March 22, 2024

Activation energy characterization of competing reactions is a costly but crucial step for understanding the kinetic relevance distinct reaction pathways, product yields, and myriad other properties reacting systems. The standard methodology activation has historically been transition state search using highest level theory that can be afforded. However, recently, several groups have popularized idea predicting energies directly based on nothing more than reactant graphs, sufficiently complex neural network, broad enough data set. Here, we revisited this task recently developed Reaction Graph Depth 1 (RGD1) set newly graph attention architectures. All these new architectures achieve similar state-of-the-art results ∼4 kcal/mol mean absolute error withheld testing sets poor performance external composed with differing mechanisms, molecularity, or size distribution. Limited transferability also shown to shared by contemporary through series case studies. We conclude an array already comparable irreducible available out-of-distribution remains poor.

Language: Английский

---

QupKake: Integrating Machine Learning and Quantum Chemistry for Micro-pK_a Predictions

Journal of Chemical Theory and Computation, Journal Year: 2024, Volume and Issue: 20(15), P. 6946 - 6956

Published: June 4, 2024

Accurate prediction of micro-pKa values is crucial for understanding and modulating the acidity basicity organic molecules, with applications in drug discovery, materials science, environmental chemistry. This work introduces QupKake, a novel method that combines graph neural network models semiempirical quantum mechanical (QM) features to achieve exceptional accuracy generalization prediction. QupKake outperforms state-of-the-art on variety benchmark data sets, root-mean-square errors between 0.5 0.8 pKa units five external test sets. Feature importance analysis reveals role QM both reaction site enumeration models. represents significant advancement prediction, offering powerful tool various chemistry beyond.

Language: Английский

---

Collaboration on Machine-Learned Potentials with IPSuite: A Modular Framework for Learning-on-the-Fly

The Journal of Physical Chemistry B, Journal Year: 2024, Volume and Issue: 128(15), P. 3662 - 3676

Published: April 3, 2024

The field of machine learning potentials has experienced a rapid surge in progress, thanks to advances theory, algorithms, and hardware capabilities. While the underlying methods are continuously evolving, infrastructure for their deployment lagged. community, due these developments, frequently finds itself split into groups built around different implementations machine-learned potentials. In this work, we introduce

Language: Английский

---