3DReact: Geometric Deep Learning for Chemical Reactions DOI Creative Commons
Puck van Gerwen, Ksenia R. Briling, Charlotte Bunne

et al.

Journal of Chemical Information and Modeling, Journal Year: 2024, Volume and Issue: 64(15), P. 5771 - 5785

Published: July 15, 2024

Geometric deep learning models, which incorporate the relevant molecular symmetries within neural network architecture, have considerably improved accuracy and data efficiency of predictions properties. Building on this success, we introduce 3DReact, a geometric model to predict reaction properties from three-dimensional structures reactants products. We demonstrate that invariant version is sufficient for existing sets. illustrate its competitive performance prediction activation barriers GDB7-22-TS, Cyclo-23-TS, Proparg-21-TS sets in different atom-mapping regimes. show that, compared models property prediction, 3DReact offers flexible framework exploits information, if available, as well geometries products (in an or equivariant fashion). Accordingly, it performs systematically across sets, regimes, both interpolation extrapolation tasks.

Language: Английский

Data-Driven Workflow for the Development and Discovery of N-Oxyl Hydrogen Atom Transfer Catalysts DOI Creative Commons
Cheng Yang,

Thérèse Wild,

Yulia Rakova

et al.

ACS Central Science, Journal Year: 2025, Volume and Issue: unknown

Published: March 24, 2025

N-oxyl species are promising hydrogen atom transfer (HAT) catalysts to advance C–H bond activation reactions. However, because of the complex structure–activity relationship within structure, catalyst optimization is a key challenge, particularly for simultaneous improvement across multiple parameters. This paper describes data-driven approach optimize catalysts. A focused library 50 N-hydroxy compounds was synthesized and characterized by three parameters─oxidation peak potential, HAT reactivity, stability─to generate database. Statistical modeling these activities described their intrinsic physical organic parameters used build predictive models discovery understand relationships. Virtual screening 102 synthesizable candidates allowed rapid identification several ideal candidates. These statistical clearly suggest that substructures bearing an adjacent heteroatom more optimal compared historical focus, phthalimide-N-oxyl, striking best balance among all target experimental properties.

Language: Английский

Citations

0
Enhancing the Predictive Kinetics of Intramolecular H-Migration Reactions of Ether Peroxy Radicals by Integrating Machine Learning with Quantum Chemistry: A Comparative Study of Generic Rate Rules and Machine Learning Techniques DOI
Jingwei Zhang, Siyu Chen, Haisheng Ren

et al.

Industrial & Engineering Chemistry Research, Journal Year: 2025, Volume and Issue: unknown

Published: May 9, 2025

Language: Английский

Citations

0
3DReact: Geometric Deep Learning for Chemical Reactions DOI Creative Commons
Puck van Gerwen, Ksenia R. Briling, Charlotte Bunne

et al.

Journal of Chemical Information and Modeling, Journal Year: 2024, Volume and Issue: 64(15), P. 5771 - 5785

Published: July 15, 2024

Geometric deep learning models, which incorporate the relevant molecular symmetries within neural network architecture, have considerably improved accuracy and data efficiency of predictions properties. Building on this success, we introduce 3DReact, a geometric model to predict reaction properties from three-dimensional structures reactants products. We demonstrate that invariant version is sufficient for existing sets. illustrate its competitive performance prediction activation barriers GDB7-22-TS, Cyclo-23-TS, Proparg-21-TS sets in different atom-mapping regimes. show that, compared models property prediction, 3DReact offers flexible framework exploits information, if available, as well geometries products (in an or equivariant fashion). Accordingly, it performs systematically across sets, regimes, both interpolation extrapolation tasks.

Language: Английский

Citations

3