Skip to content
DraconiticResearch
Search Sources API
197,768 studies31 sourcesUpdated 42 min ago
PreprintCC BY-NC-NDChemRxiv

From Vacuum Geometry to Solution-Phase Electronic Properties and Thermodynamics via Equivariant Neural Networks

  • Mathias Hilfiker1,2,3,4
  • Lorenzo D’Amore5
  • Dries Van Rompaey6
  • Marco Klähn1,2
  • Ola Engkvist1,2,7,8
  • Alexandre Tkatchenko3,4
  • Leonardo Medrano Sandonas9,10,11,12
12 affiliations
  1. 1Molecular AI, Discovery Sciences
  2. 2R&D, AstraZeneca
  3. 3Department of Physics and Materials Science
  4. 4University of Luxembourg
  5. 5Computational Chemistry, Janssen R&D, C/ Jarama 75A, Toledo, Spain, 45007
  6. 6Drug Discovery Data Sciences, Janssen Pharmaceutica NV, Turnhoutseweg 30, Beerse, Belgium, 2340
  7. 7Department of Computer Science and Engineering
  8. 8Chalmers University of Technology and University of Gothenburg
  9. 9Institute for Materials Science and Max Bergmann Center of Biomaterials
  10. 10TUD Dresden University of Technology
  11. 11Center for Advanced Systems Understanding (CASUS), Conrad-Schiedt-Straße 20, Görlitz, Germany, 02826
  12. 12Helmholtz Zentrum Dresden-Rossendorf (HZDR), Bautzner Landstraße 400, Dresden, Germany, 01328
Open at sourcePDFDOICite

Abstract

We introduce a general framework for predicting solution-phase quantum-mechanical (QM) properties of drug-like molecules directly from gas-phase geometries, without solvent-specific structural optimization or property calculations. Using current equivariant neural network (ENN) architectures, such as MACE and Allegro, we demonstrate that they can also predict extensive and intensive QM properties in implicit water at a high level of theory. Direct-and ∆-learning strategies are employed for this prediction task using the Aquamarine dataset, which provides QM properties for the same molecules in both environments. The developed models are systematically evaluated for scalability to larger molecules, transferability across chemical space, and generalization to out-of-equilibrium configurations. Our results show that, despite substantial differences between gas-and solution-phase conformational ensembles, the models can accurately reconstruct ensemble-averaged solution-phase properties from gas-phase molecular dynamics. We further integrate the learned QM properties into the QUEDsol framework (“QUantum Electronic Descriptors in SOLution”) as electronic descriptors for hydration free-energy prediction. A simple and interpretable quadratic regression model trained on the FreeSolv dataset achieves state-of-the-art performance on the challenging and chemically diverse FlexiSol benchmark, outperforming previously reported QM-and ML-based approaches. These results establish ENNs as a versatile approach for inter-phase molecular property prediction and demonstrate that environment-specific QM descriptors generated from gas-phase geometries can provide transferable inputs for complex thermodynamic modeling.

Related

  1. Geometry and electronic properties of chiral phosphorene nanotubes
    Felipe Cervantes-Sodi, Carlos Delmote, José Ángel Reyes-Retana
    arXivPreprintarXiv licenceFull textCondensed matterarXiv
  2. Equivariant Deep Learning in Direct-Product Space for Electronic-Structure Hamiltonian Prediction
    Shi Yin, Zujian Dai, Xinyang Pan +1
    arXivPreprintCC BY 4.0Full textComputational physicsCondensed matterarXiv
  3. Wasserstein-2 gradient flows and the geometry of entropy production in classical and quantum stochastic thermodynamics
    Olga Movilla Miangolarra, Ralph Sabbagh, Artemy Kolchinsky
    Physical Review ResearchPeer-reviewedPublisher termsStatistical & nonlinear physicsCrossref
  4. Emergent Loop Geometry in Gauge-Equivariant Graph Neural Networks
    Ali Rayat, Gia-Wei Chern
    arXivPreprintarXiv licenceFull textCondensed matterHigh-energy & particle physicsarXiv
  5. Charged Deformation of a Hairy Black Brane: Perturbative Solution and Thermodynamics
    Sangheon Yun
    arXivPreprintarXiv licenceFull textGravitation & general relativityHigh-energy & particle physicsarXiv

Cite

BibTeXRIS
BibTeX
@misc{hilfiker2026from,
  title = {{From Vacuum Geometry to Solution-Phase Electronic Properties and Thermodynamics via Equivariant Neural Networks}},
  author = {Mathias Hilfiker and Lorenzo D’Amore and Dries Van Rompaey and Marco Klähn and Ola Engkvist and Alexandre Tkatchenko and Leonardo Medrano Sandonas},
  howpublished = {ChemRxiv},
  year = {2026},
  month = oct,
  doi = {10.26434/chemrxiv.15010245/v1},
  url = {https://chemrxiv.org/doi/full/10.26434/chemrxiv.15010245/v1},
  note = {CC BY-NC-ND},
}
RIS
TY  - UNPB
TI  - From Vacuum Geometry to Solution-Phase Electronic Properties and Thermodynamics via Equivariant Neural Networks
AU  - Mathias Hilfiker
AU  - Lorenzo D’Amore
AU  - Dries Van Rompaey
AU  - Marco Klähn
AU  - Ola Engkvist
AU  - Alexandre Tkatchenko
AU  - Leonardo Medrano Sandonas
PY  - 2026
DA  - 2026/10/11
PB  - ChemRxiv
DO  - 10.26434/chemrxiv.15010245/v1
UR  - https://chemrxiv.org/doi/full/10.26434/chemrxiv.15010245/v1
AB  - We introduce a general framework for predicting solution-phase quantum-mechanical (QM) properties of drug-like molecules directly from gas-phase geometries, without solvent-specific structural optimization or property calculations. Using current equivariant neural network (ENN) architectures, such as MACE and Allegro, we demonstrate that they can also predict extensive and intensive QM properties in implicit water at a high level of theory. Direct-and ∆-learning strategies are employed for this prediction task using the Aquamarine dataset, which provides QM properties for the same molecules in both environments. The developed models are systematically evaluated for scalability to larger molecules, transferability across chemical space, and generalization to out-of-equilibrium configurations. Our results show that, despite substantial differences between gas-and solution-phase conformational ensembles, the models can accurately reconstruct ensemble-averaged solution-phase properties from gas-phase molecular dynamics. We further integrate the learned QM properties into the QUEDsol framework (“QUantum Electronic Descriptors in SOLution”) as electronic descriptors for hydration free-energy prediction. A simple and interpretable quadratic regression model trained on the FreeSolv dataset achieves state-of-the-art performance on the challenging and chemically diverse FlexiSol benchmark, outperforming previously reported QM-and ML-based approaches. These results establish ENNs as a versatile approach for inter-phase molecular property prediction and demonstrate that environment-specific QM descriptors generated from gas-phase geometries can provide transferable inputs for complex thermodynamic modeling.
ER  -