Multimodal Spectroscopy–Driven Reconstruction of Three-Dimensional Molecular Structures via Conditional Diffusion Model
9 affiliations
- 1School of Software Engineering
- 2University of Science and Technology of China
- 3Suzhou Institute for Advanced Research, University of Science and Technology of China
- 4Key Laboratory of Precision and Intelligent Chemistry
- 5School of Chemistry and Materials Science
- 6Department of Chemical Engineering and Applied Chemistry
- 7University of Toronto
- 8Department of Chemistry
- 9Vector Institute for Artificial Intelligence
Abstract
Three-dimensional (3D) molecular structures underpin molecular function and dynamics. Spectroscopy can encode rich structural information via energy-level transitions, but existing data-driven approaches have largely focused on predicting symbolic representations such as SMILES, or molecular topology rather than reconstructing full 3D structures. Here, we introduce SpecDiffMol, the first generative framework to directly reconstruct full molecular 3D structures from molecular formulae and multimodal spectroscopic data, including infrared, Raman, UV–vis, and NMR spectra. SpecDiffMol integrates a Transformer-based topology inference module with an E(3)-equivariant conditional diffusion model, enabling the inference of molecular connectivity, global structure, and fine-grained conformational geometry in a unified framework. Extensive experiments on large-scale theoretical spectroscopic benchmarks comprising about 400,000 molecules systematically elucidate the complementary roles of individual spectroscopic modalities in determining molecular connectivity, backbone topology, and detailed 3D conformations. By explicitly incorporating spectroscopically informed topological priors, the framework substantially constrains the chemical search space and enables robust structure generation. Meanwhile, SpecDiffMol remains effective under realistic experimental conditions. Using automated experimental spectroscopic measurements, we demonstrate its ability to track Z↔E isomerization dynamics in bridged azobenzenes, capturing conformational transitions and key dihedral angle evolution with high accuracy. As a general and extensible approach for spectroscopically driven 3D molecular structure generation, SpecDiffMol opens new avenues for autonomous structure–function discovery and real-time in situ molecular analysis across chemistry, materials science, and molecular biology.
Related
- Borosulfates Hosting Se4 2+ Ions—Crystal Structures, Chemical Bonding, Optical Spectroscopy, and Magnetism
- Solution-processed Ti3C2Tx MXene coatings for three-dimensional microelectrodes in label-free single-cell impedance spectroscopy
- Pushing the Limits of One-Dimensional NMR Spectroscopy for Automated Structure Elucidation Using Artificial Intelligence
- Model-free Reconstruction of Molecular Energy Levels by Broadband Kilohertz-accurate Cavity-enhanced Spectroscopy
- GNN-ωPBE: Range-Separation Parameters Learned from Three-Dimensional Structures
BibTeX
@misc{du2026multimodal,
title = {{Multimodal Spectroscopy–Driven Reconstruction of Three-Dimensional Molecular Structures via Conditional Diffusion Model}},
author = {Wenjie Du and Xiaohan Qin and Yixi Shen and Ledu Wang and Andreas Burger and Chuangwei Liu and Song Wang and Jun Jiang and Yang Wang and Eric Liu},
howpublished = {ChemRxiv},
year = {2026},
month = oct,
doi = {10.26434/chemrxiv.15011037/v1},
url = {https://chemrxiv.org/doi/full/10.26434/chemrxiv.15011037/v1},
note = {CC BY},
}
RIS
TY - UNPB
TI - Multimodal Spectroscopy–Driven Reconstruction of Three-Dimensional Molecular Structures via Conditional Diffusion Model
AU - Wenjie Du
AU - Xiaohan Qin
AU - Yixi Shen
AU - Ledu Wang
AU - Andreas Burger
AU - Chuangwei Liu
AU - Song Wang
AU - Jun Jiang
AU - Yang Wang
AU - Eric Liu
PY - 2026
DA - 2026/10/26
PB - ChemRxiv
DO - 10.26434/chemrxiv.15011037/v1
UR - https://chemrxiv.org/doi/full/10.26434/chemrxiv.15011037/v1
AB - Three-dimensional (3D) molecular structures underpin molecular function and dynamics. Spectroscopy can encode rich structural information via energy-level transitions, but existing data-driven approaches have largely focused on predicting symbolic representations such as SMILES, or molecular topology rather than reconstructing full 3D structures. Here, we introduce SpecDiffMol, the first generative framework to directly reconstruct full molecular 3D structures from molecular formulae and multimodal spectroscopic data, including infrared, Raman, UV–vis, and NMR spectra. SpecDiffMol integrates a Transformer-based topology inference module with an E(3)-equivariant conditional diffusion model, enabling the inference of molecular connectivity, global structure, and fine-grained conformational geometry in a unified framework. Extensive experiments on large-scale theoretical spectroscopic benchmarks comprising about 400,000 molecules systematically elucidate the complementary roles of individual spectroscopic modalities in determining molecular connectivity, backbone topology, and detailed 3D conformations. By explicitly incorporating spectroscopically informed topological priors, the framework substantially constrains the chemical search space and enables robust structure generation. Meanwhile, SpecDiffMol remains effective under realistic experimental conditions. Using automated experimental spectroscopic measurements, we demonstrate its ability to track Z↔E isomerization dynamics in bridged azobenzenes, capturing conformational transitions and key dihedral angle evolution with high accuracy. As a general and extensible approach for spectroscopically driven 3D molecular structure generation, SpecDiffMol opens new avenues for autonomous structure–function discovery and real-time in situ molecular analysis across chemistry, materials science, and molecular biology.
ER -