Context <p>The elucidation of molecular structures and properties typically relies on the integration of multiple spectroscopic techniques, which provide rich structural information from complementary perspectives. Although the synergistic analysis of multi-modal spectra can significantly enhance the accuracy of structural identification, its practical application is severely constrained by the scarcity of high-quality experimental spectral data. In particular, comprehensive datasets encompassing multiple spectroscopic modalities for the same molecule are exceedingly rare. To overcome this data bottleneck, we developed MolSpectra, a universal deep learning framework capable of predicting multi-modal spectra directly from molecular SMILES representations. MolSpectra provides a unified molecular representation architecture, enabling the synergistic input of experimental metadata, such as physical phase states, alongside molecular structural information into an end-to-end model. This framework eliminates the need to modify the core model architecture; it can predict various types of spectra simply by adjusting configuration files. We systematically evaluated MolSpectra on six datasets covering four spectroscopic techniques. These include experimental infrared (IR) spectral datasets from the NIST and Chemotion databases, a simulated IR spectral benchmark dataset based on the USPTO reaction corpus, an ultraviolet–visible (UV–Vis) spectral dataset from the NIST database, an electron ionization mass spectrometry (EI-MS) dataset from the commercial NIST 23 database, and nuclear magnetic resonance (NMR) datasets from nmrshiftdb2 and HMDB (Human Metabolome Database) for chemical shift and full-spectrum predictions, respectively. Experimental results demonstrate that MolSpectra comprehensively outperforms existing baseline models: under strict InChIKey-based splitting, it achieves maximum cosine similarities of 0.916 for IR (NIST experimental dataset) and 0.861 for UV–Vis (NIST-UV dataset) predictions; for EI-MS prediction, it reaches a maximum cosine similarity of 0.880 (NIST 23 dataset) and a Top-1 accuracy of 0.530 in spectral library matching. Regarding NMR, the framework achieves mean absolute errors (MAEs) as low as 0.153 ppm and 1.176 ppm for <InlineEquation ID="IEq1"> <EquationSource Format="TEX">$^1$</EquationSource> <EquationSource Format="MATHML"><math> <mmultiscripts> <mrow /> <mrow /> <mn>1</mn> </mmultiscripts> </math></EquationSource> </InlineEquation>H/<InlineEquation ID="IEq2"> <EquationSource Format="TEX">$^{13}$</EquationSource> <EquationSource Format="MATHML"><math> <mmultiscripts> <mrow /> <mrow /> <mn>13</mn> </mmultiscripts> </math></EquationSource> </InlineEquation>C chemical shift predictions on the nmrshiftdb2 subset, respectively.</p> Methods <p>MolSpectra takes SMILES strings as input and constructs molecular graphs using RDKit. The model employs a Message Passing Graph Neural Network (MPGNN) equipped with Sequential Signal Mixing Aggregation (SSMA) to model local and medium-range atomic interactions. To capture long-range dependencies and local structural details, the model integrates a Hierarchical Distance Structural Encoding (HDSE) and a structurally biased attention mechanism within an enhanced Transformer architecture. Node features are updated iteratively, and flexible prediction heads are employed: a node-level prediction head is used for NMR chemical shifts, while graph-level prediction heads with permutation-invariant readout functions are utilized for molecular-level spectra such as IR, UV–Vis, and EI-MS. Furthermore, the framework supports the encoding of metadata, such as physical phase states, to accelerate model convergence and improve generalization capabilities. In MolSpectra, each spectroscopic modality is trained using an independent model instance that shares the same backbone architecture, with a modality-specific prediction head adapted to the corresponding spectral dimensionality. The source code is publicly available at <a href="https://github.com/lzjforyou/MolSpectra">https://github.com/lzjforyou/MolSpectra</a>.</p>

Molspectra: a general framework for multi-spectra prediction from molecular structures

期刊论文
发表日期：

2026年5月14日

查看全文

Zhangqiang Liu,
Congcong Yang,
Runquang Lai,
Weihua Li

摘要

Context

The elucidation of molecular structures and properties typically relies on the integration of multiple spectroscopic techniques, which provide rich structural information from complementary perspectives. Although the synergistic analysis of multi-modal spectra can significantly enhance the accuracy of structural identification, its practical application is severely constrained by the scarcity of high-quality experimental spectral data. In particular, comprehensive datasets encompassing multiple spectroscopic modalities for the same molecule are exceedingly rare. To overcome this data bottleneck, we developed MolSpectra, a universal deep learning framework capable of predicting multi-modal spectra directly from molecular SMILES representations. MolSpectra provides a unified molecular representation architecture, enabling the synergistic input of experimental metadata, such as physical phase states, alongside molecular structural information into an end-to-end model. This framework eliminates the need to modify the core model architecture; it can predict various types of spectra simply by adjusting configuration files. We systematically evaluated MolSpectra on six datasets covering four spectroscopic techniques. These include experimental infrared (IR) spectral datasets from the NIST and Chemotion databases, a simulated IR spectral benchmark dataset based on the USPTO reaction corpus, an ultraviolet–visible (UV–Vis) spectral dataset from the NIST database, an electron ionization mass spectrometry (EI-MS) dataset from the commercial NIST 23 database, and nuclear magnetic resonance (NMR) datasets from nmrshiftdb2 and HMDB (Human Metabolome Database) for chemical shift and full-spectrum predictions, respectively. Experimental results demonstrate that MolSpectra comprehensively outperforms existing baseline models: under strict InChIKey-based splitting, it achieves maximum cosine similarities of 0.916 for IR (NIST experimental dataset) and 0.861 for UV–Vis (NIST-UV dataset) predictions; for EI-MS prediction, it reaches a maximum cosine similarity of 0.880 (NIST 23 dataset) and a Top-1 accuracy of 0.530 in spectral library matching. Regarding NMR, the framework achieves mean absolute errors (MAEs) as low as 0.153 ppm and 1.176 ppm for $^1$ $^{1}$ H/ $^{13}$ $^{13}$ C chemical shift predictions on the nmrshiftdb2 subset, respectively.

Methods

MolSpectra takes SMILES strings as input and constructs molecular graphs using RDKit. The model employs a Message Passing Graph Neural Network (MPGNN) equipped with Sequential Signal Mixing Aggregation (SSMA) to model local and medium-range atomic interactions. To capture long-range dependencies and local structural details, the model integrates a Hierarchical Distance Structural Encoding (HDSE) and a structurally biased attention mechanism within an enhanced Transformer architecture. Node features are updated iteratively, and flexible prediction heads are employed: a node-level prediction head is used for NMR chemical shifts, while graph-level prediction heads with permutation-invariant readout functions are utilized for molecular-level spectra such as IR, UV–Vis, and EI-MS. Furthermore, the framework supports the encoding of metadata, such as physical phase states, to accelerate model convergence and improve generalization capabilities. In MolSpectra, each spectroscopic modality is trained using an independent model instance that shares the same backbone architecture, with a modality-specific prediction head adapted to the corresponding spectral dimensionality. The source code is publicly available at https://github.com/lzjforyou/MolSpectra.