Latent Diffusion-Based 3D Molecular Recovery from Vibrational Spectra

Wenjin Wu, Aleš Leonardis, Linjiang Chen, Jianbo Jiao

TL;DR

IR-GeoDiff is the first latent diffusion model designed to reconstruct full 3D molecular geometries directly from 1D infrared (IR) spectra. By integrating spectral features into an equivariant diffusion framework, the model bridges the gap between vibrational patterns and spatial atomic arrangements, achieving over 95% molecular accuracy on standard benchmarks.

What Problem Does This Paper Address?

For decades, chemists have used Infrared (IR) spectroscopy to "fingerprint" molecules. IR spectra are generated by molecular vibrations; specific peaks correspond to the stretching or bending of chemical bonds. However, interpreting these spectra is notoriously difficult. While humans rely on empirical rules and database lookups, previous AI attempts have mostly focused on predicting 1D SMILES strings or 2D molecular graphs.

The fundamental issue is a representation gap. IR spectra are inherently 3D phenomena—they arise from the spatial movement of atoms in three-dimensional space. By ignoring the 3D geometry, previous models lost the intricate relationship between spectral features and the actual physical structure of the molecule. This paper tackles the "inverse problem": given a 1D vibrational spectrum, can we recover the distribution of 3D molecular shapes that could have produced it?

Methods

The authors introduce IR-GeoDiff, a Latent Diffusion Model (LDM) tailored for molecular geometry. The architecture consists of several sophisticated components:

  1. Equivariant Backbone: The model uses Equivariant Graph Neural Networks (EGNNs). This ensures that the physics of the molecule remains consistent; the model recognizes that a molecule is the same regardless of how it is rotated or translated in 3D space (SE(3)-equivariance).
  2. Latent Diffusion: Instead of diffusing raw coordinates, the model operates in a compressed latent space. This provides better control over the generation process and improves computational efficiency.
  3. Cross-Attention Conditioning: This is the core innovation. The model doesn't just "see" the spectrum as a flat vector. It uses a Transformer-based spectral classifier to extract features, which are then injected into the diffusion process via cross-attention. Crucially, these spectral features are linked to both nodes (atoms) and edges (bonds), mimicking how specific vibrations are physically localized within a molecule.
  4. Two-Stage Training: First, a spectral classifier is trained to identify functional groups. Then, the diffusion model is trained to reconstruct geometries, using the frozen spectral features as a guide.

Datasets

The study utilizes two primary datasets that pair 3D coordinates with simulated IR spectra:

  • QM9S: A dataset of ~130,000 small organic molecules (up to 9 heavy atoms) composed of H, C, N, O, and F.
  • QMe14S: A more challenging and diverse dataset containing ~186,000 spectra of larger molecules, covering 14 different elements (including Si, P, S, Cl, etc.).

Key Insights and Ideas

  • Edges Matter: Because IR peaks represent bond vibrations, conditioning the edges of the molecular graph on the spectrum is more physically intuitive than just conditioning the atoms. The results show that this "edge-spectral" attention significantly reduces ambiguity in the structure recovery.
  • Learned Interpretability: The model’s attention maps reveal that it "looks" at the correct parts of the spectrum for specific groups. For example, when identifying a nitrile group ($C \equiv N$), the model focuses on the 2,250 $cm^{-1}$ region, exactly as a human chemist would.
  • Distribution vs. Single Point: The authors frame this as a distribution recovery task. Since a single spectrum could theoretically correspond to multiple conformers, the diffusion model is ideal because it learns the probabilistic distribution of valid geometries rather than just a single "average" structure.

Key Results

IR-GeoDiff sets a new state-of-the-art for spectrum-to-structure tasks:

  • Molecular Accuracy: On the QM9S dataset, IR-GeoDiff achieved 95.33% accuracy (matching the exact reference molecule), compared to just 44.47% for the previous best baseline (GEOLDM).
  • Spectral Fidelity: The model achieved a Spectral Information Similarity (SIS) of 0.675, significantly outperforming standard 3D generative models (which hovered around 0.41–0.48).
  • Generalization: Even on the larger, more complex QMe14S dataset, the model maintained a high molecular accuracy of 90.70%, proving it can handle diverse chemistry beyond simple organic molecules.

Limitations and Future Work

Despite its success, the authors note a few hurdles:

  • Conformational Sensitivity: Subtle 3D changes, like intramolecular hydrogen bonding, can shift IR peaks. The model currently has limited control over these very fine-grained conformational details.
  • Intrinsic Ambiguity: Some distinct molecular scaffolds produce nearly identical IR spectra (especially in the "fingerprint" region). The authors suggest that IR data alone might not always be enough for 100% certainty.
  • Multi-Modal Future: A promising next step is integrating other spectroscopic data, such as NMR (Nuclear Magnetic Resonance), to provide complementary information about the molecular backbone and further resolve structural ambiguities.

Summary updated

Abstract

Infrared (IR) spectroscopy, a type of vibrational spectroscopy, is widely used for molecular structure determination and provides critical structural information for chemists. However, existing approaches for recovering molecular structures from IR spectra typically rely on one-dimensional SMILES strings or two-dimensional molecular graphs, which fail to capture the intricate relationship between spectral features and three-dimensional molecular geometry. Recent advances in diffusion models have greatly enhanced the ability to generate molecular structures in 3D space. Yet, no existing model has explored the distribution of 3D molecular geometries corresponding to a single IR spectrum. In this work, we introduce IR-GeoDiff, a latent diffusion model that recovers 3D molecular geometries from IR spectra by integrating spectral information into both node and edge representations of molecular structures. We evaluate IR-GeoDiff from both spectral and structural perspectives, demonstrating its ability to recover the molecular distribution corresponding to a given IR spectrum. Furthermore, an attention-based analysis reveals that the model is able to focus on characteristic functional group regions in IR spectra, qualitatively consistent with common chemical interpretation practices.
BibTeX
@misc{wu2026latent,
  author = {Wenjin Wu and Aleš Leonardis and Linjiang Chen and Jianbo Jiao},
  title = {{Latent Diffusion-Based 3D Molecular Recovery from Vibrational Spectra}},
  year = {2026},
  eprint = {2603.06113},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2603.06113},
}
Figure