Latent Diffusion-based 3D Molecular Recovery from Vibrational Spectra

Wenjin Wu, Ales Leonardis, Linjiang Chen, Jianbo JiaoSubmission to ICLR 2026 · Rejected

TL;DR

IR-GeoDiff is the first latent diffusion framework designed to directly reconstruct 3D molecular geometries from 1D infrared (IR) vibrational spectra. By injecting spectral and functional group features into both atomic node and bond edge representations, the model achieves a 90.6% recovery rate on QM9S molecules while aligning with the underlying quantum mechanics of molecular vibrations.

What Problem Does This Paper Address?

Infrared (IR) spectroscopy is a cornerstone technique for determining molecular structure. Spectral absorption peaks correspond directly to the vibrational modes of chemical bonds (such as stretching or bending), making IR inherently linked to three-dimensional molecular geometry.

Despite this physical reality, computational tools attempting to invert IR spectra typically generate 1D SMILES strings or 2D molecular graphs. This representation gap strips away explicit 3D spatial information. Meanwhile, existing 3D molecular generative models (like equivariant diffusion models) are tuned to produce broad structural diversity, making them poorly suited for the inverse problem of recovering the specific, constrained set of 3D geometries consistent with an experimental spectrum.

Methods

IR-GeoDiff couples an $\mathrm{SE}(3)$-equivariant latent diffusion process with a specialized spectral conditioning mechanism:

  1. Spectral and Functional Group Encoding: A Transformer encoder processes patch-embedded 1D IR spectra alongside chemical formulas. A Transformer decoder queries these features to predict the presence of 20 predefined functional groups, producing both global spectral embeddings and functional group representations.
  2. Geometric Latent Autoencoder: Using Equivariant Graph Neural Networks (EGNNs), molecular 3D coordinates are mapped into a roto-translationally invariant latent space ($z_x$), conditioned on atomic embeddings ($z_h$).
  3. Dual Node and Edge Conditioning: Because vibrations directly influence interatomic distances, spectral signals are injected not only into invariant node features, but critically into pairwise edge representations ($z_e$) via multi-head cross-attention. Functional group representations are also cross-attended with edge features.
  4. Sampling and Bond Assignment: Given known atom counts and types, the model samples latent coordinates via reverse diffusion, decodes them to 3D Cartesian space, and assigns bonds through a valence-guided geometric algorithm.
1D IR Spectrum + Formula ──► Transformer Encoder ──► Spectral Features
                                       │
                                       ▼ (Cross-Attention)
Latent Noise z_T ───────────► EGNN Denoising Backbone ──► Latent z_0 ──► Reconstructed 3D Geometry
                               (Node & Edge Attention)

Datasets

  • QM9S: A curated dataset of ~130,000 small organic molecules (up to 9 heavy atoms of C, N, O, F, plus H) with IR spectra simulated via density functional theory (DFT at the B3LYP/def2-TZVP level using Gaussian 16).
  • Quality Screening: The authors filtered out 603 molecules containing inconsistencies between QM9S 3D coordinates and SMILES strings. Evaluations were performed on a standardized 1,000-molecule test split, using a verified 200-molecule subset for intensive DFT recalculations.

Key Insights and Ideas

  • Edges Matter Most for Vibrational Spectra: Standard conditioning injects global context only into node features. In IR-GeoDiff, injecting spectral signals directly into pairwise edge representations proved essential. Ablating edge conditioning dropped molecular accuracy from 90.6% down to 28.0%, mirroring the physical reality that IR absorption originates from bond vibrations between atom pairs.
  • Attention Aligns with Quantum Mechanics: Cross-attention maps between edge features and spectral patches reveal that the model autonomously attends to the correct vibrational frequencies (e.g., attending near ~2250–2355 cm⁻¹ for $\text{C}\equiv\text{N}$ stretching and ~3600–3817 cm⁻¹ for $\text{O–H}$ stretching), matching DFT normal-mode displacement vectors.
  • The Graph vs. Spectral Divergence: Molecules with high 2D topological graph similarity can produce low Spectral Information Similarity (SIS) due to 3D conformational shifts, such as intramolecular hydrogen bonding. Conversely, molecules with completely different carbon backbones can yield nearly identical IR spectra if their functional groups are the same.

Key Results

Evaluating 50 sampled structures per spectrum against the reference test set:

  • Molecular Accuracy: IR-GeoDiff reached 90.6% accuracy (identifying the exact reference structure), vastly surpassing adapted 3D generative baselines EDM (2.73%) and GeoLDM (11.9%).
  • Structural Similarity: Achieved an average chemical graph similarity ($\text{sim}_g$) of 0.672 (maximum $\text{sim}_g$ of 0.969), compared to 0.227 for GeoLDM.
  • Spectral Fidelity: Achieved an SIS score of 0.644 (max 0.931) and an SIS (restricted to the functional group region) of 0.681*, outperforming GeoLDM by +0.264 and +0.291, respectively.

Limitations and Future Work

  • Conformational Sensitivity: Minor conformational deviations—especially the formation or absence of intramolecular hydrogen bonds—shift vibrational frequencies and reduce spectral match despite near-identical graph topologies.
  • Backbone Ambiguity: Saturated hydrocarbon skeletons without distinct functional groups lack sharp IR features, making skeleton discrimination difficult using IR alone.
  • Scale: Evaluated on small molecules ($\le 9$ heavy atoms). Scaling to larger drug-like molecules or biomolecules remains an open challenge.
  • Multimodal Extension: The authors identify incorporating Nuclear Magnetic Resonance ($^1\text{H}$- and $^{13}\text{C}$-NMR) as the most promising next step, which would resolve skeletal ambiguities and impose stronger structural constraints on 3D geometry.

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 integrates IR 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’s interpretation of IR spectra aligns with quantum mechanical principles of molecular vibrations.
BibTeX
@misc{wu2026latent,
  author = {Wenjin Wu and Ales 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},
}

Loading related content...

Figure