![[2602.03875] Reversible Deep Learning for 13C NMR in Chemoinformatics: On Structures and Spectra](https://arxiv.org/static/browse/0.3.4/images/arxiv-logo-fb.png){kind=logo}
[2602.03875] Reversible Deep Learning for 13C NMR in Chemoinformatics: On Structures and Spectra
Summary
This article presents a reversible deep learning model for 13C NMR in chemoinformatics, utilizing an invertible neural network to predict molecular structures from spectral data and vice versa.
Why It Matters
The research addresses the challenge of accurately predicting molecular structures from NMR spectra, a common task in chemoinformatics. By introducing a reversible deep learning approach, it enhances the understanding of molecular representation and offers a unified model for spectrum prediction and candidate generation, which could significantly impact drug discovery and materials science.
Key Takeaways
- Introduces a reversible deep learning model for 13C NMR.
- Utilizes a single conditional invertible neural network for dual predictions.
- Achieves above-chance spectrum-code prediction and meaningful structural signals.
- Demonstrates the potential of invertible architectures in chemoinformatics.
- Offers a unified approach for spectrum prediction and uncertainty-aware generation.
Computer Science > Machine Learning arXiv:2602.03875 (cs) [Submitted on 1 Feb 2026 (v1), last revised 20 Feb 2026 (this version, v3)] Title:Reversible Deep Learning for 13C NMR in Chemoinformatics: On Structures and Spectra Authors:Stefan Kuhn, Vandana Dwarka, Przemyslaw Karol Grenda, Eero Vainikko View a PDF of the paper titled Reversible Deep Learning for 13C NMR in Chemoinformatics: On Structures and Spectra, by Stefan Kuhn and Vandana Dwarka and Przemyslaw Karol Grenda and Eero Vainikko View PDF Abstract:We introduce a reversible deep learning model for 13C NMR that uses a single conditional invertible neural network for both directions between molecular structures and spectra. The network is built from i-RevNet style bijective blocks, so the forward map and its inverse are available by construction. We train the model to predict a 128-bit binned spectrum code from a graph-based structure encoding, while the remaining latent dimensions capture residual variability. At inference time, we invert the same trained network to generate structure candidates from a spectrum code, which explicitly represents the one-to-many nature of spectrum-to-structure inference. On a filtered subset, the model is numerically invertible on trained examples, achieves spectrum-code prediction above chance, and produces coarse but meaningful structural signals when inverted on validation spectra. These results demonstrate that invertible architectures can unify spectrum prediction and uncertain...
