arxiv: 2602.03875 · v4 · submitted 2026-02-01 · 💻 cs.LG · cs.AI· q-bio.QM

Recognition: no theorem link

Reversible Deep Learning for 13C NMR in Chemoinformatics: On Structures and Spectra

Stefan Kuhn , Vandana Dwarka , Przemyslaw Karol Grenda , Eero Vainikko

Authors on Pith no claims yet

Pith reviewed 2026-05-16 08:37 UTC · model grok-4.3

classification 💻 cs.LG cs.AIq-bio.QM

keywords invertible neural networks13C NMRchemoinformaticsspectrum predictionstructure generationreversible architecturesmolecular graphs

0 comments

The pith

A single invertible neural network maps molecular structures to 13C NMR spectra and generates candidate structures from spectra by exact inversion.

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The paper introduces a reversible deep learning model built from conditional invertible blocks that performs both spectrum prediction and structure generation inside one network. The forward direction encodes a graph representation of a molecule into a 128-bit binned spectrum code plus latent residuals; the inverse direction recovers structure candidates directly from that code. Because the architecture is bijective by construction, inversion is available without any separate decoder or search procedure. On held-out validation spectra the inverted outputs carry coarse but detectable structural information rather than pure noise, showing that the joint model can represent the one-to-many spectrum-to-structure relation explicitly.

Core claim

A conditional invertible network composed of i-RevNet bijective blocks can be trained to map graph-based molecular encodings to 128-bit binned 13C NMR spectra while remaining exactly invertible, so that the same trained weights produce structure candidates when the network is run backward from a spectrum code.

What carries the argument

Conditional invertible neural network built from i-RevNet bijective blocks, with the forward pass outputting a 128-bit binned spectrum code and the remaining latent dimensions capturing residual variability.

If this is right

Spectrum prediction and uncertainty-aware candidate generation become two uses of the identical trained weights.
The model produces numerically exact inverses on every training example.
Spectrum-code accuracy exceeds a random baseline on the filtered test set.
Inverted candidates on unseen spectra carry coarse structural signals that are better than chance.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

The same reversible architecture could be applied to other spectroscopies such as 1H NMR or infrared spectra without redesigning the inversion step.
Latent dimensions could be sampled at inference time to produce ranked lists of candidates ordered by how well they reconstruct the input spectrum code.
Integration with existing structure databases might allow the network to propose only chemically valid molecules during inversion.

Load-bearing premise

The combination of the 128-bit binned spectrum code and the latent dimensions must retain enough information about molecular structure that inversion still yields recognizable signals on spectra the model has never seen.

What would settle it

If the structures recovered by inverting the network on validation spectra are no closer to the true molecules than structures sampled from a random baseline, or if spectrum-code prediction accuracy falls to chance level.

Figures

Figures reproduced from arXiv: 2602.03875 by Eero Vainikko, Przemyslaw Karol Grenda, Stefan Kuhn, Vandana Dwarka.

**Figure 2.** Figure 2: A 13C spectrum of (-)-Menthol, measured at 150 MHz in CDCl3 [3]. 2 [PITH_FULL_IMAGE:figures/full_fig_p002_2.png] view at source ↗

**Figure 3.** Figure 3: F1 and loss values during training. per spectrum), a random predictor with the same probability of 1s and 0s achieves an expected F1 of about 0.0643, indicating that the obtained values are substantially above chance. In our context, the number and relative positions of 1s are more important than exact matches at all positions. In [PITH_FULL_IMAGE:figures/full_fig_p005_3.png] view at source ↗

**Figure 4.** Figure 4: The network consists of a sequence of invertible iRevNet blocks, which progressively transform the four [PITH_FULL_IMAGE:figures/full_fig_p008_4.png] view at source ↗

read the original 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 uncertainty-aware candidate generation within one end-to-end model.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Desk Editor's Note private letter to a colleague

The paper shows a conditional i-RevNet can handle both 13C NMR spectrum prediction and structure generation in one model, but the 128-bit binning and missing quantitative checks on validation make it hard to tell how much the spectrum actually drives the outputs.

read the letter

The core contribution is a single conditional invertible network built from i-RevNet blocks that maps molecular graph encodings to a 128-bit binned spectrum code in one direction and inverts to produce structure candidates in the other. Invertibility comes for free from the architecture, and the latent dimensions are meant to capture the one-to-many nature of spectrum-to-structure inference. That framing is new for this domain and avoids the need for two separate models. Training shows numerical invertibility on the training examples and spectrum prediction above chance, which is a reasonable starting point. On validation spectra the inverted outputs are described as giving coarse but meaningful structural signals, which at least suggests the model is not completely ignoring the input. The approach is clean and the choice of reversible blocks fits the bidirectional task well. The soft spot is the spectrum representation itself. Collapsing shifts into 128 bins discards the 0.1 ppm resolution that 13C NMR normally provides and drops multiplicity information. On unseen spectra this raises the risk that generated structures are mostly sampling from the training distribution rather than being tightly constrained by the specific input. The abstract gives no numbers to check this—no Tanimoto similarities, no validity rates, no comparison to a non-invertible baseline. Without those, it is difficult to judge whether the central claim holds up in practice. This work is aimed at chemoinformatics groups already using graph neural nets for spectroscopy. A reader looking for a proof-of-concept on invertible models will find something to think about, but anyone needing reliable performance numbers will want more data. The paper deserves a serious referee because the architecture idea is worth testing properly and the current gaps are fixable with standard evaluation steps. I would send it out for review and ask for quantitative metrics on validation structures plus some ablation on binning resolution.

Referee Report

2 major / 1 minor

Summary. The manuscript introduces a conditional invertible neural network (i-RevNet) for bidirectional mapping between molecular graph encodings and 13C NMR spectra encoded as 128-bit binned codes. The forward pass predicts the spectrum code from structure with latent dimensions capturing residuals; the inverse generates structure candidates from a spectrum code. Reported results include numerical invertibility on training examples, above-chance spectrum-code prediction, and coarse but meaningful structural signals on validation spectra, with the claim that such architectures unify spectrum prediction and uncertainty-aware candidate generation in one end-to-end model.

Significance. If the inversion step demonstrably produces spectrum-specific structural variability rather than samples from the learned prior, the work would be significant for chemoinformatics by offering a single bijective model that naturally encodes the one-to-many spectrum-to-structure relationship. The use of i-RevNet blocks to guarantee exact invertibility by construction is a technical strength that avoids the need for separate forward and generative models. However, the current evidence remains preliminary and does not yet establish that the coarsened spectrum representation drives chemically meaningful outputs on unseen data.

major comments (2)

[Abstract] Abstract: the claim that inversion on validation spectra 'produces coarse but meaningful structural signals' lacks any quantitative metric (e.g., Tanimoto similarity to ground-truth structures, fraction of chemically valid SMILES, or comparison against a non-invertible baseline). Without such measures it is impossible to verify that the 128-bit spectrum code, rather than the latent prior, is the dominant source of structural information.
[Abstract] Abstract and methods description: the 128-bit binned spectrum code collapses chemical-shift precision (typically resolved to 0.1 ppm) and discards multiplicity information. No ablation or sensitivity analysis is provided to show that inversion remains informative under this coarsening, which directly affects whether the reported structural signals on validation spectra can be considered spectrum-driven.

minor comments (1)

The manuscript would benefit from explicit reporting of the exact data-filtering criteria used to obtain the 'filtered subset' and the precise definition of 'above chance' for spectrum-code prediction (e.g., random baseline accuracy).

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for their constructive comments. We address each major point below and have revised the manuscript to incorporate quantitative metrics and sensitivity analysis where the original submission was lacking.

read point-by-point responses

Referee: [Abstract] Abstract: the claim that inversion on validation spectra 'produces coarse but meaningful structural signals' lacks any quantitative metric (e.g., Tanimoto similarity to ground-truth structures, fraction of chemically valid SMILES, or comparison against a non-invertible baseline). Without such measures it is impossible to verify that the 128-bit spectrum code, rather than the latent prior, is the dominant source of structural information.

Authors: We agree that the abstract's qualitative phrasing requires quantitative support to demonstrate spectrum-driven structure generation. In the revised manuscript we have added Tanimoto similarity statistics between inverted candidates and ground-truth structures on the validation set, the fraction of chemically valid SMILES produced, and a direct comparison against a non-invertible baseline that uses only the latent prior. These metrics confirm that the 128-bit spectrum code contributes measurable structural information beyond the prior. revision: yes
Referee: [Abstract] Abstract and methods description: the 128-bit binned spectrum code collapses chemical-shift precision (typically resolved to 0.1 ppm) and discards multiplicity information. No ablation or sensitivity analysis is provided to show that inversion remains informative under this coarsening, which directly affects whether the reported structural signals on validation spectra can be considered spectrum-driven.

Authors: We acknowledge that the 128-bit binning is a deliberate coarsening that sacrifices shift precision and multiplicity. While the original submission did not contain an ablation, the revised version now includes a sensitivity study that varies bin resolution (64, 128, 256 bits) and reports the resulting changes in inversion quality and structural fidelity on both training and held-out spectra. The analysis shows that the 128-bit representation retains informative signals above chance, although finer binning improves performance; we have updated the methods and abstract to reflect these findings. revision: yes

Circularity Check

0 steps flagged

No significant circularity detected

full rationale

The paper's derivation relies on the standard bijective property of i-RevNet blocks, which provides invertibility by construction as an external architectural feature rather than a self-derived or fitted result. Spectrum-to-structure mapping is achieved via explicit training on binned codes followed by inversion, with no equations or steps reducing outputs to inputs by definition, no load-bearing self-citations, and no renaming of known results. The central unification claim follows from the reversible architecture and data-driven training without circular reduction.

Axiom & Free-Parameter Ledger

2 free parameters · 2 axioms · 0 invented entities

The central claim rests on the architectural property that bijective blocks guarantee exact invertibility and on standard deep-learning assumptions about training convergence and representation power of graph encodings.

free parameters (2)

latent dimension count
Number of extra latent dimensions beyond the 128-bit spectrum code is chosen to capture residual variability but not specified numerically.
binning resolution
128-bit binning of the spectrum is a design choice that affects information loss.

axioms (2)

standard math i-RevNet style bijective blocks produce exact numerical invertibility by construction.
Invoked when stating that forward and inverse maps are available without additional training.
domain assumption Graph-based structure encoding plus latent variables suffice to represent molecular variability for NMR.
Assumed when training the conditional network and when interpreting inverted outputs as meaningful structural signals.

pith-pipeline@v0.9.0 · 5468 in / 1303 out tokens · 30944 ms · 2026-05-16T08:37:09.302350+00:00 · methodology

discussion (0)

Reference graph

Works this paper leans on

18 extracted references · 18 canonical work pages · 5 internal anchors

[1]

arXiv preprint arXiv:2105.02104 (2021)

Ardizzone, L., Kruse, J., L ¨uth, C., Bracher, N., Rother, C., K¨othe, U.: Conditional invertible neural networks for diverse image-to-image translation. arXiv preprint arXiv:2105.02104 (2021)

work page arXiv 2021
[2]

Analyzing Inverse Problems with Invertible Neural Networks

Ardizzone, L., Kruse, J., Wirkert, S., Rahner, D., Pellegrini, E.W., Klessen, R.S., Maier-Hein, L., Rother, C., K¨othe, U.: Analyzing inverse problems with invertible neural networks. arXiv preprint arXiv:1808.04730 (2018)

work page internal anchor Pith review Pith/arXiv arXiv 2018
[3]

Wiley- VCH, Weinheim, Germany (2009)

Berger, S., Sicker, D.: Classics in Spectroscopy: Isolation and Structure Elucidation of Natural Products. Wiley- VCH, Weinheim, Germany (2009)

work page 2009
[4]

Chemical Reviews125(19), 9256–9295

Das, S., Merz Jr., K.M.: Exploring the frontiers of computational nmr: Methods, applications, and challenges. Chemical Reviews125(19), 9256–9295. https://doi.org/10.1021/acs.chemrev.5c00259 9 APREPRINT- APRIL24, 2026

work page doi:10.1021/acs.chemrev.5c00259 2026
[6]

NICE: Non-linear Independent Components Estimation

Dinh, L., Sohl-Dickstein, J., Bengio, S.: Density estimation using real NVP. arXiv preprint arXiv:1410.8516 (2014)

work page internal anchor Pith review Pith/arXiv arXiv 2014
[7]

The Reversible Residual Network: Backpropagation Without Storing Activations

Gomez, A.N., Ren, M., Urtasun, R., Grosse, R.B.: The reversible residual network: Backpropagation without storing activations. arXiv preprint arXiv:1707.04585 (2017)

work page internal anchor Pith review Pith/arXiv arXiv 2017
[8]

arXiv preprint arXiv:2512.18531 (2025)

Hu, F., Tubb, J.M., Argyropoulos, D., Golotvin, S., Elyashberg, M., Rotskoff, G.M., Kanan, M.W., Markland, T.E.: Pushing the limits of one-dimensional nmr spectroscopy for automated structure elucidation using artificial intelligence. arXiv preprint arXiv:2512.18531 (2025)

work page arXiv 2025
[9]

i-RevNet: Deep Invertible Networks

Jacobsen, J., Smeulders, A.W.M., Oyallon, E.: i-revnet: Deep invertible networks. CoRRabs/1802.07088 (2018),http://arxiv.org/abs/1802.07088

work page internal anchor Pith review Pith/arXiv arXiv 2018
[10]

In: International Conference on Learning Representations (2018),https://openreview.net/forum?id=HJsjkMb0Z

Jacobsen, J.H., Smeulders, A.W., Oyallon, E.: i-revnet: Deep invertible networks. In: International Conference on Learning Representations (2018),https://openreview.net/forum?id=HJsjkMb0Z

work page 2018
[11]

Magnetic resonance in chemistry 60(11), 1021—1031,https://doi.org/10.1002/mrc.5234

Jonas, E., Kuhn, S., Schl ¨orer, N.: Prediction of chemical shift in nmr: A review. Magnetic resonance in chemistry 60(11), 1021—1031,https://doi.org/10.1002/mrc.5234

work page doi:10.1002/mrc.5234
[12]

Glow: Generative Flow with Invertible 1x1 Convolutions

Kingma, D.P., Dhariwal, P.: Glow: Generative flow with invertible 1x1 convolutions. arXiv preprint arXiv:1807.03039 (2018)

work page internal anchor Pith review Pith/arXiv arXiv 2018
[13]

Magnetic Reso- nance in Chemistry53(8), 582–589 (2015)

Kuhn, S., Schl ¨orer, N.E.: Facilitating quality control for spectra assignments of small organic molecules: nmr- shiftdb2 – a free in-house nmr database with integrated lims for academic service laboratories. Magnetic Reso- nance in Chemistry53(8), 582–589 (2015). https://doi.org/10.1002/mrc.4263

work page doi:10.1002/mrc.4263 2015
[14]

In: Proceedings of the 32nd International Conference on Neural Information Processing Systems

MacKay, M., Vicol, P., Ba, J., Grosse, R.: Reversible recurrent neural networks. In: Proceedings of the 32nd International Conference on Neural Information Processing Systems. p. 9043–9054. NIPS’18, Curran Associates Inc., Red Hook, NY , USA (2018)

work page 2018
[15]

arXiv preprint arXiv:2303.11239 (2023)

Nguyen, T.G.L., Ardizzone, L., K ¨othe, U.: Training invertible neural networks as autoencoders. arXiv preprint arXiv:2303.11239 (2023)

work page arXiv 2023
[16]

pymol.org

Schr ¨odinger, LLC: The PyMOL molecular graphics system, version 3.1.0 (2025), available athttp://www. pymol.org

work page 2025
[17]

Phytochemistry65(19), 2711–7 (2004)

Steinbeck, C., Kuhn, S.: Nmrshiftdb – compound identification and structure elucidation sup- port through a free community-built web database. Phytochemistry65(19), 2711–7 (2004). https://doi.org/10.1016/j.phytochem.2004.08.027

work page doi:10.1016/j.phytochem.2004.08.027 2004
[18]

Molecules28(3) (2023)

Wenk, M., Nuzillard, J.M., Steinbeck, C.: Sherlock—a free and open-source system for the computer-assisted structure elucidation of organic compounds from nmr data. Molecules28(3) (2023). https://doi.org/10.3390/molecules28031448

work page doi:10.3390/molecules28031448 2023
[19]

In: Gl¨uck, R., Kaarsgaard, R

Zingel, U., Kuhn, S., Vainikko, E.: Implementing reversible neural networks. In: Gl¨uck, R., Kaarsgaard, R. (eds.) Reversible Computation 2025. pp. 155–161. Springer, Cham (2025) 10

work page 2025