DiffNMR2: NMR Guided Sampling Acquisition Through Diffusion Model Uncertainty
Etienne Goffinet, Sen Yan *, Fabrizio Gabellieri · 2025
Read DiffNMR2: NMR Guided Sampling Acquisition Through Diffusion Model Uncertainty by Etienne Goffinet, Sen Yan *, Fabrizio Gabellieri (2025) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2502-05230.txt.
Introduction
Etienne Goffinet and co-authors speed up protein NMR by letting a diffusion model choose which measurements to take. Their 2025 paper attacks a chronic bottleneck: high-resolution NMR spectra take long to acquire, above all for complex proteins. Their method samples only part of the signal, rebuilds the spectrum, and uses the model's own uncertainty to pick the next most informative points.
Against leading strategies, the guided sampling lifts reconstruction accuracy by about half and cuts invented peaks sharply. Fewer fake peaks matter as much as speed, since a hallucinated line can mislead a structure. The loop of measure, rebuild, and ask-where-unsure is the core idea.
Catalog record
| CREATOR | Etienne Goffinet, Sen Yan *, Fabrizio Gabellieri |
|---|---|
| YEAR | 2025 |
| COLLECTION | PAPERS |
| FORMAT | TXT |
| FILE SIZE | 33 KB |
| VERSION | 1.0 |
| LICENSE | http://creativecommons.org/licenses/by/4.0/ |
| RIGHTS | CC BY |
| SOURCE | arXiv |
Source: https://arxiv.org/abs/2502-05230
Browse the full library · About the Tower of Babel
Full text
The opening of the work, from the library's hosted text:
DiffNMR2: NMR Guided Sampling Acquisition Through Diffusion Model Uncertainty Etienne Goffinet, Sen Yan *, Fabrizio Gabellieri Abstract Nuclear Magnetic Resonance (NMR) spectrometry uses electro-frequency pulses to probe the resonance of a compound’s nucleus, which is then analyzed to determine its structure. The acquisition time of high-resolution NMR spectra remains a significant bottleneck, especially for complex biological samples such as proteins. In this study, we propose a novel and efficient sub-sampling strategy based on a diffusion model trained on protein NMR data. Our method iteratively reconstructs under-sampled spectra while using model uncertainty to guide subsequent sampling, significantly reducing acquisition time. Compared to state-of-the-art strategies, our approach improves reconstruction accuracy by 52.9%, reduces hallucinated peaks by 55.6%, and requires 60% less time in complex NMR experiments. This advancement holds promise for many applications, from drug discovery to materials science, where rapid and high-resolution spectral analysis is critical. K eywords Diffusion Model, Nuclear Magnetic Resonance, Uncertainty Figure 1: a) Regular NMR experiment. A 2D NMR resonance map in the time-time domain (t_{e},t_{a}) is produced by recording the response of nuclei subjected to two-pulse sequences, where each row is obtained with a different time delay (called evolution time: t_{e} ) between the two pulses and t_{a} represent the acquisition time. This map is first transformed into a frequency-time representation via a 1D Fourier Transform along the acquisition time axis: (t_{e},t_{a})\to(t_{e},f_{a}) . A second 1D Fourier Transform along the evolution time axis produces the final spectrum: (t_{e},f_{a})\to(f_{e}f_{a}) . b) Non-uniform sampling (NUS) . Differing from the regular NMR experiment, Non-Uniform Sampling selects a subset of the evolution times (Row-wise Scan). After 1D Fourier Transform on the acquisition time axis, existing methods such as Compressed Sensing [ 1 ] and Low-Rank Approximation [ 2 ] are applied to reconstruct the frequency-time domain (Row Inpainting). c) Guided sampling workflow. Differing from NUS, we propose a novel dynamic sampling strategy by leveraging the diffusion model uncertainty. This is the first attempt using the diffusion model during NMR acquisition. The novelties are highlighted in blue texts. 1 Introduction Nuclear Magnetic Resonance (NMR) spectroscopy is a powerful tool for identifying the molecular structures, providing insights into the arrangement of atoms within complex molecules. It has become an indispensable tool in chemistry, biochemistry, and structural biology from its introduction in 1938 [ 3 ] . However, acquiring NMR spectra requires long acquisition times to reconstruct high-resolution information (Fig. 1 ). Traditional acquisition methods can be prohibitively slow in applications involving large biomolecules or when time is bounded, such as in drug discovery [ 4 , 5 ] or metabolomics [ 6 , 7 ] . For example, acquiring a 4D NMR spectrum (CCNOESY) of the phycobilisome linker polypeptide domain of CpcC takes up to 90 hours [ 8 ] . Non-uniform sampling (NUS) addresses this challenge by randomly sub-sampling the scanned signal while enabling high-quality spectral reconstructions (Fig. 1 ). Instead of sampling every data point in a regular pattern, NUS collects a subset of data points based on a pseudo-random or optimized sampling schedule [ 9 ] . The advantage of NUS lies in its capacity to significantly accelerate acquisition, but its effectiveness relies on robust reconstruction algorithms capable of handling incomplete data. Several methods have been developed for reconstructing spectra from NUS data, including Low-Rank Approximation (LR) [ 2 ] , Compressed Sensing (CS) [ 1 ] , and Deep Learning (DL)-based methods [ 10 , 11 , 12 , 13 ] . In this study, we introduce an AI-based approach that dynamically adapts the acquisition strategy. By integrating acquisition and reconstruction, the method leverages the uncertainty of the diffusion model (i.e., the degree of variability in the model’s reconstruction) to refine sampling in real-time, significantly reducing the acquisition burden. Using a real-life protein 2D NMR dataset [ 14 ] , we demonstrate that our method outperforms state-of-the-art sampling techniques and reconstruction methods, delivering accurate high-resolution spectral reconstructions from substantially reduced datasets. This work establishes a new paradigm in NMR spectroscopy by combining the strengths of NUS and advanced machine learning, enabling rapid, high-quality spectral analysis for time-sensitive applications. 2 2D NMR Data Two-dimensional nuclear magnetic resonance spectroscopy provides detailed structural and dynamic information about molecules by resolving spectral data along two frequency axes. This is achieved by extending the principles of 1D NMR with the additional time domain, offering insights into interactions between nuclei. 2.1 The Acquisition Process A 2D NMR experiment consists of three primary stages: preparation, evolution, and acquisition. During the preparation phase, tailored radio-frequency pulse sequences excite the nuclear spins and establish coherence pathways to encode the desired interactions. The preparation stage is followed by the evolution phase, where nuclear spin information is read during an evolution time ( t_{e} ), Revealing the chemical environment, such as chemical shifts and spin-spin couplings. The encoded magnetization is subsequently recorded during the acquisition time ( t_{a} ), generating dual time-domain data ( t_{e} , t_{a} ). These data are processed to produce a 2D frequency spectrum, where the indirect dimension ( t_{e} ) typically reflects spin-spin interactions, and the direct dimension ( t_{a} ) corresponds to the resonance frequencies of individual nuclei. Pulse sequences are carefully designed radiofrequency pulses and delays that manipulate spin states to achieve specific coherence transfers and magnetization pathways. For example, a simple COSY (Correlation Spectroscopy) experiment uses a two-pulse sequence to correlate chemical shifts of coupled nuclei, highlighting scalar (J) coupling. More complex sequences like HSQC (Heteronuclear Single Quantum Coherence) include additional steps, such as heteronuclear decoupling, to correlate proton and heteronucleus (e.g., 13 C or 15 N) chemical shifts. These advanced techniques are particularly valuable for analyzing larger biomolecules. Finally, to enhance the signal-to-noise ratio, it is common practice to perform multiple scans of the same spectrum and average them to obtain the final resonance map. 2.2 Data Processing The raw data from a 2D NMR experiment is initially represented as a 2D array of signal intensities varying over t_{e} and t_{a} . This data is transformed into a frequency domain using a two-dimensional Fourier transform (see Fig. 1 (a). Before the transform, techniques like phase correction and apodization are applied to improve spectral resolution and signal-to-noise ratio. The resulting 2D spectrum is plotted on the frequency axes of the evolution axis and acquisition axis, with cross-peaks indicating interactions between nuclei. 2.3 Non-Uniform Sampling (NUS) Non-Uniform Sampling (NUS) is a technique used to improve the efficiency of 2D NMR data acquisition by selectively sampling a subset of the evolution times ( t_{e} ) rather than collecting the full Nyquist grid. Instead of acquiring data at every increment of t_{e} , only a fraction of the points are sampled at random. One standard strategy is the Sine-Weighted Poisson-Gap sampling [ 15 ] , which assumes a Poisson distribution for the gap size between evolution times. The reconstruction of the full spectrum from sparsely sampled data requires advanced computational algorithms. One prominent method is Low-Rank Approximation (LRA) [ 2 ] , which represents the data as a Hankel matrix and approximates it with a low-rank reconstruction that best fits the sampled points. While computationally efficient, the LRA struggles to capture highly complex spectral features or systems with noise, leading to artifacts in the