A joint time-frequency domain deep learning network (JTF-Net) has been developed to improve the quality of nuclear magnetic resonance (NMR) spectra reconstruction and includes a novel method for quality assessment without needing fully sampled spectra.
Exploring protein structures, functions, and interactions is central to unraveling the molecular mechanisms of diseases and enabling targeted therapeutic innovation. Techniques like NMR, X-ray crystallography, and cryo-electron microscopy are pivotal for such investigations, with NMR offering unique advantages, such as getting insights under physiological conditions without altering the sample. While capturing multidimensional (nD) NMR spectra is invaluable, it is typically time-intensive, with sampling demand skyrocketing as dimensions increase. Non-uniform sampling (NUS) helps lessen this burden but necessitates sophisticated reconstruction algorithms to produce quality spectra afterwards.
Traditional NMR reconstruction techniques, like compressed sensing and multimensional decomposition, are limited by manual parameter settings, often leading to suboptimal results. This study introduces JTF-Net, a deep learning framework adept at synthesizing information from both time and frequency domains, enhancing reconstruction accuracy for complex NMR spectra, particularly for biomolecules such as proteins.
Using deep learning models has transformed data processing across diverse fields, including image and speech recognition. The same methodologies are now shaping the future of NMR analysis, whereby JTF-Net surpasses single-domain reconstruction models with its innovative dual approach. For example, during the reconstruction of nD protein NMR spectra, traditional algorithms may experience artifact peaks or even miss actual peaks. JTF-Net mitigates these issues by strategically integrating data across both time and frequency spectra. The JTF-Net has shown to achieve superior performances, resulting in minimal relative L2 norm error (RLNE) values with no observable peak losses—superior to existing methods.
To evaluate reconstruction quality, the study establishes the REconstruction QUalIty assuRancE Ratio (REQUIRER) metric, allowing for spectral quality assessment without prior full sampling. Not only does this provide assurance for researchers about the reliability of their reconstructed data, but it also marks significant progress against previous limitations where users had no way to gauge spectrum quality effectively. With defined threshold indicators of 0.35 for 2D spectra and 0.55 for 3D, researchers can now achieve effective evaluations from reconstructed spectral data.
Tests carried out on various protein spectra, including the 2D 15N-1H HSQC spectra of GB1 and T4L L99A, with undersampling at 12.5% proved the robustness of JTF-Net. Results demonstrated no peak loss or artifact peaks and consistently achieved superior RLNE scores, meeting quality thresholds effectively across various sampling schemes.
Researchers highlighted the automated nature of JTF-Net, circumventing the requirement for manual tuning of parameters, thereby facilitating high-speed analyses. Subsequent experiments also demonstrated the ability to reconstruct 3D spectra like the HNCO spectrum of the Azurin protein, employing various sampling rates (10-20%), thereby showcasing its versatility and effectiveness. The trend across different sampling rates consistently indicated JTF-Net's lower RLNE values, signifying its reliability as the sampling rate increases.
With deep learning's role burgeoning within the scientific sphere, JTF-Net exemplifies the potential to revolutionize conventional NMR techniques. The development of REQUIRER, forming the first reference-free quality metric, substantially enhances the feasibility of NMR usage and facilitates continued research endeavors without compromising data integrity. Enabling researchers to trust their results without the standard of fully sampled spectra fundamentally shifts the paradigm for nuclear magnetic resonance research, paving the way forward for future advancements and applications.
Through this study, JTF-Net stands as both a methodological breakthrough and tool for the research community, emphasizing the intersection of deep learning and traditional spectroscopy for enhanced scientific exploration.