Uncertainty Exploration of Deep Learning Enabled Fast Multidimensional NMR Spectroscopy of Proteins

Deep learning (DL) has shown great promise in accelerating multidimensional NMR spectroscopy with high-accuracy reconstruction. Nevertheless, its practical deployment requires not only high accuracy but also reliability evaluations along with the predictions to prevent overconfident predictions and erroneous biological interpretations. With this in mind, this study establishes a reference-free quality assessment benchmark by modeling internal uncertainty for reliable DL reconstruction of biological NMR spectroscopy. Three uncertainty quantification frameworks, namely, Deep Ensemble, Monte Carlo (MC) Dropout, and Evidential Deep Learning, are integrated into existing deep reconstruction backbones to provide uncertainty estimation along with prediction results. Experimental validations on 2D and 3D NMR spectra of proteins demonstrate that the newly established metrics enable the assessment of prediction reliability both on global reconstruction and frequency-level details without access to reference. Comparative results further indicate that Deep Ensemble exhibits advantages in reconstruction accuracy and consistency between uncertainty maps and reconstruction residuals, whereas Evidential Learning provides real-time uncertainty estimates with no additional computational cost. Its high consistence with traditional reference-based metrics allows users to pinpoint potential problematic regions, thus empowering structural chemists and biochemists to focus their attention on high-uncertainty regions in cases without ground-truth data. Therefore, our findings answer critical questions in general DL-based approaches, i.e., whether, where, and why to trust, boost the reliability of AI-assisted NMR spectroscopy, and facilitate more secure biomolecular structure determination.